• Clay has grown revenue from 5 to >30m in the past year, has taken less than 18 months to reach an eight figure run rate. They are currently hiring product and systems focused software engineers in New York City.

• Modal has grown revenue significantly in the past year and is now a unicorn. They’re currently hiring product and systems focused software engineers in New York City.

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should Join The Artificial Intelligence Underwriting Company

(Click the link to read online).

It can be hard to trust AI.

As impressive as they are, today’s models have a habit of bullshitting in unpredictable ways.

In February of last year, an Air Canada chatbot wrongly told a bereaved traveler he could fill out a ticket refund application after purchasing full-price tickets to claim back more than half of the cost. Unfortunately, no such application existed, and a small claims court had to order the company to pay the user back.

Two months later, a chatbot deployed by the NYC government was found to be dispensing illegal advice on housing policy, worker rights, and rules for entrepreneurs. It told users, among other things, that landlords don’t have to accept Section 8 vouchers or other rental assistance when discrimination on source of income is illegal.

A little after that, Virgin Money’s chatbot made the news for refusing to continue a conversation which involved the word “virgin.”

While these situations are rare (and decreasing in frequency), they and the media around them have created somewhat of an awkward environment for enterprise AI adoption. While the technology is mostly ready, it can still be hard to trust it with stuff.

This general perception is not unfounded. Foundation models and the systems built on them are inherently non-deterministic, and because of this are subject to new risks - both in regular use and in adversarial scenarios. A customer chatbot might give a bad refund, a medical support agent might share dangerous advice, and an enterprise assistant might leak confidential information.

As AI grows in capability and seeks to run more of our economy (via voice agents, computer using agents, and eventually embodied intelligence) these risks - and their consequences - will only grow in breadth and scope. Together, they present systemic challenges that hold back the adoption of AI and our realization of its benefits: nearly half of companies have cited AI accuracy and hallucinations as a major adoption risk, and more than 60% have named privacy and compliance as a top barrier as well. The percent of S&P 500 companies who reported a material risk related to AI in 10-K filings went from 12% in 2023 to 72% in 2025.

Getting this right will be important: one report forecast that by 2027, a company’s generative AI chatbot would directly lead to the death of a customer from poor information.

Historically, when new technologies introduced new risks, new standards and trust-setting mechanisms (i.e. audits, insurance) would rise in parallel to address those risks and accelerate adoption.

Think electricity: early wiring caused fires until Underwriters Laboratories’ listings and building codes set minimum safety bars. Think cars: mass adoption took off alongside seat‑belt mandates, airbag standards, and crash‑test ratings from NHTSA/IIHS which made safety legible to buyers. Think web security and online payments: e‑commerce didn’t scale until SSL/TLS certificates put a lock in the browser and PCI‑DSS set rules for handling card data. Think cloud software: enterprise deals sped up once vendors could show SOC 2 attestations so buyers knew how security and controls were managed.

In a similar way, effectively managing AI’s risks will be essential to earning trust and unlocking universal adoption.

Enter: AIUC.

‎

Background

The Artificial Intelligence Underwriting Company (AIUC) is building security standards for AI. Concretely, they offer a standard developed in collaboration with leading institutions for safely deploying AI (AIUC-1), independent audits and red-teaming against that standard, and insurance underwriting tied to audit results. Together, these three offerings provide a fully managed trust stack for AI deployments.

An AI customer support startup deploying their product with the largest companies today will be faced with a battery of custom evals, third‑party red‑teams, policy/legal reviews, and internal security sign‑offs. Post‑deployment, the startup will have to be proactive about keeping up with model updates, new jailbreak techniques, and evolving regulations. For most startups, this overall process involves making bespoke assurances, hiring third party consultants, and spinning up a number of one-off processes. It can be annoying, especially since at the end customers are still left with residual risk.

For both the vendor and buyer, AIUC offers what’s essentially “trust in a box.”

They combine a domain‑specific standard (AIUC‑1), a recurring audit/red‑team program, and insurance for the product that’s priced on the audit. In doing so, AIUC abstracts away:

• Assessment and Certification. Vendors map their system to AIUC‑1, undergo deep technical testing (safety, security, privacy, reliability) and operational/legal control reviews, and earn a certificate that enterprises can accept as a sign of diligence, similar to SOC II. The certification is renewed annually with at‑least-quarterly technical retesting.

• Risk and Accountability. AIUC’s insurance offering binds coverage for AI‑specific failure modes like hallucination‑driven loss, data leakage, IP issues, harmful/defamatory outputs, and faulty tool actions. Based on the audit, safer systems get better terms, and riskier systems must take remediation actions before they’re insurable.

The standard, which sits at the center of this, organizes trust into six domains: Data and Privacy, Security, Safety, Reliability, Accountability, and Society. Then, it turns each domain into specific, auditable requirements with a fixed update cadence so they don’t go stale. For example, clause D002 under Reliability requires companies to appoint independent third parties to evaluate hallucinated outputs at least every three months.

To make money, AIUC sells a bundled assurance + insurance program to AI startups selling into enterprises. Consider our customer-support product that’s stuck in pilot because security wants red-team results and legal wants indemnity and privacy assurances. After AIUC, the startup hands over their AIUC-1 audit report/certificate and a carrier-backed policy covering AI-specific perils (hallucinations, data leakage, IP, tool-action failures) with limits up to $50M. This combination answers the two questions every legal and IT team is asking - “is it safe?” and “who pays if it isn’t?” In doing so, AIUC doesn’t just unblock deals, it unblocks them in a scalable way - replacing the ad-hoc diligence with a repeatable motion.

The value to startups is concrete: faster enterprise sales (fewer bespoke questionnaires, a clear third-party bar), lower total cost of assurance (one integrated program instead of a patchwork of consultants), and risk transfer their customers can understand and accept. Because pricing is tied to audit results and refreshed via quarterly re-tests, safer systems earn better insurance terms over time - aligning incentives for everyone involved.

AIUC launched publicly in August of 2025 and has moved quickly to position itself as the “trust stack” for AI agents. Early customers include Cognition, ElevenLabs, Intercom, and many of the other largest names in customer support. Ada, for instance - another unicorn AI CX startup - completed AIUC-1 testing “far exceeding typical enterprise requirements,” which helped unblock deployment with one of the world’s largest social media platforms.

The company was founded late last year by Rune Kvist and Rajiv Dattani. Rune was previously the first product and GTM hire at Anthropic, as well as a board member for the Center for AI Safety. Rajiv was previously a McKinsey insurance partner and COO at METR (an AI risk evaluator).

The company emerged from stealth July of 2025 with a $15M seed round led by Nat Friedman and Daniel Gross, with support from the former CISO’s of Google Cloud and MongoDB, as well as one of the co-founders of Anthropic. They also recently announced partnerships with with CISOs and risk leaders from Google Cloud, Goldman Sachs, MongoDB, Anthropic, ScaleAI and a number of others to help shape the standard.

AIUC has a strong team, a compelling idea, and exciting traction. Their ambition is to build the trust layer infrastructure for AI agents - accelerating adoption but in a responsible way. A marriage between the EAs and E/Accs.

Can they pull it off?

We think they have a shot, let’s get into why.

‎

The Market

AIUC makes money when AI startups pay for their audits or insurance policies in order to help unblock enterprise deals.

Why is this a large market?

Well, there are a lot of companies selling AI (over $45 billion of VC funding went towards AI startups in 2024), and all of them run into trust and compliance-related friction - friction they’re willing to pay to remove. Here are some stats:

• Nearly half of companies have cited AI accuracy and hallucinations as a major adoption risk, and more than 60% have named privacy and compliance as a top barrier as well. AI-related security and privacy incidents rose by 56.4% in 2024 compared to the previous year, and a third of enterprises have reported AI-related data breaches in the past year.

• 72% of S&P 500 companies have reported a material risk related to AI in their 10-K filings this year, up from 12% in 2023. 38% of companies reported at least one reputational risk, which involves things like implementation and adoption, hallucinations and inaccurate outputs, consumer-facing deployments, and privacy and data protection. 20% of companies reported at least one cybersecurity risk, which involves things like AI-amplified cyber risk and data breaches/unauthorized access.

• One survey found AI is now the average organization’s single most cited technology risk, and they’re taking this seriously. As a result, 29% of vendors have lost a new business deal because they lacked a compliance certification, and 72% undertook a compliance audit specifically to win a client. Large companies are starting to send security and liability questionnaires for AI systems that effectively require answers around AI risk controls.

As a result of this, there’s a rising ecosystem of tools and services designed around AI governance. The AI Trust, Risk and Security Management market was valued at $2.3 billion in 2024 and is expected to grow to over $7.4 billion by 2030. While specialized AI insurance is just emerging, major insurers also see it as a growing opportunity - one estimate projects that insurers could be writing $4.7 billion in AI liability premiums annually by 2032.

There is strong historical precedent for the rise of new compliance/insurance businesses alongside new technologies and markets. Taking some of these as comparables:

• SOC 2 audits can cost a company anywhere from $20,000 to $60,000, and is often an annual expense to maintain certification. The aggregate market for these services was valued at $4.2 billion in 2024 and is projected to grow to $9.1 billion by 2033.

• Cybersecurity insurance (covering data breaches, hacks, and similar events) is also a new and rapidly growing category. The market was worth some $15 billion in global premiums in 2023 and is expected to grow to $20–30 billion in the coming years.

AI introduces a new category of insurable risk and is likely to grow quickly as well. We are already seeing the first movers besides AIUC: in 2025 Lloyd’s underwriters launched AI-specific liability policies, and Google recently partnered with Beazley, Chubb, and Munich Re to offer AI coverage for customers of their cloud AI services.

We expect growth in this category will be driven by a number of tailwinds:

• Regulatory momentum. The EU’s AI Act effectively requires “AI governance” and documentation for high-risk systems, and mandates rigorous risk management and conformity assessment for “high-risk” AI systems before they can be deployed. In the US, while regulation is still evolving, sectors like finance and healthcare are already under strict obligations for managing model risk (e.g. banks following model risk management guidelines). These public laws will normalize new standards, audits, and eventually risk transfer as standard practice.

• Capability expansion. As AI agents take on higher-stakes work, the insurable surface grows: voice/contact-center agents are becoming the first line of customer interaction, computer using agents will soon handle complete workloads, and robotic/physical systems will soon tie AI decisions to physical outcomes. Each class introduces new quantifiable loss scenarios - hallucinated guidance, bad tool calls (refunds/purchases), data leakage, unsafe code, or operational disruption - precisely the issues companies like AIUC will help digest.

• Momentum and path dependence. Once a few category leaders make audited AI controls and affirmative AI coverage a part of their sales motion, the rest of the market tends to follow (similar to how SOC 2 became a de facto gate in SaaS). For instance, once one chatbot company wins a marquee RFP by showing AIUC-1 certification, the bar effectively resets - competing vendors may find they have to match the artifact or get screened out on risk. This has already begun playing out for AIUC in the customer support category.

‎

The Product

So, now that we understand the opportunity around AI and trust, let’s have a deeper look at how AIUC will create and capture value.

Specifically, let’s dig into the three parts of their offering.

The Standard

AIUC-1 is a domain-specific standard for AI agents that turns broad “trust” goals into auditable controls. It’s organized into six risk domains - Data & Privacy, Security, Safety, Reliability, Accountability, and Society - with each domain broken into numbered requirements marked Mandatory or Optional. Each requirement comes with an audit frequency, along with a set of recommended control activities. For instance, one sub-requirement for detecting adversarial input includes implementing monitoring for prompt injection patterns and jailbreaking techniques.

The standard is explicitly designed to be kept current (it is updated quarterly) and to operationalize high-level frameworks (i.e. the EU AI Act, NIST AI RMF, ISO 42001) into concrete, testable controls auditors can check.

Going over a sampling of the standard’s categories and clauses:

• A. Data & Privacy — prevent leakage and misuse. The intent of this is to stop AI systems from leaking PII/IP, training on customer data without consent, or commingling tenants’ data. Concretely, A004 (Protect IP & trade secrets) adds safeguards to prevent leaks of internal IP/confidential info, and A006 (Prevent PII leakage) adds safeguards to prevent leaks of personal data in outputs.

• B. Security — adversarial robustness and access control. The aim here is to harden agents against jailbreaks, prompt injection, scraping, and unauthorized tool calls. Concretely, B001 (Test adversarial robustness) requires a formal red-team program aligned to an adversarial threat taxonomy (e.g., MITRE ATLAS) on a fixed cadence, and B004 (Prevent AI endpoint scraping) requires controls like rate-limiting/quotas and other defenses to blunt automated probing.

• C. Safety — prevent harmful or out-of-scope outputs. This domain requires pre-deployment testing and ongoing checks to stop toxic content, high-risk advice, deception, and policy-barred outputs. Concretely, C003 (Prevent harmful outputs) requires safeguards against distressed/angry personas, offensive content, bias, and deceptive responses, and C004 (Prevent out-of-scope outputs) enforces blocks on categories like political discussion or healthcare advice where prohibited.

• D. Reliability — reduce hallucinations and unsafe tool calls. The goal here is self-explanatory. Concretely, D001 (Prevent hallucinated outputs) requires concrete technical controls to suppress fabricated answers (e.g., fact-checking, uncertainty signalling), and D002 (3rd-party testing for hallucinations) mandates independent evaluations at least every 3 months to keep accuracy honest over time.

AIUC-1 isn’t just another framework; it’s novel because it turns broad existing standards into tests auditors can actually run on a schedule. For example, where NIST’s AI RMF tells teams to “measure” and “manage” risk, AIUC-1 sets specific pass/fail gates like independent hallucination/safety evaluations at least every quarter and prescribes named artifacts (e.g., incident runbooks with owners and timelines) that auditors can verify. In practice, that means a buyer doesn’t just see a claim that the vendor measures accuracy - they see evidence of a third-party test from the last 90 days and the remediation that followed.

The Certification

So the standard is cool, but how does it get productized?

The first way is by offering to certify individual AI products against it. This process involves hands-on red-teaming, reviews of operational and legal controls, and ongoing testing on a fixed cadence.

More concretely, the process starts with scoping and readiness. AIUC and the vendor define the agent’s capabilities (including any tools it can call), intended use cases, and data flows, then map those realities to the AIUC-1 control domains. That map determines what will be tested and what evidence is required. For example: a customer-support chatbot startup might expose tools to (a) create/modify tickets in Zendesk, (b) issue refunds via an internal payments API, and (c) read prior chats via a RAG index. In scoping, the team and AIUC enumerate these tool permissions, the data sources the RAG index can access (order history, masked PII), and the decision boundaries (what requires handoff to a human). That drives the test plan: adversarial prompts around refunds (Security/Tool-use), hallucination checks on policy answers (Reliability), harmful-content screens (Safety), and proof of privacy controls (Data & Privacy). Additionally, at this stage AIUC will specify the artifacts they will verify later (e.g., refund-action allowlist, logging of tool calls, disclosure UX to users interacting with AI, incident runbooks with named owners).

Next comes a technical battery where auditors will actively probe the system itself. On the security side, they attempt jailbreaks and prompt-injection, check rate-limiting and endpoint hardening, and assess whether the deployment environment follows least-privilege and other baseline defenses. On safety and reliability, they stress-test for harmful or out-of-scope outputs, evaluate how well hallucinations are suppressed, and verify that agent tool actions remain inside policy. On privacy and IP, they test for data-leakage routes and misuse of sensitive inputs or copyrighted material. This “real testing” approach contrasts with traditional audits that mostly check for the existence of procedures. If issues are found, AIUC may iterate with the company until they are fixed. For example: our customer-support chatbot startup fails an early prompt-injection test that coaxes the bot to issue a full refund outside policy; auditors also demonstrate scraping the model endpoint due to missing rate limits, and trigger a retrieval answer that surfaces unredacted PII from the RAG store. The startup remediates by tightening the refund tool’s pre-conditions (policy checks + human-in-the-loop for high amounts), adding rate-limiting/quotas at the API gateway, and implementing PII redaction at ingestion plus output filters. AIUC re-tests those failure cases until they pass.

In parallel, AIUC will review the organization’s operating model around the agent: change-management, logging and monitoring, access controls, disclosures, vendor diligence, and incident response plans for AI-specific failures. The aim is to confirm that the company’s processes match what the agent can actually do and that there’s a staffed plan for when things go wrong. For example: the customer-support chatbot startup formalizes a model-change approval gate (who signs off when moving from v1.7→v1.8), enables immutable logs of model outputs and tool calls to support root-cause analysis, adds an AI disclosure banner in the chat UI, and writes incident playbooks for (i) harmful outputs, (ii) privacy/data leaks, and (iii) hallucinations that cause financial loss - with owners and notification timelines. Auditors review those artifacts and sample records to confirm the controls exist in practice, not just on paper.

Once complete, the output for customers is practical and reusable: an AIUC-1 certificate and audit report that can be shared in security and legal reviews, a remediation plan where needed, and (for qualifying systems) eligibility to bind insurance with terms priced off the audit results. Ada, a real-life AI customer service startup, did just this, using its AIUC-1 certification to unblock a deployment with one of the world’s largest social-media platforms.

This certification is not one-and-done. AIUC requires an annual recertification of operational controls and at least quarterly technical testing to keep pace with model changes and new attack vectors. The standard itself is also reviewed with contributors and updated on a predictable schedule - January 1, April 1, July 1, and October 1 - so the bar moves with the threat landscape.

The Insurance

Finally, AIUC offers liability coverage specifically for AI-related failures.

The offer is straightforward: once an agent is certified against AIUC-1, the vendor can bind affirmative AI liability that pays when the agent’s behavior causes business loss. Policies are backed by established insurers and marketed to cover AI-specific losses up to $50M, including hallucinations, brand/reputation harm, data leakage, IP infringement, and tool/action failures like incorrect refunds or purchases.

Underwriting is tied to evidence produced by the certification process. The audit report and change history are the main inputs for pricing and eligibility, and AIUC’s cadence - annual re-certification plus at-least-quarterly technical re-tests - keeps those inputs fresh so terms track the agent’s evolving risk surface. In practice, safer posture (fewer/open findings, tighter controls around tool use, robust logging and incident playbooks) results in better limits/retentions. The point is to make the attestation actionable for a carrier: a live, repeatable test program that converts abstract “AI risk” into insurable loss scenarios.

Going back to our AI customer-support startup, consider a scenario where a month after deployment, a change impacts the agent’s refund tool and the agent enables several thousand dollars worth of unauthorized refunds. Because the vendor bound affirmative AI liability via AIUC’s program (papered by a partner carrier, not AIUC), the claim is adjusted under the “tool/action failure” peril and paid up to the policy limit after the retention. For the end user (the enterprise buying the support agent), this is valuable because they get a carrier-backed financial backstop for AI failures - reducing the risk around deploying the product.

How does AIUC make money? Premiums are collected by the carrier, and AIUC (acting like a program/MGA partner) is paid insurer-funded commissions/overrides on written premium, typically the majority of MGA revenue, and potentially with an additional contingent commission when the book’s loss ratio and volume targets are hit.

‎

The Long Term

AIUC’s long term mission is to become the trust layer for machine decision‑making - first for agents that type and talk, then for agents that click and code, and eventually for systems that act in the physical world.

To get there, AIUC will expand to support a variety of modalities:

• Generative media for entertainment, marketing, and education. Additional controls and audits here might revolve around brand‑safety taxonomies, deepfake/impersonation checks (likeness/voice rights), provenance and watermarking so outputs are traceable, dataset/license attestations, protections for minors, and escalation paths for borderline content.

• Voice/phone agents in the contact center (call deflection, collections, claims). These agents schedule, take PCI‑safe payments, handle multilingual triage, capture consent, and escalate on low confidence or poor sentiment. Additional controls and audits here might revolve around PII redaction in transcripts, consent/Do‑Not‑Call governance, abuse handling, and caller verification.

• Browser/computer‑use agents that take actions in SaaS (refunds, orders, ticket edits, CRM updates). These agents will operate across everything from ticketing to ERP/finance to commerce to HR/IT. Additional controls and audits here might revolve around tool allowlists and preconditions, rollbacks, and session isolation.

Additionally, AIUC can move both up the stack - working with foundation model providers - and down the stack, partnering with testing/inspection/certification giants (UL Solutions, TÜV, Intertek, SGS, Bureau Veritas) to create vertical-specific certifications and profiles. Along this second direction, you can imagine extra specific standards and policies for regulated industries like banking or healthcare. Verticalization would deepen AIUC’s hold over the broader AI underwriting space, strengthening their position.

Looking further ahead, as AI enters the physical world, AIUC may expand their testing and underwriting to support physical systems like domestic and industrial robots. The same primitives that would be required for safe deployment - safe‑stop/remote override, geofencing and speed limits, human–robot‑interaction boundaries, telematics‑backed incident logs, and environment‑specific hazard tests - map cleanly to certification artifacts and affirmative coverage for machines that move.

Now is the right moment - AI has just crossed a capability threshold where agents are doing real work, and loss events are no longer hypothetical. AIUC has an early lead on the space.

Competitive Landscape: A Sum Greater Than Its Parts

Having read all this, you may be wondering why AIUC needs to do so much at once. Does bundling the standard, audits, and insurance under one roof really create an edge?

We think it does, for a few reasons:

• Aligned incentives: binding insurance to audits fixes the “checkbox” issue. Standalone standards and audits tend to drift into “pass-the-test” theater because the tester doesn’t pay when things go wrong - an issuer-pays misalignment that infamously blew up credit ratings pre-2008. AIUC closes that gap: it prices live insurance off the very audit it runs. That gives them skin-in-the-game: if the bar is weak, claims are costly. This incentive loop is a non-obvious edge: it continuously pulls vendors toward controls that reduce loss (hallucinations, tool-misuse, leakage), something which pure auditors (no exposure) and pure insurers (no testing muscle) struggle to replicate.

• Distribution advantage: as AIUC-1 catches on and becomes industry standard, AIUC’s brand will benefit as the standard’s creator and namesake. Famous historical examples here include Adobe inventing and standardizing the PDF format, Tesla developing and open sourcing the North American charging standard (SAE J3400), and Intel inventing and prioritizing USB support.

Today, while there are a number of companies working on specific pieces of AIUC’s platform, there’s no one else doing their full combination of things.

• There are a number of alternative standards like NIST’s AI Risk Management Framework, ISO/IEC 42001 for AI management systems, and ISO/IEC 23894 for AI risk management guidance. However, compared to AIUC, they don’t prescribe agent-specific controls, run hands-on red-teaming on a fixed cadence, or tie live certification to carrier-backed AI liability. They’re valuable governance frameworks, but they stop short of a production-ready, product-level pass/fail bar.

• There are a number of independent red-teaming companies like Haize Labs, Robust Intelligence, and CalypsoAI. However, compared to AIUC, they optimize for bug-finding, not deal-closing: they hand you a report and leave the vendor to translate findings into enterprise-ready evidence, align security/legal/procurement, and answer the “who pays?” question. AIUC runs the full trust motion - scoping to real tool permissions and loss scenarios, testing against a public, versioned bar, retesting on a fixed cadence so evidence doesn’t go stale, and packaging outcomes into a reusable artifact buyers can accept.

• There are a number of insurers exploring insurance for AI agents like Beazley, Chubb, and Munich Re. However, compared to AIUC, they don’t start upstream with hands-on tests that convert model risk into procurement-ready evidence; most offerings are bolted onto cyber/cloud/tech-E&O programs and underwritten via questionnaires - not a live, pass/fail bar exercised on the actual agent. That matters because without fresh, product-level evidence, underwriting stays conservative and vendors still face bespoke diligence; AIUC closes the loop by running and refreshing the tests on a fixed cadence and feeding those results directly into pricing.

Team

AIUC wants to be the trust layer for AI agents - accelerating adoption while keeping deployments responsible. They want to help companies adopt AI as fast as capabilities advance.

It’s an ambitious vision, but the team has the right backgrounds to make this happen. Rune was previously Co-Founder and President at Momentum, a AI GTM startup, and before that the first product/commercial hire at Anthropic, where he was amongst the first people to deeply consider the problem space. Rajiv was previously COO at METR, a research non-profit developing frontier AI evaluations, and before at a partner at McKinsey focused on insurance. Both hold degrees in Philosophy, Politics, and Economics from Oxford. The team combines expertise around AI models, AI products, AI risk, and AI insurance.

Culturally, the company operates around a simple set of values. Rune summarized the team’s ethos as a group of “truth-seeking truth-tellers with skin in the game.” He elaborated:

• Truth Seeking - the team pursues reality as it is, not as hoped. They test, measure, and quantify rather than relying on abstractions, treat insurance-linked feedback loops as mechanisms to drive accurate risk assessment and continuous improvement, and hold a high bar for technical accuracy and updates standards as new evidence emerges.

• Truth Telling - the team acknowledges both AI’s potential and risks. They close the information gap by sharing facts clearly without spin or fear-based selling, build trust through substance, and communicate without jargon.

• Skin in the Game - the team demonstrates convictions through financial and reputational commitment. They back positions with capital rather than opinions alone, align incentives with positive outcomes for customers and end users, and bear costs when wrong and win trust when right. They operate with a public-goods mindset rooted in market solutions.

In addition, Rune emphasizes how they are building for durability, not exit velocity. The team is committed to building an enduring institution, one that plays a foundational role in the adoption of AI and other frontier technologies across the economy.

Conclusion

It’s not easy to trust AI.

But we shouldn’t be expected to, out-of-the-box.

Every wave of transformative technology only scaled after the right trust-setting infrastructure showed up. Electricity had UL, autos had IIHS, and the cloud had SOC 2.

AI needs it’s own version, and AIUC is building it: a live standard, real testing, and insurance that addresses residual risk.

If they pull it off, adopting an agent will feel less like a gamble and more like a running through a go‑live checklist.

The future of AI could be trustworthy by default.

They’re hiring.

‎

In case you missed our previous releases, check them out here:

• Clay - Bringing creativity back to GTM.

• Etched - The chip for transformers.

• Meter - The next modern utility.

And to make sure you don’t miss any future ones, be sure to subscribe here:

Finally, if you’re a founder, employee, or investor with a company you think we should cover please reach out to uhanif@stanford.edu - we’d love to hear about it.

• Clay has grown revenue from 5 to >30m in the past year and took less than 18 months to reach an eight figure run rate. They’re currently hiring product and systems focused software engineers in New York City.

• Modal has grown revenue significantly in the past year and has taken less than 18 months to reach an eight figure run rate. They’re currently hiring product and systems focused software engineers in New York City.

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should Join Lovable

(Click the link to read online).

Programming has always come with its set of challenges.

In the 1950s and 1960s, it would have been fitting an entire payroll system into 4KB of memory, waiting until your slot (midnight) to use the compiler, or managing physical punch cards - where debugging meant going through a thousand precisely stacked cards to search for a few misentered holes.

In the 1980s and 1990s, it would have been splitting individual programs across a dozen floppy disks, going through phone book-sized reference manuals to search for two-line examples, or manually managing memory - dealing with misused pointers, memory leaks, and vague errors like segmentation faults and core dumps.

In the 2000s and 2010s, it would have been learning the twenty ways to center a div, understanding why JavaScript decided 2 plus "2" should equal "22”, or managing the never-ending stream of new frameworks, best practices, and cloud infrastructure - everything from containerization and CI/CD pipelines to infrastructure-as-code.

Even today, it's explaining to Cursor why each "ah, I see it now!" is still wrong - while switching models.

Despite decades of progress, the challenge of software development still mostly revolves around troubleshooting errors, configuring environments, and wondering why things never quite work out of the box. This frustration is the foundation of the global IT industry, and is what keeps the trillion-dollar moat standing between people and working software filled with developer tools, Javascript frameworks, and implementation consultants.

What people have always wanted, but never quite had, is a software creation tool that "just works" - a system where you can just tell the computer what you need before it figures out the rest.

This tool is what Lovable is aiming to finally deliver.

Their aim is to build the last software product, one that will simplify software creation completely.

It’s an ambitious goal, but in their pursuit of it, they’ve created one of the world’s fastest growing companies.

‎

Background: The Landscape of AI Code Generation

Lovable is a “vibecoding” product, one that turns prompts into apps directly. Before getting too deep on them and what that means, however, it’d probably be helpful to get a sense of the broader landscape they fit into.

AI coding has been one of the hottest spaces for startups to build in, and companies in it have raised over three billion dollars in the past year.

While many of these companies - Cursor, Cognition, Windsurf, Bolt, and Lovable - often get lumped together under the same general category, there are meaningful distinctions between each, particularly around user experience and the degree to which they automate software engineering.

Companies focused on AI coding generally operate at three different levels of abstraction:

• Tools that generate code for you

• Tools that generate diffs for you

• Tools that generate applications for you

Tools that Generate Code - Models

Starting at the lowest level of abstraction, companies in this category offer models that will generate code for you.

While any language model has some code generation capabilities, a few companies have distinguished themselves by building models specifically optimized for this task through specialized pre- or post-training techniques. Notable players in this space include:

• Claude (Anthropic) - the most popular model when it comes to code generation. Cursor users pick Claude for the vast majority of their requests, and many developers associate it with strong general code generation, debugging, and architecture/design capabilities. Anthropic raised $3.5 billion in their most recent round (March 2025), valuing the company at over $61.5 billion.

• Magic - claims to have trained a frontier coding model called LTM-2 with a record-breaking 100M-token context window - significantly longer than competing models. The idea is to ingest entire codebases and provide the full context needed to complete complex tasks. While no public model has been released yet, Magic has raised $465 million to date, including a $320 million round in August 2024 led by Eric Schmidt and Capital G.

• Poolside – building code-first foundation models ("Malibu" and “Point”) fine-tuned using reinforcement learning with code execution feedback (RLCEF), a technique which lets models learn from realistic coding tasks derived from real-world code bases. Poolside has raised over $600 million total, including a $500 million Series B in October 2024 led by Bain Capital Ventures. They have not made a public release yet.

These models can output code in whatever language/framework based on the prompt you provide, but it's up to the user to include the necessary context - which might include the actual request, surrounding code, documentation, or error logs.

Tools that Generate Diffs - Editors

Moving up one layer of abstraction, companies in this category will generate semantically meaningful changes to your codebase - “diffs” - by incorporating context while servicing your requests.

AI-powered editors are the most common manifestation of this approach. Designed for technical users, these tools index your entire codebase and reference it when handling requests like "add a new pricing page that has three tiers" or "here's the error, please help me fix this." Besides generating a series of targeted changes to your codebase after each request (the “diff”) which users can accept or reject, they may also proactively suggest changes via inline completions.

These tools have become the most common adoption point for AI in coding, and accelerate development compared to raw models by handling the context engineering automatically. Notable products in this space include:

• Cursor (Anysphere) – the most popular AI coding editor, built as a VS Code-compatible IDE with deep AI integration. The product features high-context chat, autocomplete, and an "Agent" mode that can autonomously refactor code and offer suggested changes as diffs. Anysphere raised $900 million in June 2025, bringing total funding to over $1 billion and valuing the company at $9.9 billion.

• Windsurf (Codeium) – another AI coding editor that works across multiple development environments. Similar to Cursor, their product centers around "Cascade," an agent that plans, edits, and debugs across entire repositories, offering up suggested changes as diffs. Codeium raised $150 million at a $1.25 billion valuation in August 2024, then agreed to a $3 billion acquisition by OpenAI in May 2025.

• Augment – an enterprise-focused AI coding platform that targets larger, security-conscious engineering teams. Similar to Cursor, they offer chat, inline completions, self-hosted deployment, and "Next Edit" functionality that propagates changes across large codebases. Raised $227 million Series B in April 2024, bringing total funding to $252 million at a $977 million valuation.

Tools that Generate Diffs - Agents

At a similar level of abstraction to editors, agentic software engineers also contribute semantically meaningful changes to your codebase via pull requests that require human approval.

The main difference between them and editors lies in the user interface and interaction model: rather than working directly within your IDE like AI-native editors, AI software engineers operate through software development workflows reminiscent of working with people. They receive tasks through platforms like Slack, Linear, or GitHub issues, then work autonomously to break down requirements, plan their approach, implement changes across multiple files, run tests, and submit pull requests for review. Additionally, they operate over multiple steps, debugging issues as they come up before submitting a change for approval. While they’re still meant to work with technical users, the human developer's role shifts from real-time collaboration to higher-level task assignment and code review.

This interface appeals to teams who prefer to treat AI as they would a remote engineer - assigning tickets, reviewing PRs, and maintaining existing development processes. While the output (diffs requiring approval) remains the same, the workflow is quite different from working with an AI-powered editor.

Some notable products in this category include:

• Devin (Cognition) – building an agent to automate software engineering tasks end-to-end. Their agent autonomously spins up in-browser sandboxes (complete with shell, browser, and IDE) and is capable of independently writing, testing, and shipping production code as a pull request. Cognition has raised over $200 million, hitting a valuation of $4 billion in March 2025.

• Factory – offers a browser-based command center orchestrating fleets of parallel “Droids,” AI agents that autonomously handle tasks spanning the entire software development lifecycle, including planning, coding, testing, and opening merge-ready PRs. The company has raised over $20 million in total, most recently at a $120 million valuation.

• Codex (OpenAI) – OpenAI’s effort in the space, another coding agent deeply integrated within ChatGPT and Azure. It tackles individual software tasks in isolated, secure containers pre-loaded with your repositories, enabling parallel execution of multiple tasks. Codex provides detailed execution logs and diffs to support transparency and reviewability.

• Jules (Google) – Google’s effort in the space, positioned as an asynchronous, “invisible helper.” Jules clones developer repositories into secure Google Cloud VMs, autonomously writes tests, fixes bugs, implements features, and generates detailed multi-step execution plans. Jules offers audio changelogs summarizing code changes and integrates seamlessly with GitHub workflows via issue labeling.

Tools that Generate Apps - Platforms

Moving up a final layer of abstraction, products in this category go beyond generating features or code to creating full, runnable applications from a conversational spec.

While diff-generating editors can grow into a full codebase if you keep feeding them prompts, things like picking the stack, designing the schema, and deploying the project remain the job of the developer. In addition to managing code generation, text-to-app platforms handle these tasks as well, allowing non-technical or lightly technical users to go from prompt (i.e. “I need a bakery inventory tracker with a tablet-friendly UI”) to a fully fledged application in just one step. This is the newest of the above categories, and one of the fastest growing ones.

Some notable products in this category include:

• Bolt (StackBlitz) – a text-to-app platform where you describe your app in plain English, and it automatically scaffolds a complete, deployable codebase. Bolt offers an integrated browser IDE powered by WebContainers, built-in deployment via Netlify, and backend integrations including Supabase and Stripe. Bolt’s parent company, StackBlitz, raised a $105.5 million Series B in January 2025.

• Lovable – a collaborative, AI-driven text-to-app builder designed for non-technical users. A single natural-language prompt creates fully hosted, ready-to-use web applications with autogenerated frontend UIs, databases, authentication, and hosting. Its browser workspace allows real-time collaboration, Figma imports, and customization through pre-built templates.

• Replit Agent – part of Replit’s integrated cloud IDE, Agent allows users to build apps end-to-end from a single text prompt. Agent autonomously plans, writes, and tests code, manages dependencies, provides real-time UI previews, and checkpoints incremental progress. Replit previously raised over $100 million, most recently at a valuation of $1.16 billion.

Crowded as the general space is already, there are a number of still-emerging categories as well. These may offer products targeting more specific engineering workflows and tasks, and some ones we’re paying attention to include:

• Terminal-based tools like Claude Code and Warp, which - similar to the editors and agents, help generate code and commands (i.e. “help me set up my environment, help me fix my broken git state, help me run this project”) but within a different interface.

• Mobile-first text-to-app platforms like VibeCode and Rork, which focus on producing fully runnable iOS/Android applications (i.e. a version of Wordle with 6 letters), often via a mobile app themselves.

• AI security engineers like XBow and Seven AI, which automate offensive/defensive security by continuously generating attacks (i.e. identifying broken access control, taking advantage of outdated components, running SQL/prompt injections) to discover and address vulnerabilities across both AI-native and traditional systems.

• AI production and site reliability engineers like Traversal and Resolve, which automate root cause detection and remediation for production incidents (i.e. application failures, network outages, database disruptions) which reduce downtime and on-call fatigue for engineers.

Overall, the editor and text-to-app categories have been growing the fastest, and Cursor and Lovable have led each, respectively. We previously had Cursor’s lead seed investor write a guest post about them here, and now we are excited to dive into Lovable.

‎

Lovable

So, Lovable is an end-to-end platform for nontechnical users to build and host full-stack applications.

Again, while it often gets lumped under the same “vibecoding” designation as other AI coding tools like Cursor, platforms like Lovable go beyond contextual code generation in a few important ways:

• Automated backend setup – when you need a database or other backend functionalities, Lovable automatically provisions auth, generates the SQL tables it thinks you need, and exposes ready-to-call endpoints without any manual schema work.

• Automated debugging – if the generated code displays buggy behavior, Lovable automatically feeds the stack-trace back into its AI, surfaces a richer message, and proposes (or applies) a patch.

• Automated deployment and hosting – when you’re ready to go live, Lovable can automatically build, host, and get a live URL (or custom domain) in one step, with the option to export the project to Netlify/Vercel if you prefer self-hosting.

Compared with other app-building platforms though, our belief is that Lovable has distinguished themselves by building the best platform to generate production-ready products. In some sense, this has introduced a new layer of abstraction.

What separates a basic application - hosted code that runs - from a production-ready product?

While an app must simply execute, a product is expected to look polished, stay secure, and meet certain integration or enterprise requirements. Its development might involve multiple parties, and post-launch it’s something that will need growth and marketing efforts.

Lovable’s platform is specifically designed to cover these broader requirements out-of-the-box. The product helps with all stages:

• Ideation - you need to up with the idea for the product. Lovable allows you to draw inspiration from a rich template gallery and a Remix feature, which lets you fork existing apps and refine them iteratively via chat.

• Visual Polish - you need to make the product look attractive to new users. Lovable builds its apps with an opinionated stack (React, Tailwind, ShadCN components) which makes them “beautiful by default”, and lets users tweak the details with a unique visual edit mode.

• Collaboration - great products emerge when design, marketing, and engineering collaborate rather than operating in silos. Lovable provides real-time, multi-cursor editing and built-in Figma import capabilities, enabling teams to iterate quickly, share feedback, and align on a single vision.

• Security - you need to make sure the product is secure, without common errors like a lack of row-level security and publicly available API keys. Lovable offers automated security scans to proactively catch potential vulnerabilities before deployment.

• Enterprise Readiness - gaining adoption in enterprises requires navigating strict compliance standards, often a roadblock for smaller teams. Lovable simplifies enterprise adoption by offering built-in integrations (like Clerk and Supabase) for SOC 2, HIPAA, and GDPR-compliant authentication, secure data storage, single sign-on, and encryption - eliminating future compliance bottlenecks.

• Growth and Marketing - you need to grow and market the product post-launch. Lovable embeds growth into the product lifecycle, automatically generating SEO metadata, sitemaps, and listing launched apps on its built-in discovery platform. It also guides users towards vetted growth agencies and affiliates with a dedicated marketplace.

Collectively, these features position Lovable as more than just a text-to-app platform - it’s a launchpad with everything needed to build, launch, and scale market-ready products from start to finish.

Lovable launched publicly in October of last year and has experienced rapid growth since. They crossed $100 million ARR in under a year, achieving week-over-week growth of nearly 50%. Lovable now generates more revenue each hour than it did in an entire month at launch, and its customer base spans large, established enterprises as well as indie hackers and solopreneurs.

The company was founded in November 2023 by Anton Osika and Fabian Hedin. Anton previously co-founded Depict.ai (raising approximately $20 million), was an early engineer at Sana Labs, and built the widely used open-source tool GPT Engineer. Fabian, the company's CTO, previously founded a property management software startup and worked with Anton at Depict.

On the strength of their exceptional early traction, Lovable raised a €6.8 million (~$7.5 million) seed round in October 2024 led by Hummingbird Ventures. Shortly after, they closed a $15 million pre-Series A led by Creandum. Then, a few months ago, they closed a $200 million Series A at a $1.8B valuation from Accel.

Lovable has a stellar team, a product people love, and some of the fastest growth any company has ever seen. Their ambition is to build the final software application - the app you can build all other apps on top of.

Can they pull it off?

It's definitely ambitious, but they've shown more promise than anyone else we've encountered. Even if they don't achieve their full vision, it’s clear they’re building something extremely valuable.

Let’s get into why.

The Market for Prompt-to-App Platforms

You probably don’t need convincing that the broader market around software development and consumption is enormous. At the end of 2024, there were roughly 28.7 million software developers globally, with 19.6 million categorized specifically as professional developers. Global software spending continues to grow steadily by more than 10% annually, and surpassed $1 trillion for the first time last year.

To us, the real opportunity, however, lies beyond serving professional developers.

How many people today aren't software developers but would significantly benefit from easily creating and monetizing software?

We know there’s substantial demand here because of the success of tools like Microsoft’s Power Platform. This suite of no-code/drag-and-drop tools lets users build basic applications (Power Apps), automate workflows (Power Automate), create websites (Power Pages), and analyze data (Power BI) without technical skills. Today, 7.5 million casual developers have built apps for themselves or their communities, serving 33 million total monthly end users - companies like Accenture, Rabobank, Westpac, Deutsche Bahn, and Zurich Insurance have collectively created thousands of apps and automated everything from customer support to asset management.

So, let’s take a direct approach to estimating the potential market size for text-to-app tools. How many people aren't professional developers today, but could meaningfully benefit from easily creating and selling software?

Quasi-Technical Professionals

These individuals typically have strong technical familiarity but lack formal coding skills or the time to fully build out software applications. They often have deep domain expertise, clearly understand user requirements, and regularly interact with technology teams. Text-to-app tools allow them to independently create software solutions, test prototypes quickly, and iterate efficiently, removing the usual bottlenecks and accelerating their workflow.

• Product Managers: Globally, there are about 2.4 million product managers. These people routinely define product specifications and success metrics, making them natural adopters of prompt-to-app platforms to rapidly prototype and launch initial app versions.

• UX/UI Designers: Estimates suggest there are between 1.5 and 3 million UX/UI designers globally. Designers regularly create detailed wireframes; vibecoding tools allow them to directly convert these into production-ready applications, speeding up development cycles significantly.

• IT Professionals: There are 9.4 million IT specialists in the EU and 9.9 million in the U.S. alone, with millions more globally. These roles frequently involve developing internal tools, dashboards, and automation scripts - exactly the kind of applications suited for rapid creation with AI.

• Scientists, Researchers, and Data Analysts: there are roughly 8.85 million researchers worldwide, and data-analysis roles are among the fastest-growing tech occupations globally. These professionals often prototype data apps and analytical pipelines, making them ideal candidates to leverage vibecoding tools to accelerate discovery and streamline workflows.

Net New Entrepreneurs and Small Businesses

These potential users typically already have distribution channels, domain expertise, or clear business needs. They understand their markets deeply, recognize unmet customer demands, or have specific operational problems they could solve with tailored software. Text-to-app platforms empower these individuals by drastically lowering the technical barrier, enabling them to rapidly prototype and launch custom digital solutions without needing extensive coding skills or significant upfront investment in software engineering.

• Small-business owners: There are approximately 400 million small businesses worldwide. Many lack the resources to hire professional developers, yet they frequently require software solutions to boost revenue, streamline operations, or enhance customer engagement. Prompt-to-app platforms can empower these businesses by enabling them to create tailored software without technical hurdles.

• Content Creators: The creator economy continues to grow, with 165 million new creators joining social platforms since 2020 and YouTube alone hosting 64 million active channels. Many of these creators already monetize digital products; vibecoding tools naturally extend their capability to create and sell niche-specific micro-apps or plugins.

These are all first-degree estimates of people who could directly find uses for and benefit from text-to-app tools today.

Though, when the costs of creating, hosting, and distributing any form of content drop dramatically - as we’ve seen with images on Instagram or videos on TikTok - innovation tends to follow in unpredictable ways. In this case, teachers might create personalized learning apps or classroom management tools. Local governments could quickly deploy apps for reporting civic issues and tracking responses. Nonprofits might build fundraising platforms or tools to coordinate community services.

In short, the market for text-to-app tools isn't just large - it could open up entirely new ways for hundreds of millions of people to solve problems and improve their daily lives.

Lovable’s Approach

So, out of all the text-to-app companies out there, why do we think Lovable stands out?

In our view, Lovable is the best solution for creating production-ready applications - real products (not toys) with the style, functionality, and polish that real users expect.

Lovable accomplishes this by innovating in three important areas: technology, product experience, and distribution.

Technology

Code generation with current models has become relatively straightforward. The real challenge is turning raw AI outputs into functioning, secure, and maintainable applications.

From a technology perspective, Lovable delivers here by being highly opinionated across the entire stack. This deliberate limitation helps the model produce focused, robust results:

• An opinionated backend. Lovable insists on using Supabase as the canonical data layer. Users can connect their databases in a few clicks by logging in and selecting their Supabase project. Because Supabase packages SQL, row-level security, real-time listeners, storage, and authentication behind a single Postgres endpoint, the agent always has a predictable interface to reason about. It can easily query schemas, infer relationships, and analyze logs to debug efficiently. Lovable has integrated Supabase more deeply than competitors, supporting advanced features like real-time functionality and edge functions (edge functions in particular aren’t deployable directly via API, meaning Supabase had to do more extra work - perhaps running the CLI server-side in temporary containers or VMs - to support this). While a "bring-your-own-database" approach might seem flexible, it would force the model to juggle different permission models and migration formats, complicating reasoning and multiplying the number of possible failure modes. By standardizing on Supabase, Lovable can treat migrations, secrets and logs as first-class primitives instead of unstructured text, and that tight coupling is exactly what lets the AI debug live issues rather than merely generate code snippets.

• An opinionated frontend. Lovable constrains the front-end to ShadCN/Radix components styled with Tailwind CSS. For example, when the AI generates a confirmation dialog, it simply imports the Radix AlertDialog wrapped in ShadCN’s styled Dialog, applying Tailwind classes like bg-primary text-white px-4 py-2 rounded-md shadow to buttons. This approach instantly creates an accessible, dark-mode-friendly UI compliant with WCAG standards. While this restriction might initially appear limiting compared to free-form CSS generation, it dramatically simplifies the decision space for the AI, eliminating common errors like inaccessible color schemes while ensuring visual consistency across regenerations. With a predefined palette of components, the model can focus on application logic rather than style, ensuring visual polish and consistency without manual intervention.

• An opinionated editing experience. Lovable isn't just prescriptive about generated code; it clearly defines how that code moves from prompt to production. Large-scale structural changes happen in AI Chat Mode, while detailed visual adjustments are managed in Visual Edit Mode. This clear division streamlines development, resulting in cleaner diffs and polished production-quality apps. Features typically start in AI Chat Mode, where users specify high-level goals, such as “Add a usage-based billing dashboard with monthly summaries.” Here, Lovable’s AI acts as a proactive partner and virtual CTO as opposed to a passive code printer - asking clarifying questions, exploring options, writing an implementation plan, editing files, compiling code, checking logs, and iteratively refining until it works. Instead of generating isolated snippets, Lovable ensures every loop either results in a working build or precise debugging guidance, avoiding cryptic errors. Once the basic feature is complete, the process moves to Visual Edit Mode, a unique Lovable offering. Here, designers (or developers) make pixel-perfect adjustments visually—moving buttons, changing text colors, or editing copy directly on the live app. This avoids the inefficiency of repeatedly prompting the model for minor changes, significantly reducing costs and unintended side-effects.

Lovable’s approach extends to other native integrations as well - GitHub for code hosting, Stripe for payments, and Netlify/Vercel for deployments with custom URLs. Additional recommended integrations include Resend for email, Clerk for authentication, Make for workflow automation, and Replicate for model hosting.

Together, these opinionated constraints - a fixed backend, mandatory agent loop, prescriptive design system, and targeted visual edits - combine into a tightly-integrated system optimized specifically for AI-driven development. While broader compatibility might sound appealing, it dilutes the AI's context and shifts complexity onto users.

Product

Building a successful product involves more than just creating an app—it typically requires careful navigation of each step from the initial idea to a polished, shipped product. Lovable clearly understands this distinction and supports users with thoughtfully designed workflows at every stage of the product lifecycle, making the overall journey smooth and efficient:

• Ideation with Remixing and Forking: Many great products start by building on existing ideas, iterating and experimenting quickly. Lovable’s Remix feature supports this workflow by allowing users to clone existing projects - both public projects and their own - into clean, fresh workspaces without carrying over database data. Because remixes retain stable IDs and Git history, teams can confidently test ideas, merge successful experiments, and integrate improvements, streamlining the path from initial idea to viable product.

• Sophisticated Debugging for Iterative Refinement: Real-world products often encounter issues, making effective debugging an important step in ensuring quality. Lovable treats debugging as a core workflow, clearly surfacing runtime errors and offering users an AI-driven “Try to Fix” option. The AI analyzes detailed Supabase, Vercel, and browser logs, identifies the issue, and proposes targeted fixes. If the initial fix isn't sufficient, users easily transition into Chat Mode, where the AI iteratively clarifies problems and generates minimal, Git-backed diffs for straightforward review and debugging. This structured approach helps efficiently refine the product while minimizing potential regressions.

• Visual Polish for Market Appeal: Visually appealing interfaces often help products gain adoption. Lovable’s unique Visual Edit mode provides a Webflow-like experience directly within the app, enabling precise, pixel-level adjustments. Each JSX element is linked to specific source code lines via unique and immutable JSX IDs, allowing designers and developers to quickly refine visuals without triggering full code regeneration. Hot-reloading offers immediate feedback, supporting quick and precise UI improvements that enhance overall user experience.

  

• Collaboration to Align Product Vision: Effective collaboration between design, engineering, and marketing teams frequently leads to better products. Lovable uniquely facilitates this collaboration with real-time, multi-cursor workspaces featuring integrated Git versioning and precise permission controls. For more complex tasks, Dev Mode offers an in-platform IDE experience, allowing experienced developers to manually edit alongside the AI. This unified environment helps all team members stay aligned, productive, and coordinated in delivering a cohesive product.

• Production Readiness for Enterprise Adoption: Gaining adoption in enterprises typically means meeting stringent compliance and security standards. Lovable proactively addresses these requirements through built-in AI-driven security scans before each deployment, checking for common issues like incorrectly configured Supabase RLS policies and hard-coded secrets. Additionally, Lovable’s login flow provides an SSO option, enabling larger organizations to integrate with existing identity providers. Comprehensive admin roles, automated secret management, and detailed audit trails further reduce risk, ensuring products created with Lovable are secure, compliant, and enterprise-ready.

• Seamless Code Export and No Lock-in: Supporting growth and scalability often involves keeping products flexible and transferable. Lovable ensures users never feel locked in by offering seamless GitHub integration that continuously syncs projects for easy deployment on platforms like Vercel or internal Kubernetes clusters. Teams can rapidly prototype within Lovable and confidently hand off their work to larger development teams, supporting long-term product evolution and maintainability.

• Built-in Distribution Support: Effective product distribution post-launch is important for long-term success. Lovable includes an integrated SEO engine that automatically generates metadata, XML sitemaps, and robots.txt files through simple prompts. Additionally, Lovable’s Partner marketplace helps non-technical founders easily access vetted growth agencies and professionals, facilitating efficient product marketing and expansion within the Lovable ecosystem.

Collectively, these targeted features and workflows form a structured, integrated framework that guides users effectively from initial concept to shipped and polished product.

Distribution

Finally, Lovable’s distinctive distribution approach directly aligns with its core philosophy: empowering users to create polished, externally-focused products rather than just basic applications. By specifically targeting aspiring entrepreneurs and teams building businesses or products for external users, Lovable leverages accelerators, hackathons, and other user-driven promotional tactics designed explicitly for this audience. This focused strategy has driven Lovable’s early market leadership, organically attracting users motivated to build successful software products:

• Structured Programming for Product Launches: Lovable’s flagship accelerator, "Shipped," explicitly guides participants from initial ideas to fully launched, market-ready products and opened its first season in June of this year. Participants benefit from $3 million in cloud and tooling perks, a structured weekly schedule of lectures and check-ins, and an end-of-program demo day in San Francisco. Accepted participants join a dedicated Discord cohort channel, with progress updates shared publicly - transforming each in-flight project into effective marketing for future cohorts.

• Hackathons Showcasing Product Potential: Lovable regularly organizes focused hackathons tailored for entrepreneurs building real, market-ready products. For instance, the June 2025 AI Showdown offered full platform access free for forty-eight hours, complemented by a $65,000 prize pool across distinct tracks for OpenAI, Anthropic, and Google Gemini integrations. The event's announcement in We Are Founders detailed the prize structure and emphasized the model-versus-model competition theme, followed by a subsequent article highlighting the three winning apps and confirming prize distributions. Earlier that month, Lovable also co-hosted a USA vs. Canada hackathon, awarding a $10,000 prize and a slot at Lovable’s in-person demo day.

• Educational Content Supporting Product Builders: Lovable consistently provides practical educational content tailored to aspiring entrepreneurs, including live streams, AMAs, and workshops such as "Build and Monetize Your App in 45 Minutes" or "Landing Page Challenges." These resources clearly illustrate the journey from concept to monetizable product, appealing directly to users focused on practical product development rather than abstract demonstrations.

  

• UGC and Collaborative Discovery: Lovable's most effective promotion comes from showcasing real projects built by its users. "Launched," a Product-Hunt-style platform, prominently displays recently deployed products, driving discovery through community-driven upvotes. Lovable’s documentation explicitly positions "Launched" as a major channel for traffic generation, and the company amplifies featured products via their own social media accounts for additional visibility. Additionally, deployments feature clear branding with "Built with Lovable" default link preview images. Integrated multi-user editing further encourages team collaboration, expanding Lovable’s adoption internally as teammates introduce it to each other.

• Targeted SEO for Product Discovery: Lovable currently offers over 130 highly-specific landing pages optimized around precise use cases such as Customer Feedback Tools, CRM systems, Interactive Storytelling Apps, and other solutions entrepreneurs might seek to build and distribute. This targeted approach captures users as they search for something to build an idea in their head.

  

Lovable’s founder and CEO, Anton, even reinforces this ethos by angel investing in businesses built using Lovable.

Competitive Landscape

Now that we have a deeper understanding of Lovable’s strengths as a product and company, let’s take a deeper look at how they compare against other products with text-to-app functionality.

Last Generation Low/No-Code

Each of these products offers low/no-code tools for building full-stack applications, but they emerged before generative AI became mainstream, so their foundational approaches are fundamentally different. Broadly, these earlier-generation tools rely heavily on proprietary ecosystems, closed runtimes, and restrictive design paradigms that limit flexibility, scalability, and portability.

In practice, this means that while they're great for quickly prototyping or visually assembling basic applications, they fall short once your app grows more complex or requires deeper integration with third-party services. This lock-in often leads to issues down the line, such as difficulty exporting code, integrating new features, or scaling effectively to handle increased traffic or complexity. Additionally, expanding beyond built-in capabilities usually requires cumbersome hacks or expensive workarounds.

In contrast, the new wave of generative-AI-driven tools like Lovable start from a fundamentally different place. Instead of restricting users to proprietary environments, they generate clean, fully portable, and standards-compliant code from day one. This makes it straightforward to add custom business logic, integrate external tools, deploy on any platform, and scale without hitting vendor-imposed limitations.

Taking some of the larger players in this category one-by-one:

• Bubble offers a drag-and-drop visual app-building experience built around a proprietary runtime and database. It’s great for quick prototypes, but severely restricts flexibility due to its closed ecosystem. It has no straightforward way to export code or integrate third-party tools, leading to lock-in and scalability issues once apps grow beyond basic hobbyist use. In contrast, Lovable generates clean, standards-based code from the start, enabling easy export, integration, and scaling within the broader ecosystem.

• Webflow shines at creating visually polished front-end websites and interfaces with the most sophisticated visual editor available. However, it lacks important full-stack features like authentication, backend logic, and relational databases, making the development of full-stack applications difficult and costly with external integrations. In contrast, Lovable directly generates complete, deployable applications - including backend logic, authentication, and APIs - from simple text prompts, combining ease of use and robust functionality.

• Airtable relaunched in June 2025 as an “AI-native app platform.” Their new agent, Omni, builds apps and automations from prompts but remains constrained within Airtable’s closed runtime. There's no direct code export, and extending functionality beyond Airtable’s built-in capabilities involves custom scripting. In addition, Airtable primarily targets internal, B2B use cases at large established companies, with a focus on optimizing existing workflows. In contrast, Lovable empowers builders to create entirely new, externally-facing products with fully open-source and exportable code, providing complete flexibility in integration, customization, and deployment without vendor lock-in.

• Retool pivoted towards generative development in April 2025 with an AI App Generation preview, enabling quick assembly of internal applications using Retool’s predefined component blocks and integrated development environment. However, Retool specifically caters to internal use cases within larger enterprises, emphasizing B2B solutions aimed at improving internal processes (i.e. building dashboards) as opposed to focusing on people who want to launch completely new and often externally-facing products like Lovable. The two products have different use cases and end users.

Next Generation No-Code

Like Lovable, these products launched after generative AI took off, generating apps end-to-end directly from simple prompts. Unlike traditional no-code platforms relying on custom ecosystems or proprietary frameworks, these tools often produce real, usable code built on open-source foundations. But compared to Lovable, they generally don't handle the complete product lifecycle as smoothly - whether it’s backend integration, collaboration with teammates, or growing beyond quick prototypes.

Taking some of the larger players in this category one-by-one:

• Bolt: Bolt, built by the StackBlitz team, is the closest competitor to Lovable, offering full-stack app generation directly in the browser using StackBlitz’s WebContainers. Technically, this is quite impressive - WebContainers allow you to spin up a full Node.js runtime locally without servers, offering instant reloads and offline development. However, Bolt lacks Lovable's features geared towards makers and entrepreneurs. It doesn't have a visual editor, remixing capabilities, SEO optimization tools, or strong community-building features. As a result, despite Bolt's early advantage (it was about 10x Lovable’s size revenue-wise at the end of last year), Lovable has rapidly caught up and surpassed Bolt by offering a more comprehensive and growth-friendly experience.

• Vercel v0: Vercel v0 is designed to rapidly generate frontend React components and polished UI elements using generative AI, deeply integrated within the Vercel hosting ecosystem. Yet, it's strictly frontend-oriented—there’s no built-in support for databases, authentication, or APIs. It also lacks the visual editing and remixing features provided by Lovable. Given Vercel’s core business revolves around hosting rather than end-to-end app generation, the limitation is somewhat understandable. However, users needing more than frontend scaffolding will quickly find themselves handling manual backend integrations.

• Replit Agent: Replit Agent uses AI to automatically generate and deploy apps on Replit’s custom cloud environment—essentially a smart coding assistant that handles coding, debugging, and deployment from a simple description. Besides using a proprietary ecosystem and being less flexible with exports like other last-generation tools, Replit’s interface remains more code-centric than other tools in this category due to its origins as a collaborative IDE for technical users (similar to Google Docs for developers). While this makes it great for developers comfortable with terminals, IDEs, and traditional software engineering workflows, non-technical users and designers might prefer Lovable.

  

• Canva Code: Canva, primarily known as a user-friendly design tool, recently introduced Canva Code. This feature enables users to effortlessly embed interactive HTML/CSS/JS snippets directly into Canva designs, perfect for quick landing pages or interactive marketing campaigns. However, Canva Code remains limited to frontend features; backend integrations (such as authentication or databases) must be implemented manually by advanced users. Additionally, it doesn’t allow exporting complete projects—only individual snippets of code—which requires manual integration outside the Canva ecosystem.

• Figma Make: Figma, another widely-used design tool, introduced Figma Make to turn visual design frames directly into interactive web apps, allowing designers significant frontend control. While excellent for design-centric workflows, Figma Make currently lacks backend support—no built-in databases, authentication solutions, or options for exporting to Git or other hosting services. As a result, Figma Make excels in prototyping and visual demos but falls short for comprehensive, production-ready full-stack app development.

Team

Lovable wants to build the last piece of software - the app you can build all other apps on top of.

It’s an ambitious vision, but they’ve put together an excellent team to make it happen. Consider a sampling of their roster:

Anton Osika — Co-Founder, CEO

Anton was previously the Co-Founder and CTO of Depict.ai, a YC-backed startup that raised $20 million. Before that, he was a founding engineer at Sana. He holds a MS in Engineering Physics and Applied Mathematics from the KTH Royal Institute of Technology.

Fabian Hedin — Co-Founder, CTO

Fabian was previously Frontend Lead at Depict. Before that, he was Co-founder and CTO of TenFAST, a property management startup. He holds a BS in Industrial Engineering and Management from the KTH Royal Institute of Technology.

Elena Verna - Growth

Elena was previously Interim Head of Growth at Amplitude. Before that, she was Interim CMO at Miro and SVP, Consumer Product at Malwarebytes. Before that, she was SVP of Growth at SurveyMonkey. She holds a BS in Statistics from UC Berkeley.

Tiger Abrodi - Member of Technical Staff

Tiger was previously a Senior Product Engineer at ENLYZE. Before that, he was a Senior Software Engineer at Bobsled.

Marten Wiman - Member of Technical Staff

Marten was previously Co-Founder of Optimeringsfabriken, a consulting company. Before that, he was a Software Engineer at Google and Veridict. He holds a MS in Engineering, Computer Science, and Physics from the KTH Royal Institute of Technology.

Tomas Nordström - Member of Technical Staff

Tomas was previously Co-Founder and CTO of PowerMate. Before that, he was a Data Scientist at Nrlyze AB. He holds an ME in Computational and Applied Mathematics from the KTH Royal Institute of Technology.

Emil Ahlbäck - Member of Technical Staff

Emil was previously a Founding Engineer at Godmode. Before that, he was a Frontend Engineer at Depict. He holds an MS in Computer Science from the KTH Royal Institute of Technology.

Amir Salim - Member of Technical Staff

Amir was previously Founder of Education Geeks, a bootstrapped edtech startup he bootstrapped to $5m in revenue.

Viktor Eriksson - Member of Technical Staff

Viktor was previously an Engineering Manager at Stripe. Before that, he was a Software Engineer at Pair Team. He holds a MS in Engineering, Industrial Engineering, and Management from the KTH Royal Institute of Technology.

Alexander Wikström - Member of Technical Staff

Alexander was previously the Founder of Darwin, an AI sales agent. Before that, he was the Head of Business Development and GTM for Forloop. He holds a BS in Business and Economics from the Stockholm School of Economics.

Nad Chishtie - Product Designer

Nad was previously VP of Design at Element. Before that, he was Lead Designer at CMUNE.

Mindaugas Petrutis - Growth

Mindaugas was previously Director of Technical Programs at On Deck. Before that, he was Head of EMEA Community at InVision, and the Founder of Sketch Labs.

Felix Haas - Product Designer

Felix was previously Designer in Residence at Creandum. Before that, he was Co-Founder of Blossom Design, a design studio, and the first designer at Mapify Travel.

Caroline Wong - Business and Operations

Caroline was previously Co-Founder of ZOPA 254, a startup building AI solutions for African retailers. Before that, she was a consultant at BCG. She holds an MS in International Political Economy from LSE.

Lovable has assembled a strong team with some of the most ambitious talent across Europe, including many former founders and hustler-solopreneurs-types - exactly the audience of builders Lovable now serves.

Conclusion

Building software products has always come with some set of challenges.

These challenges, though, have come down with time: punchcards became bits, low level programming languages became high level ones, and our infrastructure learned to scale itself.

Lovable wants to be the final and most significant step on this staircase: the app which can build all other apps, to make software development truly effortless.

The future of software might finally be simple.

They’re hiring.

‎

In case you missed our previous releases, check them out here:

• Clay - Bringing creativity back to GTM.

• Etched - The chip for transformers.

• Meter - The next modern utility.

And to make sure you don’t miss any future ones, be sure to subscribe here:

Finally, if you’re a founder, employee, or investor with a company you think we should cover please reach out to uhanif@stanford.edu - we’d love to hear about it :)

• Clay has grown revenue from 5 to >30m in the past year, has taken less than 18 months to reach an eight figure run rate, and in the 9 months since we’ve started writing this has announced two new funding rounds. They’re currently hiring product and systems focused software engineers in New York City.

• Modal has grown revenue significantly in the past year and has taken less than 18 months to reach an eight figure run rate. They’re currently hiring product and systems focused software engineers in New York City.*

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information from partnered venture firms. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should Join Clay

(Click the link to read online).

Cold email.

The last great art.

Some say literature peaked with Joyce. Storytelling with Woolf, and satire with Vonnegut. Harold Bloom, railing against the “school of resentment” that to him defined modern literature, once quipped we were witnessing the “final phase" of literary criticism. Lamenting the state of modern writing, critic Joseph Bottum mourned that “The novel didn't fail us. We failed the novel.” In A Reader’s Manifesto, B.R. Myers bemoaned “the growing pretentiousness of American literary prose” on behalf of readers everywhere.

Heartbreaking.

But Bloom, Bottum, and Myers must not have been spending time in their spam folders, where the great authors of our time have been doing their real work.

It takes vision to write the right cold email.

How many emojis should you use? Is using two exclamation marks in the same paragraph okay? How do you optimize the first 40 characters (the ones visible before truncation on mobile gmail), and how soon should you follow up? On subject lines - should you go lowercase for casual vibes, caps for urgency, or try starting with a “Re:” so they think it’s a response?

And on the endless amount of personalization you can do - how many LinkedIn posts or talks should you reference? Is mentioning the dog in their profile picture too much? Should you congratulate them on that work anniversary from 3 weeks ago?

If not an art, there’s at least a craft to developing the right outreach strategy - to finding the right people, writing the right message, and sending it at the right time. Getting creative with who, how, and when you hit prospects is the métier millions of of SDRs, recruiters, and call center employees junior VCs spend the first decade of their careers refining.

At its worst, this can be a rather soul-sucking task, the corporate equivalent of hazing. At its best, however, much of this work - outbound pipeline generation as the industry calls it - can be creative and rewarding. Many of us can probably remember the best cold call or email we’ve ever gotten, and the conversation it led to after.

Clay is promising to bring creativity and craft back to go-to-market, without the sacrifice of scale.

In the process of doing so, they’ve become one of the fastest growing companies in the world, and a promising one to consider joining.

Background: The Landscape of GTM Tools

Before getting too deep into Clay, it’d probably be helpful to understand the broader landscape of sales tooling they fit into.

Sales teams use many different types of tools. There are prospecting/pipeline-generating tools they use to help generate leads. There are CRMs that they use to help track and manage leads. There are sales enablement tools to help train sellers, and there’s an emerging category of sales agents to augment or even replace sellers entirely.

There are tools to help with every stage of the sales funnel.

In this section, we want to dive deeper into the world of prospecting tools specifically, as that’s the function most companies initially start using Clay for. To start, we found it helpful to break the process - usually done by Sales Development Representatives (SDRs) and involving everything up until the first meeting - into three stages.

Finding Leads

Prospecting usually begins by finding some list of people (prospects) or companies (accounts) to reach out to. These lists can come from lots of different sources - they might be subscribers to the company newsletter, members of a Slack community, developers who have starred a Github repository, or, most commonly, a filtered set of targets drawn from LinkedIn or some other data provider. Alongside LinkedIn, some popular vendors of prospects and their contact information include:

• Apollo - a database of 275M+ contacts and 73M+ companies.

• RocketReach - a database of 700M+ contacts and 60M+ companies.

• ZoomInfo - a database of 321M+ contacts and 104M+ companies.

• Lusha - a database of a 150M+ contacts and 50M+ companies.

As a sales rep with a territory and customer profile to target, you would usually begin by pulling a list of all the prospects and accounts that fit your criteria. For instance, if you were selling AI-enabled customer support software to midsized B2B tech companies in the Bay Area, you might begin by using Apollo to pull a list of companies who are in San Francisco, have between 50 and 500 employees, and are in the Computer Software industry. Then, you would pull the name and email of each company’s corresponding customer support leader.

Researching Leads

Once you have a basic list of prospects, you often use additional information around each company and person to filter it down further. Some more advanced data points to incorporate here might include where the person sits in the company’s org chart, what other tools the company uses today, how the company is doing in terms of revenue or web traffic, or how many job listings the company has posted.

Gathering these data points - a process called enrichment - can be done by aggregating information from specialized vendors. For instance:

• Crunchbase is a source on a target account’s size. It has information on a company’s investors, funding amount, and funding stage.

• Crystal is a source on source on buyer personalities. They’ve scraped public data to create personality profiles for anyone with a social media presence.

• BuiltWith is a source on company tech stacks. They tell you what tools and technologies a company is using, like if they’re on Zendesk, Shopify, BigComerce, or Klaviyo.

• Semrush is a source on company web traffic. They provide estimates of monthly traffic for specified domains, as well as guesses around who a company’s competitors are.

To continue our above example, an SDR selling customer support software to midsized B2B tech companies might further filter their list down by targeting accounts who are both using legacy support software and experiencing inflecting growth. To do so, they might filter their list down to companies who are current customers of Zendesk or Freshdesk and who have also recently raised a round of funding or seen significant headcount growth.

This is also where you would spend time manually researching prospects, which might include looking at the account’s website, listening to a recent interview with the CEO, or skimming through the company’s blog. For instance, if your prospect has a LinkedIn post advertising a new support opening on their team, you could eventually reference that in your outreach and explain how your product helps reduce the number of people they need to hire.

Outbounding Leads

Finally, once you have a filtered list of contacts and the information needed to personalize things, the final step is to actually reach out. This means writing and sending emails, making cold calls, and following up until you receive a response.

Copywriting and sequencing are the main aspects of this stage, which mean writing the email and defining the outreach/follow up strategy. Tools for these tasks include:

• Copy.ai - a tool for AI-powered copywriting. They help generate marketing copy, ad headlines, and social media content.

• Jasper - a tool for AI-driven content creation. They specialize in long-form content, brand voice consistency, and marketing automation.

• Mailgun - a tool to help deliver emails. They provide APIs for sending, receiving, and tracking emails with high reliability.

• Lavender - a source for AI-assisted email personalization. They optimize cold emails with real-time coaching, personalization, and data insights.

To continue our above example of selling support software to midsized tech companies who are both growing quickly and on a legacy platform, you might try referencing their rapid growth and asking if their current system - Zendesk or Freshdesk - is doing enough for them. If their head of customer service is trying to hire several people, you might explain how your product can help reduce headcount needs.

Optimizing, coordinating, and automating these three basic functions - finding people to contact, researching them, and reaching out - is what most of modern sales tooling is built around.

As the space has matured, many of the companies we mentioned (i.e. Apollo, ZoomInfo) have expanded into supporting more than one of these functions, which is why the market today can feel somewhat soupy. Many of the newer “AI x Sales” or “AI-native GTM” companies who support top-of-funnel operations have bundled or abstracted away a number of these functions as well, which hasn’t helped. Some emerging categories of products in this space you might’ve heard of include:

• AI SDRs. Companies like 11x, Artisan, and AiSDR are building AI sales representatives with the goal of automating full, end-to-end pipeline generation - their agents they will pull the list, research the prospects, and reach out for you.

• Inbound pipeline generation. Unify, Koala, Warmly, Common Room leverage first-party data (i.e. who’s visiting which page of your website when) to help exploit warm leads. They still help drive top-of-funnel, but focus more on inbound than outbound and are so are their own category.

• Intelligent CRMs. Rox, Day, Folk, and other AI CRMs help users monitor and research prospects in the background, while tools like Actively bring intelligence to existing CRMs by helping users prioritize existing accounts.

While it can be challenging to make sense of how everything fits together, broadly speaking, each of these companies has approached by the market by either trying to own either a key data source (i.e. third-party data like Apollo/Crunchbase or first-party data like Unify/6Sense) or a key workflow (i.e. writing emails like Copy.ai, researching leads like Rox, or the entire pipeline-generation process like 11x). That is, they define themselves by having either some advantage on data or some advantage in automation.

To us, what makes Clay unique is how while they’re world-class at helping with each of the functions above, they’ve approached things by fundamentally owning no data source or workflow.

Clay’s Approach

Like the companies mentioned above, Clay supports businesses with workflows related to outbound pipeline generation - companies usually start using them to surface, research, and contact relevant leads for their business.

Unlike the above products, however, which each aim to own some data source or workflow, Clay is more of an integration and orchestration layer on top of a bunch of other data sources (i.e. Apollo, Crunchbase, 6Sense, Salesforce) and automation primitives (i.e. ChatGPT, Zapier, Salesforce, Sendoso). People use Clay as a no-code tool to stitch these data sources and tools together into longer workflows, allowing Clay to support a broader and more personalized collection of automations than other tools. As we will explore, it’s this approach that’s made them so uniquely powerful in an increasingly crowded market.

Concretely, Clay's product takes the form of a table. In each Clay table, each row represents an account or prospect, and each column represents either a data point on that prospect (i.e. their name, website, or LinkedIn profile) or some transformation on a prior column. For example, if the first column is the account’s name, the second column might use Apollo to pull the company’s website based on the name, and the third column might use Crunchbase to pull the company’s funding based on the website.

Going back to our example of a company selling customer support software, a basic Clay workflow might involve automating the core find, research, and outbound loop:

1. Using Clay’s company and people database, you could assemble a list of leads by pulling every software company located in the Bay Area with over 50 people and a head of customer support.

2. Next, by creating a few columns and using the relevant integration for each, you can enrich each lead with data points around growth rate, capital raised, the size of their support team, the tools they’re currently using, and relevant contact information. As an orchestration layer, Clay centralizes all of these disparate data sources and even has a waterfall system where if one source doesn’t have the information you need, a number of backup sources are automatically queried.

   

3. Next, you can research each lead using the Claygent tool - a web research agent that can answer custom questions about each prospect. For instance: Who are their customers? What's their product’s value proposition? What's a typical support question they might encounter? You can create a Claygent column for each of these questions and generate answers.

   

4. Finally, you can draft an email using the OpenAI tool, which incorporates the data you’ve pulled in prior columns along with any templates you might want to use. You can then use one of the sequencing tools Clay is integrated with to send the email.

   

The waterfall enrichment scheme and web research agent were two early innovations that made Clay great for this basic use case, which ended up becoming a wedge for the product. They also demonstrate the fine balance Clay has struck - being verticalized enough to excel at bread-and-butter GTM workflows, yet flexible enough to enable completely new ones. This second point is where Clay’s strength and staying power has come from: what users find charming about the product is how creative you can get with it.

• Getting creative. Clay's real power comes from chaining its columns and transformations in new ways, which is how users can create new GTM and GTM-adjacent workflows. With the right combination of columns, you can automate capturing a screenshot of a prospect’s webpage, scoring a prospect’s Typeform submission and automatically personalizing a follow-up, generating transcripts from a prospect’s YouTube videos, and generating content from a customer’s support conversations.

• Staying cohesive. Despite this diversity, what keeps things cohesive is the simple table format underlying everything. Beyond naturally enabling massive parallelism, this commonality makes it easy for people to share and understand different workflows, as they all take the same form. Each workflow and campaign feels consistent in its touch, and this keeps Clay usable as a product.

The long tail of workflows and automations you can build only with Clay is what makes the product so sticky and powerful.

Clay launched publicly in early 2022 and has has grown swiftly ever since. The company went from 0 to multiple 8 figures of revenue in just over two years, and today serves many of the most notable companies in tech: customers today include Ramp, Anthropic, Square, Dropbox, Notion, Verkada, Brex, and DBT. Users like the product:

The company was founded in 2017 by Kareem Amin and Nicolae Rusan. Kareem was previously VP of Product at the Wall Street Journal, a cofounder of Frame (acquired by Sailthru), and a Program Manager at Microsoft. Nicolae was previously VP of Product at Dow Jones, Kareem’s cofounder at Frame, and a Program Manager at Microsoft. In 2021, they were joined by Varun Anand, a third cofounder who was previously Director of Operations at Newfront, Head of Expansion at Candid, and the Special Projects & Communications team at Google. Clay originally spent 4 years building a more general-purpose no-code tool and initially launched with that. After getting no traction and spending some time wandering through the desert, they pivoted to focusing on outbound sales and revops workflows specifically. PMF quickly followed, and they were then off to the races.

Throughout all this, Clay has raised a $2.5M seed round led by First Round capital, a $13.5M Series A led by Sequoia Capital, and a $46M Series B led by Meritech Capital. They recently became a unicorn and raised a further $40M Series B expansion, also led by Meritech.

Clay has had an incredible trajectory over the past two years, but we’d like to go a bit deeper when examining whether they might be worth joining as a company.

There are two points in particular that matter when answering this question:

• Is Clay operating in a great market?

• Is Clay well-positioned within that market?

The Market for Go-to-Market Tools

Clay makes money when companies use them for different GTM-related workflows and automations.

You probably don’t need convincing that GTM teams buy a lot of software - as a revenue generating function, their tooling needs are usually a top priority. Additionally, since sales and marketing teams typically have substantial budgets (around 38% of revenue for software companies), any efficiency boost or cost-saving they can find is taken seriously. There’s a reason companies spend $107.5 billion globally on CRMs, and why US businesses spend over $800 billion a year on sales force compensation.

Even looking solely at the market for prospecting tools, the market opportunity is impressive:

• A lot of companies prospect. The US alone has roughly 5.7 million sales professionals, of which nearly half (47.2%) primarily work remotely as inside sales reps—meaning they're looking for prospects online.

• Prospecting takes a lot of time. Over 40% of salespeople cite prospecting as their biggest challenge, ahead of even closing (36%) and qualifying (22%) leads. Just drafting emails consumes an average of 21% of their workday.

If Clay can meaningfully improve the efficiency or effectiveness of these prospecting efforts—by improving personalization (personalized emails see 2.5x higher response rates), saving time, or enabling greater scale — companies will gladly pay for that.

But what's exciting about Clay is that it's more than just another prospecting tool—it's a no-code workflow builder deeply integrated with AI. This positions Clay right at the intersection of some compelling trends:

• The adoption of no-code and low-code tools. The market for no and low-code tools is expected to reach $50 billion by 2028, with some 75% of businesses projected to adopt some combination of them. Already, nearly 60% of new applications are being built outside of IT, often by employees with limited or no technical skills. Companies embracing these tools report significant improvements in speed, with 29% of organizations seeing their app delivery timelines cut in half. The rise of companies like Bolt and Lovable, which each reached $10 million in revenue in under two months, show the sizable demand for these sorts of tools.

• The growing importance of personalization in sales. 79% of B2B marketers agree that personalized content is the best way to engage buyers, with companies using these strategies seeing revenue growth jump by 10-15%. With AI now powering personalized outreach, this is becoming standard practice — 92% of businesses already use AI-driven personalization, making advanced personalization tablestakes to stay competitive.

• The rise of Generative AI. Gartner projects that by 2028, 60% of B2B sales tasks will be executed generative AI—up from less than 5% today. Already, 65% of businesses use generative AI in at least one function, primarily marketing and sales. Teams leveraging AI in sales overwhelmingly report better outcomes: 83% experienced revenue growth and 68% saw headcount growth, compared to only 66% and 47%, respectively, among teams not using AI.

  

Perhaps most importantly, though, startups selling into revenue-focused teams have an especially favorable market dynamic:

• The ROI is straightforward to quantify: Are you generating more leads at a lower cost? Clear ROI helps internal champions justify spending and close deals faster. This contrasts with selling something like HR analytics software, where the ROI can be less obvious or take longer to measure.

• Budgets for revenue-generating teams are usually protected during downturns, because these teams directly impact the company's bottom line. In comparison, something like employee engagement or culture software might be first to go, because their impact is harder to directly tie to short-term financial outcomes.

• Revenue teams tend to be open to experimenting (you always need more leads), allowing startups to scale quickly without hitting common roadblocks like the slow-moving procurement cycles in government or highly regulated industries.

All of this adds up to shorter sales cycles, larger potential contracts, and opportunities for rapid revenue growth. In other words, Clay’s market is structurally more attractive than most.

Clay’s Positioning

So, of all the intelligent pipeline-generation tools out there, why do we (and so many others) like Clay’s the best? How have they become a "default" tool that seems to show up in nearly every startup’s toolkit — alongside the likes of Notion, Slack, or Ramp?

We'll analyze this question from a few angles. First, we'll start at the product level to understand what specifically makes Clay unique. Next, we'll move up a level to explore the principles guiding their approach to building the product. Finally, we'll zoom out further to examine how Clay has approached its own go-to-market strategy and market positioning.

Product

At the most basic level, Clay stands out because of a handful of novel features. Consider the typical prospecting workflow we discussed earlier — Clay was the first to market with meaningful improvements at each step:

• Enrichment: via "Waterfall Mode"

• Research: via "Claygent"

• Copywriting: via AI-driven prompts

Individually, each of these features represented a specific upgrade over previous methods. Waterfall mode, for instance, saved users money because they no longer had to subscribe individually to multiple data services — Clay gave them just what they needed, à la carte. Together, these targeted improvements gave Clay its initial wedge into prospecting, which drove much of its early adoption and growth.

If you speak with users, though, Clay’s true uniqueness and magic emerge when you start chaining these and other individual capabilities together. Additional steps you might leverage to create workflows include cleaning or modifying data, running it through custom JavaScript, cross-checking and syncing with your CRM, scraping websites, performing automated Google searches, hitting arbitrary APIs, or executing custom AI prompts. Some sophisticated workflows we’ve seen people build with Clay:

• You can set up a webhook to automatically import meeting data—dates, attendees, emails—directly into a Clay table, then use Clay’s HTTP API column to trigger a Zapier integration that formats notes and syncs them directly into Notion.

• You can paste a LinkedIn profile link into a Clay table, instantly creating a profile with the user’s recent posts, personality insights, business type, and decision-making authority. Clay can then send you a Slack message with a neatly formatted Google Doc containing detailed company research—recent news, acquisitions, product launches, customer case studies—all automatically organized for you.

  

• You can create a table of LinkedIn influencers and have Clay automatically enrich it with follower counts, recent post topics, and average engagement metrics (likes, comments). You can set this up so that whenever you add a new influencer, their recent posts automatically populate another Clay table, ready to export to Google Sheets for deeper analysis.

This ability to chain basic actions together unlocks entirely new automations that simply weren’t possible before — exactly what a robust no-code or low-code platform should provide. This makes sense once you understand Clay’s origins: the product was originally billed as a general-purpose, no-code spreadsheet tool — “a spreadsheet that can fill itself.”

Over time, Clay layered deeper specialization around go-to-market workflows, moving away from recruiting, accounting, and other general use cases. Today, it's evolved into a purpose-built, no-code platform tailored specifically for that vertical. Its tabular format gives users the granular control, visibility, and massive parallelism required for outbound-related workflows at scale.

While many SaaS tools handle the basics of prospecting, very few serve the "long tail" of niche tasks that salespeople frequently want to automate — things like taking automated screenshots of prospect websites, scraping founders’ LinkedIn posts, or moving data into and out of Salesforce/Hubspot. Clay addresses these previously unmet needs in a polished and flexible way.

Fundamentally, its popularity comes down to the fact that Clay enables users to build workflows that they just simply couldn’t before.

Branding and Distribution

The novelty to Clay’s product explains part of their early adoption, but there's more to their momentum. So, what’s set them on a path to becoming one of those "default" tools every startup has heard of and tried — like Notion, Rippling, Ramp, or Slack?

It’s actually kind of an intriguing question, and we spent a good amount of time thinking about it. Besides building a strong and differentiated product, we’ve noticed a few attributes common among startups that seem to achieve that rare “it” factor:

• Craftsmanship: Builders, not bureaucrats, typically run startups. These people deeply appreciate well-crafted, thoughtfully designed products. Companies like Figma, Linear, and Stripe have all benefited from this mindset.

• Creativity, Customizability, and Personalization: Startups are full of people who love experimentation. Growth hackers, product engineers, and “founding” employees of any sort like modular but “compound” tools they can adapt and personalize to their needs — think Airtable, Notion, or Rippling.

• Ecosystem and Community: Startup employees are constantly talking, sharing, and recommending tools to each other. If there are specific influencers dedicated to sharing tips about your product, you are probably in a good spot. Companies that do a great job cultivating and leveraging communities tend to gain momentum quickly — i.e. Notion, Vercel, and Cursor.

Clay shows up well along each of these axes. It’s a thoughtfully designed, highly customizable product with a thriving community of power users who frequently share tips, hacks, and creative workflows with one another. Clay intentionally fostered this sense of community early on through an unconventional approach: they channeled all customer support exclusively through a single Slack community, eliminating traditional email and chat support. The community started with just 200 people, and has now grown to over 11,000. In addition, the company has actively nurtured a small ecosystem of "claygencies," specialized consulting firms who help new users adopt and implement Clay, by showcasing these partners on their social media channels and website. Collectively, these moves have positioned Clay well among brand-name customers known for being tastemakers in enterprise software.

This is important, because there’s also often some degree of path-dependence at play when becoming a “hot” company. Not all revenue is created equal - given the choice, it’s far better to have your revenue coming from brand-name companies than unknown ones or a random collection of SMBs. This is because future customers will often look to the vendor well-known companies use to inform their own decisions around what’s the industry standard. A big part of the enterprise sales cycle is your champion selling your tool internally, and it’s a lot easier for them to do this when they can point to a leading company in their vertical (with a more sophisticated procurement process) using it also. If you’re selling to other startups, like Clay has been doing, having customers like Ramp, Verkada, Notion, Brex, and Anthropic are usually the early signal future customers will look for.

Clay’s early customer list put them on this trajectory, and this has given them a leg up.

The Long Term

In the limit, what does something like Clay become?

Varun described their goal as becoming "the final growth tool" — something used if you're involved in sales, marketing, or anything growth-related.

There’s a reasonable roadmap to achieving this: Clay can always expand its scope by integrating more tools and data sources. Our main question revolves around the table UI they’ve committed to — at what point does it eventually become limiting? The current interface excels at parallel tasks but might struggle with workflows that run in the background or that rely on more complex triggers. Perhaps, longer-term, Clay could begin moving back toward its roots, evolving into a broader, pure-play automation platform.

One perspective we have more conviction on is around the future of AI-driven sales. As research agents, cold email generators, and similar tools become commoditized, the novelty and effectiveness of traditional, personalized outbound will diminish, turning once-meaningful signals into noise. When hyper-personalized outreach becomes the norm, how will the best go-to-market teams differentiate themselves? To us, the frontier will always involve some form of creativity and hustle — real human effort that can be scaled but difficult to automate out-of-the-box. Because of this, tools like Clay, which support flexible experimentation and keep a human in the loop, will remain important. They'll enable teams to continuously evolve their go-to-market workflows, the fluctuations of which will remain an important source of competitive advantage.

Competitive Landscape

Now that we have an understanding of Clay, we can take a deeper look at how they stack up against their competitors.

While there isn't a company offering exactly what Clay does — a no-code tool specifically verticalized for sales — there are several adjacent categories of companies vying for similar territory and competing for the most promising talent and mindshare in the space.

The two that you’ve probably heard of, and we’d like to discuss in more detail, are AI SDRs and AI CRMs.

Clay versus AI SDRs

We’ll start by focusing on AI SDRs, which include the likes of 11x, AI SDR, and Artisan, among others. These platforms aim to automate the entire top-of-funnel process, from prospecting and lead generation to outreach and engagement, with the aim of replacing human SDRs entirely. 11x’s Alice, for instance, will automatically find a list of leads, filter them down to the highest-intent prospects, draft personalized emails to each, email them, and follow up until a meeting is booked. It does so autonomously and at scale, allowing for high-volume outreach.

In contrast, Clay isn't looking to replace human sales teams. Instead, it’s fundamentally designed to empower and augment human-driven workflows. Looking into the future, there are several reasons this may position Clay more advantageously:

• Avoiding the "personalization arms race": As AI-driven SDRs lead to a flood of automated cold emails, competition intensifies around who can deliver the most convincingly personalized message at scale. What we consider effective personalization today may quickly degrade into spam tomorrow, as email as a channel degrades and prospects become overwhelmed by indistinguishable AI-generated outreach. Clay avoids this trap by keeping humans at the center, leveraging their creativity to craft genuinely differentiated outreach strategies.

• Humans aren’t going anywhere: Sales fundamentally involves relationships and navigating nuanced human dynamics, especially in enterprise scenarios. While many tasks can be automated, people will always buy from people. Human go-to-market professionals won’t disappear, and at moments will continue to remain responsible for both strategic oversight and intricate, tactical interactions.

Ultimately, the key question is this: What should be the human role in sales and SDR workflows moving forward? If you believe in fully automating sales from start to finish, then joining an AI SDR company might be more exciting for you. But if you think humans will remain essential to the process — and that the future lies in enhancing human potential rather than replacing it — then you may be happier at Clay.

Clay versus AI CRMs

Next, let’s have a look at the AI CRMs, which include the likes of Rox, Day, and Folk, among others. These products offer a CRM with a number of AI-driven bells and whistles built in or on top: Rox, for instance, will research prospects in the background, identify the optimal time to ping them, draft an email, and offer to do so for you. Another common feature is automated record cleaning and data ingestion.

While AI CRMs share some similarities with Clay, particularly in their goal to empower human sellers, we don't view them as direct competitors. For one, AI CRMs generally target mid-funnel activities like nurturing and converting leads already within the sales pipeline. Clay, by contrast, generally focuses on top-of-funnel tasks like supporting initial outreach and prospecting efforts. In addition, Clay doesn't aim to replace the underlying CRM — instead, it serves as a flexible application layer complementing existing CRM systems.

We quite like the innovations AI CRMs bring to the table (and will explore one in more depth sometime in the future), but if you can only join one company here are a few reasons to prefer Clay:

• Supporting diverse, customizable workflows: Clay’s no-code nature allows it to support the long tail of workflows. It supports specialized, user-developed tools for enrichment, personalization, and workflow customization — capabilities that AI CRMs won’t have built into them.

• Complementary, not competitive: Clay doesn't compete directly with foundational CRM tools like Salesforce; instead, it enhances their utility. While Salesforce might be a garbage product, it remains incredibly entrenched and durable as the core system of record in sales organizations.

Ultimately, the key question is this: Will Salesforce and Hubspot still be as dominant in a decade? If you believe AI has the potential to meaningfully disrupt the core CRM landscape significantly within the next ten years, then joining an AI CRM-focused company might be compelling. But if you think the current systems of record are likely to persist, then Clay may be more interesting as a layer built on top of them.

Team

Clay wants to be the starting point for anything outbound-related.

It’s an ambitious vision, but they’ve put together an excellent team to make it happen. Consider a random sampling of their roster:

Kareem Amin — Co-Founder, CEO

Kareem was previously the VP of Product at The Wall Street Journal. Before that, he was the Co-Founder of Frame, which was acquired by Sailthru. He holds a BEng in Electrical Engineering from McGill University.

Varun Anand — Co-Founder, Head of Operations

Varun was previously founded the venture fund 89 Christopher, which he now does as a side gig in addition to Clay. Before that, he was the Director of Operations at Newfront, Head of Expansion at Candid, and worked on special projects at Google. He holds a BA in Political Science from the University of Pennsylvania.

Osman Sheikhnureldin— Head of GTM Operations

Osman was previously a Senior Marketing Operations Manager at Rippling. Before that, he was the Head of Growth at Shelf. He holds a BA in Political Science & History from George Washington University.

Nefaur Khandker — Head of Product Design

Nefaur was previously a Founding Designer at DoubleLoop. Before that, he was a Principal Product Designer at Arcadia and at Khan Academy. Earlier in his career, he was a Software Engineer at Apple. He holds a BS in Computer Engineering from Virginia Tech.

Puneet Sabharwal — Head of People and Brand

Puneet was previously the Co-Founder and CEO of Horti. Before that, he was the Head of Product Design at Olo (valued at over $1 billion). He holds a BFA in Commercial Art from the College of Art, Delhi.

Alex Miller — Machine Learning Engineer

Alex was previously a Machine Learning Engineer at Yahoo, Cove, Reddit, and Twitter. He holds a BA in Mathematics from Wesleyan University.

Luna Ruan — Engineering Manager

Luna was previously a Tech Lead at Meta. Before that, she was a Software Engineer at Pinterest. She holds a BS in Computer Science from Carnegie Mellon University.

Cara Silverstein — Engineering Manager

Cara was previously an Engineering Manager at Materialize. Before that, she was a Senior Software Engineer at Attentive and at Google. She holds a BSE in Computer Science from Princeton and an MBA from Harvard.

David Stevens — Engineering Manager

David was previously an Engineering Manager at SevenRooms. Before that, he was an Engineering Manager at Peloton and a Data Scientist at Uber. He holds a BA in Mathematics from Williams College.

Nancy Dong — Partnerships Lead

Nancy was previously the Founder and CEO of Punchcard. Before that, she was an investor at FJ Labs. She holds a BA in Economics and Political Science from Northwestern and an MBA from Harvard.

Scott Tousley — Head of Content

Scott was previously the Head of Marketing at Persefoni. Before that, he was the Head of Startup Growth at HubSpot and a Senior Content Marketing Manager at Siege Media.

Alex Kloeti — Enterprise Growth Strategy

Lex was previously a Senior Revenue Operations Strategist at RevPartners. Before that, he was an Account Executive at Gong. He holds a BS in Economics and Computer Science from McGill University.

Collin Fijalkovich — Senior Software Engineer

Collin was previously a Software Engineer at Snap. Before that, he was a Software Engineer at Google. He holds a BE in Electrical Engineering and Computer Science from MIT.

Mark Hahnenberg — Software Engineer

Mark was previously a Software Engineer at Meta. Before that, he was a Software Engineer at Airtable, a Staff Software Engineer at Nylas, and a Software Engineer at Facebook. He holds a BS in Computer Science from Carnegie Mellon University.

Larry Li — Software Engineer

Larry was previously a Senior Software Engineer at Google, where he spent seven years. He holds a BS in Computer Science and Mathematics from Rensselaer Polytechnic Institute.

Anna Wang — Product Engineer

Anna was previously a Product Engineer at Retool. She holds a BS in Computer Science from Harvard University.

Josh Hanson — Data Scientist

Josh was previously a Data Scientist at Eppo. Before that, he was a Senior Product Data Scientist at ChowNow and a Product Data Scientist at Loom. He holds a BA in Economics from the University of Kansas.

Brian Liang — Software Engineer

Brian was previously a Software Engineer at RockSet (acquired by OpenAI). He holds a BA in Computer Science from Northwestern University.

Conclusion

There’s an art to writing a great cold email -

- just as there is for the rest of your GTM activities.

Finding that perfect balance of creativity and efficiency in your work — whether that's sending pizzas to your customers’ offices, Venmo-ing them your contact details, or including personalized screenshots of their website in an email — is the most important thing in standing out and getting ahead.

That spark of creativity is exactly what's missing from most next-gen sales tools, but has been baked into Clay from the start.

To recap:

• Clay is unique. Every other piece of sales/growth software (Apollo, Artisan, Rox, 6Sense, etc.) tries to own some data source or workflow. Uniquely, Clay tries to own neither; it is an integration and orchestration layer on top of other data sources and automation primitives. This has made Clay better at executing on core workflows, better at attacking the long tail of workflows, and made the product both stickier and easier to adopt.

• Clay is creative. Clay is centered around creativity, something baked deep into the history, culture, and DNA of the company. Beyond being a value we believe deeply in, this creativity is central to the future of effective GTM; in a world where millions of AI SDRs are sending emails and making cold calls, effective GTM teams will leverage Clay’s balance of creativity and scalability to separate themselves from the noise.

• Clay is powerful. You can run a simple outbound email generation pipeline, but you can also use it to import meeting data from Zoom, generate a set of notes, and add the notes to a Notion table. The composability built into the platform makes it uniquely capable of automating GTM workflows no startup will build bespoke software for.

They’re hiring.

‎

In case you missed our previous releases, check them out here:

• Etched - The chip for transformers.

• Meter - The next modern utility.

• Modal - The future of GPUs is serverless.

And to make sure you don’t miss any future ones, be sure to subscribe here:

Finally, if you’re a founder, employee, or investor with a company you think we should cover please reach out to uhanif@stanford.edu - we’d love to hear about it :)

• Modal has grown revenue >50x in the past year and has taken less than 18 months to reach an eight figure run rate. They’re currently hiring product and systems focused software engineers in New York City.

• Etched has built a chip that can run transformers >10x faster than GPUs. They’re currently hiring Machine Learning Researchers, Compilers Engineers, Verification Engineers, Firmware Engineers, RL Design Engineers, interns, and more in Cupertino.

• Partnership: our friend Billy joined Rippling ($13B+) before its Series A. The best advice he ever got was to think like a VC when looking for a startup job, so he built a free tool to help others do the same. Prospect uses the same structures and data sources as the best data-driven VCs (i.e. Coatue, Tribe) to help candidates find the highest potential startups to join.

  If you’re looking for more promising companies to join, we highly recommend checking them out.

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information we’ve received from venture firms we work with. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should Join: Inside Edition

July 2024

As you know, the goal of this newsletter is to highlight promising early-stage companies to join, and to do so in a way that’s thorough, unbiased, and (occasionally) entertaining.

We do this because startups are risky business, and we think people deserve a quality resource to help figure out which ones are actually worth joining — one that’s less shallow, biased, and hype-driven than most of what’s out there.

Our main format is a series of monthly deep dives. By picking and deeply analyzing just one company each month (all with inflecting/breakout traction), we hope to provide a better starting point than traditional resources like the LinkedIn Top Startups list — to us, it’s important to understand why a company might succeed before joining it, and we felt most resources recommending companies didn’t go deep enough to be helpful with that.

In a similar spirit, since the start of this year we’ve been running a secondary series where once a quarter, we invite a few investors we respect to share a piece on a portfolio company they’re excited about. So far, it’s been a good way to share more opportunities — and perspectives beyond our own.

You all enjoyed the first two editions, so now we’re back with a third.

This entry features four companies in total, each backed by a different firm and each operating in a different industry. All are at an early stage, and all are growing quickly. At the bottom of each piece, we’ve added a few editorial notes on the company and why we respect the investor who’s recommending it.

This took a good amount of work from everyone involved, and we’d like to thank our contributors for participating.

We hope you enjoy :)

Why You Should Join Anterior

Copilot for health insurance approvals

When your doctor orders a scan, prescribes a new medication, or tells you you need a surgery, you usually want to move as quickly as possible. But you also want to know it will be paid for, which for many patients in the U.S. means first getting your insurance company to sign off. In this “prior authorization” process, the insurer’s clinicians review your medical records - often a massive amount of different kinds of information, from test results to doctors notes, PDFs to imaging files - and compare them with a set of care guidelines to make a decision, often taking hours to tackle.Meanwhile, those patients have to wait for even the simplest approvals; in some cases, care is delayed as much as a month. And the problem is not just about speed, but also cost: One-quarter, or $950 billion, of U.S. health care spending goes to administrative expenses, with as much as $570 billion lost to inefficiencies that don’t improve patient outcomes. We at Sequoia have seen again and again, across industries, that AI can be a powerful tool for solving just such challenges. But in this case, the trouble has been that while people with expertise in health care or AI are relatively easy to find, people with both are quite rare. Luckily for us, though, Anterior (fka, Co:Helm) co-founders Abdel Mahmoud and Zahid Mahmood applied to Arc, Sequoia’s company building immersion for early stage companies.

Born in Libya, Abdel came to the U.K. as a refugee at 8 years old. By 15, he had his own software business—and by 18, he had joined the prestigious Royal Military Academy Sandhurst, where he became the youngest infantry officer in the U.K. Then Abdel became a doctor. He loved helping patients, but grew frustrated with the mountains of paperwork that kept him from spending more time with them, and eventually left medicine, returning to University College London for a master’s in computer science and going on to work in product at Facebook and Google. Zahid, too, founded a company as a teenager—the first of his now-four founder journeys followed by a period building payment platforms for large insurance companies, which gave him deep expertise in the industry. Anterior is an LLM-powered co-pilot for administration, handling all the work that comes before the clinical decision-making. Where it might take a nurse several hours to track down, organize and summarize hundreds of pages of records they need, Anterior can do it in seconds, allowing clinicians to increase their output tenfold and focus on what matters most: helping patients. With partners including one of the top U.S. health plan administrators, Abdel, Zahid and their team will soon be making a difference for millions of patients, and the waitlist is growing quickly. And while they are starting with prior authorization, there is a long list of pain points Anterior will ultimately be able to address, from case management to underwriting and beyond.

— George Robson, Partner at Sequoia

Anterior is hiring talented AI engineers, product engineers, QA automation engineers, and more in New York and London - if you’re interested, reach out to their team here. We also recommend checking out the rest of George’s awesome portfolio at Sequoia.

‎

Why You Should Join Valar Labs

Building AI to identify the right cancer treatment

There are nearly ~2M new cancer diagnoses annually in the US, many more globally. To select a therapy, oncologists rely on clinical trials and treatment guidelines, which typically provide data on an overall response rate (ORR), or the fraction of all patients with a given cancer diagnosis who respond to a specific therapy. But this data is still hard to translate to individual patients: for a therapy approved with an ORR of 35%, there are unfortunately two patients who won’t respond for every one patient who will. Identifying the patients who are unlikely to respond could save them valuable time (to try other therapies)—and spare them harsh side effects.

In order to make every single one of these cancer diagnoses, a biopsy of the patient’s tumor is laid onto a slide, stained in standard ways, and histologically classified as cancer upon review under a microscope by a pathologist. During the COVID pandemic, a huge number of these slides were digitized – meaning that every cancer patient already has images of their tumor histology readily accessible for their care team.

What if these same histology images contained far more information than what they are telling us today? What if the slides have hidden insights buried within them, invisible to human pathologists, but accessible to AI? What if these signatures could be used to predict how a patient’s disease would progress? Or whether he or she would respond well to a particular therapeutic regimen?

To answer these questions, the Valar team has partnered with oncologists, cancer centers, and research teams all over the world to build several large registries of thousands of patient specimen images, carefully paired with clinical response data. They have rigorously trained AI algorithms on these retrospective data—and published results showing that AI can indeed learn what humans simply cannot see. Valar’s AI diagnostic tests can predict how aggressive a bladder cancer patient’s individual cancer will be—or whether a pancreatic cancer patient is likely to respond to the FOLFIRINOX treatment regimen—or if an ovarian cancer patient is a high- vs low-likelihood responder to platinum-based chemotherapy.

The strength of these early data have made clinical research partners excited to engage in an early access program for Valar’s first clinical-grade, CLIA-validated diagnostic product (Vesta) in bladder cancer. There is still a lot more to learn, but we’ve heard the feedback again and again from oncologists: if this information was knowable, every patient and provider would want to know.

The Valar co-founders—Anirudh Joshi, Damir Vrabac, and Viswesh Krishna—met as researchers in Andrew Ng’s AI Lab at Stanford. They were intensely motivated to use AI to do something big for patients, and were moving fast to learn everything they could about cancer care.

Perhaps even more impressive than their technical speed has been their extraordinary depth of understanding of the clinical scenarios that they are trying to impact. To earn this understanding, they have cold-emailed and ultimately mind-melded with dozens of oncologists and pathologists. One leading oncologist we spoke with said he loved collaborating with the Valar team because of their “complete humility and willingness to learn about what we do clinically.” These are rare traits in a team that is also so cutting-edge in their own AI, software, and execution capabilities. And these are exactly the combination of skills that will be required to develop, validate, distribute, and scale AI to every oncology mind, and to every care team in the world.

— Vineeta Agarwala, General Partner at Andreessen Horowitz

Valar is hiring talented software engineers, machine learning engineers, and more in Palo Alto - if you’re interested, reach out to their team here. We also recommend checking out the rest of Vineeta’s portfolio at Andreessen Horowitz. As a former physician, researcher, and operator at tech-driven healthcare companies, Vineeta (who holds an MD/PhD from Harvard/MIT) has a uniquely deep understanding of the healthcare space.

‎

Why You Should Join Armada

Bringing AI to the edge

Cloud computing has enabled a revolution in how businesses operate, with 65% of corporate data now in the cloud and over $35B spent on datacenter buildouts last year alone. Now, companies make their decisions with data, and AI is rapidly changing how valuable work gets done. But this wave has hit a wall: the kind of work that builds, carries and defends the world around us—physical, industrial work—has stayed largely the same.

On oil rigs, mines, and defense installations across the globe, faulty connectivity struggles to pull terabytes of data per site back to HQ, process it, send it back, and change business operations in anything close to real time. The largest companies in the world don’t even bother. Meanwhile, cloud bills stack up to the tune of $100s of millions per company, and the hardest workers in America continue business as usual.

Armada is changing that paradigm. Leveraging a deep partnership with Starlink, Armada is building the hardware and software to take AI to the edge: real work, on-site. Its modular datacenters (picture a shipping container full of GPUs) can operate anywhere on land or sea, process data, train and deploy AI models, and deliver results in real time. An oil rig or Navy ship in the Arctic can have the same speed and processing power as if they were stationed inside an AWS datacenter.

A revolution this big demands an incredible team. Founders Dan Wright and Jon Runyan have built and led multi-billion dollar businesses—Dan at AppDynamics and DataRobot, and Jon at Okta. Their CTO, Pradeep Nair, built Azure Global from a skunkworks project to a world-conquering juggernaut. And Prag Mishra, the Chief AI Officer, led the rollout of AI efforts for robotics in Amazon warehouses.

Armada has the opportunity to build a new Cloud giant from the outside-in—this is not an easy-does-it B2B SaaS company. Armada builds hardware, software, and AI applications, and lands them with the largest companies and governments in the world. Already, Fortune 500 companies, US State Agencies, and government programs have joined the armada. This could be a generational company.

— Shahin Farshchi, General Partner at Lux

Armada is hiring talented software engineers, account executives, and more - if you’re interested, reach out to their team here. We also recommend checking out the rest of Shahin’s portfolio at Lux. He has a remarkable track record and has previously partnered with MosaicML, Zoox, and Relativity Space.

‎

Why You Should Join Prophet Security

Autonomous threat-hunting in minutes, not months

Every day, security operations teams at companies all over the world are barraged with hundreds of threat alerts, warning them of possible attacks from nation-states and malicious actors. The alert volume is far greater than any team of threat hunters could ever process on their own, so most alerts go untriaged. This is one reason why the average time to detect a breach was 204 days last year.

Even when attacks are detected, security analysts spend hours investigating: How did the bad guys get access? How long have they been in our system? What did they take? Did they leave anything behind? Security has often been described as a game of cat and mouse, but in practice, it’s trying to find the breadcrumbs that could signal an attack in an impossibly large haystack, before the haystack grows so large that it crushes you.

Automating threat hunting and response is not a new idea. For the past decade, SOAR (security orchestration, automation, and response) tools have tried to eliminate manual work and accelerate security teams. Many have fallen short, requiring analysts to manually map brittle workflows that aren’t always applicable to the unpredictable alerts that swarm a security operations center.

Prophet Security uses language models to detect, investigate, and remediate alerts in minutes instead of months, enabling security engineers and analysts to focus on higher level work. Prophet uses read-only access to other security tools like Okta, Crowdstrike, Splunk, and more to synthesize an alert graph. Because each tool formats alerts differently, the graph is normalized and embedded in vector space to feed a planning and reasoning engine.

Prophet checks contextualized alerts against investigation tools and data sources to decide if the alert is likely a true positive and assess severity. The planning agent produces a step-by-step plan, prompting an instruction-tuned action model to execute and amend the plan recursively as new information is discovered.

The human threat analyst receives a summary of the alert, key findings, a detailed timeline, and a recommendation on what to do next. In early customer deployments, Prophet has found their solution reduces the time to resolution by 10x on real-world alerts, and empowers security analysts to focus on protecting customers, users, and employees.

The Prophet team brings extensive depth in security and AI. Co-founders Kamal Shah and Vibhav Sreekanti worked together as CEO and VPE at StackRox, a container security company that was acquired by Red Hat in 2021. There, they had a key insight: giving security teams more threat visibility doesn’t actually make them more secure without them having a better way to manage threats in the first place. The broader team consists of early engineers from security unicorns like Abnormal Security, Mandiant, Expel, and more.

Prophet emerged from stealth earlier this year with a set of well-known companies as early customers, and has been receiving impressively positive feedback. As malicious actors use generative AI to attack companies more often and with greater ferocity, Prophet is becoming the crystal ball that security teams rely on to power their security operations.

— Rak Garg, Partner at Bain Capital Ventures

Prophet Security is hiring talented machine learning engineers and product designers - if you’re interested, reach out to their team here. We also recommend checking out the rest of Rak’s portfolio at BCV — he’s one of the sharpest and most thoughtful investors we know.

Conclusion

If you’re open to new roles or just curious about their products, Anterior, Valar, Armada, and Prophet Security are all worth checking out. In case you missed it, last quarter’s edition featuring suggestions from Benchmark’s Chetan Puttagunta, Greylock’s Saam Motamedi, Sequoia’s Josephine Chen, General Catalyst’s Trevor Oelschig and Vedant Suri, and A*’s Kevin Hartz is worth reading through too.

We hope you’re enjoying the start of summer. See you next month!

‎

Thanks to George, Vineeta, Saam, Shahin, and Rak for their help with this piece.

In case you missed our previous releases, check them out here:

• Warp - Revolutionizing the most foundational developer tool.

• Pinecone - Vector databases index the AI revolution.

• WorkOS - The API to increase your TAM.

And to make sure you don’t miss any future ones, be sure to subscribe here:

Finally, if you’re a founder, employee, or investor with a company you think we should cover please reach out to us at ericzhou27@gmail.com and uhanif@stanford.edu - we’d love to hear about it :)

• Modal has grown revenue >50x in the past year and has taken less than 18 months to reach an eight figure run rate. They’re currently hiring product and systems-focused software engineers in New York City.

• Etched has built a chip that can run transformers >10x faster than GPUs. They’re currently hiring Machine Learning Researchers, Compilers Engineers, Verification Engineers, Firmware Engineers, RL Design Engineers, interns, and more in Cupertino.

• Partnership: our friend Billy joined Rippling ($13B+) before its Series A. The best advice he ever got was to think like a VC when looking for a startup job, so he built a free tool to help others do the same. Prospect uses the same structures and data sources as the best data-driven VCs (i.e. Coatue, Tribe) to help candidates find the highest potential startups to join.

  If you’re looking for more promising companies to join, we highly recommend checking them out.

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information we’ve received from venture firms we work with. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should Join Modal

(Click the link to read online).

At this point, you’re probably familiar with the important applications of generative AI.

The Chrome extension that’ll do your homework for you.

The app where you can reanimate dead celebrities.

The AI wingman for charming people on Tinder, the AI girlfriend for when things fall apart, and the AI therapist for when she eventually dumps you too.

You know, the game-changing stuff.

You might be surprised, then, that despite searches for “AI Girlfriend” being up 2400% last year, almost all the 1.3 trillion dollars generative AI will eventually generate is set to come from boring, enterprise-y things.

Assistants and productivity tools for white collar workers. Agents to write code and fix outages for software engineers. Copilot for Office 365, which over 60% of the Fortune 500 are already using.

Did you know that Accenture (the consulting firm) has already booked over $1 billion in Generative AI engagements?

That might be more than like every AI startup combined.

Ok, maybe you’re not surprised.

But if you’re like us, it’s a little disappointing.

A year and a half into generative AI, it’s worth asking — is the future of AI just big companies building copilots for existing products? What’ll it take for more AI-native apps and products to hit the mainstream?

Much has been written about the different paths and challenges here, but the small piece we’d like to explore today is around the accessibility of compute.

While access to cutting edge models has been democratized through open source, what hasn’t been is the infrastructure required to train and run them. For most hackers, GPU-based workloads (i.e. inference pipelines, training runs) still aren’t as accessible because of how difficult and expensive GPUs are to reserve, even for small amounts of time.

Most developers can’t justify reserving a large block of GPUs just to experiment with their own models or to prepare for production scale that may or may not come.

Of course, working around pricy instances is nothing new.

In the past, developers solved a similar problem with the concept of cloud functions - individually packaged pieces of code that could be scaled independently and automatically. Instead of renting a traditional server for a fixed rate (regardless of usage/idle time), offerings like AWS’s Lambda or GCP’s Firebase Functions would let you host code and pay per function call instead.

This paradigm would be particularly useful for GPU-based workloads. You can imagine how it would make working with AI more accessible, which is why there’s been strong demand for exactly this sort of product.

Unfortunately, none of the cloud providers have figured out how to build it.

Fortunately, the team at Modal has.

Background: The Landscape of ML Infrastructure

Before getting too deep into Modal, it’d probably be helpful to get a sense of the broader landscape they fit into. What is a “serverless GPU”, and how does it compare to all the other ML hosting solutions out there?

These days, there are lots of companies looking to make deploying and scaling machine learning models easier. Many of them — from GPU re-sellers to end-to-end models as a service — get bundled under the same general umbrella of “model hosting infrastructure”.

These products are all quite distinct, but their value propositions sound similar enough that it’s easy to get confused if you’re not technical.

At a high level, we’ve found it helpful to split companies and their offerings into three general categories, each operating at a different level of abstraction:

• Companies that will host and autoscale a model for you.

• Companies that will host and autoscale any piece of code for you.

• Companies that will host hardware for you.

Models as a Service

Starting off at the highest level of abstraction, companies in this category will host and autoscale a model for you. They’ll expose a simple endpoint for your application to call, automatically scaling the model up and down to meet demand while avoiding wasted GPU time. These offerings often have large collections of community-published models you can call out-of-the-box — think text generation models (i.e. Llama 7B), image generation models (i.e. SDXL), or image restoration models (i.e. GFPGAN). In most cases, you end up paying per second of GPU time used.

Besides being the easiest way to run popular models, they’re also the easiest way to host a custom model. If you have a model’s weights and a script to run it, these services will package, deploy, and host the entire thing for you. For instance, when containerizing your model, Replicate’s Cog will automatically generate a FastAPI endpoint to call it and a queue worker to batch its requests.

Companies with offerings that fit in this category include Replicate, Baseten, BentoML, Mystic, Hugging Face’s Serverless Inference API, and AWS’s Sagemaker. While there are certainly nuances between these offerings (i.e. Sagemaker’s autoscaling endpoints require you to define your own scaling policy based on different utilization metrics, for a flat fee Mystic offers this functionality to companies using their own cloud), at a high level what’s being offered is an endpoint users can call without worrying about anything else.

Functions as a Service

Moving down one layer of abstraction, companies in this category will host and autoscale any arbitrary piece of code for you. Also known as “serverless”, offerings here let you upload a function and run it as a one-time script, a regularly scheduled cron job, or an endpoint/microservice that will automatically scale up and down to meet demand.

You can kind of think of this as a superset of the “model as a service” category, where you can run any workload (i.e. training models, scraping data, rendering teapots) and with wider variety of triggers while still only paying for the time your code is running. You generally won’t get pre-hosted models you can call out of the box, but you will get more control over the exact code that’s being autoscaled.

Like we mentioned earlier, serverless offerings have been around for a while. For CPU-based workloads, AWS’s Lambda functions, GCP’s Cloud Functions, and Azure’s Functions are all examples of offerings in the space. For GPU-based workloads, though, the paradigm is new: only a few startups have begun offering them in the past year, driven by demand from the rise of generative AI.

Companies with offerings that fit in this category include Modal, Beam and Inferless. We will go deeper into the differences between companies in this category later.

Hardware as a Service

At the lowest level of abstraction are companies that essentially rent out hardware. These companies provide “instances” (virtual machines) with specific parts (i.e. 8x A100s) that users can use for their application. Users are responsible for SSH’ing in to set up and run their workloads themselves, and they’re also responsible for acquiring more instances and distributing a workload if additional scale is needed. Generally, these instances are rented out on-demand or in advance for large blocks of time (i.e. by week, by month, by year). They’re the cheapest per hour, but you end up paying a fixed rate regardless of actual utilization.

This is the bread and butter of all the major cloud providers. AWS, GCP, Azure, Oracle, and IBM all have offerings in this category. The recent surge in GPU demand has also spawned a number of startups doing the same: Lambda Labs, Coreweave, Vast, and Brev are all players in this space as well.

Of course, there are still a number of companies that don’t fit neatly into one of the categories above. For instance:

• Runpod has offerings in all three categories (a serverless offering, an endpoints service, and a way to rent entire instances).

• TogetherAI sells instances (Together GPU Clusters) and also endpoints (Together Inference) for running popular open source models. In addition, they have a proprietary inference engine which makes running things faster and cheaper.

• Foundry has an orchestration layer which dynamically splits workloads across instances based on factors like demand, capacity, and how long users are willing to wait.

• Anyscale has both a platform letting users deploy Ray applications serverlessly and an endpoints offering competing directly with some of the companies mentioned above.

In a general sense, though, we’ve found this framework helpful in making sense of the large and growing number of companies in the ML hosting space.

If things are still fuzzy, here’s a small chart we made which simplifies things further:

Modal

So, Modal is a serverless GPU offering, meaning they host and autoscale GPU functions for users. Think of them as AWS Lambda, but with support for GPUs as well as CPUs.

With Modal, users can deploy an endpoint to run inference on Llama-70B, but they can also render a Blender scene, run a simulation, scrape data, or generate embeddings for a million different images before running a K-means clustering workload across all of them. Moreover, they can do so while paying per second of active CPU/GPU time.

Concretely, this creates value in a few ways:

• Lower cost - developers don’t have to pay for time their machines are idle.

• Better developer experience - developers don’t have to worry about provisioning, configuring, or scaling their workloads.

More than that, though, Modal has built the best serverless GPU offering on the market. Instead of building on Docker and Kubernetes, Modal’s system is powered by a custom serverless runtime, a custom file system, a custom image builder, and a custom job scheduler.

The superior performance this has enabled creates value in a few ways:

• Lower cost - Modal’s faster scale-up times ultimately mean less idle time and lower costs for users.

• Better developer experience - Modal’s lower cold start times mean a faster development loop, letting users run code on remote machines as if they were developing locally.

‎

Modal launched publicly in October of last year and has grown swiftly ever since. Besides growing 50x in the past year and hitting an 8 figure run rate, they’re now doing more revenue each hour than they were doing all month just a year ago. Customers include large companies like Ramp and Scale, high profile startups like Cognition Labs, Suno, Cursor, and Substack, and other sizable entities we won’t reveal.

The company was founded in 2021 by Erik Bernhardsson and Akshat Bubna. Erik was previously CTO of Better.com and one of the first 30 employees at Spotify, where led their early data science/machine learning teams and built the core of their music recommendation systems. Akshat was previously a Staff Software Engineer at Scale, where he joined as an early employee after studying Computer Science and Mathematics at MIT. Both have a gold medal from the IOI.

On the back of their impressive team and traction, Modal raised a $16 million dollar Series A in October of last year led by Redpoint Ventures. Before that, they had raised a $7 million seed round led by Amplify Partners.

Modal has a remarkable team, a product developers love, and the ultimate vision to solve every engineer’s infrastructure problems, once and for all.

Will they be able to pull it off?

We think they’ve got a shot.

Let’s get into why.

‎

The Opportunity

To understand whether Modal might be a good business, there are really two questions we have to answer:

• Is there a large market for serverless GPU functions?

• Does Modal have a shot at leading that market?

The Market for Serverless GPU Functions

Modal makes money when other companies run functions on their GPUs.

You probably don’t need to be persuaded that lots of people want to run code on GPUs - it’s why companies are spending so much on datacenter buildouts and why server providers are raising so much money. The real question is whether a sizable portion of these people want to do it in a serverless manner.

One helpful proxy here is how much people are spending on serverless CPU functions. If a lot of people are finding value in serverless CPU functions, it’s reasonable that a similar portion will find value in their GPU counterparts as well.

The numbers here are promising:

• Today, over 70 percent of AWS customers, 60 percent of GCP customers, and 49 percent of Azure customers use one or more serverless solutions. Last year, serverless adoption for organizations running on Azure grew 6 percent, serverless adoption for organizations running on GCP grew 7 percent, and serverless adoption for organizations running on AWS grew 3 percent.

• There have been a number of sizable companies built on vertical-specific applications of serverless functions — think Vercel and Netlify, both of which are serverless frontend/CDN offerings. 7 percent of organizations running serverless workloads are also using at least one of these emerging cloud platforms. Of these, 62 percent are using a frontend development platform like Vercel or Netlify and 39 percent are using an edge compute offering like Cloudflare or Fastly.

Since both general (i.e. AWS Lambda, Google Cloud Functions) and vertical-specific (i.e. Vercel, Netlify) versions of serverless functions have found meaningful adoption, we would guestimate that a meaningful number of companies would find the same benefits — easy scalability, built-in redundancy/uptime guarantees, generally better developer experience — attractive for their growing GPU/AI workloads as well.

If anything, we could expect the proportion to be even higher for GPU-based workloads:

• GPU instances are more expensive than CPU instances. Idle time is more pricey, and the only way to avoid it is by using a serverless offering.

• GPU instances are more supply-limited than CPU instances. It’s less likely you can rent a rack of them and provision more on the spot as you need to scale — for many companies, serverless will be their only option.

Modal’s early traction points to this being something developers have wanted for a while.

‎

Modal’s Approach

So out of all the serverless GPU offerings out there, why do we like Modal’s the best?

Compared to its peers, Modal is better on two fronts: technology and developer experience.

Technology

Instead of using Docker and Kubernetes like most serverless offerings, Modal has built a custom serverless runtime, a custom file system, a custom image builder, and a custom job scheduler. Building everything from scratch has given their platform better performance — think lower cold start times, faster scale ups, and ultimately lower costs for users.

How does it all work?

For illustrative purposes, let’s explore an early, simplified version of Modal’s architecture.

When you run a containerized function on a remote machine, the worker usually begins by pulling down the container image you’ve picked. Once that’s done, it’ll start the container and run your function inside.

However, this process is pretty slow. Pulling down the image, especially the first time, can take a while — it involves downloading all the necessary files, dependencies, and configurations needed to run your app. This includes a lot of stuff, much of which isn’t needed at runtime — containers often contain a lot of bloat!

To get around this, Modal has built what’s essentially lazy loading for files within container images.

Early on, they would take their users’ images and put them on NFS - a distributed file system protocol which lets clients access files over networks as if they were stored locally. Then, instead of pulling the container image in its entirety before running it, worker machines connected to NFS could simply use runc - a Docker subroutine which launches a container from an arbitrary directory - to run said container from it’s image’s location on NFS directly. Since NFS makes it as if all the necessary files are already on the worker’s machine, runc is able to proceed. The files get fetched as needed under the hood (handled by the NFS protocol), something which helps prevent much of the bloat from getting downloaded.

To address the remaining latency this introduces (i.e. when starting a container, reading each required file over NFS now takes several milliseconds), Modal has built a custom filesystem which makes significant usage of caching.

See, container images have a lot of overlap:

Modal has built a content-addressable filesystem which stores each unique file just once — even if it’s being used across thousands of different containers and stored at thousands of different locations.

Besides being better from a storage point of view, this structure makes caching easier as well. When runc needs a file, Modal maps the requested filepath (i.e. /lib/x86_64-linux-gnu/libc-2.31.so) to a hash of the requested file’s contents, which serves as the address of the original file. Then, if the file hasn’t been cached on local disk already, it’s fetched (either from Modal’s file server or some remote, external server if it’s completely new) and returned before being cached for the next run. If it’s already there, it’s returned immediately.

Since there’s so much overlap between unrelated images, there’s a good chance that by the time you run your function, a prior run (by a completely different user) will have resulted in your worker caching most of the files your container needs. This minimizes the necessary reads over NFS, which cuts down on latency even further.

Now, there are a number of deeper implementation details we won’t go into here. For instance, how does the filesystem and its index get updated? What if the container image doesn’t already exist on NFS? If you’re interested in learning more, there’s a great talk Erik gave on this topic, much of which we’ve tried to summarize here.

The upshot, however, is that Modal containers have much faster start times than Docker containers:

• Using runc + NFS to get around downloading the entire image means files are lazy-loaded — only what’s used in the actual function is downloaded onto the worker machine.

• Files are cached on worker machines across containers and across container runs, so when files are needed they can often be accessed right away. Caching by file is a lot more granular than caching by layer, which is what Docker does out of the box.

Modal has since optimized their system a lot more. For instance, they now have a custom protocol where instead of using NFS, they serve blobs using Cloudflare’s R2 service. Instead of hacking things with runc, they use gvisor, which is generally more secure. They’ve also made a number of optimizations to make subsequent (not just initial) startup times faster - think storing "traces" of workloads and prefetching data that’s likely to be read, snapshotting CPU memory, etc.

The result is the same: every time you spin up a new container on one of Modal’s thousands of machines - either because you’re developing and want to test something or because you’re serving and need to scale - that container will come up much faster than if you were somewhere else. This is important not just for cost or performance, but also, as we’ll see, for ensuring the best possible developer experience.

‎

Developer Experience

Modal’s core technology, the feature set they’ve built around it, and the platform they’ve built on top of it have collectively enabled a better developer experience than anything else we’ve seen. There are a few reasons behind this:

• For certain users, Modal’s core technology (quickly spinning up a container and running a function in it) enables a uniquely efficient development loop. Developers working on data-heavy applications often find it hard to work locally — local environments are usually different than the cloud environments their work will be productionized in (i.e. different hardware constraints), so developers who work locally often find themselves with confusing deployment issues and a great deal of re-written code. As a result, many data teams follow some version of the following remote development loop: they’ll build a container, push it, trigger the job, and download the logs for debugging purposes after. This process can take minutes to hours each time and can be a real pain. In contrast, with modal run it takes less than a second to spin up a container and run your code on a remote machine with the right environment and hardware. This compresses that loop into something as efficient as developing locally.

• When it comes to feature completeness — out-of-the-box support for typical features and integrations — Modal is just more mature than everything else. For instance, beyond the basics they also support websockets, concurrent inputs on a single container, streaming endpoints, tunnels (live TCP ports), mounts for cloud buckets, and more.

• As a platform and business, Modal is also the easiest to work with. They have the most generous free tier. They were the first to offer support for H100s. Even on customer service, their response time on Slack is unparalleled. It’s not often that the CEO and CTO are supporting individual developers late at night, but we’ve seen their team’s commitment ourselves:

  

The common thread we’ve noticed across all this is a general commitment to craftsmanship and polish. We’ve seen this in everything Modal has built — from their core technology/documentation/APIs down to their support philosophy/error messages/blog post visuals. They clearly care a lot about what they’re building and are willing to take the time and effort to do it right, even if it’s not always the easiest path forward.

This is particularly rare in infrastructure companies, and something we admire deeply.

‎

Competitive Landscape

Now that we have a deeper understanding of Modal’s strengths as a product and company, let’s take a deeper look at how they compare against other “function-as-a-service” offerings.

Runpod Serverless

Runpod is a platform which offers endpoints-as-a-service, functions-as-a-service, and instances-as-a-service. They recently raised a 20 million dollar seed round from Dell Technologies Capital and Intel Capital. Runpod Serverless, their functions-as-a-service offering, competes directly with Modal.

In comparing the two products, our overall impression (as users) was that Modal had a better product and developer experience.

Tech-wise, to our knowledge Runpod only supports Docker images, which suffer from the cold start problems discussed above. Modal has addressed this with a custom stack while maintaining compatibility with Docker-based workflows: while Modal uses a custom container builder, they have a custom implementation of the Dockerfile specification which allows developers to use environments previously defined in that manner.

From a DX perspective, Runpod is also a bit harder to work with. If you don’t already have a Docker image you’d like to use, its on you to define and push it to the container registry before getting started. In contrast, Modal doesn’t require any infrastructure at all — users just create an account, install the modal Python package, and set up a token. They can then start running jobs right away.

Beyond that, Runpod Serverless doesn’t have a free tier, making it more difficult for people to try. They also don’t support some of the more advanced features that Modal does (i.e. job scheduling).

Beam

Like Modal, Beam is a platform which only offers functions-as-a-service. The company (formerly known as Slai) raised a 3.5 million dollar Seed Round led by Tiger Global in 2022.

From a technical perspective, Modal and Beam are actually quite similar: Beam also has a distributed file system (they’ve stuck with NFS), a custom image format, and a content-addressable storage system, all of which they built this past summer.

However, we found Beam’s offering to be generally less mature than Modal’s — they’re still missing many of the features, integrations, and compliance-related checkmarks most serious and production-facing developers would expect.

For instance, to our knowledge you can’t build a Beam container from a Dockerfile like you can with a Modal container. Beam doesn’t support concurrent inputs within containers, streaming endpoints, or the ability to expose live TCP ports on containers. Beam isn’t SOC2 compliant like Modal is, they don’t offer integrations (i.e. with Vercel or Okta) like Modal does, and they don’t have any H100s like Modal does. Worst of all, they require a credit card to try their product.

The upshot is that if you were an early employee, investor, or customer picking between the two, it’s not clear why you wouldn’t just go with Modal.

Inferless

Inferless is an Indian platform which only offers functions-as-a-service. The company was founded in 2019 and is backed by Peak XV (formerly Sequoia Capital India). The product remains in private beta — you have to join a waitlist to get access.

While we haven’t tried their product, their documentation suggests a platform that’s generally less mature than Modal’s. For instance, they don’t support streaming endpoints, scheduled jobs, or H100s.

Hyperscalers

If you’ve gotten this far in the piece, there’s probably one big question still lurking at the back of your head:

Why haven’t GCP, AWS, or Azure built this already?

The hyperscalers pioneered the concept of functions as a service, and they have the world’s largest supply of GPUs — it only makes sense that they’d move into this market.

While we can only speculate, there are two likely reasons they haven’t yet:

• Virtualization challenges. Exposing GPUs to short-lived, ephemeral VMs is not trivial technically. Each serverless function may have different GPU resource requirements, and allocating GPU resources dynamically while ensuring optimal utilization/performance is challenging. In addition, a serverless platform needs to ensure compatibility between GPU drivers, runtimes, and the underlying virtualization layer. Different GPU models and architectures may have specific driver requirements, and maintaining compatibility across different serverless instances can be challenging.

• Security and isolation challenges. Ensuring strict data isolation between serverless functions executing on the same physical GPU, mitigating side-channel attacks that might exploit shared GPU resources, and managing access control and permissions for serverless functions accessing GPUs are all non-trivial technical challenges.

Modal has worked hard to address many of these issues, and they have a head start as a result.

Ultimately though, it’s probably only a matter of time before the hyperscalers release competing offerings of their own.

When they do, though, it’s not a guarantee that their offering will be as performant or well-crafted as the platform Modal has built. Modal has always focused on developer experience, something which the hyperscalers don’t have the best track record with.

Team

Modal’s long-term vision is to build the world’s best serverless platform. They want to abstract away every developer’s challenges with infrastructure. In the same way the cloud allowed people to get around buying and hosting physical servers, Modal envisions a world where they can do the same with the instances we all use today.

To put it another way, Erik told us his ambition is to someday replace Kubernetes.

It’s a bold vision. If anyone’s going to pull it off, though, it’ll probably be this team. Modal’s first 7 employees had 5 IOI gold medals between them — consider a sampling of their roster.

Erik Bernhardsson - Co-Founder, CEO

Erik was previously CTO of Better.com. Before that, he was one of the first 30 employees at Spotify, where he led their early data science/machine learning teams and built the core of Spotify’s music recommendation systems.

Akshat Bubna - Co-Founder, CTO

Akshat was previously a Staff Software Engineer at Scale AI, where he joined shortly before their Series B. He holds a degree in Computer Science and Mathematics from MIT.

Jonathan Belotti - Software Engineer

Jonathan was previously Team Lead for the Data Platform at Canva. Before that, he was a Software Engineer at Zendesk. He holds a BS in Computer Software Engineering from RMIT and an MS in Mathematics and Statistics from the University of Technology Sydney.

Michael Waskom - Software Engineer

Michael was previously a Staff Machine Learning Scientist at Flatiron Health. Before that, he was a Research Scientist at NYU. He began his career as a Doctoral Researcher at Stanford, where he holds a PhD in computational psychology.

Eric Zhang - Software Engineer

Eric was previously a Software Engineer at Convex, where he was their first engineering hire. Before that, he was a Quant Research Intern at Jump Trading. He started his career as a Software Engineering Intern at Scale, and he holds a BA and MS in Computer Science from Harvard.

Irfan Sharif - Software Engineer

Irfan was previously a Software Engineer at Cockroach Labs. Before that, he was an ML Infrastructure Engineering Intern at LinkedIn. He holds a B.A.Sc. in Electrical and Computer Engineering from the University of Waterloo.

Ro Arepally - Software Engineer

Ro was previously a Tech Lead at Asana. Before that, he was a Software Engineer at RedCircle. He started his career as a Software Engineer at Uber, and he holds a BS in Mathematics from CMU.

Elias Freider - Software Engineer

Elias was previously the Analytics Data Lead at Kry. Before that, he was Co-Founder of EdQu, an edtech startup. He began his career as an Analytics Software Engineer at Spotify and holds an MS in Computer Science from the KTH Royal Institute of Technology.

Daniel Norberg - Software Engineer

Daniel was previously a Senior Staff Engineer at Spotify. Before that, he was a Software Engineer at Treasure Data. He holds an MS in Computer Science from the KTH Royal Institute of Technology.

Staffan Gimåker - Software Engineer

Staffan was previously a Senior Software Engineer at Detectify. Before that, he was a Senior Software Engineer at Spotify. He holds a BS in Computer Engineering from the KTH Royal Institute of Technology.

Richard Gong - Software Engineer

Richard was previously a Software Engineering Intern at Scale AI, a Trading Intern at Jane Street, and a Perception Intern at Cruise. He holds a BS in Electrical Engineering and Computer Science from MIT.

Margaret Shen - Head of Business Operations

Margaret was previously an Associate at Accel. Before that, she was a Product Manager at OakNorth and Dropbox. She holds an MBA from HBS and a BS in CS from Stanford.

Rachel Park - Software Engineer

Rachel was previously a Software Engineering Intern at Warp. Before that, she was a Software Engineering Intern at Mem. She holds a BS in CS from Stanford.

Sona Dolasia - Design

Sona was previously a Senior Product Designer at Scale AI. Before that, she was a Product Designer at Spotify. She holds a BA in Cognitive Science from UC Berkeley, where she was a Regents Scholar.

‎

Conclusion

Working with GPUs is expensive and challenging today, but it won’t be if Modal succeeds.

One can but imagine the wave of industry-defining, category-creating, entrepreneurial activity this will unleash.

The John D. Rockefeller of AI girlfriends is out there somewhere, and he’s been waiting for a platform like Modal to build upon.

They’re hiring.

‎

In case you missed our previous releases, check them out here:

• Warp - Revolutionizing the most foundational developer tool.

• Pinecone - Vector databases index the AI revolution.

• WorkOS - The API to increase your TAM.

And to make sure you don’t miss any future ones, be sure to subscribe here:

Finally, if you’re a founder, employee, or investor with a company you think we should cover please reach out to us at ericzhou27@gmail.com and uhanif@stanford.edu - we’d love to hear about it :)

• Meter just raised a sizable round from Sam Altman and Lachy Groom after more than tripling revenue, square footage, and deployed networks. They are actively hiring for several functions, including a number of engineering, product, and GTM roles that aren’t yet publicly listed. Reach out to jobs@meter.com if you’re interested in any roles not on their website.

• Etched has built a chip that can run transformers >10x faster than GPUs. They’re currently hiring Machine Learning Researchers, Compilers Engineers, Verification Engineers, Firmware Engineers, RL Design Engineers, interns, and more in Cupertino.

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information we’ve received from venture firms we work with. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should Join: Inside Edition

April 2024

As you know, the goal of this newsletter is to highlight promising early-stage companies to join, and to do so in a way that’s thorough, unbiased, and (occasionally) entertaining.

We do this because startups are risky business, and we think people deserve a quality resource to help figure out which ones are actually worth joining — one that’s less shallow, biased, and hype-driven than most of what’s out there.

Our philosophy in doing this has been to been to offer the opposite of one of those “Top Startups” lists — to us, it’s important to understand why a company might succeed before joining it, and we felt most resources recommending companies didn’t go deep enough to be helpful with that. By picking and deeply analyzing just one company each month (all with inflecting/breakout traction), we hoped to provide a better starting point for people exploring potential career moves. So far, this approach has worked well: over 20,000 of you have subscribed, and most of the companies we’ve covered are growing quickly — The Browser Company, Hadrian, and SmarterDx have all recently raised more money, Hebbia, Traba, and Etched have all doubled headcount in the past year, and Loop is set to more than 10x their payments volume between the start of last year and the end of this one.

But of course, there have also been tradeoffs.

The main one is that we’re unable to cover every company we want. At 6,000 words per piece and one piece per month, we just can’t get to every biotech, climate, aerospace, AI, and software company we’d like to. As a result, finding ways to share more companies while upholding our values is something that’s been top of mind for a while now.

This new series was one result of this.

Last November, we began turning to a few investors we respected to share their perspectives on some companies we were both excited about. The idea was that investors could offer a more direct perspective on the companies they work with, and that we could all benefit as a result.

People liked the first edition, so now we’re back.

This second edition features five companies in total, each backed by a different firm and each operating in a different industry. All are at an early stage, and all are growing quickly. At the bottom of each piece, we’ve added a few of our own thoughts on the company and why we respect the investor who’s recommending it.

This took a good amount of work from everyone involved, and we’d like to thank our contributors for participating.

We hope you enjoy :)

‎

Why You Should Join Orb

The Modern Pricing Platform

In the evolving landscape of B2B SaaS, one truth has emerged with striking clarity: the rigidity of traditional pricing models no longer serves the dynamic needs of modern businesses. This realization isn't just theoretical; it's backed by the substantial shifts in strategies by companies like Cloudflare and AI companies pivoting towards usage-based pricing to more accurately reflect value delivery and manage compute costs.

At the heart of this transformation is the critical role of billing infrastructure. Historically, billing systems were constructed in a bygone era for static, straightforward pricing plans – the evolution being limited to licensed seat increases and decreases. However, as the frontier of SaaS pricing has expanded, these legacy systems have increasingly become a blocker to innovation. They not only stall a company's ability to iterate and innovate on pricing strategies but also compromise transparency and flexibility for customers.

With the adoption of usage-based and more nuanced pricing strategies nearly doubling within B2B SaaS in the past five years, it's clear that agility in billing systems is not just an added advantage; it's an imperative.

Against this backdrop enters Orb, born from a vision shared by Alvaro Morales and Kshitij Grover, who, through their experience with billing systems at Asana, understood the pivotal role a dynamic billing infrastructure plays in a company's growth. Their firsthand experiences echoed an increasingly loud narrative in the tech industry—that the tooling to evolve pricing and packaging can be a significant determinant to a company’s growth trajectory.

Greylock led Orb's Seed round and increased our investment in their Series A. The platform stands out not just for its technical capability but for its foundational philosophy—it's built by developers for developers, emphasizing ease of use, autonomy, and rapid value delivery.

The result is an end-to-end revenue solution that makes pricing and billing more agile for every stakeholder. It empowers enterprises to experiment iteratively with pricing, customizing plans, and gleaning actionable insights from real-time customer usage data. This is not just about providing a billing solution; it's about fostering a culture of innovation and adaptability.

This culture resonates deeply with my perspective on technology's role in addressing contemporary challenges. The ability to innovate, adapt, and respond to market demands is what sets apart the leaders from the laggards. The realms of billing and pricing require an understanding of data that’s at the heart of every modern business. And a team that approaches the problem with a first-principled, developer-first approach.

Orb has already begun to redefine the billing infrastructure for modern companies such as Perplexity, Vercel, Replit, and Pinecone. Their journey signals a transformative shift in how businesses approach billing—a shift from static, opaque systems to transparent, flexible, and growth-enabling data infrastructure.

In championing Orb, we're not just investing in a platform; we're investing in the future of how businesses unlock their revenue potential. It's a testament to what's possible when innovation meets necessity, echoing my belief in the power of technology to fundamentally improve how we operate and grow in the digital age.

For those intrigued by the prospect of shaping the next generation of software businesses focusing on dynamic pricing and billing, Orb's journey represents an exciting frontier. You should consider joining the Orb team—they are hiring.

— Saam Motamedi, General Partner at Greylock

Orb is hiring talented software engineers and product managers - if you’re interested, reach out to their team here. We also recommend checking out the rest of Saam’s portfolio at Greylock. He has a remarkable track record and has previously partnered with Abnormal Security, Snorkel AI, Adept AI, and Cresta.

‎

Why You Should Join MindsDB

Customize AI with Enterprise Data

MindsDB is one of the fastest growing open source projects that is enabling developers to integrate AI and their enterprise data into their applications. MindsDB Cloud, which will be released to the public soon, will allow developers to bring the power of MindsDB open source to the cloud and will allow developers to move into production quickly.

AI will help every company thrive, but off-the-shelf, generic AI usually doesn’t completely meet their needs. The core value proposition of MindsDB is that with MindsDB’s nearly 200 integrations, any developer can create AI customized for their purpose, faster, and more securely. Their AI systems will constantly improve themselves — using companies’ own data, in real-time.

MindsDB empowers organizations to harness the power of AI by abstracting AI models as Generative AI Tables. AI tables, introduced by MindsDB, abstract AI models as virtual tables so you can simply query AI models for predictions. These tables are capable of learning from the input data and generating predictions from the underlying model upon being queried. With MinsdDB, you can join multiple AI tables (that abstract AI models) with multiple data tables (that provide input to the models) to get all predictions at once. This abstraction makes AI highly accessible, enabling development teams to use their existing SQL skills to build applications powered by AI.

Founded by Jorge Torres and Adam Carrigan, MindsDB has been deployed 100,000+ times and is backed by a large community with over 700 contributors. Some of the largest companies in the world to the fastest moving startups use MindsDB to get into production with AI quickly.

— Chetan Puttagunta, General Partner at Benchmark

MindsDB is hiring talented engineers and designers for their product and marketing teams - if you’re interested, reach out here. We also recommend checking out the rest of Chetan’s portfolio at Benchmark. Chetan has a stellar track record and has previously partnered with MongoDB, Elastic, and MuleSoft.

‎

Why You Should Join Privy

Giving All Users Simple Access to the Decentralized Web

Privy is now a key part of the default stack for developers building delightful consumer applications in crypto. Crypto will only realize its potential if we can make it more accessible and useful to mainstream users. What stands out most about this team is the level of ambition they bring to the space. They are a hard-charging, fast-moving team and have been able to create real value in a tough market because they believe (and we do too) that the opportunity to change the web through crypto is unique. Working at Sequoia, we see a lot of ambitious founders, Privy is a world-class team building a leadership position in real-time.

Privy now powers access to on-chain systems and onboarding for OpenSea, Zora, Blackbird, Friendtech and Courtyard among many others. In the past 13 months, their systems have onboarded 2.5M users and handled more than $100M in transactions.

The company’s broader mission is to help empower users on the web: making it easier for devs to build intuitive, user-respecting apps, and for users to control their identity online. Crypto offers core features that help shape this future and easing onboarding into this space is just the start of co-founders Henri Stern and Asta Li’s ambitions—they want to help build a stack that enables the future of user identity.

At the heart of their current product is a simple library that makes it easy to add beautiful authentication flows and powerful embedded wallets to any app. The setup takes only 9 minutes and helps get more users onboarded to web3 products faster. As a member of the Privy team, you get to bring web2 convenience—like sign in with email and social—to web3 products with strong privacy guarantees. This not only increases conversions dramatically, but also increases an app’s potential user base. Privy makes the hard things easy, allowing developers to focus on their product.

The Privy library integrates easily with ethers.js, web3.js and wagmi, to help scale products across all devices, wallets and browsers. Customers use Privy to deliver a world-class experience with highly customizable modals, deep web3 integrations, simple interfaces and easy-to-use out of the box flows.

Security is a top priority at Privy. Threat modeling and defense in depth are built into product requirements. Its robust infrastructure has passed rigorous security reviews and pentests, and is retested on an ongoing basis. Privy’s self-custodial architecture puts users in control of their keys, without sacrificing usability.

The Privy team has diverse experiences across web3, fintech and self-driving and have reshaped user experiences and data infrastructure on the web. They are keenly attuned to the trade-offs between convenience and decentralization online, and strive to build opinionated yet nuanced software that creates a path forward for developers. In that sense, Privy offers a rare blend of deep, technical work building a developer SDK for a highly technical audience, and true real-world impact and UX considerations, since that SDK ultimately delivers user-facing experiences. It’s a unique blend that requires you to flex muscles across technical excellence and product intuition. Privy is a global, distributed company with a core hub in New York City. Their hybrid model is remote-friendly with a culture that believes great work can happen anywhere, but also values creating a vibrant in-person culture.

I have had the privilege of working with Henri and Asta since their Seed round in 2021. Last November, Privy raised an $18M Series A led by Paradigm, with participation from Sequoia Capital, BlueYard Capital, Archetype and others. Privy is using this funding to grow the team and continue to drive the development, security and scalability of their onboarding toolkit and identity platform. If you are passionate about reshaping the crypto stack and the future of the web, Privy is hiring!

— Josephine Chen, Partner at Sequoia

Privy is hiring talented fullstack, backend, and security-focused software engineers - if you’re interested, reach out to their team here. We also recommend checking out the rest of Josephine’s portfolio at Sequoia — she’s one of the sharpest and most thoughtful investors we know.

‎

Why You Should Join General Biological

The Future of Biomanufacturing

The chemicals industry is responsible for nearly everything you touch – from plastics and solvents, to food additives and more. It is also massive: annual sales are roughly $5T worldwide, with north of 50 companies doing $10B+ in annual revenue.

But the business is fraught with structural inefficiencies. Traditional manufacturing processes suffer from low yields – half the input carbon ends up as waste or in the atmosphere – and require extreme temperatures and pressures. The sector is the largest user of industrial electricity, and consumes 1⁄3 of every barrel of crude oil. Toxic factories, plagued by costly explosions and health scandals, are located far from users, driving massive distribution costs (~20% of revenue). Yet the industry has seen virtually no disruption in its 100 year old history.

General Biological — GB — is rewriting this. The team has developed a first-of-kind technology stack that combines engineered microorganisms, novel reactor hardware, and continuous manufacturing to produce chemicals from renewable resources such as sugar, not oil. Their processes run at room temperature, with nearly pound-for-pound yields, and no hazardous waste. The high yields and ultimate process simplicity allows for low CapEx, modular plants that can be built fast and placed close to users. Taken together, GB’s breakthrough manufacturing approach – safe, simple, distributed – nearly triples margins in some of the world’s largest markets, outcompeting fossil incumbents who have been unwilling to innovate. Indeed, GB’s most important products may never be the chemicals they make, but rather the factories themselves.

In short: GB is ushering forward the era of the biorefinery, doing for bioproduction what ExxonMobil did for petroleum a century ago. The team combines experiences in chemistry, biotech, hardware, and software to introduce the world to what biomanufacturing will be.  If you’d like to be on the forefront of the biggest shift in the material economy since crude oil and work with a team with world-changing chemistry (!), reach out.

— Kevin Hartz, General Partner at A*

General Biological is hiring talented molecular biologists, chemists, and chemical engineers - if you’re interested, reach out to abhi@generalbiological.com. We also recommend checking out the rest of Kevin’s portfolio at A. Besides being a seed investor in Airbnb and Pinterest, Kevin is himself an experienced operator - he was previously Co-Founder and CEO of both Eventbrite and Xoom (acq. Paypal).

‎

Why You Should Join Tractian

Empowering Industry Workers with AI-Assisted Maintenance Solutions

Manufacturing is the heart of the global economy. While technology provides huge, compounding advantages to those with access to software, hardware and data, manufacturing professionals have been largely left behind. Once overlooked, innovation in this space is accelerating as more and more attention is paid to the ‘real economy’ and for good reason, as manufacturing alone constitutes 12% of US GDP.

The question – “when should I take my machine offline to conduct maintenance on it?” seems simple, but in reality involves several complex trade-offs. Manufacturing downtime includes both planned downtime due to scheduled asset maintenance as well as unplanned downtime due to human error, equipment failure etc. The impact of unplanned downtime is massive – for Fortune 500 firms alone, the cost is ~$1.5T or 11% of their annual revenues. Today, the majority of plants rely on reactive maintenance meaning they only perform repairs after equipment fails which further exacerbates the issue of unplanned downtime.

The unique combination of an inspired team with domain expertise, deep customer empathy, and world-class engineering talent led the Tractian team to set out to eliminate unplanned downtime and provide technicians with comprehensive oversight of their operations. Always a step ahead, Tractian’s founders were immersed in the world of maintenance far before the founding of the company in 2019. As the son of a factory worker and maintenance technician, Igor Marinelli set out with co-founders Gabriel Lameirinhas and Leonardo Vieira to revolutionize the role of the maintenance professional by equipping them with technology.

Through intelligent sensors affixed to industrial assets, Tractian can collect real-time data such as vibration, temperature and current that is used to identify operating issues and predict mechanical and electrical failures. Rather than simply alerting operators when values deviate from the norm, the company leverages AI to understand exactly why values have deviated and articulates prescriptive actions operators can take to fix issues and prevent downtime.

Tractian counts Kraft Heinz, Procter & Gamble, Caterpillar, American Wood Fibers and several others as happy customers. The company has seen incredible growth over the past year and has no signs of slowing down anytime soon. In fact, they’re just getting started. With nearly 300 employees spanning offices in three countries, Tractian’s streamlined hardware-software solutions are no doubt making a global impact. By harnessing the power of data and AI, Tractian is not only revolutionizing the role of the maintenance professional but also reshaping the future of manufacturing, ensuring a more efficient, productive, and resilient global economy.

— Trevor Oelschig, Managing Director at General Catalyst and Vedant Suri, Investor at General Catalyst

Tractian is hiring talented software engineers, hardware engineers, and GTM professionals for their teams in the United States, Mexico, and Brazil - if you’re interested, reach out to to their team here. We also recommend checking out the rest of Trevor and Vedant’s portfolios at General Catalyst. Trevor has a great track record and has previously partnered with Aura, Fivetran, and Contentful.

‎

Conclusion

If you’re open to new roles or just curious about their products, MindsDB, Orb, Privy, General Biological, and Tractian are all worth checking out.

In case you missed it, last quarter’s edition featuring suggestions from Sequoia’s Bogomil Balkansky, Index’s Shardul Shah, 8VC’s Alex Kolicich, Neo’s Ali Partovi and Suzanne Xie, and Pear’s Arash Afrakhteh is worth reading through as well.

We hope you’re enjoying the start of spring. See you next month!

‎

Thanks to Josephine, Chetan, Saam, Trevor, Vedant, and Kevin for their help with this piece.

In case you missed our previous releases, check them out here:

• Warp - Revolutionizing the most foundational developer tool.

• Pinecone - Vector databases index the AI revolution.

• WorkOS - The API to increase your TAM.

And to make sure you don’t miss any future ones, be sure to subscribe here:

Finally, if you’re a founder, employee, or investor with a company you think we should cover please reach out to us at ericzhou27@gmail.com and uhanif@stanford.edu - we’d love to hear about it :)

• Meter just raised a sizable round from Sam Altman and Lachy Groom after more than tripling revenue, square footage, and deployed networks. They’re actively hiring for several functions, including a number of engineering, product, and GTM roles that aren’t yet publicly listed. Reach out to jobs@meter.com if you’re interested in any roles not on their website.

• Etched has built a chip that can run transformers >10x faster than GPUs. They’re currently hiring Machine Learning Researchers, Compilers Engineers, Verification Engineers, Firmware Engineers, RL Design Engineers, interns, and more in Cupertino.

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information we’ve received from venture firms we work with. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should Join Meter

(Click the link to read online).

Do you remember the last time you got mad at your computer?

Like really mad.

Maybe you accidentally refreshed a web form and lost all your progress. Maybe you were on an important Zoom call and it kept cutting out. Maybe you were watching a movie on Netflix and it wouldn’t stop buffering. Maybe you were trying to withdraw from FTX.

If you’re like most people, though, chances are it had something to do with slow internet.

There isn’t much that gets more people worked up.

A man in Singapore recently tried to strangle his landlord over slow Wi-Fi. A man in the Philippines used a hammer to smash his router for the same reason. A 90-year old in America recently spent $10,000 on 2 ads in the Wall Street Journal to tell the CEO of AT&T his internet sucked, complaining that streaming a movie on Roku was like watching a “slideshow.”

Enough people have complained about Comcast, the largest home internet service provider in the country, that they at one point had a lower customer satisfaction rating than the IRS.

While we can’t condone shooting your computer or throwing it out a window when your video keeps buffering, we also can’t downplay the importance of good Wi-Fi these days. The way things work today, having crappy internet impacts everything from

• Education - students without home internet access are usually half a GPA point below students with access. This ripples out to longer-term impacts on academic ability, college admissions, and career opportunities.

• Economics - a 2010 study of OECD countries found doubling broadband speed has the effect of boosting GDP growth by 0.3% a year. In the US, average broadband speeds going from 10 Mbps in 2010 to 174.2 Mbps in 2020 had the effect of boosting GDP by $1.3 trillion.

• Health - 10% of UK adults experience stress due to slow or unpredictable internet connections. One study found delays when streaming video can increase heart rates by 38%, and another found almost half of people lose their temper over laggy (>60 second delays) technology each week.

With the internet now serving 66.2% of the global population and handling 456 million terabytes of traffic a month (a figure that’s growing 20.5% each year), we rely on it for enough things that last year, the UN High Commissioner for Human Rights said “it may be time to reinforce universal access to the internet as a human right.” A few months later, the chair of the FCC agreed, proposing rules that would reclassify broadband access as an essential service like power or water.

But if there’s a reason networks aren’t as reliable as power or water, it’s because they’re not nearly as easy to set up and maintain.

Especially for the larger ones you’ll find in schools, offices, and warehouses, networks are harder to manage for a few reasons:

• Nothing comes pre-installed. Wi-Fi isn’t like power or water, which comes ready with any space you’re moving into. The complexity, cost, and customizability of networks means they’re almost always left to individual tenants to set up. The choice of provider, network design, capacity requirements, security requirements, hardware, management tools, etc. are all things companies have to decide on, implement, and maintain themselves.

• There are more moving parts. Running a network means picking an ISP, acquiring the right hardware, running cable, installing/configuring that hardware, troubleshooting issues, and upgrading/scaling things as needed. There are many more vendors, parts, and stakeholders involved than with traditional utilities.

• They are dependent on your specific space. The design of your network depends on the physical layout of your building and where your devices will be located. A warehouse with racks and robots will need a different network than an office with desks and meeting rooms, even if the square footage is similar. You have to account for things as specific as not putting access points near microwaves, elevators, or bathrooms.

If you’ve ever set up a network yourself at home, you know how much fiddling is needed to get things working right. Imagine how much worse it is when you’re doing it for orders of magnitude more people across orders of magnitude more space, especially if you still don’t have much networking experience.

For something so important, it’s surprising that the experience of setting up and managing a network hasn’t been more abstracted away — the way every other essential utility has.

This, in a nutshell, is what Meter is trying to do.

Background

Meter offers a fully-managed network as a service.

In the status quo, moving into a new office, warehouse, or school means setting up a network from scratch. This means finding and negotiating with ISPs based on your bandwidth requirements, designing a network based on your floor plan, picking and procuring routers/switches/access points based on your design, and installing/configuring/testing everything at the end. After setup, maintenance means troubleshooting issues as they come up (i.e. why did my Zoom call drop?), tuning and optimizing your network in response (i.e. realizing it’s because there are too many devices on that AP’s frequency band), and adding security measures as needed (i.e. DNS filtering or Firewalls). For organizations without network engineering expertise, this means hiring a constellation of consultants, Managed Service Providers (MSPs), and subcontractors — or a lot of pain for their own IT team.

For these organizations, Meter offers what’s essentially a network in a box. For a flat monthly fee, they’ll handle everything from setup to maintenance to support, using their own hardware, software, and staff to make it happen. In doing so, Meter abstracts away everything related to:

• Setup. This includes ISP procurement & installation, custom network design, hardware provisioning and installation, low voltage cabling, full IT closet/server room buildouts, testing and configuration, site surveys and project management - everything needed to get the network running.

• Maintenance. This includes preemptive network management/optimization, cloud dashboards for manual troubleshooting, and ongoing support/maintenance via a 24/7 operations team which helps with network outages, growth/scaling requirements, and config management/updates - everything needed to keep the network running.

Enabling this is a proprietary hardware (security appliances, switches, routers, and access points), software (dashboard, security tools), and firmware stack developed by Meter in-house. By achieving complete vertical integration — both producing their own technology and installing/managing it — Meter has more control, visibility, and understanding of their client networks than a vendor who just made the hardware or an MSP who just installed it ever could.

Concretely, this has allowed them to create value over traditional service providers in two ways:

• Better service. Meter isn’t just there when you file a ticket - their complete understanding/control over your network means they’re continuously optimizing it based on usage patterns to prevent issues in the first place. Having one party behind your hardware, network design, and maintenance also makes it easier in case you do need to escalate - there are no information siloes to overcome.

• Better pricing. Beyond eliminating upfront capital costs, Meter is actually cheaper than buying hardware separately and working with MSPs to install/maintain it. The fewer middlemen, incident prevention, and more efficient incident resolution enabled by vertical integration allow Meter to offer their flat pricing model while still maintaining strong margins.

Meter’s flat monthly fee is truly flat - they charge per square foot per month, making them more similar to S3 than traditional hardware vendors. Everything from hardware upgrades to network design to installation is included in their subscription — there are no one-off, special costs. In many cases, their pricing is even easier to understand than traditional utilities.

Meter was founded in 2015. After spending a few years developing their own hardware and software stack, the company began rolling out their first few deployments in 2019. The company has grown rapidly since - Meter was working with hundreds of customers by 2022 and has more than tripled revenue, square footage, and deployed networks since. Today, the firm serves tens of millions of square feet and supports customers across all industries: startups like Brex, real estate companies like Tishman Speyer, and entire school systems like Higher Ground Education’s 165 locations all rely on Meter for their networking. Customers love them: GoBolt, a fulfillment and logistics company, reduced their total networking costs across 14 fulfillment centers by 25% with Meter. Other IT teams have reported an 80% decrease in support tickets. A Director of IT & Security told us switching to Meter for his unicorn company’s new headquarters was a “no-brainer.”

The company was founded by brothers Anil and Sunil Varanasi. Anil studied network engineering in college, and while in school started a sizable business building network systems with his brother. After selling the business, the two realized there was an opportunity to re-imagine the networking experience from the ground up, stemming from personal pain points setting up all those networks. Deciding to start with the hardware, the two moved to China for a month to begin experimenting with that. They ended up staying for a full year.

To fuel their recent growth, Meter just raised a fresh $35 million dollars from Sam Altman and Lachy Groom. Previously in 2022, they had raised a $38 million dollar Series B from Sequoia and Lachy Groom. Before that, they had raised a $12 million dollar Series A, also from Sam Altman and Lachy Groom. Along the way, they’ve also assembled quite the roster of angels: Reid Hoffman (LinkedIn), Diane Green (VMWare), Sanjit Biswas (Meraki/Samsara), Tobi Lutke (Shopify), David Solomon (Goldman Sachs), Egon Durban (Silverlake), and the Collison brothers (Stripe) are all investors. As it turns out, the pain of network engineering is something almost anyone can empathize with.

Meter has a strong team and a sophisticated offering that, for the first time, promises to take all the pain out of running an enterprise network. Their mission is to make Wi-Fi as much of a utility as power or water — something you can just turn on or off when you move into a new office.

Will they be able to pull it off?

Networking is one of the biggest markets in all of technology. It’s a tough, crowded space with a lot of players, none of which have been able to offer the experience Meter’s promising so far.

Well, we think they’ve got a shot.

Let’s dive deeper into why.

‎

The Opportunity

We’ll begin our exploration of Meter like we do for all of our companies, from a first principle:

Large companies lead large, growing markets.

Many startups with great teams and solid growth don’t get very big because there just aren’t enough customers to grow into. As a result, we’ll want to show two things:

1. Meter operates in a large, growing market.

2. Meter can become a major player in that market.

The Market

Meter operates in the market for enterprise network hardware and services. They make money when companies who need help with an office, warehouse, or in-store network outsource those needs to Meter. Since every company needs a network but relatively few have the network engineers to support it, Meter must operate in a large market.

Simple, right?

Well, we can be a bit more precise.

While Meter serves businesses of all sizes and industries today, the customers who most strongly feel a need for their services are mid-market to enterprise firms with small IT teams but multiple site locations - think a large university system, clothing brand, or tech company, all of whom we’ve seen Meter serve. These are the companies that want their internet to just work, but don’t have the resources to work on it.

Today, most of these organizations are working with one or several managed service providers (MSPs) and hardware vendors to help design, implement, and maintain their network. As the purest proxy for Meter’s market, then, we might ask how much businesses are spending on such managed service providers today.

Empirical estimates vary, but people generally agree it’s a lot:

• One estimate suggests global annual spend on Managed Network Services will grow from $65.7 billion in 2023 to $89.9 billion in 2028, at a CAGR of 6.5%. Another suggests global annual spend on the broader Infrastructure Implementation and Managed Services category was $318.2 billion in 2022. Another valued the even broader global IT Managed Services market at $472 billion in 2023, up 12.7% from from $419 billion in 2022.

• A survey of 1,400 IT professionals across North America, EMEA, and Latin America found managed services was the only budget category that grew between 2021 and 2023, rising from 16% to 18% of total IT spend - hardware, software, and cloud-based service spend were all proportionally down.

• In 2024, 84% of MSPs predict revenue to increase, with 77% predicting profit to increase also. A survey of 500 MSPs found average MSP revenue grew from $8.93 million in 2021 to $12.25 million in 2022, with the average one forecasting further growth to $16.29 million in 2023.

At a higher level, overall spend on IT services continues to grow, jumping 8.7% from 1.38 trillion in 2023 to 1.5 trillion in 2024. Notably, this is greater than the 6.8% overall IT spend is projected to grow at - across nearly every industry, CIO spending for managed services is growing faster than internal spending on IT talent, which is approximately 3%. With salary growth expectations closer to 6%, it’s clear that companies are spending more money on fewer employees and outsourcing the rest: Gartner predicts spending on consulting for IT work will grow by 10% to 15% across almost every industry this year. The market for outsourced IT help was already big, but broader industry trends mean MSPs (and Meter) are absolutely feasting right now.

But Meter is far more than your typical MSP: they also make their own hardware.

As a result, Meter is competing with traditional vendors (Cisco, Juniper, Aruba) in the network hardware market as well. This, too, is a sizable opportunity: one estimate has the global Enterprise Network Equipment market growing from $72.44 billion in 2024 to $124 billion in 2029, at a CAGR of 11.35%. Another estimate has it growing from $101 billion in 2023 to $133 billion in 2027, at a CAGR of 9.1%.

Network as a Service

More broadly, the trend Meter has helped define and lead is that of “Network as a Service” — offering customers a complete, consumption-based package of hardware and services that abstracts everything else away. Beyond being lower-overhead for organizations without networking expertise, NaaS’s benefits include better pricing, quicker deployment, and easier scaling.

The numbers on NaaS are early and less well-defined, but still promising nonetheless:

• One report suggests NaaS will be adopted by 15% of all enterprises by the end of 2024, up from less than 1% in 2021. Another found that 49% of IT leaders are planning or considering a NaaS transformation for their organization, with 69% of interested groups planning to invest within 12 months. A third found that 47% of enterprise technology leaders planned to adopt NaaS, with 10% saying they had transitioned already and just 14% saying they had no plans to transition at all.

• With regards to benefits, one survey found that 76% of companies thought NaaS could help reduce operations costs, 60% thought it could enable a shift from CapEx to OpEx, 65% thought it could help enhance security, and 57% thought it would free up time for innovation and strategic initiatives. Another survey found 62% of IT leaders were considering it because of enhanced network security, 58% because of optimized network performance, and 48% because of guaranteed service levels. Interestingly, while IT is the department that would most benefit from NaaS, 46% of operations and 40% of engineering teams surveyed were pushing for it also.

As a result of all this, one estimate has the NaaS market growing at a CAGR of 32% to $78.38 billion by 2028.

Tailwinds

The rise of NaaS is a part of this broader growth in managed services/outsourcing we observed earlier.

Why has this been happening?

There are many reasons, but a major one has been necessity - the proliferation of cloud deployments means there just aren’t as many people with networking experience as there used to be:

• 86% of US-based CIOs expect at least 25% of their network engineers to retire within the next five years. 95% of CIOs have reported that a shortfall of network engineers had led to an inability to manage networks, something which 91% of network engineers agreed with.

• More broadly, 59% of IT teams have found it “somewhat” or “very” challenging to find skilled IT talent. IT employment has been gently declining for a while now, at one point contracting for 23 consecutive months between January of 2022 and January of 2024.

Almost by necessity, companies have been outsourcing their network-related responsibilities - it’s often too difficult to hire or retain the people they’d need otherwise. It makes sense why something like Network as a Service would be taking off right about now.

‎

The Strategy

So we know companies are looking for help managing their networks. The question now is whether Meter will be their solution of choice.

Meter’s customers — mid-market to enterprise companies with small IT teams but lots of network to manage — judge Meter based the quality of network they provide and their service in doing so.

In this, Meter competes with both the MSPs these organizations would’ve otherwise outsourced their networking needs to and the vendors those MSPs would’ve otherwise sourced their network hardware from.

Compared with them, Meter’s vertically integrated approach allows them to offer a better service at a better price. Let’s look at how Meter is usually better on both fronts, and why much of it is uniquely enabled by their vertical integration.

Pricing

Working with traditional MSPs usually means paying sizable upfront costs for network design and setup (i.e. for hardware, site surveys, network design, low voltage cabling, etc.), as well as a number of recurring and one-off fees for support, maintenance, and upgrades.

Meter, in contrast, charges per square foot per month, inclusive of hardware, setup, and upgrades — their model involves zero upfront costs.

Why is this approach better?

For one, it’s more predictable. Meter’s pricing allows companies to move network-related costs from CapEx to OpEx, making it easier for IT teams to plan, budget, and procure upfront. Since everything is included, customers don’t have to worry about finding budget for one-off repairs or upgrades going forward either.

More importantly, though, this model also ends up being cheaper.

Most MSPs would charge a premium for the sort of turnkey, project-managed bundle Meter is offering, but Meter has never had a customer where the traditional vendor + MSP combo would’ve been cheaper. Meter has always been more cost effective.

For instance, we went through the numbers with one IT director who was deciding between Meter and some MSPs for their new, 40 thousand square foot office in San Francisco. Compared with what the MSPs were quoting for setup alone — not even including annual fees for support, maintenance, and upgrades — Meter’s first year price was already 5x lower. Faced with a >5 year breakeven and the prospect of avoiding a sizable upfront capital commitment, going with Meter was a “no-brainer”.

How is this possible? Is Meter just another startup getting subsidized by VCs?

It’s because they handle everything themselves.

Meter’s able to beat everyone on price while maintaining margins because with them, you aren’t dealing with a large number of vendors, subcontractors, and middlemen, each with their own hidden costs/fees/margins to take care of — since Meter builds their own hardware/software/firmware stack, they don’t need to work with anyone else. Additionally, since they design their own networks and deploy their own hardware, support (upgrades, resolving tickets, etc.) can be offered far more efficiently — Meter is already an expert on the networks of all their clients, and there are no knowledge silos to be broken down.

Another reason is the growing way they’ve been acquiring customers.

When Meter customers move to new spaces, they leave their old Meter setups behind. Then, when new companies move in, instead of going through the pain of setup, all they have to do is call to turn things on — just like a traditional utility. Meter has already acquired dozens of customers this way, and it’s brought customer acquisition costs down significantly.

In some deeper sense, when setting things up Meter acquires both the customer and the space itself. Owning the space is a durable advantage compounds with time — we see it becoming an interesting moat of sorts in the future. Physical network effects (i.e. railroads, telephone lines) are the strongest moats in business, and Meter is starting to build one.

Services

Working with traditional MSPs usually means hiring separate groups or sub-contractors for different network-related subtasks — think network design, cabling and installation, upgrades, or network security. Meter, in contrast, offers an integrated service with installation, support, optimization, and management all covered — instead of dealing with several parties, you only deal with one.

Why is this approach better?

Having one owner for everything makes resolving issues much more efficient. If the MSP helping maintain your network isn’t familiar with a design or hardware choice made by a different party, it becomes much harder for them to troubleshoot an issue potentially stemming from it. With good MSPs this will result in a longer resolution time, but with bad ones this can result in blame games between the vendor, the configurator, and the ISP, where no one party wants to take ownership.

Since there are no such knowledge or ownership silos in a network managed by Meter though, this isn’t an issue.

More concretely, Meter’s vertically integrated approach improves support on three levels:

• Preventative maintenance. Traditional MSPs operate reactively, making changes and improvements when they receive complaints or network alerts from the tools they use — a.k.a when something breaks. In contrast, Meter operates proactively, preemptively optimizing network designs and infrastructure with software to avoid failures in the first place. For instance:

  • If an access point goes out, the system is immediately alerted to increase the strength of a neighboring access point.

  • Wi-Fi channels are automatically adjusted to prevent overlap, and transmit power/bandwidth are automatically adjusted to optimize connectivity.

  • When upgrading the network from WiFi 5 → WiFi 6 → WiFi 7 (as Meter customers will have experienced from 2022 → 2023 → 2024), Meter’s system automatically readjusted the physical layout and software configurations of access points to optimally use the benefits of the new standards and hardware.

  Meter is constantly improving their network’s “operating system” in this way. As a result, some customers have seen an 80% drop in tickets since switching over.

• Self-maintenance. When there’s a need for it, Meter’s software also makes it easier for customers to adjust and fix things themselves. Since there’s a direct line of feedback between Meter’s end-users and Meter’s engineers (in contrast to traditional vendors, who are separated from end-users by MSPs), Meter can build features and adapt things in response to common feedback. For instance:

  • When expanding into the education market, Meter heard from many users about the importance of content filtering (i.e. preventing students from using TikTok). In response, Meter worked with Cloudflare to launch a DNS Security feature. Previously, IT teams would’ve had to use a separate vendor for this, but now, Meter customers could simply can turn it on within their dashboard, no extra contracts or payments required.

  • Similarly, many Meter customers manage a number of sites remotely — think a large brand with many storefronts, each with their own networks. For these customers, the ability to understand, reconfigure, and root cause issues while remote was very important. In response, Meter developed their “digital twin” concept - a fully virtualized version of a network that’s updated in real-time, down to individual ports and switches. In addition, they went and added backup LTE antennas to their controllers so that even if ISP circuits went down, things could still be accessed remotely. If you’ve ever wanted to visit an IT closet through a web browser, this is the closest you’re going to get.

  

  

• Support-led maintenance. Between their ownership of the entire service stack (ISP relations, network design, network setup, network maintenance) and the entire product stack (hardware, firmware, software), Meter is more familiar with the networks of their clients than any MSP or vendor ever could be. As a result, this makes it more efficient for Meter to diagnose and fix issues that have been escalated to them. For instance:

  • If a laptop has an issue related to Wi-Fi, Meter customers typically just give Meter that computer’s address. Since Meter’s support team has access to unified logs across every level of the stack, they’re able to troubleshoot what happened with the laptop’s connection relatively easily. This is much faster than the IT team parsing log events and stitching together interpretations themselves.

  • In the Fall of 2022, Meter had an uptick in customers complaining about slow internet connection and dropped Zoom calls. Meter’s QA and engineering teams weren’t able to find any issues in Meter hardware and software. Since Meter manages hundreds of networks though, they were able to quickly diagnose that the issue was specific to Macbooks and push an interim fix for their customers. The root cause was an update Apple pushed causing Macbooks to use AWDL (Apple Wireless Direct Link), a Wi-Fi interface which caused connections to periodically reset.

The end of the last point is worth double clicking on: MSPs offer a lot more than just network services and often don’t specialize in them. Instead, they help clients with many different IT-related tasks, from managing security to SaaS tools to compliance. In fact, 85% of MSPs planned to add between 1 and 8 of services last year, with the average adding 5 overall.

In contrast, Meter only does networking.

‎

Competitive Landscape

Networking is one of the largest markets in all of tech.

Now that we have an understanding of Meter’s unique strengths, let’s see how they compare against potential competitors.

Hardware Vendors and MSPs

We’ll start by focusing on the vendors and MSPs people today are outsourcing their networking needs to.

• Vendors are the firms that provide the hardware — think Cisco, Juniper, Aruba, Arista, Fortinet, and Ubiquiti. These firms produce the routers, switches, access points, and controllers needed for networks to run, as well as the software needed for IT teams to manage them. They usually sell their products directly or via channel partners, leaving their configuration and implementation up to the MSPs or IT teams actually in charge of designing and installing the network.

• MSPs are the firms that provide the service. They generally operate more locally — the SF Bay Area, for instance, has TruAdvantage, ZAG Technical Services, Parachute Technology, ComputerLand of Silicon Valley, and OneClick Solutions Group, among others. One estimate suggests there are up to 20,000 profitable MSPs in North America.

Since Meter provides both hardware and services, they end up competing with both. To recap how Meter’s vertical integration gives them an edge:

• Offering both hardware and services under one roof is more economical - there are no subcontractors or partners with their own fees and margins to take care of. Think Tesla vs Ford.

• Compared with vendors specifically, Meter’s vertical integration results in a better alignment of incentives. Traditional hardware vendors are focused on selling boxes, whereas Meter is focused on selling working Wi-Fi. As a result, Meter’s hardware, management software, network designs, and service playbooks all work together to create better outcomes, which isn’t something in scope for traditional vendors.

• Compared with MSPs specifically, Meter’s vertical integration results in less siloing of information and ownership. This makes it easier for the provider (Meter) to both proactively optimize networks and to resolve issues that have been escalated to them.

NaaS Offerings

NaaS offerings present a more direct competitor since they (in theory at least) also offer everything under one roof. This isn’t a super crowded market since it takes a lot to offer everything needed to run and continuously support a network — it took Meter several years before they were ready for their first deployment, and several more before they were mature enough for some of their larger customers today. As a result, most of the players here are hardware vendors who already had extensive product portfolios. Let’s have a look at the major offerings.

• Nile - this is the main other startup in the space. They’ve raised $300 million in total, including a $175 million Series C last year from March Capital and Sanabil Investments to begin focusing on the Middle East and Europe. Nile is a promising competitor in this new and growing market, and like Meter is growing quickly.

  

• Cisco+, Juniper NaaS, Greenlake - these are the NaaS offerings of traditional vendors. Cisco runs Cisco+, Juniper Networks runs Juniper NaaS, and HPE/Aruba runs Greenlake. Compared with Meter, these offerings:

  • Aren’t vertically integrated. These firms produce their own hardware, but contract the service component out to MSPs instead of using their own staff. As a result, these NaaS offerings aren’t much more than different payment models — since they’re not vertically integrated like Meter, they lack the pricing and service advantages Meter has. They still have knowledge silos, fragmented ownership, and subcontractor margins to worry about.

  • Aren’t widely available. For instance, Cisco+ is actually only available for datacenter applications and isn’t actually available for networking yet. Anecdotally, HPE is being “selective” about who they’re offering Greenlake and its NaaS-based pricing structure to.

  

• The issue? All these companies already have sizable market share and are making good money upfront with the current model. As a result, they have no incentive to switch their customers over to a NaaS-based payment plan. Additionally, services are a skillset they don’t have much experience with, which is why they’ve been forced to partner with MSPs. As a result, none of these offerings are as integrated as Meter.

‎

Execution

Meter has a strong, diverse team working together to build the future of networking. Consider a sampling of their roster:

Apurv Bhartia - Head of Wireless

Apurv was previously a Senior Engineering Manager at Cisco Meraki. Before that, he was a Research Intern at Microsoft and a Software Engineer at Huawei. He holds a Ph.D. in Computer Science from UT Austin.

Nilesh Hirve - Head of QA & Technical Support Engineering

Nilesh was previously a Director of Software Engineering at Cisco Meraki. He holds an M.S. in Telecommunication and Network Management from Syracuse University.

Abhijeet Kumar - Head of Software Engineering

Abhijeet was previously an Engineering Lead at Nextdoor. Before that, he was a Staff Engineer at Workspot and a Software Engineer at Juniper Networks. He holds degrees from IIT Guwahati and the University at Buffalo.

Sarah Ahmad - Head of Customer Success

Sarah was previously Director of Enterprise Account Management, West Coast at WeWork. Before that, she was an Account Manager at Vendormate. She holds a degree in Marketing from the University of Georgia.

Sarah Quon — Head of Marketing

Sarah was previously a Product Marketing Lead at Opendoor. Before that, she was a Product Manager at Sephora. She holds a B.S. in Business Administration from USC.

James McKenna — Head of Finance

James previously held Finance and Strategy roles at Stripe. Before that, he was a Vice President at Serent Capital and an Associate at FFL. He holds an MBA from Stanford GSB and a B.A. in Economics from Princeton.

Danyal Kothari - Head of Operations

Danyal was previously a Product Manager at Uber. Before that, he was a Senior Associate Consultant at L.E.K. Consulting. He holds a B.A. in Chemistry from the University of Pennsylvania.

Pike Hughes - Head of Product Operations

Pike previously led charging infrastructure for North America at Tesla. He holds a B.A. in Economics & International Relations from Tufts.

David Richards - Head of Product Marketing

David previously held Product Marketing roles at Moveworks and Domo. Before that, he was a Strategy Consultant at Accenture. He holds an MBA from Northwestern and a B.A.Sc. in Bioengineering and Biomedical Engineering from USC.

Jeff Leopando — Head of Supply Chain Management

Jeff was previously a Senior Technical Program Manager at Facebook. Before that, he was a Project Manager at Apple. He holds a B.A. in Anthropology from Harvard and an MEng in Logistics, Materials, and Supply Chain Management from MIT.

Greg Cooper - Product Designer

Greg was previously the first product designer at Stripe, where he was for ten years. Before that, he was a YC founder.

Sean Rose - Product Manager

Sean was previously a Product Manager at Slack. Before that, he was a Product Manager at Box. He holds a B.S. in Symbolic Systems from Stanford.

Samrat Gupta - Software Engineer

Samrat was previously a Member of the Technical Staff at Cisco Meraki. Before that, he was a Software Engineer at Cisco. He holds an M.S. in Embedded Software Engineering from Carnegie Mellon.

Andrew Thorp - Software Engineer

Andrew was previously a Software Engineer at Stripe. Before that, he was a Software Engineer at Pure Charity. He holds a B.S. in Computer Science from Liberty University.

Eric Cheng - Software Engineer

Eric was previously a Senior Software Engineer at Uber. Before that, he was a Program Manager Intern at Microsoft. He holds a BASc in Systems Design Engineering from the University of Waterloo.

Andrew Kramolisch - Software Engineer

Andrew was previously a Senior Engineering Manager at EasyPost. Before that, he was a Senior Software Engineer at Airbnb. He holds a B.S. in Computer Science from Virginia Tech.

Isaac Diamond — Software Engineer

Isaac was previously a Software Engineer at Stripe. Before that, he was a Production Engineer Intern at Facebook. He holds a B.S. in Computer Science from UC San Diego.

Paul Ugolini — Software Engineer

Paul was previously a Firmware Engineering Intern at Samsara, a System Software Intern at Nvidia, and a Hardware Design Engineer Intern at Thimble.io. He holds a BS in Computer Engineering from Rochester Institute of Technology.

Shubham Saloni - Wireless Engineer

Shubham was previously a Wireless Software Developer at HPE Aruba. Before that, she was an R&D Embedded Software Engineer at Broadcom. She holds an MS in Computer Systems Networking from NC State.

Simon Symeonidis — Embedded Systems Engineer

Simon was previously a Software Engineer at AdGear. Before that, he was a Software Engineer at Aeras Inc. He holds an M.S. in Computer Science and a B.Eng. in Computer Software Engineering from Concordia University.

Hiram L. — Embedded Systems Engineer

Hiram was previously a Platform Firmware Architect at Avast. Before that, he was a Software Engineer at Remotium. He holds a B.S. in Computer Engineering from Cal Poly SLO.

Abhinav A. - Platform Engineer

Abhinav was previously a Staff Software Engineer at Samsara. Before that, he was a Senior Staff Engineer at Argosy Labs and a Member of the Technical Staff at Cisco Meraki. He holds an M.S. in Computer Science from Indiana University Bloomington.

Drew Sutton — Sales

Drew was previously an Account Development Executive at Salesforce. Before that, he was a Data Analytics Manager at IRA Group. He holds a B.S. in Business Administration from Georgia Tech.

Girish Gopalan — Strategic Partnerships

Girish was previously Head of Strategy and Operations for Stripe Issuing at Stripe. Before that, he was an Investor at King Street Capital Management. He holds a B.S. in Economics and a B.A. in Biology from Wharton/UPenn.

‎

Conclusion

You’ll probably never stop getting mad at your computer.

If Meter’s successful, though, in the future it won’t be because of your internet.

If the 19th century brought us water as a utility and the 20th century brought us electricity, the 21st will probably bring us connectivity.

Meter’s hiring.

‎

Thanks to Lachy and Frank for their help with this piece.

In case you missed our previous releases, check them out here:

• Warp - Revolutionizing the most foundational developer tool.

• Pinecone - Vector databases index the AI revolution.

• WorkOS - The API to increase your TAM.

And to make sure you don’t miss any future ones, be sure to subscribe here:

Finally, if you’re a founder, employee, or investor with a company you think we should cover please reach out to us at ericzhou27@gmail.com and uhanif@stanford.edu - we’d love to hear about it :)

• Etched has built a chip that can run transformers >10x faster than GPUs. They’re currently hiring Machine Learning Researchers, Compilers Engineers, Verification Engineers, Firmware Engineers, RTL Design Engineers, interns, and more in Cupertino.

• Sponsored - Our friend Anthony recently announced the third cohort of NEXT, a curated (~10% acceptance rate) community of entrepreneurial mid-career and exec-level folks thinking about their next roles. NEXT offers events with top VCs, workshops with career coaches, and a collaborative search experience with well-networked peers. We recommend joining the upcoming cohort, which starts in late Feb - mention us in your application for an expedited review.

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information we’ve received from venture firms we work with. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should Join Etched

(Click the link to read online).

If you’ve worked at a tech company of any sort this past year, you’ve probably heard about the recent GPU shortage.

Demand for generative AI has grown way faster than the supply of chips enabling it, and early on this left everyone wanting try it waiting in long lines.

At a private event last May, Sam Altman noted that OpenAI’s short-term product plans were severely GPU-limited, delaying the release of features like their fine-tuning API or GPT-4’s 32k context window. At an interview last September, the chairman of TSMC (the manufacturer of Nvidia’s H100 and AMD’s MI250) said they were only filling 80% of customer needs and that full demand couldn’t be met for another year and a half. As a result of all this, Nvidia was taking two to three months to fulfill new orders, companies were waiting several months to rent large blocks of GPUs, and Azure was even offering refunds to customers with unused GPU capacity so they could re-sell it to people who needed it more.

At one point, a16z estimated that demand for compute outstripped supply by a factor of 10.

There had never been a better time to be a chip company.

So naturally, every company became a chip company.

Two VCs bought 3000 H100s (at that point more than what Microsoft’s entire research arm had) for their portfolio companies to use. A crypto mining company raised 200 million of debt to buy H100s to rent, using the H100s themselves as collateral. A Colorado-based cannabis consulting company merged with a 2-month old, 0-employee data center firm called HyperScale Nexus to form a GPU hosting business that had somehow secured 30,000 H100s.

Satya Nadella, the visionary CEO of Microsoft, once observed that “every company is a chip company” (or something close).

Crucially, all these chips have been rented, purchased, or stolen to run one thing: transformers.

Transformers are the deep learning architecture used by GPTs, Llamas, Mistrals, Codexes, and Whispers — all of today’s most popular models. The architecture has become dominant in every natural language processing task (i.e. text generation, machine translation, summarization) and critical in problems as varied as computer vision (i.e. being used in stable diffusion’s encoder or in GPT-4’s vision module), audio processing (i.e. with Whisper), and protein folding (i.e. with Alphafold). Whereas before people were using specialized architectures for different tasks (i.e. CNNs for image recognition, LSTMs for text generation), transformers have sort of found their way into everything — by 2022, 70% of AI-related arXiv papers were mentioning the term.

Why this has happened is beyond the scope of this piece, but in broad (oversimplified) strokes it’s because transformers are exceptionally:

• Generalizable. The attention mechanism transformers use excel at capturing relationships and long-range dependencies between elements in a sequence, something which a large class of machine learning tasks benefit from (i.e. text can be modeled as a sequence of words, images as a sequence of patches/features, audio as a sequence of MFCCs).

• Scalable. Transformers scale more naturally (i.e. stacked encoders/decoders, multiple attention heads) and parallelize better than traditional architectures. This has allowed them to grow very large very quickly, compensating for their lack of task-specific specialization.

• Trainable. Most importantly, transformers can learn the relationships between said elements in sequences without supervision (i.e. without labeled training data). Because of this, transformers have access to orders of magnitude more training data than their traditional counterparts, most of which require labeled examples to learn from.

The result of all this?

If software ate the world and AI is eating software, then transformers are eating AI.

The chip shortage won’t last, but its existence showed a clear demand for hardware that can run AI — particularly transformers — efficiently.

If you could build a chip that’s significantly faster and more efficient at running these models, you’d have a profound impact on the trajectory of AI, the products being built with it, and its adoption across hundreds of different industries.

You’d also make a lot of money - there’s never been a better time to be a chip company.

That’s exactly what Etched is doing.

(And yes, they are a real chip company).

Background

Etched is building a chip specifically for transformer inference.

While GPUs are great for any program that’s massively parallelizable (i.e. rendering engines, particle simulations, deep neural networks), Etched’s chip can only run one thing — transformers.

The chip, called Sohu, can run many different types of transformers (i.e. ones with different sizes, different modalities, different attention mechanisms, different activation functions, different masking systems, different encoding schemes, different sampling methods), but it can’t train them or run unrelated models like LSTMs or DLRMs. As a result, while it’s not as specific as chip that’s been hardcoded to run a specific program (i.e. an ASIC that can only run Llama 70B or SHA256), it’s also not as flexible as a GPU that can be programmed for general computation (i.e. Nvidia GPUs using CUDA).

Why is this level of abstraction in particular valuable?

Well, transformers and their applications might be the fastest growing category in software today, but existing chips are hardly optimized to run them — as we’ll explore later, only roughly 4% of an H100’s transistors are being used for actual matrix math during inference. At the same time, new and better models are coming out so frequently it’d be unwise to optimize your design around any one of them in particular. As a result, optimizing at the architectural level makes the most sense - by doing so, you can keep enough flexibility to run today’s best models while also getting a major performance boost over today’s best chips.

Big emphasis on performance boost.

While we’re unable to share performance figures publicly, we’ve gone into the weeds and have seen that on models like Llama 70B in full FP16 precision, Sohu is at least an order of magnitude better than today’s chips. This is with respect to things like:

• Throughput, generating more tokens per second.

• Latency, reducing the time-to-first-token.

• Cost, lowering the cost per token.

We’ll go into more depth on where this performance comes from, but the upshot is that Sohu can run today’s most popular models faster than anything else. We’ve seen Sohu’s verified throughput metrics on Mistral 7B, CodeLlama 34B, Whisper V3, OPT-175B, CodeGen 16B, and many other models, and they are dramatically better than what’s out there today.

Beyond serving existing models faster, the remarkable three-pronged throughput/latency/cost improvements Sohu enables will unlock entirely new classes of applications, both technologically and economically. Imagine:

• Real-time voice assistants with far lower latency and no compromise on output quality.

• Better pair programmers that can search and compare many possible code paths in parallel.

• More sophisticated agents and even swarms/ensembles of agents that can solve problems together.

To make developing and deploying such applications easy, Etched plans to release a complete server system and software stack with Sohu. This is to make it easy to port models on to their hardware.

We’ll go deeper on their product and GTM strategy below, but the upshot is they’ll be selling this package to companies that run their own servers (i.e. hyperscalers like AWS, foundation model companies like Anthropic, or recently-pivoted cannabis companies like Hyperscale Nexus).

Etched was founded in 2022 and, while we can’t share specifics, has accomplished quite a bit in their brief 18 months of existence. According to the Information, they’ll be announcing their chip later this year — we’re very excited.

The company was founded by Gavin Uberti and Chris Zhu, two Harvard dropouts. Gavin previously developed the ARM Cortex-M backend for Apache TVM (an open source deep learning compiler stack) and Chris previously did research in combinatorics and high performance computing. Joining them are a team of highly experienced chip designers:

• CTO Mark Ross was formerly CTO of Cypress Semiconductor (acq. 9.4 billion) and has shipped 5 systems that did over a billion in revenue.

• VP Hardware Engineering Ajat Hukkoo was previously VP Engineering for Intel’s Custom Silicon Group and through his career has shipped over 300 million chips to production.

• Chief Architect Saptadeep Pal was previously Co-Founder/Principal Engineer at Auradine (one of America’s largest bitcoin mining system producers) and did chip architecture research at Nvidia.

On the back of their impressive traction, vision, and team, the company raised a 23 million dollar seed round last year from Primary Venture Partners and a small group of angels.

Etched has a remarkably strong team and a groundbreaking, technically sophisticated product. Their mission is to build the hardware that will enable AGI and eventually superintelligence.

Will they be able to pull it off?

Chip companies aren’t exactly easy to build — we’re not talking about a CRUD app or a GPT wrapper here.

Well, we think they’ve got a shot.

Let’s dive deeper into why.

The Opportunity

We’ll begin our exploration of Etched like we do for all of our companies, from a first principle:

Big companies lead large, growing markets.

Many startups with great teams and solid growth don’t get very big because there just aren’t enough customers for them to grow into. As a result, we’ll want to show two things:

1. Etched operates in a large, growing market.

2. Etched can become a major player in that market.

Market Size

Companies are spending lots of money on chips that can run the latest models efficiently. As we saw earlier, demand for these chips is almost insatiable:

• Last quarter, Nvidia sold 14.5 billion dollars worth of datacenter chips, up 279% from a year ago. Jensen Huang, Nvidia’s founder and CEO, predicts $1 trillion will be spent upgrading data centers for AI over the next 4 years.

• Last year, one out of every three dollars invested in startups went to an AI company. Despite startup funding being down 30% overall, AI startups still raised a collective $50 billion, 9% more than the $45.8 billion raised in 2022. Much of that has been earmarked for compute - for some companies, up to 80% of total capital raised.

• Hyperscaler cloud capex continues to grow, now largely driven by buildouts of AI capacity. By 2025, Microsoft/Amazon/Google are projected to spend a collective $120 billion annually on new data centers. Meta is aiming collect some 350,000 H100s by the end of 2024, which at a conservative 25k each is almost $9 billion from them alone.

  

Given this, the promise of any business that can build a more performant chip is clear. ChatGPT and the products built on it are some of the fastest growing pieces of software ever — any chip that can run them faster is going to see enormous demand.

The real question is whether such a chip can be built.

We’ll keep this section short and spend most of our time answering that question instead.

Execution

To better understand whether Etched can build the business they’re envisioning, we'll need to answer three questions, each on a different topic:

• Science - is a chip like Sohu possible?

• Engineering - what progress has Etched made building it?

• Go-to-market - once ready, how will adoption work?

We’ll examine these questions one at a time, looking at everything from the principles behind hardware acceleration to common practices for simulating chip performance.

As a quick disclaimer, remember that all of this will remain heavily oversimplified — due both to the constraints of our medium and the confidential nature of certain details. We’ve done our best to be thorough, but you’ll have to read a textbook (or join Etched) if you want to know how this stuff really works :)

The Science

Most of us have some intuition that you can trade flexibility for performance, but we’d like to start by building a more concrete understanding of how this works.

Given a fixed amount of silicon, how do you make a chip that executes a program as quickly as possible?

Well, every program is just a sequence of instructions. These instructions are processed by processors (cores), each of which have three parts:

• A fetch/decode unit that determines which instruction to run next.

• An execution unit (ALU) that performs the operation described by that instruction.

• A context unit (registers) that stores the state (inputs/outputs) needed to perform that operation.

The simplest processors execute one instruction per clock cycle.

How do we speed this up?

While you can get fancy with systems to minimize the number of instructions executed or the latency per instruction (i.e. branch prediction, out-of-order control logic, memory pre-fetching), for programs with parallelism, the simplest answer is also the most effective: process more instructions at once.

The central principle behind GPUs is trading fast workers (i.e. a few cores with fancy behavior) for more workers (i.e. many cores with simple behavior). For instance, compared with CPUs, GPUs have:

• More cores. The 10th Generation Intel “Comet Lake” i9-10900K CPU has 10 cores, each with their own fetch/decode, execution, and context units. In contrast, the Nvidia V100 GPU has 80 SM modules (which you can think of as cores), while the GeForce RTX 4090 has 144. You can actually see these in die shots of the chips.

  

  

  

• More execution units. Imagine you’d like to execute the same instruction on 8 different pieces of data. Instead of fetching and executing that instruction 8 different times over 8 different clock cycles, a core with 8 execution units could simply broadcast that instruction to those units where it can be executed in parallel. This pattern is called single instruction, multiple data (SIMD), and here it lets us go from 8 clock cycles to 1.

  GPUs leverage this to further improve throughput per core. For instance, each SM “core” on a V100 has 4 16-wide f32 execution units, meaning it can handle a total of 64 f32 MUL-ADD instructions per clock cycle. This means that across 80 SM modules, we have a total of 64 * 80 = 5,120 execution units (CUDA Cores, using Nvidia’s terminology) handling 5,120 pieces of data per clock cycle. The H100 has 14,592 such execution units across 114 SM modules. That’s a lot!

  

There are other things GPUs do to keep things running quickly. For instance, they use sizable caches and significant multithreading (having multiple execution contexts that execution units can hotswap between) to hide memory access latency — dead time while the right data is being loaded into the right execution context. They also have specialized execution units for certain types of computations — the V100, for instance, has specialized tensor core execution units alongside its more general CUDA core execution units.

The upshot, though, is that GPUs trade more sophisticated behavior per core for having more execution units overall (cores, ALUs per core). This makes them much faster for applications with sufficient parallel work to utilize all available execution units. In cases where there isn’t though, CPUs and their fancier cores are actually faster (i.e. w/ branch prediction, memory pre-fetching, broader instruction sets) and more versatile (i.e. capable of handling interrupts, virtualization, different privilege levels). You wouldn’t want a GPU to run your operating system, but they’re hard to beat for applications like deep learning or fluid simulations.

Sohu takes these principles up a notch.

Sohu

GPUs are optimized for parallel computation, but they aren’t perfect. When it comes to running transformers, some quick back-of-the-envelop math (original credit to Chipstrat) shows there’s still opportunity for improvement:

• Taking a top-down approach, let’s estimate how much we can optimize the runtime of a popular model like Llama 70B. Running Llama has two steps: loading it into memory and doing a forward pass (the math).

  • Llama 70B has 70 billion parameters, which quantized at FP8 (one byte per parameter) is 70 gigabytes in total. Running the model on 4 H100 NVLs (each with a memory bandwidth of 3.9 TB/s), it should take 70 / (3900 * 4) = ~4.5 ms to load the model into memory.

  • Most neural network operations are memory limited rather than math limited, meaning it takes longer to access memory than to perform the actual computations. To be generous, we can upper bound it and say the forward pass takes the same amount of time as the loading step, or ~4.5 ms.

  In theory, then, it should ideally take ~9 ms to run Llama 70B from scratch with this setup.

  In practice, however, Nvidia reports it takes 109ms (time-to-first-token, FP8 quantized Llama 70B on 4 H100 NVLs with 2048-token input length). There’s an extra 100ms of latency being added somewhere - but where?

• Taking a bottoms-up approach, let’s see if we can optimize the design of today’s chips directly. Gavin, Etched’s CEO, has said only ~4% of an H100’s transistors are being used for the matrix multiplications driving LLM inference. Is this true?

  Humor us for a second. If 100% of an H100’s transistors were being used for matrix multiplications, how many FLOPS would it produce? Nvidia’s tensor cores use fused multiply-add circuits (FMAs) for matrix multiplication, each of which does two operations per clock cycle — an addition operation and a multiplication operation. Running at a base clock speed of 1590 MHz, a single FMA would produce 1590 MHz * 2 FLOPS = 3.18 GFLOPS. Since state-of-the-art FMAs take roughly 10k transistors to construct, converting all 80 billion of the H100’s transistors to FMAs would give us 8,000,000 FMAs in total, generating a collective 8,000,000 * 3.18 GFLOPS = 25,440 TFLOPS.

  In practice, however, Nvidia reports the output of the H100 at 990 TFLOPS (without their assumption of sparsity), meaning we only have 990 / 25440 = ~3.8% of the FMAs we’d ideally like. This suggests that much of the H100’s silicon area is stuff besides execution units - but what?

This “bloat” comes from many places. Software-wise, things like thread management and resource contention, CUDA kernel setup and teardown, inefficient memory access patterns, and superfluous control logic add overhead. Hardware-wise, space we’d otherwise use for execution units must be saved for things like network infrastructure, which allows data to move from any one SM module to any other.

The upshot is that since GPUs are optimized for general parallel computation, there’s a lot of hardware/software overhead that’s not relevant to transformer inference. Since Sohu is designed specifically to execute transformers, it’s able to replace much of that with more relevant compute units, driving that 4% up and that latency down.

More concretely, Sohu leverages several attributes of transformer inference workloads to improve performance. These include:

• Deterministic data flow. In contrast to say a particle simulation, AI models are deterministic — they have no “if statements” or conditional branches.

  As a result, with transformer workloads you know when and where each piece of data needs to be ahead of time. You can orchestrate memory access patterns so that the right data is in the right place at the right time, which removes the need for a lot of flexibility-related overhead. For instance, you can cut branching and other control flow related blocks, the L1-L3 cache hierarchy (you know in advance when you’ll need each piece of data), infrastructure to pass data between different cores, and much more. In their place, Sohu has added many more execution units.

• The nature of attention. Attention is a computation between two operands dependent on prior sequence data (i.e. query matrix Q and key matrix K are both functions of input sequence x), and Etched has proprietary hardware to do this without using standard GEMM (general matrix multiplication) engine architectures. Sohu uses less surface area to compute attention than traditional chips, which allows it to include many extra execution units.

• Relevant data types. Transformers only use low precision numbers so there’s no need for fp64 compute units like the H100 has. Sohu has removed execution units of data types that aren’t commonly used in transformers, allowing it to include more execution units for the ones that are.

Etched also leverages other magical things like activation approximation (allowing Sohu to compute popular and fully custom activation functions with less silicon area than GPUs/TPUs), improvements around large scale batching, and cleverly optimized memory access patterns around data re-use to make things more efficient.

Believe us, they’ve filed lots of patents.

The Engineering

Now that we understand the general principles driving Sohu’s performance, how much progress has Etched made actually building it?

We’ve gone into the weeds with them and while we can’t share the details on their progress, two things are clear: they’re pretty far along and they’ve been exceptionally rigorous about verifying Sohu’s performance/functionality.

Digging deeper on the second, you might still be wondering where the performance figures from earlier came from. How can we treat them with any degree of confidence when Etched hasn’t actually produced a chip yet?

Well, contrary to popular belief, you can actually simulate the performance of a chip (to a high degree of confidence, within several percentage points) before fabricating it for real. This is standard practice in the chip industry, and at a high level involves a few steps:

• Design. Chip designers will describe the digital circuit they’d like to make in code using a hardware description language (HDL) like Verilog or VHDL. At the register-transfer level (RTL) of abstraction, the circuit is described in terms of registers (storage elements) and the transfer of data between these registers, which includes logical operations, arithmetic operations, and control flow. This code doesn't represent the physical circuit components directly but describes how data moves and is transformed within the system.

• Simulation. Once the RTL description is complete, simulation tools are used to verify the logical correctness of the design. Tools like Verilator, Synopsys VCS, or Cadence Xcelium are used to compile the HDL into a cycle-accurate behavioral model, which you can then apply different sets of inputs against. That way, designers can ensure their design behaves as intended before moving forward.

• Synthesis. Finally, designers can use a synthesis tool like Synopsys Design Compiler or Cadence Genus to translate the RTL description into a gate-level representation, which maps the RTL constructs to actual hardware elements like logic gates and flip-flops. This process also involves optimizations to meet the desired performance, area, and power specifications. After this, a timing analysis can be done to ensure that the circuit meets the required speed performance. For instance, critical path analysis might be performed, where the longest paths data might travel within a clock cycle are examined to ensure they can be traversed within the given clock period.

While there might be other steps after (i.e. formal verification to mathematically prove specifications are met under all possible inputs, mapping the synthesized design to a physical layout on silicon, checking for possible physical issues like electromigration or antenna affects), at a high level this is how chip designers verify the functionality and performance of their designs before fabrication.

Etched has done similar things to verify the functionality and performance of their designs, but they’ve also used a few less conventional testing methods on top. While we can’t detail them here, here’s one example of just how meticulous they’ve been.

Transformers include several numerical operations beyond FP16 multipliers — think activation functions, up/downcasting blocks for conversion, accumulators for keeping track of layer norm means/standard deviations, and sine and cosine blocks for rotary embeddings. Not all of these functions are covered in the IEEE-754 standard, and as a result many can be computed in slightly different ways. To ensure that Sohu behaves the same as GPUs at the block level, Etched has run exhaustive simulations of each and every one of these math blocks to make sure the computed numerics are equivalent to those of popular GPUs.

This team means business.

The Business

Once Etched’s chip is ready, how will they go about taking it to market?

Sohu might be the best chip in the world for transformer inference, but companies wanting to use it will need more than just that. Most aren’t interested in building their own boards to mount the chip, servers to hold the boards, or firmware to run the servers. They definitely aren’t interested in writing their own compilers so their models can run on a new piece of hardware.

Because of this, Sohu will be sold as a complete system, accompanied by a software stack letting users port their models over in just a few lines of code.

From the end-user’s perspective, this software part is critical.

One reason Nvidia’s chips have been so popular for so long is the dominance of CUDA, which is the API Nvidia developed allowing their GPUs to be programmed for non-graphics workloads. Since being released in 2007, CUDA has become something of a moat for Nvidia - its popularity means lots of developers are familiar using it and that many tools/libraries have been built on top of it.

As a result, the switching cost of adopting any non-CUDA chip is quite high. Since CUDA is proprietary, developers would have to learn whatever their new chip’s version of CUDA is so they could port their model over. As you can guess, this is quite difficult.

Etched, however, doesn’t suffer from this problem. Since Sohu only runs transformers, they’ve been able to write a simple transformer-specific compiler making it easy for users to port models over to their chip. This would be a lot harder for chips capable of general computation, but constraining the problem from running arbitrary code to running just transformers makes developing said software easier.

As a result, adoption from the thousands of companies running transformer inference should be easier. The sales cycle is no longer a pure hyperscale play: there are plenty of startups, scaleups, clouds, and enterprises doing on-prem compute that are looking to optimize their cost and latency.

Long-term, as models become more powerful and the cost to get state of the art increases, the time between major models will increase also. As this happens, Etched may develop ever-more specialized chips, eventually becoming the model-specific AI company as well.

Team

Etched wants to build the world’s best chip for transformer inference. More than that, they want to create the hardware enabling AGI and eventually superintelligence.

It’s an ambitious mission, but they’ve built a team with the incredible software and hardware engineering experience needed to pull it off.

Consider a sampling of their roster:

Gavin Uberti - Co-Founder, CEO

Gavin was previously a Software Engineer at OctoML. Before that, he was an Algorithms and Backend Engineer at Coursedog. Earlier, he was a Software Engineering Intern at Xnor.ai. He dropped out of Harvard while pursuing a BS and MS in Computer Science.

Chris Zhu - Co-Founder, Director of Compilers

Chris was previously a math and HPC researcher at Harvard University. Before that, he was a Software Engineering Intern at AWS. Earlier, he was a Software Engineering Intern at AvantStay. He dropped out of Harvard while pursuing a BS in Computer Science and Mathematics.

Mark Ross - CTO

Mark was previously the Co-Founder and VP of Engineering at Striiv. Before that, he was the CTO at Cypress Semiconductor (acq. 9.4 billion). Earlier, he was a Senior Director at Cisco. He started his career as an Architect at Sun Microsystems and holds a BS and MS in Electrical Engineering from Stanford University.

Robert Wachen - COO

Robert was previously a Co-Founder at Prod, a startup accelerator. Before that, he was the Co-Founder and CEO at Mentor Labs, which was acquired by Crimson Education. He dropped out of Harvard while pursuing a BS in Decision Science.

Ajat Hukkoo - VP Hardware Engineering

Ajat was previously the VP of Engineering for Intel’s Custom Silicon Group. Before that, he was a Director of Engineering at Broadcom. He holds a B.Tech in Electrical Engineering from IIT Bombay.

Saptadeep Pal - Chief Architect

Saptadeep was previously a Co-Founder and Principal Engineer at Auradine (one of America’s largest bitcoin mining system producers, valued of 500 million). Before that, he did architecture research at Nvidia. He holds a B. Tech in Electrical Engineering from IIT Patna and an MS/Ph.D. in Electrical and Electronics Engineering from UCLA.

Abhishek Chinchalpet - Principal Verification Engineer

Abhishek was previously a Verification Engineer at Elastic.cloud. Before that, he was an ASIC Design Verification Engineer at Waymo. Earlier, he was a Senior Silicon Engineer at Google. He started his career as a Hardware Engineer at Oracle and holds an MS in Computer Engineering from Texas A&M.

Reynold Leong - Founding Principal Engineer

Reynold was previously a Staff Verification Engineer at Cruise. Before that, he was a Principal Verification Engineer at Quanergy. Earlier, he was a Senior Verification Engineer at Violin Memory. He holds a BS in Electrical and Computer Engineering from UC Berkeley and an MS in Electrical and Computer Engineering from UC Irvine.

AppaRao Challagundla - Founding Principal Engineer

AppaRao was previously the SOC Chief Architect (Confidential Compute) at Intel. Before that, he was a Senior Design and Verification Manager at Intel.

Prasad Bolisetty - Director of Software

Prasad was previously a Software Engineering Manager at Intel. Before that, he was the Director of Software at Soft Machines. Earlier, he was a Sr. Principal System Software Engineer at Broadcom. He holds a BE in Electronics & Communication Engineering from Andhra University.

Chang-Ming Lin - Principal Engineer

Chang-Ming was previously an ASIC Design Manager at XPENG. Before that, he was a Sr. ASIC Design Manager at Silicon Motion Technology Corp., and earlier, a Design Lead at Intel Corporation. He holds an MS in Electrical and Electronics from the University of Utah.

Carson Robles - ASIC Design Engineer

Carson was previously a Hardware Engineer at Optiver. Before that, he was a Hardware Engineering Intern at Astranis. He holds a BS in Computer Engineering from Cal Poly SLO.

Hsu-Kuei Lin - ASIC Design Engineer

Hsu-Kuei was previously a Digital Circuit Designer at Meta. Before that, he was a Staff Engineer at Qualcomm. Earlier, he was a Member of the Technical Staff at NUVIA. He started his career as a Senior Hardware Engineer at Microsoft and holds an MS in Electrical Engineering from the University of Michigan.

Competitive Landscape

Etched is a transformative player in the world of AI chips, but at first glance they’re not the only player in the space. If you google “chips for AI”, you’ll find a laundry list of companies appearing to make similar bets.

Now that we’ve got an understanding of Etched’s unique strengths, let’s see how they compare against some of these competitors.

Incumbents

Nvidia controls 87% of the discrete GPU market. Second place goes to AMD, who controls about 10%. Several hyperscalers have built internal silicon teams to try and commoditize Nvidia (i.e. Google’s TPUs, AWS’s Trainium and Inferentia, Azure’s Maia), but so far all have trailed Nvidia by 6-18 months when it comes to raw performance.

Each of these incumbents produces general-purpose chips for both inference and training. As a result, they’re much slower than Etched when it comes to transformer inference. Not only that, compared to our predictions around the performance of the next few generations of GPUs, Etched should still be more performant.

So why won’t Nvidia just clone Sohu?

You can never fully rule out this possibility (startups, man), but we can think of a few reasons they won’t:

• Etched is making a bet that a transformer-like architecture will remain dominant for the next few years. This is a risky proposition: model architectures have changed before, and while transformers have become industry standard, potential rivals emerge all the time. Nvidia’s dominance in the status quo means they have zero incentive to take a similar risk.

• CUDA, Nvidia’s API to program their GPUs for general computation, is one of their strongest moats. It’s durability stems from its sophistication, and its sophistication stems from the general purpose nature of Nvidia’s GPUs. Building an architecture-specific chip would negate much of CUDA’s importance/utility, eating into their own moat. Nvidia’s dominance in the status quo means they’re disincentivized to take a similar risk.

Not to mention, building a chip like Sohu isn’t exactly trivial — Etched has made a number of (patented) technical breakthroughs, so even if an incumbent tries they’ll be starting from behind.

Plus, if Etched is successful, general-purpose chips won’t go anywhere — H100s and TPUs will still be used for training and for non-transformer workloads.

Startups

There are a number of startups that are also building general-purpose AI chips for both inference and training. These include:

• Cerebras, which has raised over 720 million. Instead of slicing a silicon wafer into lots of smaller, individual chips like most companies, Cerebras has turned the entire wafer into one giant chip with lots of extra processing power and on-chip memory. It’s an amazing feat of engineering — the idea was to reduce data transfer/latency bottlenecks by fitting the entire model into on-chip memory (i.e. the WSE-2 has 50GB of on-chip SRAM). Unfortunately, they didn’t anticipate how large modern LLMs would get, which has led to some challenges.

• SambaNova Systems, which has raised over a billion. Their latest chip, the SN40L, is a general purpose AI accelerator with a unique part-HBM, part-GDDR memory system. It leverages Sambanova’s Reconfigurable Data Unit (RDU) architecture, which essentially allows you to configure the physical flow of data across the chip as a means of minimizing unnecessary data movement. It’s sold as part of a full stack AI platform, complete with server systems and software.

• Graphcore, which has raised over 760 million. Their flagship product is a chip they’re calling the Intelligence Processing Unit (IPU). The IPU unfortunately hasn’t lived up to its promise, and the company is struggling financially. Microsoft, one of their key investors and early customers, pulled out of a deal for their processors last year.

While these companies have raised lots of money and shipped chips, they’ve struggled to achieve adoption. NVIDIA’s GPUs often ended up being more performant and the software switching costs often ended up being too high. As a result, several of these companies have pivoted to a “model-as-a-service” approach, abstracting the software and hardware away from the end customer and offering more of a white-glove ML service.

There are also a number of startups building general-purpose chips for inference. These include:

• Groq, which has raised over 360 million. Their main product, the Language Processing Unit (LPU), is a smaller, cheaper chip that can be chained together in pods of 264 (in the future, 4,128) to create servers for inference.

• d-Matrix, which has raised 154 million. Their product, Corsair, is an inference platform that’s based on chiplets - smaller, specialized chips that can be combined to form a comprehensive integrated circuit. The Corsair’s chiplets, called Jayhawk IIs, further feature a digital in-memory compute (DIMC) architecture designed to minimize unnecessary data movement.

• Kneron, which has raised over 190 million. They’re building AI chips optimized for edge devices, particularly self-driving cars. They’re not focused on data centers or LLM inference and aren’t directly competitive with Etched.

• Sima, which has raised over 200 million. Like Kneron, they’re also building AI chips for edge devices. They’re not focused on data centers or LLM inference and aren’t directly competitive with Etched.

By only supporting inference, these chips will have better performance than more flexible AI accelerators. However, they’ll still be significantly worse than chips which have specialized on an architecture.

Etched is building such a specialized chip. By betting on transformers specifically, they run the risk that their chips will be obsolete by the time they get to mass production. If they’re right though, they’ll be the most performant chip for transformers by far.

There are a few companies that are making bets similar to Etched, but they all seem to be earlier in the process. These include:

• MatX, who’s building a chip focused specifically on LLM inference. They’ve made a different bet than Etched, however, and have optimized their chip around quantization — specifically models using int4 precision. Whether models can be quantized to that level without degrading performance remains an area of active research.

• Positron, who’s building a “Transformer Inference Appliance” by chaining existing GPUs together. They aren’t building their own chips yet.‎

Conclusion

If Etched succeeds, running transformer-based models will never be the same again. Why wait in line for GPUs when you can get >10x the performance with Sohu instead?

They’re hiring.

‎

Thanks to Philip Bogdanov for his help with this piece.

In case you missed our previous releases, check them out here:

• Warp - Revolutionizing the most foundational developer tool.

• Pinecone - Vector databases index the AI revolution.

• WorkOS - The API to increase your TAM.

And to make sure you don’t miss any future ones, be sure to subscribe here:

Finally, if you’re a founder, employee, or investor with a company you think we should cover please reach out to us at ericzhou27@gmail.com and uhanif@stanford.edu - we’d love to hear about it :)

If you’re interested in roles with any of the companies we’ve covered (or the unreleased ones we have in our pipeline 👀), please fill out this form to join our talent network.

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information we’ve received from venture firms we work with. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should Join: Inside Edition

January 2024

As you know, the goal of this newsletter is to highlight promising early-stage companies to join, and to do so in a way that’s thorough, unbiased, and (occasionally) entertaining.

We do this because startups are risky business, and we think people deserve a quality resource to help figure out which ones are actually worth joining — one that’s less shallow, biased, and hype-driven than most of what’s out there.

Our philosophy in doing this has been to been to offer the opposite of one of those “Top Startups” lists — to us, it’s important to understand why a company might succeed before joining it, and we felt most resources recommending companies weren’t going deep enough to be helpful with that. By cherrypicking and deeply analyzing just one company each month (all from the portfolios of great investors), we hoped to provide a better starting point for people exploring their next roles.

In our quest to achieve this, we’ve naturally ended up talking with lots of VCs.

We do this to gather information on which startups are doing well, rumors on which startups aren’t doing so well, and (most importantly) suggestions for overpriced coffee shops.

Now, we know investors get a bad rap on Twitter

…and it’s true, we’ve left a few meetings scratching our heads.

But most of the investors we’ve met have actually really impressed us with their thoughtfulness, curiosity, and insight.

We recently began thinking it’d be worth sharing some of their perspectives more directly — many of you have reached out for additional company suggestions, so we thought something here might be a good way to help out (especially since it’s hard for us to write more than one piece ourselves each month).

With that in mind, we invited several investors we respected deeply to share their perspectives on some companies we were both excited about.

This new series is the result.

Our first edition features five companies in total, each backed by a different firm and each at an early stage. At the bottom of each investor-contributed piece, we’ve added a few editor’s notes on the company and why we admire the investor who’s recommending it.

Not everyone has a VC they can DM for advice when they’re thinking about their next role.

We hope this will be the next best thing :)

‎

Why You Should Join Chkk

Preventing Kubernetes Problems Before They Start

Kubernetes is designed for flexibility. That’s part of what has made it the de facto standard for cloud-native applications, with a recent Cloud Native Computing Foundation survey finding 96% of organizations are using or evaluating the platform. But that flexibility also comes with a cost: complexity. To take full advantage of K8s’ vibrant ecosystem of add-ons, teams must navigate multiple configs and track intricate dependencies across a maze of open-source and commercial products—and one seemingly minor mistake can cause a major chain reaction.

For organizations with business-critical apps on Kubernetes, availability and uptime are paramount, but most of the tools for uptime and availability only alert you about problems after they have occurred. Platform Engineering, SRE, and DevOps teams are forced to wait and see, risking credibility and wasting time on reactive fire-fighting after even routine upgrades.

But now there is a better solution: Chkk, a platform for proactively identifying and addressing infrastructure errors, disruptions and failures before they ever lead to downtime. Just as an X-ray scans the human body to spot health issues before they land you in the ER, Chkk finds and diagnoses availability risks in Kubernetes environments.

Co-founders Awais Nemat, Fawad Khaliq and Ali Khayam previously worked together on the AWS team that built Elastic Kubernetes Service (EKS), where they had a key insight: if one enterprise experienced a disruption, the root cause was likely to affect other enterprises, too. Rather than all of those organizations responding reactively, one by one, why not share their knowledge and prevent problems in the first place?

At the heart of this Collective Learning Technology is Chkk’s Availability Risk Signature Database, which tracks hundreds of Kubernetes issues sourced from publicly available information such as incident reports, release notes and forum discussions—along with data from Chkk’s own users—to populate a Knowledge Graph of risks. It is very similar to how the security industry tracks CVEs. The platform then streams that information to customers and contextualizes it against their own infrastructure, detecting and remediating issues before they cause disruptions. In other words, teams get to learn from each others’ mistakes—so they don’t have to repeat them.

Chkk counts Alef Edge, Fairtiq, Nexoya, Phi Labs, Yoti and many others among their happy customers, and they’ve been getting impressive feedback since the launch of their Kubernetes Availability Platform in late October. As the new year begins, many more organizations will no doubt be on the hunt for enhanced Kubernetes resiliency and an insurance policy against outages in 2024—and Chkk may be just the solution they’re looking for.

— Bogomil Balkansky, Partner at Sequoia

Chkk is hiring talented software engineers for their North America team - if you’re interested (and especially if you have experience with Go, Kubernetes, and AWS/GCP/Azure), please reach out to their team here. We also recommend checking out the rest of Bogomil’s portfolio at Sequoia (i.e. Mutiny, Temporal, and Wiz). Bogomil was previously VP Cloud Recruiting Solutions at Google and Senior VP of Cloud Infrastructure Products at VMware. He also runs a most excellent cooking blog.

‎

Why You Should Join Footprint

Solving for Fraud and Friction

The conventional wisdom has been businesses face a tradeoff when they onboard a customer: onboard more users while accepting more fraud, or add friction to processes to prevent bad actors at the cost of good actors. Both options are bad. Businesses lose billions of dollars a year due to fraud, while missing out on customers due to the friction put in place to avoid the fraud. Footprint is the first company I met that had a plan to solve for both fraud and friction.

They do this with a fundamentally different philosophical approach to onboarding. It starts with perhaps a paradox: they don’t believe you solve fraud by solely looking for bad actors. This is especially true in an age of Generative AI giving fraudsters even more tools to generate an infinite amount of real-looking identities. This is because the Footprint team knows we will never find every bad actor. Instead, they are creating a closed-loop ecosystem that begins to firmly place proven good actors in their own bucket.

To do this, Footprint has built a compound platform. They built fully integrated KYC, Security, and Auth suites that could beat industry incumbents standalone, and together are magical. Companies have trusted their vaults with millions of PII records, their flows to onboard thousands of people a day, and their auth to prevent ATOs and phishing. Their product suite sits on top of novel technology they are pioneers in bringing to market. Footprint leverages App Clips and Passkeys to stop duplicate fraud in real-time and cryptographically verify the same person who created an account is the same person logging back onto a platform. Meanwhile, they are one of a handful of companies built on AWS Nitro Enclaves to securely vault the data.

The result is a flawless developer experience. Five lines of code to integrated a hosted onboarding flow which will collect PII, verify it, vault it, and re-authenticate the subject. Traditionally these were point solutions that did not work together. Footprint has realized the 1+1=3 nature across these tools and the result is the future of identity—one where it is portable. They realized if you solve security at onboarding you can solve friction going forward, and once you solve friction you can solve fraud.

Footprint is a deeply technical team with deep expertise in the problem area. Alex, the CTO, studied cryptography at MIT before starting a mobile authentication company that would be acquired by Akamai. Their fraud and risk team includes leaders from Stripe and Sardine, while their frontend and design helped scale early products at companies such as Personio, Kyte, and Observe.

Footprint counts a variety of delighted customers. These range from investment platforms with over a million users like Bloom, to leading car rental marketplaces like Flexcar, to tenant screening platforms servicing hundreds of thousands of units such as Findigs.

— Shardul Shah, Partner at Index

Footprint is hiring talented frontend and backend engineers - if you’re interested, please reach out to their team here. We also recommend checking out the rest of Shardul’s portfolio at Index (i.e. Twelve Labs, Expel, and Wiz). He has a remarkable track record and is an expert on all things security and cloud infrastructure.

‎

Why You Should Join Ollama

Run Foundation Models on Your Machine

In the world of software development, it’s critical to have an efficient and fluid local development workflow. This axiom is reflected in the world of dev tooling, where the most exceptional projects & products often are the ones that introduce the greatest developer delight. When developers are bottlenecked by package management, version conflicts, merge & branching issues, and have to frequently leave their IDE and CLI, all they really want to do is close their laptop, smash a keyboard, and/or unplug a monitor. However when DevEx is optimized, the beatings no longer need to continue until morale improves.

As developers want to incorporate LLMs into their projects and applications, they’ve now incurred a new set of headaches. The majority of model usage is vendor-hosted, and the developer experience is tied to the cloud – devs in the experimentation phase now need to deal with latency overheads, rate limiting by model providers, and a whole host of information security concerns (let’s maybe not send personal emails over the wire). However, the advent of open-source models with strong reasoning capabilities and near-performance-parity has paved the way for new experiences with local LLMs, both in experimentation and in production.

Unfortunately, this pain also extends itself beyond the development cycle. For application programmers that would like to package their products to run locally (e.g. a Rewind) they would have to build their own solution to handle model downloading & config management, execution on CPU / GPU / heterogeneous underlying hardware, and performance optimization.

Enter Ollama, the best way to get started with local LLMs. They make it seamless for developers to run models like LLaMa 2, Mistral 7B, and even multi-modal models like LLaVA locally. It’s available today for macOS and Linux (with GPU-based acceleration), with Windows compatibility on the horizon.

Under the hood, Ollama leverages llama.cpp to run quantized models that can fit in memory and a distribution framework similar to Docker to enable downloading & storing models. They support importing models from formats like GGUF/GGML and offer the flexibility to customize models with the beautiful abstraction of a Modelfile (akin to a Dockerfile) – enabling users to configure prompts, base models, and in the future potentially fine-tunes / LoRAs. In creating a Modelfile, Ollama users can now distribute metadata & data along with models in a self-contained way.

Additionally, Ollama has a wealth of integrations, including terminal clients, UI wrappers (e.g. for chatbots), and libraries like Langchain to enable local RAG. They also have a handful of community-contributed plugins that allow for the seamless incorporation of Ollama models into everyday software applications like Discord, Telegram, and Obsidian.

The Ollama team, led by Jeffrey Morgan and Michael Chiang – the former architects of Docker Desktop, brings unparalleled expertise in local development. They ship fast, and it shows with their ever-growing library of available models. Ollama has polish & deliberate abstractions that fundamentally help accelerate the software development lifecycle in the era of LLMs. If you’d like to join the 23K+ people who’ve starred the repo & downloaded Ollama over 1.3 million times, check it out for yourself here.

— Alex Kolicich, Founding Partner at 8VC

If you’re a talented software, systems, or ML engineer interested in learning more about Ollama, check out their GitHub and Discord communities, or reach out to jeff@ollama.ai / michael@ollama.ai. We also recommend checking out the rest of Alex’s portfolio at 8VC. Prior to being a founding partner at the firm, Alex was an engineer and early-product advisor at Clarium, Palantir, and Google.

‎ ‎

Why You Should Join Cursor

The AI-powered Code Editor

An incredible team, a large market, ambitious vision, and sky-rocketing traction are why you should join Cursor.

Cursor’s founders are a dream team of technical leaders from MIT who’ve won IMO/IOI medals, led one of the first LLMs deployments at Google, and created two of the world’s most popular programming games. We knew them well before funding them: CEO Michael Truell and cofounder Aman Sanger were among the most talented and entrepreneurial Neo Scholars. We’d have backed either of them in a heartbeat, let alone together. They were joined by two of their smartest friends, Arvid and Sualeh, making one of the most potent young startup teams anywhere.

Their vision and market are extremely ambitious: to automate all software engineering. This is an opportunity that doesn’t just cap out at a few billion dollars and could lead to a generational company. The vision is personally exciting to us at Neo given our community's strong CS roots. The Cursor team has been programming since middle school and working on AI for almost as long. Redesigning programming with AI feels like what they were born to do.

There are many goliaths competing for the same prize, and Cursor has surprising advantages over other players. We're betting on their three-pronged approach of growing virally down-market, owning the entire editor, and building a tool that’s already useful today – it’s reminiscent of Figma's winning strategy. A year after our investment, they raised an $8M war chest from OpenAI, securing them special model access. Remarkably, Cursor is already becoming one of the most-used code editors in the world, a testament to both their team and their strategy.

Cursor is used by tens of thousands of engineers every day, including those at companies like Ramp, OpenAI, and Midjourney. Their rapidly-growing paid subscriber base is already generating over $2M in ARR, despite launching just eight months ago. Notably, this growth is fueled organically by word-of-mouth, a hallmark of some of the most successful companies in tech history. Their recent releases of faster AI-powered edits, linting, terminal commands, and image support have only accelerated growth.

Cursor is a lean team where each engineer gets lots of responsibility for research and product decisions, and even interns get to own projects end-to-end. This often involves very challenging problems in ML, product engineering, performance, and infrastructure. For instance, their upcoming “Copilot++” release has involved training a custom model from proprietary data, clever inference caching, careful networking, and UI experimentation. Cursor’s team takes pride in offering an intense learning environment, with a mantra: “You will learn more here than you will anywhere else for the rest of your life.”

— Ali Partovi and Suzanne Xie, Partners at Neo

Cursor (Anysphere) is hiring talented founding engineers and researchers in San Francisco - if you’re interested, please reach out to their team here. We also recommend checking out the rest of Ali and Suzanne’s portfolio at Neo (i.e. Warp, Pavilion, and Vanta). Both are experienced operators - Suzanne was previously Head of Product for B2B Payments at Stripe and Ali was previously Co-Founder of Code.org and LinkExchange (acquired by Microsoft). If you’re a college student studying CS, we also highly recommend checking out Neo Scholars.

‎

Why You Should Join Orby AI

Revolutionize Your Workflow with Tailored Automation

In today's fast-paced business environment, organizations are bogged down with an overwhelming amount of manual document processing, validation, and organization tasks. This reality spans across various roles – from HR associates managing employee onboarding to sales professionals tracking leads, and finance managers handling invoices. These repetitive tasks, although seemingly brief, cumulatively consume a significant portion of an employee's time over months or even years.

While existing Robotic Process Automation (RPA) tools like UiPath and Automation Everywhere have made strides in addressing these challenges, they often stumble when facing complex, dynamic workflows. They typically require extensive initial setup and programming, limiting their scope to tasks defined by strict rules – representing merely 20% of potential automation opportunities.

Orby AI emerges as a groundbreaking solution, going beyond the limitations of traditional RPA. By integrating advanced AI with observational learning, Orby AI streamlines the automation process, removing the need for tedious programming. Upon installation, Orby AI intelligently observes and learns from a user's interactions, understanding the intricacies of their job and the applications they use. This "observe, learn, and automate" approach rapidly identifies inefficiencies, offers optimization suggestions, and, over time, independently executes repetitive tasks, continually refining its performance based on user feedback.

What sets Orby AI apart is its ability to adapt to each user's unique workflow. Every Orby AI instance is trained to mirror the specific work habits of its user, evolving into a personalized assistant. This not only frees up time for more creative and high-impact tasks but also proposes innovative workflow enhancements. At its core, Orby AI is the first to offer a complete end-to-end automation platform, developed by training its own models on observed user workflows.

The vision for Orby AI was conceived by co-founders Bella Liu and Will Lu in 2022. Bella, with her experience in product leadership at UiPath, and Will, with his background in AI innovation at Google, recognized the limitations of current RPA tools. Their collaborative effort has resulted in Orby AI – a comprehensive solution designed to tackle the entirety of repetitive tasks, far beyond the conventional 20%, positioning it as an indispensable partner in workflow automation.

— Arash Afrakhteh, Partner at Pear

Orby AI is hiring talented software engineers, research scientists, and account executives - if you’re interested, please reach out to their team here. We also recommend checking out the rest of Arash’s portfolio at Pear. Arash is highly technical and was previously Co-Founder and CTO of Cariden Technologies, a bootstrapped infra startup acquired by Cisco.

‎

Conclusion

If you’re open to new roles or curious about their products, Chkk, Footprint, Ollama, Cursor, and Orby AI are all promising companies and worth checking out.

Happy New Year, and until next month!

‎

Thanks to Bogomil, Shardul, Alex, Vivek, Ali, Suzanne, Arash, Pejman, and Jill for their help with this piece.

In case you missed our previous releases, check them out here:

• Warp - Revolutionizing the most foundational developer tool.

• Pinecone - Vector databases index the AI revolution.

• WorkOS - The API to increase your TAM.

And to make sure you don’t miss any future ones, be sure to subscribe here:

Finally, if you’re a founder, employee, or investor with a company you think we should cover please reach out to us at ericzhou27@gmail.com and uhanif@stanford.edu - we’d love to hear about it :)

If you’re interested in roles with any of the companies we’ve covered (or the unreleased ones we have in our pipeline 👀), please fill out this form to join our talent pool.

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information from partnered venture firms. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should Join Loop

(Click the link to read online).

In March of 2021, a 400 meter long container ship called the Ever Given got stuck in the Suez Canal. It got stuck in such a way that the entire canal was blocked, crippling global trade for a week and causing between 6 and 10 billion dollars worth of damages.

You probably remember this - it was all over the news.

It spawned lots of memes.

But did you know that just two years earlier, the exact same ship ran into and severely damaged a tiny, 25 meter pleasure ferry in Germany?

Or that in 2022, the Ever Forward (a different ship by the same company) ran aground in Chesapeake Bay, where it got stuck for over a month?

Or that in 2019, the Ever Summit (yet another cargo ship by the same company) managed to crash into a port crane in Vancouver?

The clear lesson here is to avoid any ship whose name starts with “Ever”.

A secondary lesson might be that freight’s just a really tough business.

As an industry, logistics can be immensely complex — coordinating the intricate flow of goods between ports and destinations or shippers and carriers can be extremely difficult, especially at scale.

Small errors can have large consequences, and large errors can break the global economy.

But, taming that complexity and replacing it with efficiency can mean big opportunities. You might recall us finding this a few months ago when we first explored logistics with Traba (who by the way just raised another round) — as it turns out, there’s lots of room for ambitious startups to improve how ports and warehouses approach temporary staffing.

This month, we’d like to explore another company doing something similar, but in a different part of logistics:

Freight payments.

Payments in freight are just as complex as the distribution networks they serve.

Each invoice a shipper (someone with stuff to ship) receives from a carrier (someone with the truck to ship it) involves lots of different documents (packing lists, weight and inspection tickets, etc.) and lots of different variables (fuel surcharges, gate fees, etc.). This information is generally unstructured, and its format varies widely across the different contracts, invoices, and billing systems used by America’s 1.1 million different carrier companies. This means that for a large shipper like Coca-Coca (who works with 36,000 different carrier partners), handling invoices means reconciling tens of thousands of different payment-related emails, PDFs, CSVs, JPEGs, and EDI 210 transaction sets.

Since there’s no single place where every payment-related document and data source has been cleaned and centralized, it’s nearly impossible for large shippers to get accurate and real-time visibility into their transportation finances. Essentially anything you’d want to do here is difficult:

• It’s hard to make sure the invoices you’re receiving are accurate and that you’re not overpaying.

• It’s hard to estimate how much you’ll actually spend (given variable/unexpected charges) with a specific route or carrier.

• It’s hard to identify bottlenecks and inefficiencies in your transportation network.

Last year, truckers alone carried 11.46 billion tons of primary freight, sending invoices worth 940.8 billion dollars in the process. Because of how convoluted freight billing is, it’s estimated almost 25% those invoices had some kind of error — it was just too overwhelming to check them all.

For an industry that represents 3 percent of US GDP, you’d think we’d have a way to make sure all our spend here were accurate and optimized.

The financial layer of freight needs a source of truth.

Enter: Loop.

‎

Background

To recap, when a shipper (i.e. Coca-Cola) needs to send some product from a bottling center to a distribution center, they’ll contract a carrier (i.e. a trucking company) to help them do so. During the delivery process, that carrier will generate a good amount of data used to compute the final price. This might include:

• Basic figures like freight rates, shipping weights, fuel surcharges, and distance traveled.

• Contract specifics like discounts, delivery guarantees, insurance, and gate fees.

• Accessorial charges like inside pickup fees, lift gate fees, wait time fees, protection from freezing fees, and weight and inspection fees.

Each of these variables can come with their own supporting documents and details, which themselves can be spread across different document types (bill of ladings, packing lists, etc.), formats (PDFs, photos, EDI 210 transaction sets, etc.), and sources (mail, email, TMSs, ERPs, etc.).

The complexity and fragmentation associated with even one of these invoices can make it difficult for shippers to do something like a simple audit, where you check each charge for accuracy. For instance, investigating an invoice that’s higher than expected, discovering it was caused by an unexpected liftgate fee, determining whether a liftgate was needed, and potentially contesting the charge might mean combing through several different files across several different sources.

As you can imagine, doing this for many different invoices across many different carriers can overwhelm even the largest of transportation and finance teams.

Carriers often charge for fuel, but the amount they charge can be contingent on a specific fuel price index. The example above is taken from a 72 page public Fedex contract, which shows the relationship between the LTL/TL Fuel Surcharge and the current state of the US Energy Information Administration’s Gasoline and Fuel index. Doing an audit of the fuel surcharge (a subpart of a subpart of the final price) means making sure it matches the current fuel price in the index you agreed to tie your pricing relationship to. Naturally, different carriers might use entirely different indices, rates, and file formats. You see why this might be hard to automate?

‎

To solve this problem, Loop has built the first platform enabling shippers and third-party logistics providers (3PLs) to centralize all their payments-related data. By leveraging AI to automatically ingest, standardize, and link shipment data across different carriers, file types, file formats, and file sources, Loop is able to structure all the raw and unstructured data shippers and 3PLs normally receive.

This industry-first system of record would be valuable enough on its own, but Loop builds on top of it with a series of features that help automate every other finance-related workflow. By automating everything from invoice audits to dispute resolution to scenario planning to data visualization to rules-based cost allocation to the actual payments themselves, Loop creates lots of value for their customers:

• Short term, Loop helps shippers reduce unnecessary spend. By automatically auditing every invoice for errors and discrepancies, Loop makes identifying and resolving that unexpected liftgate fee from earlier is much easier.

• Long term, Loop helps shippers optimize their entire transportation budgets. With the analytics uniquely enabled by their data layer, shippers can forecast spend per carrier or product far more accurately, which can be crucial for forecasting, contingency planning, network scenario planning.

Collectively, using Loop results in a 2-7% net reduction in transportation spend and an 80% increase in back office productivity, both incredible metrics considering the vast budget/team sizes of their customers. Unlike most of their competitors, who take a portion of the savings identified, Loop uses a SaaS-based model that’s offers a flat fee based on payment volume processed.

One way to understand Loop’s game is by comparison to another popular fintech, Ramp. Like Ramp, Loop’s initial wedge was helping customers spend less - Ramp via automated insights and partnerships, and Loop via automated freight audits and analytics. Like Ramp, Loop is leveraging this wedge to centralize all of their customers’ finance and payments related-data, something which uniquely positions them to build downstream products on top. As we’ll see, this is exactly what they’re doing — Loop wants to be the ultimate platform for everything at the intersection of supply chain and finance.

Loop launched publicly in 2021 and booked $25 million worth of invoices their first month in production. In the two years since, they’ve continued to grow swiftly — they now book $4 billion worth of payments annually. Their customers include some of the largest shippers and third-party logistics providers (3PLs) in the world, including J.P. Morgan Chase, Nationwide, Loadsmart, and several others we can’t share.

Customers love them: GILLIG, a large manufacturer of transit buses, shared that they went from manually “gut checking” 30% of their freight invoices to rigorously auditing 100% of them, saving 6% of their total transportation spend as a result.

The company was founded in 2021 by Matt McKinney and Shaosu Liu. Previously, Matt was a Data Science Manager and a Senior Data Scientist at Uber. Shaosu was a Software Engineering Lead on Uber Freight, where he launched the product and scaled it to 2 billion in annual revenue. Before Uber Freight, Shaosu was on the team that built surge pricing, everyone’s favorite feature.

On the back of their impressive traction and team, Loop raised a $6 million dollar Seed round in 2021 led by 8VC, a $24 million dollar Series A in 2022 led by Founders Fund, and a $35 million Series B this year led by Index Ventures and J.P. Morgan.

Loop has a strong product, excellent traction, and a committed team. They’ve got a vision to become supply chain payments’ system of record and a mission to change how we spend 3% of GDP.

But will they be able to pull it off? Especially given how challenging the logistics industry is?

Well, we think they’ve got a shot.

Let’s dive deeper into why.

‎

The Opportunity

We’ll begin our exploration of Loop like we do for all of our companies, from a first principle:

Large companies lead large, growing markets.

Many startups with great teams and solid growth don’t get very big because there just aren’t enough customers for them to expand into. As a result, we’ll need to show two things:

1. Loop operates in a large, growing market.

2. Loop can become a leader in that market.

‎

Market Size

Loop makes money by helping companies process and optimize their freight payments. Since so many companies are spending so much on freight, Loop must operate in a large market.

Simple, right?

While we could stop here, we recognize that wouldn’t make for a very satisfying section.

To us, it makes sense to consider the question of Loop’s opportunity in two parts: how large is the market they’re starting in, and how large are the markets they can expand into?

Initial Market

Shippers initially come to Loop for help optimizing their freight payments. Loop helps them achieve this by:

• Automating freight audits, which eliminates extraneous fees and unnecessary spend.

• Analyzing payments data, which is important for planning and forecasting purposes.

• Surfacing network insights, which helps shippers negotiate more favorable carrier contracts.

How large of a market is this?

One proxy would be the size of the market for freight audits today, largely served by consulting firms and point solutions. Estimates vary, but it seems pretty large: one source puts it at 3.1 billion dollars annually (growing 6.8% each year), and another has it growing to $4.8 billion by 2024.

Another, more principled approach would be to measure the opportunity for for value creation. Loop creates value by helping companies save on their freight operations, a sizable opportunity for a few reasons:

• Audits represent a sizable opportunity since freight payments still contain so many errors. The former CEO of Maersk (a far better shipping company than Evergreen) once observed that nearly 12% of container industry invoices are inaccurate, and some have estimated that up to 25% of freight invoices contain an error.
  ‎Some of these errors (i.e. having the wrong billing address or rate per mile) can be simple to track down and fix, but others are more complicated. Unplanned detention, demurrage, or per diem fees, unapplied discounts or affiliation savings, freight class errors, customs errors, undiscussed accessorial fees, missed tariff discounts, miscalculated fuel surcharges, miscalculated weights, and incorrect exchange rates can stack up in the background — it’s estimated shippers could recover 2-5% of their total transportation spend through auditing alone.

• Analytics represent a sizable opportunity since without a clean data layer, they’re so difficult to generate today. There’s an industry trend here: Chief Supply Chain Officers (CSCOs) identified advanced analytics as the second highest priority emerging technology investment, with only 9% reporting no plans to invest. 79% of CSCOs are developing training to drive the adoption of these analytics within their organizations.

  Very few supply chain organizations have the data foundations needed to accurately represent their transportation networks. The one Loop has built can help shippers better understand which carriers to prioritize in a region, which areas to expand into, and which bottlenecks exist in their transportation networks. All of these things can have a huge impact.

If this sounds familiar, you might be reminded of the last company we covered, SmarterDx (who also might’ve just raised another round). They do something similar, leveraging domain-specific artificial intelligence to correct errors in hospital documentation. Similar to Loop, this eventually helps hospitals recover lost revenue, improving their bottom lines.

Adjacent Markets

For their audit and analytics products, Loop had to build a first-of-its-kind, robust and extensible data layer allowing shippers to digitize, centralize, and organize all their payments-related data.

It’s this data layer that uniquely positions Loop to eventually expand into other finance and operations-related workflows. Being a shipping organization’s financial system of record makes it easy to expand into the broader, 20.24 billion dollar logistics and supply chain software management market, which just happens to be the fastest growing market in enterprise application software.

For instance, Loop might expand into:

• Supply chain financing. In the status quo, invoices take 30 to 90 days to pay. Supply chain financing enables carriers to get paid early in exchange for a small discount (usually 2-3%) on their bill. How does it work? In an example case, a financier might pay the carrier 98% of their invoice upfront. The shipper will then pay the financier 99% of the carrier invoice on the day it’s due, allowing the financier to keep the 1% difference as profit. This is a win-win-win for carriers, shippers, and financiers:

  • Carriers are in a capital-intensive business. Getting paid early is huge for their ability to operate and expand.

  • Shippers get to hold their working capital for longer and also get 1% discount on their bill.

  • Financiers make 1% of the carrier invoice by floating the shipper the funds.

  This itself is a swiftly growing market: global supply chain finance volumes are up 21% this year to $2,184 billion, funds in use are up 20% to $858 billion, and revenue is up 23% to $97 billion. This growth has been consistent, and isn’t because of some pandemic-induced surge in freight demand:

  

  Since Loop already handles payments for their customers, they’re perfectly positioned to expand into this market down the line.

• Procurement workflows. Shippers span a wide variety of industries (i.e. industrial manufacturing, consumer packaged goods, apparel, financial services, etc.), and many need to procure raw materials to produce their own products. Just as shippers receive invoices from carriers for transportation, they also receive invoices from other shippers for raw materials. Like with carriers, these invoices must be audited and paid, and like with carriers, this is a laborious and antiquated process. Since Loop is already helping shippers with their freight audits, they’re perfectly positioned to start helping shippers with their procurement invoices and workflows also.

Loop’s ambition is to eventually handle everything at the intersection of freight and finance — to be wherever money flows in the supply chain.

Again, there’s an interesting parallel to another compound fintech startup here — Ramp. Both companies leveraged products helping customers spend less into platforms centralizing all finance and operations-related data, uniquely positioning each to build downstream finance and ops products for their respective verticals. Similar to how Ramp expanded from cards to help startups with expense management, procurement, working capital, and vendor management, Loop will do the same with shippers.

Aside: Macro Considerations

Before going forward, it’d be unfair not to address the ongoing freight recession and it’s impact on Loop. Logistics startups have had a hard time this year: Flexport’s revenue is down 70%, and Convoy (a former Loop customer) recently shut down. Is it possible that Loop’s market is actually… shrinking?

While the market isn’t exploding like it was during the pandemic-era freight boom, there are a few structural reasons why Loop has continued to grow and why they haven’t been as affected as other shipping startups:

• Flexport and Convoy are different types of businesses:

  • Flexport is a freight forwarder, meaning they arrange the transportation of goods on behalf of shippers. They handle tasks like storage, customs clearance, and documentation.

  • Convoy is a freight broker, meaning they act as an intermediary between shippers and carriers. They help negotiate transportation rates and coordinate the logistics of pick-up/delivery.

  Both were hurt by the drop in rates caused by the freight recession, which directly impacted their margins. In contrast, Loop is more of a pure software company with a software business model and software margins.

• Loop’s primary value propositions are cost and labor savings, both of which are top-of-mind for shippers and 3PLs during recessionary environments. For 3PLs in particular, the pressure to drive operating efficiency and increase working capital amid compressing margins actually leads them to solutions like Loop, which is why they’ve been growing quickly amongst this market segment.

Freight boom and busts have been around since the Motor Carrier Act of 1980, which removed federal entry controls on interstate trucking and made it easier for carriers to set their own rates. The current freight recession is nothing new, and will pass like each of the previous ones.

What won’t change, however, is how crucial freight and freight payments are to our economy. They’re not going anywhere — just like the massive shippers and 3PLs that comprise Loop’s core customer base. If Loop can stay the course, now’s an excellent time to rebuild a pretty foundational part of the freight system and larger economy.

‎

Market Share

So we know there’s demand from shippers and 3PLs to help optimize their freight payments. The question now is whether Loop will be their solution of choice.

As a startup, market leadership is a question of gaining market share and defending it. Loop looks promising when it comes to both.

Gaining Market Share

Although product-wise, most of Loop’s functionality is centered around audits and analytics, what really differentiates them is their mastery of data, which happens under the hood.

Like we mentioned earlier, payments-related data in freight is complex and fragmented. The documents involved are often spread across different file

• Types like bill of ladings, proof of deliveries, weight and inspection tickets, and packing lists.

• Formats like PDFs, JSONs, CSVs, JPEGs, emails, and EDI 210 transaction sets.

• Sources like APIs, EDIs, TMS integrations, ERP integrations, treasury partners, and various external data sources (i.e. fuel rates from the Department of Energy, mileage from PC Miler).

As a result, shippers often outsource the audits and research they need to different consulting firms or specialized SaaS tools. For most, the complexity of the data involved means it’d be too cost and labor-intensive to do in-house.

But this sort of outsourcing is a surface-level solution to a much more fundamental issue: it’s way too difficult for shippers to access their own data .

By centralizing and structuring this information for their customers, Loop solves this underlying issue. By

• Digitizing shipment data (allowing users to store any document from any system)

• Linking shipment data (allowing users to standardize, assign, and group data across carriers and documents)

• Analyzing shipment data (allowing users to search, explore, and export enriched data across carriers and geographies both in aggregate and at the line-item level)

Loop empowers shippers to do anything with their data more easily, not just audits. Compared to outsourcing, this more foundational approach is like curing the disease instead of just treating its symptoms.

The ability to drill into spend at the network, carrier, accessorial, and invoice levels is on it’s own a killer feature, but Loop builds on top of it with best-in-class audit and analytics functionality:

• Audits - on top of automatically identifying/root-causing discrepancies between invoiced and expected costs, Loop also automates exception management - addressing missing documents, mismatched rates, and unexpected accessorials.

• Analytics - on top of allowing users to analyze their data on their own, Loop also automatically surfaces insights to help with carrier performance or contract negotiations. They even offer a scenario planning tool where you can simulate different network scenarios for cost-effectiveness and efficiency.

A consulting firm or point solution might be able to tell you why an invoice is off, but it won’t address why that audit was so difficult to do in the first place. And it definitely won’t help you search for the same mistake across other invoices for use in next month’s contract negotiations.

On top of these two core functions, Loop has begun vertically integrating a series of other products. This works since customers already trust Loop as a source of truth.

For instance:

• Loop automates the end-to-end payment process for shippers, handling everything from cost allocation based on accounting rules/scheduling/aging to the actual payments and invoices.This means when a $1000 freight invoice is processed, Loop's system automatically allocates specific amounts to distinct account codes in the company's general ledger (i.e. $100 to INBOUND.013.4044, where INBOUND means inbound freight, 013 is the facility location, and 4044 is the business unit.

• Loop has begun offering supply chain financing with a Quickpay product. The control they have over shipper data gives them a major advantage when it comes to underwriting these sorts of arrangements.

• Loop is building functionality to help 3PLs automate accounts receivable workflows, the process where they generate invoices for shippers based on the invoices received from carriers. This is also a relatively complex and manual process today, but Loop is in the process of automating it.

So to recap, how does Loop get shippers to choose them over the consulting firms and point solutions they’re using today? It’s a few things:

• Loop is the only solution helping shippers organize their underlying data. By giving shippers control over it, they’re actually addressing the root issue.

• Loop is more than just a point solution. Beyond audits, their sophisticated data layer enables everything from rules-based cost allocation to scenario planning to automated insight generation.

More than just helping shippers save money, Loop is fundamentally improving how they consume and use data.

‎

Defending Market Share

Now that we understand how Loop’s been gaining market share, let’s consider how they can keep it. It’s still early days, but Loop has several moats they’ve been building into their business:

• Switching costs. By centralizing/structuring all this previously unstructured data and by owning core workflows like payments, audits, and forecasting, Loop becomes the default system of record for anything finance-related in freight. It becomes really hard for shippers to move off of Loop since every function (i.e. finance, accounting, operations) relies on the data they uniquely control.

• Technology. Bringing structure to all this data is a difficult technical problem. Parsing is hard since computer vision and natural language processing only recently got good enough to ingest all the different files and file types Loop deals with. Domain modeling is hard since each carrier generates their own agreements and rate contracts, meaning there’s an infinite number of rules around how much a shipment should cost.

  Loop has technical depth along each of these axes, leveraging proprietary models trained on proprietary data and a sophisticated domain representation for their core products. This moat gets wider with time: as Loop collects more data for their models and discovers more rules to include in their system, it gets harder and harder for new players to catch up.

• Platform effects. Loop can share anonymized insights around carrier contracts, rates, and performance across shippers to help users improve their own transportation networks. The more shippers join the platform, the more valuable Loop becomes in this regard.

‎

Competitive Landscape

Loop is a transformative player in the world of freight payments, but they’re far from the only company in the space. Now that we’ve got an understanding of Loop’s unique strengths as a product and company, let’s see how they stack up against their competitors.

Freight Audit Firms

Audit firms are the traditional solutions in the space. Like Loop, they help shippers with payments, audits, and analytics, but they do so in a consulting capacity.

There are many players here, but a few examples would include ShipSigma, Trax Technologies, and ARTC Logistics. Each of these firms relies on a dedicated team of consultants for tasks Loop has automated with AI — ShipSigma, for instance, assigns personnel to audit freight invoices for clients while ARTC Logistics can be commissioned to conduct “What-If” rate analyses assessing the impact of carrier rate changes.

Compared with these players, Loop has several things going for it:

• Automation makes them faster, more comprehensive, and more accurate than any human audit team. No traditional firm could ever offer near real-time audits or insights like Loop.

• The data layer they’ve built addresses the root cause for outsourcing in the first place. With Loop’s data, that “What-If” analysis could be done with a spreadsheet in-house.

• Loop isn’t limited by their employee count. Software allows Loop to scale with zero marginal costs, unlike the companies here.

Freight Audit Software

Next, there are a number of companies offering more technically oriented products. Like Loop, they help shippers with payments, audits, and analytics, but they do so with a number of point solutions that don’t address the underlying problem of data integrity.

Some notable players in the space include Shipware, OpenEnvoy, and Intelligent Audit. Some have grown quite large: Intelligent Audit, for instance, is almost 30 years old and processes over 20 billion in payments each year.

While these firms also provide tools to help shippers audit their invoices, they fall short of Loop in several important ways:

• Product-wise, they’re brittle point solutions. While they might automate audits for a specific contract from a specific carrier whose data they have a specific integration with, they don’t generalize like a human or AI auditor would. Why?

• Because product-wise, they lack the robust data foundation Loop has. Like consulting firms, this prevents them from addressing the root issue of data integrity/accessibility. This also prevents them from expanding into more sophisticated offerings, like analytics or automated insight generation. Why?

• Because technology-wise, these companies don’t have much going on. A lack of machine learning talent (we stalked each company’s LinkedIn profile) makes it unlikely they could build an ingestion system like Loop. An outdated stack and probable mountain of technical debt means it’s unlikely they’ll ever get that talent.

Financial Services Companies

Finally, there are a few banks and financial services companies that bundle audits into the freight finance services they primarily offer. Some major players here include US Bank (the fifth largest bank in the country) and Cass Information Systems. These firms also operate as consultants, giving them the same drawbacks as the traditional audit firms we looked at earlier. In addition, they also suffer from a lack of focus: since their main business isn’t freight (and even in freight their main business isn’t audit or analytics), they have little pressure to improve their data, audit, or analytics-related offerings.

These will be more direct competitors when Loop expands their footprint in freight financing, but even then Loop’s control of data will give them the advantage when it comes to things like underwriting.

Execution

Loop has redefined how shippers work with payments data, but they eventually want to handle everything at the intersection of freight and finance.

It’s an ambitious vision, but Loop’s built a team with both the engineering and operational chops to pull it off.

Consider a sampling of their roster (they have lots of people from Uber Freight and Flexport):

Matt McKinney - Co-Founder, CEO

Matt was previously a Data Science Manager and a Senior Data Scientist at Uber. Before that, he was a Founding Data Scientist and Software Engineer at MakerSights. He holds a BS in Industrial & Systems Engineering from USC.

Shaosu Liu - Co-Founder, CTO

Shaosu was previously a Software Engineering Lead at Uber, where he launched and scaled Uber Freight to over $2 billion in annual revenue. He holds a BS in Mathematics and Computer Science from UIUC.

Anastasia Fullerton - Head of Marketing

Anastasia was previously Head of Marketing at Extend (last valued at $1.6 billion). Before, she was a Senior Product Marketing Manager at Amplitude and a consultant at Accenture. She holds a BS in Management Science & Engineering from Stanford.

Heidi Yang - Head of Product Design

Heidi was previously Product Design Lead for Load Logistics at Uber Freight. Before, she was a Product Designer at Intuit. She holds a BS in Entrepreneurship and an MS in HCI from Carnegie Mellon.

Lauren Oliver - Head of Customer Success

Lauren was previously GM, Money Operations at Uber Freight. Before, she was a Key Account Manager at C.H. Robinson. She holds a BA in Political Science from UIUC.

Max Heinritz - Software Engineer

Max was previously a Senior Staff Software Engineer at Flexport, where he was tech lead for transportation automation. Before, he was a Software Engineer and Product Manager at Google. He holds a degree from UPenn.

Cody Krainock - Software Engineer

Cody was previously a Software Engineer at WhatsApp and a Software Engineering Manager at Flexport (engineer #13).

Sean Nicolay - Software Engineer

Sean was previously a Staff Software Engineer at Flexport and a Software Engineer at Facebook. He holds a degree from UIUC.

Matthew Ebeweber - Software Engineer

Matthew was previously a Software Engineer at Instagram and a Senior Software Engineer at Flexport. He holds a degree from UT Austin.

Paul Scheid - Software Engineer

Paul was previously a Senior Software Engineer at Flexport. Before, he was a Software Engineer at Wolverine Trading. He holds a degree from WashU.

Kevin Cai - Software Engineer

Kevin was previously a Senior Software Engineer at Flexport. He holds a degree from the University of Waterloo.

Sam Eastwood - Business Operations

Sam was previously Chief of Staff and manager of the Shipper Operations team at Uber Freight. He holds a degree from Northwestern.

Zach Negroni - Senior Account Executive

Zach was previously an Enterprise Account Execute at Samsara and a Corporate Account Executive at Box. He holds a degree from Boston College.

Tom Zidar - Senior Data Scientist

Tom was previously a Senior Data Analyst at Uber Freight. Before, he was a Data Scientist at Motive (aka KeepTruckin, last valued at $2.85 billion). He holds a degree from UMich.

Simon Demertzis - Solutions Engineer

Simon was previously a Senior Solutions Architect at MongoDB. Before, he was an Account Partner at Slack and a Director of Solutions Engineering at Decision Engines. He holds a degree from Northeastern Illinois University.

Will Taft - Software Engineer

Will was previously a Software Engineer at Composable Analytics. He holds a degree from Yale.

Evan Richards - Software Engineer

Evan Richards was previously a Senior Software Engineer at Uber. He holds a degree from MTSU.

Jack Taylor - Software Engineer

Jack was previously a Software Engineer at Rakuten. He holds a degree from Davidson.

Shin Thu San - Product Manager

Shin was previously a Product Manager at Block, Vanta, and Flexport. She holds a degree from UC Berkeley.

Shuwei Chen - Product Manager

Shuwei was previously a Product Manager at Isometric Technologies and Flexport. He holds a degree from UCLA.

Christina Danita - Software Engineer

Christina was previously a Software Engineering Intern at Athelas. She holds two degrees in CS from Stanford.

Grace Lee - Software Engineer

Grace was previously a Software Engineering Intern at Gusto. She holds a degree from Brown.

Conclusion

If Loop succeeds, the world of freight will essentially have its first automated clearing house — paying for a truckload of Coke will be as simple as buying a bottle of it at the store yourself.

Freight might be one of the most complex industries on Earth, but paying for it doesn’t have to be.

They’re hiring.

‎

Thanks to Wes Friedman and Vivek Gopalan for their help with this piece.

In case you missed our previous releases, check them out here:

• Warp - Revolutionizing the most foundational developer tool.

• Pinecone - Vector databases index the AI revolution.

• WorkOS - The API to increase your TAM.

And to make sure you don’t miss any future ones, be sure to subscribe here:

Finally, if you’re a founder, employee, or investor with a company you think we should cover please reach out to us at ericzhou27@gmail.com and uhanif@stanford.edu - we’d love to hear about it :)

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information from partnered venture firms. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should Join SmarterDx

(Click the link to read online).

In 2022, 97.6 million Americans lived in a federally-designated Health Professional Shortage Area — any location with less than 1 doctor for every 3,500 residents.

That’s 15 million more than in 2021, the year prior.

With 38% of physicians expressing some desire to retire in the next year and our country’s 65 or older population growing 47% by 2032, this problem is only going to get worse: by 2034, we’ll be short between 54,100 and 139,000 healthcare practitioners, up from the 17,000 we need today.

If you’re in medical school, prepare to never retire.

If you’re a doctor, prepare to see a lot more patients :)

So it’s a growing problem that instead, doctors are seeing a lot more paperwork.

In 2005, physicians spent a third of their time on administrative work and clinical documentation — summarizing a patient’s visit and activities for future reference. These days, it’s much higher - doctors today spend twice as much time here as actually meeting with their patients.

Between 2009 to 2018, the chart notes physicians write and review doubled in length. One review of 100 million notes containing 33 billion words found the average patient record had more than half the words of Shakespeare’s Hamlet. This means a physician seeing 10 patients in a day would be responsible for reading around 85 pages of single-spaced text across 691 notes — all before writing a good amount more in their own documentation afterwards.

All this contributes to physician burnout and takes away valuable time doctors could be spending with their patients.

The trouble is, this paperwork (documentation) is actually quite important.

As the canonical record of a patient’s hospital visit, clinical documentation is used to communicate across visits, healthcare providers, and insurance providers. It’s the fundamental source of a patient’s medical history — making sure it’s thorough and accurate is critical for ensuring continuity and quality of care. Small, secondary/tertiary diagnoses can become important with time and it’s important that they don’t slip through the cracks because someone was too busy to describe them.

If anything, doctors should be spending more time on documentation - about 70% of patient records contain incorrect information.

In one survey that had 22,889 patients read their own records, 21% found errors that were between “very serious” and “somewhat serious”. One patient found errors describing their medications and allergies, another with BRCA-1 saw they were labeled non-BRCA-1, and another got a PT referral for the wrong body part.

So should doctors be spending more time or less time on clinical documentation?

Well, in an ideal world, the responsibility for documentation would be completely separate from doctors.

Imagine walking into court as a defendant and discovering the judge presiding over your case was also responsible for typing out a transcript of your case and checking it for accuracy. The solution wouldn’t be to give the judge a better stenograph machine, no matter how good it is — it’d be to disconnect the responsibility from the judge.

Even when doctors are in a room with a patient, they’re only able to spend 52.9% of the time talking to or examining them: 37.0% of the time, they’re doing paperwork.

Improving clinical documentation and the process around would be one of the highest-leverage ways to can improve our healthcare system. Removing physicians as the bottleneck for every patient’s system of record with a more accurate, more automated approach to clinical documentation would improve workflows for doctors, outcomes for patients, and data integrity for hospitals.

That’s the future SmarterDx is building.

Background

Each time you visit the hospital, every test taken, scan done, and action prescribed gets summarized into a set of clinical documentation for that visit.

This documentation is prepared by your physician and is important to the healthcare process — it’s used as a summary of prior visits so future doctors looking through your file can get an understanding of your medical history without having to look through the raw charts and results themselves. This documentation is also important to the business of healthcare — it’s what eventually what gets formalized into standardized codes for billing purposes.

The challenge is, summarizing all that’s wrong/all that’s been done for a patient can be a lot — the average patient is diagnosed with 7 secondary conditions on top of their primary one and the average Medicare user on a high severity visit is diagnosed with 18. With 70 percent of doctors receiving more information than they can manage, it’s no surprise not every condition for every patient is documented very thoroughly - things slip through the cracks.

As a result, relying on doctors for record-keeping isn’t great for anyone involved:

• Doctors are more stressed and have less time to spend with their patients.

• Patients suffer worse outcomes — poor records hurt quality and continuity of care.

• Hospitals end up having to hire Clinical Documentation Integrity (CDI) teams, who work to manually identify and correct errors in documentation.

SmarterDx has developed an AI-based solution for clinical documentation improvement.

They use the same raw inputs as physicians — over 30,000 data points from lab results to radiology reports to pathology reports to vital signs to respiratory therapy to medication ordered to physical therapy notes — to fill in gaps a physician might have missed. For instance, if a patient admitted with a blood pressure of 170/100 gets prescribed medication for hypertension but the doctor forgets to record this as a secondary condition, SmarterDx can fill in that gap while also citing the clinical evidence to support that correction. So far, SmarterDx has identified inaccuracies in 80% of the charts they review.

The accuracy required for this “second pass” is accomplished via a combination of custom features, traditional machine learning techniques, deep learning, and LLMs - it’s much more than a fine-tuned version of GPT-4. You can think of SmarterDx’s product as a highly domain-specific application of retrieval-augmented generation, a Copilot for clinical documentation that’s able to cite its sources.

The mission is to begin offloading the burden of documentation from physicians to machines, something which would create value for everyone involved:

• Doctors worry less about documentation and can spend more time with patients.

• Patients receive better outcomes — better records improve quality and continuity of care.

• Hospitals save money and don’t have to hire as large of a CDI team.

As helpful as SmarterDx is in improving the accuracy of clinical documentation, its their impact on downstream revenue that’s really gotten hospitals excited.

Since most serious documentation errors are errors of omission (it’s more common to leave things out than to write too much in), correcting them in practice often means collecting on revenue hospitals would’ve otherwise missed (i.e. recording a missed secondary condition or a condition whose complexity was otherwise understated).

Of the errors SmarterDx discovers, roughly 15% end up affecting hospital revenue. SmarterDx then takes a cut of that, which is today how they monetize.

So far, this has been an incredibly effective business model. Each year, the average hospital using SmarterDx has recovered $800,000 in otherwise missed revenue per 100 beds, or $2 million per 10,000 discharges. Based on this, SmarterDx grew revenue 6x in the past year and are on track to 25x over a two year period. The company scaled to over $1 million in contracted revenue before raising their first round of funding last year and will end this year at a $6 million run rate. By the end of the first half of next year, they will be at $15 million.

The company was founded in 2021 by Michael Gao, M.D., Josh Geleris, M.D., and Nathan Perilo. Michael previously oversaw applied AI at New York-Presbyterian, was an Assistant Professor in Medicine and the Medical Director for Transformation at Cornell, and did his Residency in Internal Medicine at New York-Presbyterian after receiving his M.D. from the University of Michigan. After receiving his M.D. from Israel’s Technion, Josh also did his Residency at New York Presbyterian before joining Columbia’s Irving Medical Center as an Assistant Professor. Nathan was previously Director of Quant Libraries and Global Engineering at CitiBank.

Joining them are Floodgate and Flare Capital, who co-led SmarterDx’s $6 million dollar seed round in 2022.

SmarterDx has a strong product, excellent traction, and (as we’ll soon see) a rather stacked team. They’ve got a vision to transform the healthcare system’s system of record and a mission to free doctors from the burden of documentation.

But can they build a lasting business off of it all? Especially given how challenging it is to change anything in healthcare?

Well, we think they have a shot.

Let’s dive deeper into why.

The Opportunity

We’ll begin our exploration of SmarterDx like we do for all of our companies, from a first principle:

Large companies lead massive, growing markets.

Many startups with product-market fit don’t get very big (and aren’t worth joining) because they don’t meet this condition. Thus, we must establish two things:

1. SmarterDx operates in a massive, growing market.

2. SmarterDx will become a leader in that market.

Market Size

SmarterDx makes money when hospitals use their software to correct inaccuracies in clinical documentation. In other words, they operate in the market for clinical documentation integrity (CDI) solutions.

If there are 1) lots of inaccuracies in clinical documentation and 2) lots of hospitals willing to pay to address those inaccuracies, then the market for SmarterDx’s technology is large: there’s an opportunity to create a lot of value and an opportunity to capture a significant part of it.

We think there’s reason to believe both of these conditions:

• There are lots of inaccuracies in medical documentation.

  We’ve already discussed how despite chart notes doubling in length between 2009 and 2018, 70% of patient records still contain incorrect information. These facts might be related: electronic medical records have a serious problem with duplicate text.

  One study of 104,456,653 routinely generated clinical notes containing 32,991,489,889 words found 50.1% of it was copied-and-pasted from prior documentation. Another study of 24,000 notes from 460 clinicians at UCSF found more than 80% of it was copied-and-pasted from a prior document. Both studies found the copying to be prevalent across physicians of all levels of training, and that physicians were equally likely to copy from prior documentation they wrote themselves and prior documentation written by their colleagues. All this copying makes it more likely outdated/inaccurate information is accidentally left in and updated/important information is left out.

  Handwritten notes are worse: they’re often physically unreadable. Before the rise of electronic medical records, up to 25% of medication errors were related to illegible handwriting, contributing to an estimated 7,000 deaths each year.

  

  Ultimately, the root cause of all this is simple: physicians don’t have the bandwidth to make documentation as good as they’d like it to be. As we’ll see next, this is something hospitals have had to work around.

• Hospitals are willing to pay to fix documentation.

  Today, hospitals hire teams of Clinical Documentation Integrity (CDI) specialists who do manual second passes over documentation, also with the goal of fixing errors and recovering missed revenue. These teams have meaningful impacts on patient outcomes: 88% of hospitals confirmed documented quality improvements within 6 months of implementing a clinical documentation integrity program.

  CDI as a practice is widespread today, but it only really picked up in the last two decades with the rise of easily-revisable electronic medical records. For instance, while the total number of healthcare job listings increased just 9% from 2007 to 2011, the number of listings for clinical documentation specialists increased by 223%, which made it the second fastest growing title in health informatics. By 2013, 66% of hospitals had already implemented a CDI program, with about half of those who hadn’t planning on creating one by the following year. It continues to grow today: the Bureau of Labor Statistics classifies CDI professionals under “Medical Records Specialists”, a profession which continues to grow roughly three times as fast as the average occupation.

Despite both of these facts, however, hospitals are facing increasing pressure to automate or outsource functions like CDI. There are several reasons behind this:

• People are expensive. Each year, administrative expenses account for over 25% of total national health care expenditures. This is both dramatically higher than other industrialized countries and growing faster as a proportion of overall costs than actual clinical expenses.

  Hospitals spend 22.6 billion each year maintaining medical records, and revenue cycle (the department CDI teams most commonly fall under) is the number one function hospital administrators are looking for outsourcing solutions in.

• There aren’t enough people. Even if they aren’t looking to outsource, some 48% of revenue cycle management departments are experiencing labor deficiencies, with one in four finance leaders needing to fill over 20 positions to be fully staffed. More generally, over 57% of health systems and hospitals have over 100 open roles to fill.

Put this all together - significant inaccuracies in clinical documentation, a willingness to pay to fix them, a general pressure to outsource/automate things - and it’s clear why demand for a tool like SmarterDx has been strong.

Given broader trends like the incoming shortage of physicians or the growing adoption of AI-powered tools in healthcare, there’s reason to believe this demand will grow also.

Market Leadership

So we know hospitals are adopting solutions to improve their clinical documentation. The question now is whether SmarterDx will be their answer of choice.

As a startup, gaining market share is a function of having a great product (good reason to adopt) and great distribution (little reason not to adopt). SmarterDx has strong answers to both.

Product

Compared to the teams of clinical documentation specialists hospitals employ today, SmarterDx’s end-to-end, AI-driven approach offers several benefits:

• Greater accuracy. SmarterDx identifies inaccuracies in 80% of the charts they review, providing all the clinical evidence needed to support corrections in each case. These errors range from very minor to very serious, but about 15% of them end up affecting hospital revenue.

  This impressive lift in accuracy is accomplished in two ways:

  • Better inputs. SmarterDx uses the same clinical results physicians review as inputs for their algorithms — over 30,000 ground truth data points including lab results, radiology reports, pathology reports, vital signs, respiratory therapy, medication ordered, and physical therapy notes. Starting with raw clinical data is critical for the accuracy, interpretability, and thoroughness of SmarterDx’s outputs. If you were to train a model solely on existing documentation instead (i.e. how Github Copilot was trained on Github repositories), you’d do a lot worse because of how many inaccuracies the existing corpus contains.

  • Better models. SmarterDx uses a combination of custom, domain-specific features, traditional ML techniques, and deep learning approaches alongside LLMs to solve for high recall/precision while maintaining high explainability. It’s this hybrid approach that allows SmarterDx to surface clinical evidence when correcting errors — relying primarily on LLMs (even tuned ones) wouldn’t work since LLMs suffer from hallucinations and aren’t readily explainable, both unacceptable traits in medical settings.

  It’s worth re-emphasizing how the corrections surfaced by SmarterDx are discovered after human CDI teams and the software tools they use have already made their own suggestions. SmarterDx operates as a final check for all clinical documentation produced.

  The impressive domain expertise and engineering underpinning all this constitutes a growing technical moat for the company.

• Greater affordability. SmarterDx’s automated solution is more efficient than entire teams of human CDI specialists. We were shown how at one hospital, SmarterDx was recovering more revenue than the entire team of CDI specialists they were working alongside.

• Greater accessibility. SmarterDx’s “second pass” approach comes in after physicians and doesn’t directly interrupt their work. This makes their product easier to use and also begins to address the root issue - the burden of documentation-related work on physicians. This is in contrast to weaker, more traditional AI “copilot” systems that still rely on direct inputs from physicians. We dive deeper into these below.

Distribution

Compared to most health tech startups, SmarterDx has found their product much easier to distribute. Consider one hospital’s testimonial:

Why is this the case?

Well, it’s worth considering why hospitals are so difficult to distribute into normally. There are two systemic forces working against any budding healthtech company:

• High adoption friction. Hospitals are often using relatively old software (read: no simple APIs) and operate without strong IT departments. The technology used isn’t really standardized across hospitals, and the processes around these technologies — data pipelines, user habits, etc. — are complex and difficult to update. Together, this means that integrating a new product (getting it to work with existing tools/data schemas) and adopting it (training teams to use it effectively) is a high-friction process, both for the hospital and the company.

• Low adoption incentives. As businesses, hospitals benefit from monopolistic market positions and thus face less pressure to innovate and invest in new technologies. You’ll never see hospitals browsing for new products to try on Product Hunt. This means that to even be considered, products have to solve major pain points or seriously benefit a hospital’s bottom line.

This combination of high adoption friction and low adoption incentive means it’s harder for new products to gain traction with hospitals than with most markets.

Thankfully, SmarterDx has crafted both their product and business model to address these issues:

• Reverse ETL makes their product easier to integrate from a technical perspective. Unlike competitors who often require live data, specific formats, or integration into certain electronic medical record systems (all things hospital IT departments and EMR systems aren’t great about), SmarterDx has no technical barriers to adoption — you can get up and running with a regular flat file drop. The secret is a suite of reverse ETL systems built in-house allowing SmarterDx to take data in whatever format they want before mapping it to a unified patient schema across hospitals, which their models use. From the hospital’s perspective, this reduces integration friction.

• Zero upfront cost makes their product easier to adopt from a budget/procurement perspective. The unique accuracy and efficiency of SmarterDx’s system enables them to charge via a revenue split instead of a more traditional subscription model, something which offloads financial risk from hospitals. From the hospital’s perspective, this reduces procurement friction.

• End-to-end automation makes the product easier to adopt from a workflow perspective. Since the product sits at the very end of the clinical documentation process, SmarterDx doesn’t require physicians to change their habits or workflows. Since it generates corrections complete with clinical evidence, it’s very easy for CDI professionals to accept the output. From the hospital’s perspective, this reduces adoption friction.

• Revenue recovery as a quantifiable value proposition makes it easier for hospitals recognize the value of the product. This both aligns and increases adoption incentives.

Competitive Landscape

SmarterDx isn’t the only business leveraging AI to help hospitals improve the integrity of their documentation and revenue cycles. Now that we have an understanding of SmarterDx’s unique strengths as a product and company, let’s see how they stack up against their competitors.

Below, we consider a few categories of software products, from most competitive to least competitive.

Documentation Improvement

There are a large number of other companies also selling software to help hospitals improve their documentation. Let’s review some major players in the space.

• 3M is a public company that sells many different industrial, worker safety, healthcare, and consumer products. Among their offerings are Engage One and CDI Collaborate, two software tools that help CDI teams with their work. Engage One offers AI-enabled “nudges” (reminders) as doctors write their documentation, and CDI Collaborate is a cloud-based CDI workflow application.

• Iodine is a unicorn that sells AwareCDI, an AI-enhanced CDI workflow platform. The primary value prop is a built-in AI system called CognitiveML which ranks files based on the likelihood they contain errors. This helps CDI teams prioritize the charts they review, improving efficiency. AwareCDI also offers a query system making it easier for CDI specialists to ping physicians for feedback.

• Nuance Communications is a Microsoft-owned, formerly public company offering CDE One, another cloud-based CDI workflow application. Similar to Iodine, CDE One also leverages AI to generate prioritized worklists for CDI teams.

  

Unlike SmarterDx, these products are primarily focused on workflow management and collaboration, things that help CDI teams organize their work. When they do incorporate AI, it’s generally a weaker version than what SmarterDx has deployed: instead of directly offering corrections and supporting evidence like SmarterDx has deployed, they operate indirectly through prioritized lists or “nudges” telling physicians to write extra details down.

The upshot is that compared to these products, SmarterDx goes a step further and automates the end-to-end process independently of human input, only requiring confirmation at the end. This makes them far more powerful with regards to accuracy and efficiency.

Coding Automation

If you look up medical billing and AI, there’s a good chance a medical coding startup will be the first company to pop up. These businesses help automate the process where documentation is converted to the standardized codes used to bill patients. Some players here include:

• Fathom Health - AI to help automate medical coding. Last raised a $46 million Series B from Alkeon Capital and Lighspeed.

• CodaMatrix - AI to help automate medical coding. Last raised a $55 million Series A from SignalFire.

• Nym Health - AI to help automate medical coding. Last raised a $25 million Series B from Addition.

These aren’t competitors, but it’s worth explaining how they differ from CDI companies like SmarterDx.

Documentation is what ultimately gets converted to medical codes, meaning companies like SmarterDx operate a stage before the companies above. Since you can’t code or bill for something that was never documented, these success of these companies ultimately relies on the integrity of the clinical documentation they’re converting.

Revenue Cycle Management

There are a large number of companies offering software to help hospitals manage the broader function of revenue cycle. These aren’t really competitors either, but some offer features related to revenue recovery (SmarterDx’s value proposition) as part of their platforms. We’ve included a few here for thoroughness.

• Vitalware is a revenue integrity platform for hospitals. They help hospitals avoid losing revenue to poor charge capture performance (missed charges, incorrectly coded services, delays in billing) and changing reimbursement guidelines. They also help hospitals manage their chargemasters.

• Cerner is another revenue cycle management platform for hospitals. They help manage scheduling, financial clearance, and registration for patients. They also automate many of the billing workflows hospitals go through.

• Craneware is another revenue cycle management platform for hospitals. They help hospital administration manage pricing, revenue, cost & decision support, and chargemaster updates.

Execution

Changing how the healthcare system thinks about clinical documentation means challenging a large number of traditional healthcare workflows and software providers head-on.

That requires a strong team.

Fortunately, SmarterDx has such a team, one with a lots of experience at the intersection of AI and healthcare. Consider a sampling of their roster:

Michael Gao - Co-Founder, CEO

Michael previously oversaw applied AI at New York-Presbyterian Hospital. Before that, he was an Assistant Professor in Medicine and the Medical Director for Transformation at Cornell. He holds an M.D. from the University of Michigan.

Josh Geleris - Co-Founder, Head of Product

Josh was previously an Assistant Professor at Columbia University Irving Medical Center. He holds an MD from Israel’s Technion and did his residency at New York Presbyterian Hospital.

Nathan Perilo - Co-Founder, Director of Technology

Nathan was previously the Director of Quant Library and Global Engineering at CitiBank. Before that, he was a consultant with Infusion Development.

Wayne Grodsky - Chief Commercial Officer

Wayne was previously Senior Vice President at PointClickCare (last valued at $5 billion), Chief Growth Officer at Access TeleCare (acquired for $300 million), and Enterprise Growth Officer at ZirMed (acquired by Waystar).

Henry Su - Head of Finance and Operations

Henry was previously VP of Business Operations at Embark Trucks (IPO’d at $5 billion) and a Management Consultant at McKinsey.

Sandra Frykman - Head of Sales

Sandra was previously Vice President of Sales at Elemento Health and Vice President of Sales at Vivian Health. Before that, she was Head of Sales at Incredible Health.

Brendan White - Director of Growth Operations and Enablement

Brendan was previously Director of Growth Operations and Enablement at PointClickCare (last valued at $5 billion). Before that, he was Director of Sales Operations at inContact (acquired for $940 million).

JP Korelc - Regional Vice President of Sales

JP was previously Director of Business Development at OncoLens. Before that, he was Director of Health Systems Sales at Current Health and Regional Vice President of Sales at Access TeleCare (acquired for $300 million).

Janita Adams - Regional Vice President of Sales

Janita was previously Regional Vice President of Sales at Incredible Health (last valued at $1.6 billion). Before that, she was Regional Vice President of Sales at Access TeleCare (acquired for $300 million).

Ning Liang - Principal Software Engineer

Ning was previously Co-Founder and CTO of HealthSherpa, a $100M ARR healthtech startup. Before that, he was a Software Engineer at Twitter, Rapleaf, and Microsoft. He holds a BS in Mathematics and an MS in Physics from Yale.

Alexander Yankov - Principal Software Engineer

Alexander was previously a Staff Software Engineer at Google. Before that, he was a Founding Engineer at Alkymi, a Machine Learning Engineer at HyperScience, and Software Engineer at Goldman Sachs.

Jason Zeiber - Principal Product Designer

Jason was previously a Senior Principal Product Designer at Clari (last valued at $2.6 billion). Before that, he was a Product Architect at DealPoint (acquired by Clari), a Director of UX at Bumped, and a UX Lead at HP.

Marie Klosterman - Lead Data Scientist

Marie was previously Director of Data Science at Olive (last valued at $4 billion). Before that, she was a Data Scientist at Optum (UnitedHealth Group).

Ruth Nancy - Director of Product Management

Ruth was previously Director of Product Management at Olive (last valued at $4 billion). Before that, she was Director of Product Management at HealthVerity (last valued at $720 million).

Scott Fleming - Research Engineer

Scott was previously a Graduate Student Researcher at Stanford Medical School. Before, he was a Machine Learning Research Intern at Apple and a Research Scientist Intern at Verily. He holds a BS in MCS, an MS in CS, and a PhD in Biomedical Data Science, all from Stanford University.

Richard van Dys - Senior Software Engineer

Richard was previously a Senior Software Engineer at Truepill (last valued at $1.6 billion). Before that, he was a Senior Platform Engineer at Moovweb.

Bates Jernigan - Senior Software Engineer

Bates was previously a Senior Software Engineer at Carbon Health. Before that, he was a Software Engineering Manager at SpaceIQ and Full Stack Developer at Speed Digital.

Steven Reed - Software Engineer Steven was previously a Senior Software Engineer at Mutiny. Before that, he was a Senior Staff Software Engineer at Paypal.

The team today is 17 people and growing swiftly.

If you’ve ever wanted to work alongside the founding CTO of a $100M ARR healthtech startup, a global director for quant engineering at Citi, or a Stanford PhD in machine learning, now’s your chance.

Conclusion

American physicians will face major challenges in the coming years, but more paperwork doesn’t have to be one of them.

SmarterDx is building a future with better work for physicians, better outcomes for patients, and better data for the healthcare system.

They’re hiring.

In case you missed our previous releases, check them out here:

• Warp - Revolutionizing the most foundational developer tool.

• Pinecone - Vector databases index the AI revolution.

• WorkOS - The API to increase your TAM.

And to make sure you don’t miss any future ones, be sure to subscribe here:

Finally, if you’re a founder, employee, or investor with a company you think we should cover please reach out to us at ericzhou27@gmail.com and uhanif@stanford.edu - we’d love to hear about it :)

If you’re interested in roles with any of the companies we’ve covered (or the unreleased ones we have in our pipeline 👀), please fill out this form to join our talent network.

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information from partnered venture firms. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should (Still) Join Pinecone

(Click the link to read online).

About a year ago, we picked Pinecone as the second company to feature in this newsletter.

A few months ago, they announced their 100 million dollar Series B led by Andreessen Horowitz. Besides Andreessen, other new investors included ICONIQ Growth. Prior investors Menlo Ventures and Wing Venture Capital also participated in the round.

We wanted to take this chance to reflect on the original piece while exploring how Pinecone has progressed in the year since. How are they doing? Where are they now?

Let’s get into why you should (still) join Pinecone.

‎

Background

Vectors are the backbone of modern artificial intelligence.

As the universal input and output format for neural networks, you can’t use AI without also using vectors. For instance:

• Sentences are converted to vectors so they can be used by language models. These power applications like semantic search, machine translation, and conversational chatbots.

• Images are converted to vectors so they can be used by vision models. These power applications like object detection, facial recognition, and generative fill.

• Recordings get converted to vectors so they can be used by audio models. These power applications like speech transcription, anomaly detection, and covers of classic songs by South Park characters.

If you want to work with the inputs and outputs of machine learning systems at any scale, you’ll have to get comfortable working with vectors.

The problem is, vectors are kind of hard to work with.

Unlike traditional data types, which we could scalably store/update/query with an arsenal of traditional databases, there was no such system for vectors. Doing something as foundational as finding the closest vector to some query vector (something chatbots often do when looking for a source to cite) was hard when you needed a result in constant time — and even harder at scale.

Traditional databases didn’t work with vector data because vector data required a completely different set of indexing algorithms:

Efficiently searching through a mountain of vectors for close matches to one of interest (what a lot of the applications above boil down to) is a very difficult task. Vectors aren’t standard data - you can’t stick them into standard databases and expect standard runtimes on standard queries. **Rather, they require special indexing protocols based on special algorithms.

In a very real sense, you could think of vectors as the first data type native to machine learning: if traditional apps were sequences of manually-defined transformations on manually-defined data structures (classes, objects), AI-native apps were sequences of learned transformations on learned data representations (vectors, tensors).

This novelty meant we needed a different set of tools to work with vectors properly, a set of tools that didn’t yet exist.

Pinecone was trying to change that.

They had built the first fully managed database for vectors, making it super easy for developers to work with this foundational new data type. Out of the box, Pinecone offered:

• Simple Querying - instead of implementing your own tree, neighborhood graph, hash, or quantization-based approximate nearest neighbors indexing protocol for constant-time vector search, you could query Pinecone with a simple REST API.

• Simple Scaling - instead of sharding and distributing your database yourself as you scale (a non-trivial task), Pinecone does everything for you. They also automatically parallelize request handling and spread replicas across different availability zones to improve throughput and system robustness.

Thousands of developers were already using Pinecone, ranging fast-growing startups storing a few thousand vectors to Fortune 500 companies storing over 50 billion. Pinecone was already powering the semantic search inside of Mem, the feed ranking inside of Clubhouse, the cloud infrastructure alert management system inside of Expel, and various systems within Gong, Course Hero, and Klue.

Most notably, it was being used to power the facial recognition system for cows on a farm.

It was obvious that Pinecone had built a useful product.

What was less obvious was whether they could build a massive, “category defining” business off of it.

Well, we thought they had a shot:

The AI revolution is in full swing and much of it is powered by vector data. As a result, vector databases index the AI revolution: literally by indexing the collective output of the world’s models, and figuratively since so many AI applications are built on top of vector search. As the strongest vector database company, Pinecone will lead that revolution.

Our chain of reasoning was roughly as follows.

A Large Market

Pinecone operated in the market for vector databases — they made money from other companies looking to build on top of vector databases. If the value generated by applications uniquely enabled by vector databases was large, then the market for vector databases would also be large.

We thought this was a large, growing, and important market because the low-latency vector search these databases uniquely enabled powered tons of important applications, each with massive markets of their own. These included:

Image Recognition. Applications leveraging image recognition (Instagram, Pinterest) work by using neural networks to convert images to vectors. You can then search that vector against the vectors of known images to find close matches, telling you what the original image was. Image recognition has applications in security and surveillance, medical imaging, visual geolocation, facial or object recognition, barcode reading, gaming, social networking, and e-commerce. The image recognition market is set to be worth 109.4 billion by 2027.

Recommendation Engines. Applications leveraging recommendation engines (TikTok, Twitter, Tinder) work by representing content as vectors. You can then search that vector against the vectors of other content to find close matches, giving you similar content to recommend. The recommendation engines market was worth 2.12 billion in 2020 and is set to reach 15.13 billion by 2026.

Chatbots. Applications leveraging chatbots (Ada, Wysdom) work by converting user questions to vectors. You can then search that vector against the vectors of known questions to find close matches, telling you which answers to provide. The chatbot market was worth 2.9 billion 2020 and is set to be 10.5 billion by 2026.

We clearly undershot the market for chatbots, but a win is a win.

After all, our broader thesis was that vector databases stood to index all of AI development activity — we knew that as the universal input/output format for modern AI systems, vectors were a data type anyone doing anything with AI would have to work with. That this was so difficult meant meant there was a huge opportunity for vector infrastructure startups like Pinecone to create and capture value — betting on vector databases was betting on AI development activity itself.

A Leading Product

Pinecone was the first fully managed vector database, but they weren’t the only vector database on the market. At the time of publication, the other options were Milvus (Zilliz), Vespa, Weaviate, Qdrant, Vald, and GCP’s Vertex AI Matching Engine. We wrote detailed, direct comparisons of each of these against Pinecone (as well as with vector search libraries like ScaNN/FAISS/ANNOY, traditional search libraries like Elasticsearch/Opensearch, and self-hosted vector databases like Vald) which you can read in the original piece if interested.

The upshot was that Pinecone had the best product, both technically and with regards to developer experience. For instance:

• Pinecone was the first to offer hybrid search in their database, allowing developers to apply metadata filters on vector queries. Pinecone was the first to offer partial updates on vectors and metadata - other solutions required you to delete and re-upload vectors in a non-atomic fashion. Pinecone offered best-in-class integrations with Apache Spark, OpenAI, Deepset, and Cohere, which made connecting your database with existing models/data sources as easy as possible.

• Pinecone was the only managed service on the market with a free tier. While everyone else pointed you to their open source project, trying Pinecone was as easy as logging in with Github, spinning up an index, and uploading your vectors. For developers wanting to test solutions quickly and without paying (i.e. the high-growth startups and teams most often outsourcing vector search), this was critical. By providing a fully managed free tier, Pinecone’s time to value was an order of magnitude lower than anything else on the market, meaning you could start 10x faster with them.

Plus, as the first company to market with a managed vector database, they had a two year head-start on the intricacies around building a managed service — the technical details, the developer experience, the enterprise-readiness.

As late as July of 2022, Pinecone remained one of only two fully managed offerings (serving databases is hard) and the only option with a free tier. It was clear they were best positioned to capture any new demand for this new class of infrastructure.

A Strong Team

How did Pinecone build the best product in the space? They had the best team building it.

Besides comparing Pinecone’s product against their competitors, we also compared the strength of their respective teams. Beyond generally being higher quality than their competitors, there were two things we really liked about the folks at Pinecone:

• They had lots of experience with machine learning. This gave them knowledge of the kinds of products/workflows Pinecone would be used for and empathy for their end-users.

• They had lots of experience with databases. This gave them the insight and intuition needed to build a database product and take it to market.

For instance:

• Founder and CEO Edo Liberty was a Director of Research at AWS and Head of Amazon AI Labs, where he was behind SageMaker, Kinesis, QuickSight, and Amazon ElasticSearch.

• VP of Engineering Ram Sriharsha was VP of Engineering at Splunk, head of the genomics product vertical at Databricks, and a Principal Research Scientist at Yahoo.

• Board member Peter Wagner was previously a seed investor in Snowflake. Board member Tim Tully was previously CTO of Splunk and VP Engineering at Yahoo.

• Advisor Bob Muglia was previously CEO of Snowflake. Advisor Bob Wiederhold was previously CEO of Couchbase.

In essence, we saw a market poised for enormous growth, a leading product in that market, and a stacked team focused on beating the competition.

What wasn’t to like?

For the full analysis, read the original piece here.

‎

Growth

So how has Pinecone progressed in the year since?

Well, before even getting to Pinecone, it’s worth appreciating how much vector databases as a whole have grown. The category has raised a collective $350 million at this point, with almost 70% of that being in the last 12 months:

• Zilliz raised their $60 million Series B Extension (August 24th, 2022).

• Chroma raised their $18 million Seed Round (April 6th, 2023).

• Qdrant raised their $7.5 million Seed Round (April 19th, 2023).

• Weaviate raised their $50 million Series B (April 21st, 2023).

And of course, Pinecone announced their $100 million Series B on April 27th.

With regards to product quality, team quality, capital raised, valuation, and (most importantly) traction, Pinecone remains a clear leader in the market. Let’s look at how both Pinecone the product and Pinecone the company have grown in the year since.

The Product

Pinecone’s userbase has grown dramatically in the past year.

At this time last year, Pinecone was serving several “thousands” of users, all accrued roughly within a year of launch.

Today, Pinecone has over 10,000 signups every day. Their core product — a vector database — has seen some major updates. These include:

• High-throughput vector indexes. Pinecone’s vector indexes run on pods, pre-configured units of hardware optimized for either storage (s1) or throughput (p1). Pinecone’s next-generation p2 performance pods use a new graph-based index to achieve search speeds 10x faster than p1 and significantly higher throughput per replica (up to 200 queries per second). p1 and s1 pods were also upgraded to see 50% faster queries and higher throughput. Shipped 12/8/22.

• Sparse-dense embeddings. Dense vectors have fewer values (100-1000), but nearly all of their values are non-zero. Sparse vectors have more values (i.e. 100,000), but nearly all of their values are zero. Sparse and dense vectors are used to represent different things (i.e. term-level importance vs semantic meanings), but it’s useful to query over both if you’re building something like hybrid lexical-semantic search or multimodal search. You couldn’t previously store both types in one database for performance reasons, but Pinecone’s new sparse-dense hybrid search technology has solved that problem. Shipped 2/23/23.

• Vertical Scaling. Previously, an index growing above available capacity meant you had to scale horizontally by creating a new index with more pods and switching over manually, which took engineering time. Now, vertical scaling means you can grow your existing index up to 8x with zero downtime by adjusting the capacities of individual pods. Shipped 8/16/22.

• Upgraded Free Plan. Storage-optimized s1 pods are now available on the free plan, giving users 5x greater capacity while trying Pinecone. Shipped 8/16/22.

Pinecone has also picked up some serious new customers: Klarna, HubSpot, Yelp, Bain, GoDaddy, Accenture, ClickUp, Zapier, Shopify, CVS, Midjourney, YCombinator, BambooHR, and Microsoft now all use the service.

A suite of new enterprise-facing features has made this possible:

• SSO. Pinecone now supports a centralized access point, streamlined user management based on roles, and secure authentication methods.

• Usage dashboards. A new usage dashboard allows developers to start projects, add team members, and view usage/spend without leaving the console. Users can break things down with filters by timespan, project, and service to visualize and predict future usage.

• Multi-cloud support. Pinecone is now available for deployment on the GCP, AWS, and Azure marketplaces. For large customers, this helps to consolidate spend/billing while also reducing procurement time (i.e. licensing and configuration).

• Integrations. A suite of new partnerships and integrations with popular tools (i.e. OpenAI, LangChain, Databricks, Datadog, TruLens, ElasticSearch) has made it easier for developers to integrate Pinecone.

The Company

The company itself has also grown significantly in the past year. Pinecone has maintained a high hiring bar and has managed to recruit some impressive new people. They include:

• Bob Wiederhold - President and COO

  Bob was previously CEO and Executive Chairman of Couchbase (NASDAQ: BASE). Before that, he was CEO of Transitive (acquired by IBM) and CEO of Tality, a design services and intellectual property company with $200M of revenue.

• Elan Dekel - VP of Product

  Elan was previously Product Lead for Core Data Serving at Google, where he was product lead for the indexing/serving systems powering Google search, YouTube search, Maps, Assistant, Lens, Photos, and more. Before that, he was Founder and CEO of Medico, which was acquired by Everyday Health.

• Chuck Fontana - VP of Business Development

  Chuck was previously SVP, Corporate & Business Development at SentinelOne (NYSE: S). Before that, he was VP, Corporate & Business Development at Okta and Director, Business Development & Product Marketing at Cisco.

• Michael Proia - VP of Sales

  Michael was previously Vice President Americas at Kong (last valued at 1.4 billion). Before that, he was Regional Vice President of Sales at Hortonworks.

• Travis Donia - Director of Engineering

  Travis was previously VP of Engineering at Clarivate (NYSE: CLVT). Before that, he was CTO/Head of Product at Context Matters, which was acquired by Clarivate.

• Hien Phan - Head of Product Marketing

  Hien was previously Director of Product Marketing at Amplitude (NYSE: AMPL). Before that, he was Director of Product Marketing at Mode Analytics, which was acquired by ThoughtSpot for $200 million.

• Kyle Himmelwright - Director of Revenue Operations

  Kyle was previously Head of Revenue Operations at Birdeye, which just crossed $100m ARR and 100,000 customers. Before that, he was Director of Revenue Operations at Yelp.

• Amie Ernst - Director of Talent Acquisition

  Amie was previously Global Talent Leader - Core Engineering at Goldman Sachs. Before that, she was a Senior Manager, Recruiting at Amazon Web Services.

‎

The Future

As far as they’ve come, Pinecone still has a long way to go. AI is moving very quickly, and staying a market leader isn’t a given. To stay on top product-wise, Pinecone will have to keep executing along several axes:

• Building a great database. At the center of everything, Pinecone can’t stop innovating on their core product - this means making it faster, more reliable, more flexible, and more efficient with improvements around indexing, scaling, querying.

• Building a great service. One layer out, Pinecone can’t stop improving the developer experience, production readiness, and enterprise readiness that separates their managed offering from open-source ones. This means features around data management, metadata handling, security, compliance, and more.

• Building a great ecosystem. Another layer out, Pinecone can’t stop nurturing the ecosystem that’s driven so much of their growth. This means continuing with integrations, partnerships, meetups, educational content, and community-driven events.

If they can keep it up, Pinecone’s future is promising.

Vectors are a cornerstone of machine learning because they’re how computers make sense of unstructured data (text, images, video). Caching and organizing them is how machine learning systems today maintain state between inference runs.

In other words, a long-term store for vectors is the closest thing AI systems today have to long-term memory.

Vector databases are long-term memory for AI.

‎

Conclusion

Vectors are the universal input and output format for neural networks. Every image generated, chatbot queried, or forward pass ever started with someone converting their input into a vector. In a very real sense, vectors were the first data type native to machine learning.

Pinecone was the first database giving them the respect they deserved.

They’re (still) hiring.

‎

If you’re interested in the investing side of things, our friends at NFX (the largest seed-stage fund in the world) are looking for a new member to join their investment team. Spending a few years at a venture firm can be a great launchpad for an operating career, especially if you have an engineering background. Learn more here.

In case you missed our previous releases, check them out here:

• Warp - Revolutionizing the most foundational developer tool.

• Pinecone - Vector databases index the AI revolution.

• WorkOS - The API to increase your TAM.

And to make sure you don’t miss any future ones, be sure to subscribe here:

Finally, if you’re a founder or investor with a company you think we should cover please reach out to us at ericzhou27@gmail.com and uhanif@stanford.edu - we’d love to hear about it :)

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information from partnered venture firms. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should Join The Browser Company

(Click the link to read online).

If you were to stack rank the most important pieces of software from 20th century, an operating system would probably be near the top.

For developers, early platforms like MS-DOS abstracted away large amounts of hardware, preventing them from worrying about things like disk storage, network interfaces, or process management. For users, they offered a more convenient and efficient way to run and manage all their programs (i.e. Microsoft Word, Excel, and Powerpoint).

With time, operating systems centralized the tools and workflows of office workers.

You can see their impact every day on your desktop.

If you were to stack rank the most important pieces of software from 21st century, a web browser would probably be near the top.

For developers, early browsers like Internet Explorer offered an easy way to distribute software globally and instantly to millions of people. For users, they offered a more convenient and efficient way to run and manage all their web apps (i.e. Google Docs, Sheets, and Slides).

With time, web browsers centralized the tools and workflows of office workers.

You can see their impact every day on your desktop.

You might not have noticed, but browsers today handle a lot of what operating systems used to do. Ever since our workflows began moving from the desktop to the cloud, we’ve been relying more on our browsers to manage web apps and less on our operating systems to manage their desktop counterparts. For users and developers, this meant tabs were replacing windows, network requests were replacing downloads/installations, and Javascript threads were replacing processes.

In other words, browsers have surpassed operating systems as the dominant application environment: even before the pandemic and remote work, 37% of information workers were spending 60% of their time completing tasks on a web browser. Remote/hybrid work accelerated that shift: if you’re a student, engineer, or creative, you know what it’s like to spend all day switching between Drive, Notion, Slack, Jira, Figma, Github, Spotify, and Youtube.

By consolidating all the separate apps we used to use under one, browsers have become this natural successor to operating systems. As a result, they’re easily the most widely and deeply used piece of software today.

But was the browser designed for this massive responsibility?

If you’re one of 45% of people juggling at least 20 tabs (or 21% juggling at least 30), you probably have your reservations.

Like early operating systems, the browser as it exists today wasn’t meant to be more than a basic portal for interfacing with apps. Windows 1.0 had no desktop, no taskbar, and no user accounts — it was one screen letting you open and use apps by clicking on them. This worked well when you were only using your computer for VisiCalc once or twice a day, but it broke down when people began using their computers for more time and more tasks.

If you’re managing 10-20 different programs and windows, an interface like this holds you back:

That’s why as more people became power users of their computers, later operating systems introduced things like desktops, taskbars, profiles, searchbars, recycling bins, and wallpapers to help users organize, optimize, and personalize their machines. As computers became more central to peoples lives and the workflows they handled more complex, these abstractions and personalizations were necessary to help manage that complexity.

Browsers have become similarly central, the workflows they handle similarly complex.

But unlike modern operating systems, they don’t do anything across applications to help organize and manage that complexity. Like Windows 1.0, your browser remains essentially a giant list of all your tabs.

Browsers have become the natural successor to operating systems, adopting a similarly central place in the lives of both developers and users.

But unlike operating systems, they haven’t grown beyond a thin shell around the applications they host — a problem because users need something to help manage the internet’s ballooning complexity.

There’s a sizeable opportunity in modernizing the most used piece of software in the world.

The Browser Company has made the most promising attempt so far.

Background

As you might guess, The Browser Company is building a next-generation web browser.

Called Arc, the product was designed to feel almost more like an operating system than something like Google Chrome. For instance, a “favorites” section for your tab bar serves the same purpose as pinned apps on your Macbook’s dock. Spaces and profiles function similarly to spaces and profiles on your computer, making it easy to group tabs and accounts specific to different parts of your life.

Arc has been extremely popular since release: the product saw ~10% week over week growth across the first 9 months, a further 2x growth in Q2 2023, and now serves “hundreds of thousands” of users five days a week.

If you hang around tech or design circles, you’ve almost certainly heard about the product. Arc’s userbase is considerably more passionate about their browser than the average Chrome user:

But why?

For anyone who’s never used Arc, it’s not obvious what the product’s killer features are, which exact chord they’ve struck with so many different people.

To us, the best way to understand The Browser Company’s product philosophy and Arc’s appeal would be the parallel we’ve drawn so far to operating systems:

• Web browsers today (i.e. Chrome) are like early operating systems (i.e. Windows 1.0). Both became dominant application environments, which naturally led to significant workflow complexity. Both were also too simple to help users manage that complexity.

• Web browsers of the future (i.e. Arc) are like modern operating systems (i.e. macOS Ventura). Both were designed for heavy usage and offer a suite of abstractions, personalizations, and optimizations across apps to help manage that complexity.

These parallels exist more concretely across Arc’s feature set and, as Arc’s users have discovered, are invaluable when you spend lots of time in the browser. For instance:

• macOS offers a dock, a desktop, and separate spaces to help group and organize your desktop applications (windows). Arc offers a “favorites” section, a “pinned” section, and separate spaces to help group and organize your web applications (tabs).

• macOS offers Spotlight Search to help you to search across your apps, documents, emails, and everything else. Arc offers a Command Bar to help you search across tabs, actions, history, and everything else.

• macOS offers different user profiles on the same computer to help separate different accounts/contexts (i.e. work vs personal). Arc also offers profiles to help you separate different browser accounts/contexts (i.e. work vs personal).

• macOS offers built-in utilities like Notes, Stickies, and Messages so you don’t have to find and download as many separate apps. Arc offers built-in utilities like Easels and Notes so you don’t need to leave the browser as often.

• macOS offers stylistic personalizations around your background, screensaver, and accent colors. Arc lets you set custom icons for tabs, background colors for spaces, and HTML/CSS/JS for entire websites using a more advanced version of Chrome extensions called Boosts.

On top of these principles drawn from traditional operating systems, Arc has built out a number of optimizations unique to their web-native environment. For instance, a built-in media player lets you adjust music no matter which tab you’re on. Picture in picture mode keeps calls/videos visible even if you’re on a different tab. Site search lets you search within specific sites (i.e. Youtube) from the Command Bar, and built-in support for split screen lets you tile different windows right next to each other.

So why has Arc been so popular?

It’s because for their users, Arc is the operating system for the internet they’ve always needed, offering features to help manage/organize/personalize their web apps like their operating system used to do for their desktop apps. In doing so, Arc has filled a lot of gaps/pain points people used to feel when dealing with web-first workflows.

To put it another way, Arc is to Chrome what Windows 10 was to Windows 1.0.

Although the company isn’t charging for anything yet (we’ll dive into potential business models later), you can imagine there’s probably some way to make money if they can maintain their trajectory and re-invent the most popular piece of software in the world.

The Browser Company was founded by Josh Miller and Hursh Agrawal in 2019. The two had previously worked together to found Branch, a link-sharing app acquired by Facebook. After working at Facebook together for a few years, Josh went on to be a Director of Product at the White House and an investor at Thrive, and Hursh founded a software development consultancy called Enso Code.

The company’s ambitious vision helped them to raise a $5 million Seed Round in 2020 from Thrive Capital and individual investors like Jeff Weiner (LinkedIn), Ev Williams (Medium), Dylan Field (Figma), Jason Warner (GitHub), and Akshay Kothari (Notion). In 2021, they followed up by raising a $13 million dollar Series A from operator-investors like Patrick Collison (Stripe), Eric Yuan (Zoom), Bret Taylor (Salesforce), Mike Krieger (Instagram), Emilie Choi (Coinbase), and Reid Hoffman (LinkedIn).

The Browser Company has stellar traction, clouted investors, and (as we’ll see in a bit) one of the most stacked teams ever assembled in startups. They’ve got a vision to transform how we interface with the internet and a mission to build the operating system of the future.

But can they build a sizable business off of it all? Especially when no one’s paying for their web browsers today?

We think they have a shot.

Let’s have a deeper look at why.

The Opportunity

In 2022, there were 5.3 billion internet users worldwide, representing 66% of the global population. 3.3 billion of them accessed the internet via Google Chrome, and over a billion of them accessed it from Safari.

Browsers are the most popular piece of software in the world.

Unfortunately for browser startups, almost no one’s paying for their browsers today. Every notable web browser is free: Chrome was mostly a strategic push to help Google maintain dominance over search, and Safari was just a way for Apple to keep up with Microsoft.

So for the Browser Company to become a successful business, they need to achieve two non-trivial things:

• Get lots of people to switch to Arc.

• Get lots of people to pay for Arc.

In other words, they’ve got to show Arc can create value (people will switch) and that they as a business can capture value (someone will pay). If they can pull this off, the opportunity is almost unbounded: browsers are the most popular piece of software in the world, so there’s almost no limit on their total user base.

Let’s start with the first task.

Creating Value

Like any product, Arc can get lots of users to switch if they create value for lots of users.

What kinds of people get value out of Arc?

Power users.

Arc’s various optimizations and organizational features are most helpful if you’re juggling Slack, Linear, Rippling, Github, Outlook, and Drive like a modern office worker — they’re not nearly as helpful if all you’re doing is using Facebook a few times a day. Its the same reason your operating system wouldn’t be useful if you were only using Microsoft Excel — you might as well just leave that open 24/7.

Breaking things down further, we think Arc will create value for lots of people because:

• More people are becoming power users.

• Existing browsers aren’t equipped to handle power users.

Powerful Users

More people are becoming browser power users not because they love Chrome, but because they have to: every app is shifting to the browser.

You know Google Docs, Salesforce, and Asana are all web apps, but did you know the real-time transit maps on BART trains were also? Or that the crew interfaces on the Dragon 2 spacecraft were written with HTML/CSS/Javascript?

If you’re building something with a 2D user interface, the odds you’ll do it with something besides browser technologies are almost none. This is for four reasons:

• Standardized UX. The HTML DOM and Javascript have reached critical mass in developer mindshare/experience, making web technologies a de-facto standard for anything with a user interface. It’s gotten to the point where even mobile (React Native) and desktop (Electron) apps are being built with web technologies. If you’re designing, it’s easiest to leverage the browser.

• Expansive Ecosystem. A bottomless stack of frameworks, modules, and APIs means it’s faster to build a web app than any other at this point. A deep pool of web developers means it’s easier to build a team also. If you’re building, it’s easiest to leverage the browser.

• Standardized Distribution. Everything can run browsers now, from laptops to cell phones to Samsung refrigerators. With browsers, you can write your software once and deploy it seamlessly to billions of people, all of whom can access it without a download. If you’re distributing, it’s easiest to leverage the browser.

• Structural Advantages. Making your app cloud-native enables real-time collaboration, infinite storage, and continuous updates, all table stakes for modern software. These unique features mean more apps are being built or refreshed with the browser in mind. If you’re scaling, it’s easiest to leverage the browser.

At this point, only the most compute-intensive of applications (i.e. video games, video editing) are still run locally - everything else lives in the browser. But even these apps are slowly shifting also - constantly improving bandwidth and new web standards (HTML5, WebGL, WASM) are making web apps more powerful all the time. Figma’s collaborative editor, for instance, was uniquely enabled by a custom rendering engine written in WebGL.

As a result, more people are using more webapps all the time, and the number of browser power users is increasing.

Weak Browsers

Unfortunately, this poses a problem: as we’ve explored, most browsers aren’t great at handling lots of apps and lots of tabs.

In particular, the lack of organization holds them back: having tons of indistinguishable tabs open at once can be stressful and inefficient for users. That’s why operating systems come with spaces, docks, and desktops — by helping you prioritize, organize, and search through apps, you only see/use what you need to. This reduces mental load and increases efficiency, which is huge for anyone who uses their browser a lot.

As a result, it’s easy to understand why power users are switching to browsers like Arc. For power users, Arc offers three value propositions to make the experience working across apps easier and more pleasant:

• Organization - features like favorites, spaces, pinned tabs, and profiles help you organize/prioritize your tabs, reducing mental load.

• Optimization - features like split screen, in-website search, a built-in media player, picture-in-picture mode, and the command bar help you work more efficiently.

• Personalization - features like custom backgrounds, icons, and boosts (website reskins) help make your browser feel more “yours”.

Unlike traditional browsers (designed when most people were casual users of the internet), Arc was built for power users of the modern internet. It’s easy to see why so many of them have been switching over.

Capturing Value

Arc is a popular product, but it’s also a free product.

This is potentially a problem because The Browser Company is a company (😲), and companies have to make money.

So what will they do?

Well, Josh himself has acknowledged how they don’t have a set plan here yet, which we’ll also acknowledge at this point is the biggest (highkey existential) risk to the business.

Even so, we have reason to believe they’ll figure something out.

If you believe the prior analysis, that Arc will continue growing and that more and more power users will switch over, The Browser Company will be in such a uniquely leveraged position that it’ll be tough for them not to make money.

Compared to literally any other software tool, a successful Arc will enjoy:

• Unparalleled scale. Browsers are the most popular piece of software in the world, so even capturing 1% market share gives Arc a larger user base than almost any other software tool.

• Unparalleled leverage. Browsers are the most foundational piece of software, impacting how every piece of software running on top is discovered, used, and perceived.

• Unparalleled engagement. Browsers are the most used app in the workspace, and Arc’s users are more engaged than any other browser.

Startups are a game of what can go right, and we think that if things go right for The Browser Company, they can go very right.

Even so, we realize this is all kind of a cop-out answer. Before continuing, we’ll leave you with three more concrete ways they might be able to monetize.

Arc as an Enterprise Tool

Josh has said he thinks “Arc for Teams” is one way The Browser Company might monetize. Here are three potential wedges a “browser for enterprises” might lean into:

• Security

  The browser is the tool most used by enterprises, but it wasn’t built with security in mind. Arc might add security-focused features like:

  • Granular policies on networks requests/locations/assets.

  • Visibility over browser behavior (logs, incident traces).

  • Control over browser behavior (copy/paste/download/upload).

  Other features might include built-in web isolation, web filtering, or zero-trust access. This would be useful in preventing cyberattacks for all employees, but especially for contractors or other workers who use their own devices.

• Artificial Intelligence

  The browser is an ideal insertion point for AI assistants since it has access to all the stuff you encounter online. Arc might add AI-focused features like:

  • An infinite memory assistant like Rewind.ai.

  • A system for running AI agents/action transformers natively in the browser. In contrast to the many AI-helpers being built into specific apps, a browser-based agent would be uniquely positioned to handle tasks across apps. This is way more helpful for approximating real, human workflows.

• Collaboration

  The browser is the most used tool in the workplace, but it remains relatively siloed. Arc might drive productivity by adding collaboration-focused features like:

  • Team-specific boosts and extensions.

  • Browser-native integrations with popular SaaS tools.

  • Shared easels, notes, and comments across websites.

Arc as an Individual Tool

While the core browser will always remain free, The Browser Company might offer a premium version of Arc with extra features for interested individuals. These might include:

• Special boosts (website reskins/remixes), perhaps available on a marketplace for boosts/extensions.

• Special features (i.e. an AI assistant, stronger ad/tracker blocking) that improve the core experience.

Arc as a Distribution Channel

Although less desirable, Arc can always be monetized as a distribution channel for other companies. For instance, Google is known for paying Apple 15 billion a year to remain the default search engine on Safari. The Browser Company could go down a similar path, perhaps being paid to integrate/recommend popular web apps to their users.

Competitive Landscape

The market for web browsers is highly competitive and includes some of the biggest players in the world. Now that we’ve dug into the specifics of Arc’s product and philosophy, let’s look at how they compare against their potential competitors. Like many markets, this space is developing rapidly, so please use this as a starting point for your own analysis.

Traditional Browsers

Chrome, Edge, and Safari are the three most popular browsers in the world, each backed by a major tech company (Google, Microsoft, and Apple). Because of this alone, we can’t really write them off - if any decide to put serious effort into building something resembling Arc, they’ll be tough to compete with.

But of course, the real question is will they do it?

We think it’s possible but less likely for a few reasons:

• Browsers aren’t a core part of any of their businesses. Google’s focus is on ads, Microsoft’s is on their enterprise suite, and Apple’s is on their hardware. Beyond a lack of prioritization/focus, this can have a material impact on their products: Google, for instance, won’t do anything that harms their ads business (which is why Chrome is getting rid of ad blockers). The Browser Company has no such restrictions.

• Arc represents a dramatic departure in design/philosophy from traditional browsers, who mostly render stuff for you and get out of the way. Popular browsers wouldn’t be able to make a leap in design as drastic as Arc without potentially confusing a large portion of their users. Even if the big players wanted to clone Arc, this would put them in an innovator’s dilemma-style bind.

• Least satisfying but still important, building something as nuanced and avant-garde as Arc is just difficult at large companies. This isn’t because they don’t have great people (they do!) but because it’s hard to build conviction on something this new and intricate when you have lots of people.

If we had to throw a guess, Microsoft would be the most dangerous of the three. Besides having a history of ripping popular SaaS products (i.e. Teams vs Slack, Loop vs Notion), they’ve already demonstrated interest in improving browsers by introducing AI to Edge. Chrome would likely be the second, since they stand to lose the most from an insurgent web browser.

Even then, however, they might find it hard to steal Arc’s users. Switching costs with Arc are far higher than with most browsers: Arc users will have customized their browser with favorites, pinned tabs, spaces, custom backgrounds, and various boosts.

Startups

Startups building a next generation browser have generally taken one of two angles.

Security-focused browsers have traded Arc’s broad ambitions for a laser-like focus on improving cybersecurity. They don’t do much UI/UX-wise, but they do have a clearer value proposition and business plan. Some players here include:

• Island - a security-focused, Chromium-based web browser for the enterprise. It comes with built-in security features like web filtering, web isolation, smart network routing, zero-trust access, and safe browsing. They came out of stealth last year with a $115 million Series B and a $1.3 billion valuation. If you’re interested in later-stage companies, this would be a promising one to join.

  

• Talon - another security-focused, Chromium-based web browser for the enterprise. It comes with built-in security features (i.e secure authentication, data loss prevention, zero trust controls) and offers security teams deep visibility into browser and application activity.

There are companies in this camp that are strong and growing quickly, which is worth recognizing. While they don’t compete directly with the more UX/productivity-focused Arc (who’s trying to disrupt the idea of a browser itself), you can imagine them potentially going after the same enterprise users and budget someday.

Productivity-focused browsers are more similar to Arc in how they’re also trying to improve the fundamental browsing experience. Some players include:

• SigmaOS - a WebKit-based, productivity-focused browser. Like Arc, it also allows you to rename, save, and group tabs in a more convenient way. They raised a $4 million seed round led by LocalGlobe in 2022.

  

• Sidekick - a productivity-focused browser focused on enterprise workflows. It offers a dock-like sidebar, “AI based tab suspension”, and universal search across all your open tabs. They raised a $2 million seed round led by Kleiner Perkins in 2021.

  

We haven’t seen any products in this camp with the same traction, brand, community, or engagement as Arc. On most of the important dimensions, it looks like Arc is the leader in this group.

Execution

As discussed above, successfully re-imagining the browser effectively means competing with Google/Microsoft/Apple — and winning.

That takes a strong team.

Thankfully, The Browser Company has one, especially on design and especially at the senior level. If you’ve ever wanted to work directly with a former VP at Google or SVP at Slack who chose to join as ICs, now’s your (probably only) chance. Consider a sampling of their team outside the founders:

• Peter Vidani - Product Designer

  Peter was previously SVP of Design at Slack, SVP of Product at Salesforce, Head of Product at Quip (acquired by Salesforce for $750 million), and Head of Design at Tumblr.

• Darin Fisher - Software Engineer

  Darin was previously VP of Engineering at Google, where he helped build the Chrome browser and Chrome OS. He was also a Software Engineer at Neeva, IBM, and AOL (Netscape). He holds a BS and MS in Mechanical Engineering from UC Berkeley.

• Ben Cunningham - Design Engineer

  Ben was previously Head of UI Design at Tesla. Before that, he was a Senior Staff Software Engineer at Tesla and a Tech Lead at Facebook. He holds a BS in CS from Stanford.

• Tara Feener - Product Engineer

  Tara was previously VP Product and Senior Director of Engineering at Vimeo. Before that, she was Head of Engineering at WeTransfer.

• Jane Kelly - Data

  Jane was previously Head of Growth at Retool, where she was one of the first 5 employees. Before that, she was Director of Data Products at Buzzfeed. She holds a BA in Geography, Art, and Art History from Colgate.

• Dustin Senos - Head of Design

  Dustin was previously Head of Design at Medium and Founder of Small Animal Studios, a digital product studio. He was also a Software Engineer at Disney, where he worked on Club Penguin.

• Sébastien Marchand - Software Engineer

  Sébastien was previously a Staff Software Engineer at Google, where he was for 11 years.

• Dolapo Falola - Head of Engineering

  Dolapo was previously a Director of Engineering at Slack, a Senior Engineering Manager at Instagram, and a Senior Engineering Manager at Facebook. He holds a BS in Computer Engineering from UT Austin.

• Rebecca Grunberg - Data Scientist

  Rebecca was previously a Data Science Lead at Quora and a Software Engineer at Stripe. She holds a Ph.D. and an MS from MIT in Management Research.

• Dara Oke - Product Designer

  Dara was previously a Product Design Lead at Netflix and Program Manager at Microsoft. She holds a BS in CS from UT Austin.

• Brian Michel - Engineering Manager

  Brian was previously a Senior Engineering Manager at Cash App, Twitter, and Tumblr. He holds a BS in Information Systems from Drexel.

• Renato Mangini Dias - Software Engineer

  Renato was previously a Senior Staff Software Engineer at Google, where he worked on Fuschia, Android, and Chrome.

• Saleem Abdulrasool - Software Engineer

  Saleem was previously a Senior Software Engineer at Google and Facebook. Before that, he was a Senior Member of Technical Staff at VMWare. He holds a BS in CS from UIUC.

• Hiroshi Yamauchi - Software Engineer

  Hiroshi was previously a Staff Software Engineer at Google, where he was for 16 years. He holds an MS in CS from Purdue.

• Matias Seijas - Software Engineer

  Matias was previously a Software Engineer Lead at Block (Square) and a Staff Engineer at Betterment. He holds an MS in Interaction Design from NYU.

• Vivek Galatage - Software Engineer

  Vivek was previously a Senior Principal Architect at Visteon, a Senior Chief Engineer at Samsung, and a Senior Software Engineer at Nokia.

• Andrew Monshizadeh - Software Engineer

  Andrew was previously a Senior Software Engineer at Slack, a Software Engineer at Facebook, and a Software Engineer at Instagram.

Having used Arc ourselves for a year this point, we can attest to how quickly this team ships. It’s almost annoying how often you have to update the product for new features.

Conclusion

Browsers are the most important piece of software in the world today.

They’ve surpassed operating systems as the dominant application environment, centralizing and amplifying the world’s collective software output. The internet as we know it wouldn’t exist without them.

For The Browser Company, meaningfully improving the browser is just the beginning. The broader vision is to build the first of what they call “internet computers” - imagine a computer living in the cloud, naturally syncing activity across your phone, computer, AR headset, refrigerator, and Tesla.

The future of the internet awaits.

They’re hiring.

In case you missed our previous releases, check them out here:

• Warp - Revolutionizing the most foundational developer tool.

• Pinecone - Vector databases index the AI revolution.

• WorkOS - The API to increase your TAM.

And to make sure you don’t miss any future ones, be sure to subscribe here:

Finally, if you’re a founder or investor with a company you think we should cover please reach out to us at ericzhou@stanford.edu and uhanif@stanford.edu - we’d love to hear about it :)

Welcome to “Why You Should Join,” a monthly newsletter highlighting early-stage startups on track to becoming generational companies. On the first Monday of each month, we cut through the noise by recommending one startup based on thorough research and inside information from partnered venture firms. We go deeper than any other source to help ambitious new grads, FAANG veterans, and experienced operators find the right company to join. Sound interesting? Join the family and subscribe here:

Why You Should (Still) Join Warp

(Click the link to read online).

About a year ago, we launched this newsletter by picking Warp as our first company to feature.

Last month, they announced their 50 million dollar Series B led by Sequoia Capital. Besides Sequoia, other new investors included Sam Altman (OpenAI) and Tobi Lutke (Shopify). Prior investors Dylan Field, GV, Neo, Contrary, BoxGroup, Jeff Weiner, and Marc Benioff also participated in the round.

We wanted to take this chance to reflect on the original piece while exploring how Warp has progressed in the year since. How are they doing? Where are they now?

Let’s get into why you should (still) join Warp.

Background

The terminal is the most foundational developer tool.

Without it, developers wouldn’t be able to run, test, or deploy their software. It’s the interface allowing programmers to talk to their computers.

Every startup ever launched, every program ever written, every budget 80s hacker scene — they all started with a command from the terminal.

But if the terminal looks like a Cold War-era relic, it’s because it is. While other developer tools have become more collaborative and user-friendly, the terminal remains a siloed eyesore:

The terminal hasn’t changed since 1978, when the VT100 was introduced.

Modern software teams are far more collaborative and agile than teams 40 years ago. Every other tool they use reflects that. Modern project management (Jira, Asana), development (Github, VSCode), deployment (AWS, CircleCI), and monitoring (Datadog, Sentry) tools were all designed with collaboration and speed in mind.

When we found Warp last year, they were trying to change that.

They were building a modern, cloud-native terminal for teams. Out of the box, Warp included features like

• AI-powered command search and autocomplete.

• Blocks that naturally grouped commands and their outputs.

• Modern input editing (i.e. you can click to position your cursor).

It was a terminal for the 21st century — as fast and capable as the original, but re-imagined for modern users.

Most of Warp’s private Beta users were using it five days a week. Mike Krieger (Co-Founder of Instagram) was using it daily while building Artifact, his new startup. As devs ourselves, we used Warp every day and found it incredibly polished.

People really liked the product:

But could Warp be a great business?

We thought they had a shot:

The terminal occupies a unique position at the bottom of every developer’s stack. There’s a multi-billion dollar opportunity in making the most important of developer tools as agile and collaborative as modern software development processes demand it to be.

Our chain of reasoning was roughly as follows.

A Large Market

Every developer uses the terminal. By building a better terminal, Warp’s TAM was effectively every developer on earth. We thought this was a large, growing, and important market:

Every company hires software engineers. As of 2021, there were roughly 1.8 million software developers in North America working in industries ranging from technology to healthcare to financial services.

Every company is interested in maximizing the productivity of their software engineers. Software engineers are expensive, both to recruit and retain. Companies will pay top dollar to maximize the productivity of their developers, either through SaaS tools (Jira, CircleCI, IntelliJ) or other means (pallets of RedBull, 8pm dinner service, free bananas at Amazon).

Global spend on software R&D is increasing. “Digital Engineering spend” will constitute over 50% of overall global engineering R&D spend by 2024, growing at a CAGR of 19% to reach over 1 Trillion USD.

A Valuable Product

The terminal hasn’t meaningfully changed in 40 years. By adding modern features like cursor positioning, autocompletion menus, and an easy way to copy outputs, Warp could easily 10x the user experience and capture a sizable part of that market.

At a high level, we thought Warp could leverage the modern UX, cloud connectivity, and data-rich environment unique to their product to build powerful features driving productivity.

For instance, to drive productivity for individual developers:

• AI command search and autocomplete would make it so any individual could be productive on terminal-related tasks, not just CLI gods who know every command and flag by heart.

• A terminal-based app store would allow individuals to leverage the creativity of the Warp community. Similar to Chrome extensions, developers could write GUI-based extensions for Warp around individual tasks like source control, debugging, port scanning, cluster management, or certificate/key management.

• Native integrations developed by the Warp team with common tools like Slack, Github, or Postman would help de-silo Terminal-related work.

And to drive productivity for teams of developers:

• Shared workflows, notebooks, environments, sessions, secrets, credentials and other terminal-native primitives would allow teams to centralize, share, and iterate on knowledge and processes more easily.

• Block sharing, search, and notebooks would allow teams to collaborate asynchronously on tasks (i.e. getting feedback on output, working out kinks during setup, or running experiments and tests with command line tools).

• Multiplayer would allow teams to collaborate synchronously out of the same terminal session, speeding up debugging or incident-related tasks.

• Infinite history would offer teams valuable insights around anything from incident resolution (i.e. play-by-play reconstructions for debriefs after an uptime firefight) to general activities (i.e. which tasks are most commonly run, which bugs/warnings are most frequently encountered).

Crucially, these value propositions were uniquely accessible to Warp and not other devtool companies because:

1. Terminals are that foundational and used by everyone for everything.

2. Warp had done the hard work of building a performant, extensible terminal from scratch.

What did we mean by this?

A company building a simple autocomplete plugin for a terminal, for instance, would fundamentally be limited to changes that sit on top of existing terminal UX rather than things that live natively in the terminal. They wouldn’t be able to offer the same quality of UX (i.e. built-in menus and buttons) or depth of functionality (i.e. shared terminal environments or sessions).

Here, the second point was particularly important — supporting all these features without sacrificing performance was a real feat of engineering. Warp’s first version was actually a simple Electron app, but it was scrapped for performance reasons.

The system was entirely re-written in ✨Rust✨ with a custom-built UX framework interfacing directly with the GPU. It’s very fast as a result:

An Open Landscape

The terminal hadn’t changed in 40 years, and no one else was really trying.

Besides Warp, the only other terminals were open-source emulators like Alacritty, Hyper, Kitty, iTerm2, or Wezterm. While many shipped with a similar set of “offline” improvements (i.e. autocomplete, split panes, hotkey windows, etc.), their open-source nature meant they’d never have the cloud-based value propositions (AI assistants, shared workflows/notebooks/environments) Warp would have.

They [Alacrity, Hyper, iTerm] didn’t have companies behind them paying for the development of such features or the server bills required to support them and won’t be able to access the unique cloud-based value propositions described above as a result. They don’t pose a competitive risk to Warp, like how Pencil Project never posed a threat to Figma.

There were smaller companies (i.e. Fig, Charm) who were building extensions and add-ons for terminals, but no one else doing the hard work of re-imagining the terminal from the ground up.

The landscape was completely open.

A Strong Team

Beyond universally having impressive backgrounds and experiences (rare for an early-stage company), there were two things we really liked about the people at Warp:

• They had an extremely strong engineering team, which gave them the insight and intuition needed to build for developers.

• They had lots of experience applying cloud concepts to legacy tools (i.e. through Google Docs/Sheets), which was relevant to Warp’s vision.

For instance:

• Founder and CEO Zach Lloyd was a Principal Engineer at Google, where he earned 10 patents and was tech lead on Google Docs and overall head of Google Sheets.

• Design lead Shikhiu Ing previously led a team of 30 at Google as the design lead for Google Docs.

• Software engineers David Stern, Chuck Pierce, and Aloke Desai were previously Staff Software Engineer/Tech Lead at Youtube, Staff Software Engineer at LinkedIn, and Senior Software Engineer/Tech Lead at Google, respectively.

• Growth lead Michelle Lim previously worked at Robinhood, Facebook, and Slack as a KPCB Fellow while also co-founding a startup focused on doctor-patient communications.

We were also impressed by the amount of thought Zach put into company design/culture, even at their early stage. Warp maintained a detailed public wiki on company practices and culture, which we encourage checking out if you’re a founder yourself.

In essence, we saw a large and growing market, a foundational but legacy product in that market, and a stacked team building a highly technical, significantly improved version of that product.

With no real competitors.

What wasn’t to like?

For the full analysis, read the original piece here.

Growth

So how Warp has progressed in the year since?

They certainly haven’t stood still. Let’s look at how both the product and the company have grown.

The Product

Warp isn’t publicly sharing numbers, but we’ve had a peek and can confirm that their user base has grown dramatically in the past year. Both engagement and retention metrics are high and growing also.

Much of this has been driven by two major new features Warp has launched:

• Warp AI is an embedded Copilot for the terminal. It allows users to get explanations for errors and outputs, ask for help on complicated workflows and scripts, and easily execute the suggested commands, all without leaving the command line. Shipped 3/16/2023.

  

• Warp Drive is a space in the terminal where teams can securely save, annotate, and share commands as workflows. In the future, Warp Drive will also include shared notebooks, environments, secrets, sessions, and other terminal-native primitives. Shipped 6/21/2023.

  

In addition, Warp has shipped a large number of “quality-of-life” features, improving the base product while adding support for edge cases.

Some notable ones we’ve pulled from their changelog include:

• Support for subshells in zsh, bash, and fish for a better experience with Docker, GCP, Poetry, etc. You can configure which commands you’d like to “Warpify”, which makes all of Warp’s IDE features available in that subshell. Shipped 5/18/2023.

• Command corrections - automatic command suggestions to correct errors in prior console commands. Shipped 11/10/2022.

• Syntax highlighting and error underlining in the input, available out-of-the-box. Shipped 10/13/2022.

• Opt-out of telemetry or use the network log to view all of Warp’s network activity. Shipped 12/6/2022.

• Drag and drop re-ordering of Tabs (12/15/2022), renaming tabs (3/20/2023), color-customized tabs (9/8/2022), dimming inactive terminal panes (1/26/2023).

Again, it’s worth emphasizing how none of Warp’s major features would be possible if the product wasn’t built the way it was.

If Warp wasn’t an entirely new terminal but rather a simple plugin or extension, it wouldn’t have the accounts, cloud connectivity, or modern UX primitives (i.e. buttons, inputs, menus) required to enable something like Warp Drive or Warp AI. And if Warp wasn’t a performant ✨Rust✨ app, it wouldn’t be able to support those features without seriously hindering performance.

All those features (and more in the future) were only possible because Warp solved the technical challenges of building a performant, modern, and extensible terminal experience from scratch.