A neolab is an AI startup organized around a fundamental research bet rather than a product roadmap. The team is typically a small group of researchers out of a frontier lab or a top university, the pitch is a technical thesis, and the funding round often closes before there is revenue, a shipped product, or in some cases even a model. The label spread through 2026 in press and venture coverage and it is applied loosely: some writers mean only the very large generalist labs chasing superintelligence, others include any research-first AI company, including ones aimed at a single domain like law or biology. The distinction that matters for a founder is not the label but the funding logic. A neolab is underwritten on the belief that a breakthrough will arrive and create a market, which is a different risk than a company underwritten on customers it already has.

How is a neolab different from an AI application startup?

The difference is what the company is betting on and where its learning accumulates. A neolab bets that its research produces a capability nobody else has, and its asset is the model or the method. An application company bets that a workflow is valuable and hard to serve, and its asset is the accumulated knowledge of what breaks in that workflow and how to fix it. That second asset is the one Factory CTO Eno Reyes argues is durable, because he expects the technical advantage itself to be temporary: in his words, "the technology is just no longer the moat". The practical consequence is that the two company types fail differently. A neolab fails when the frontier catches up to its research for free. An application company fails when the workflow it serves turns out not to matter, or turns out to be identical everywhere.

Will most neolabs fail?

Reyes puts it higher than most: asked about predictions that 70 to 80% of neolabs would die in three to five years, he said "it could be 80 to 90% of Neolabs die in the next 18 months". He is careful that dying is the wrong frame, because for many of those teams the outcome will be an acquisition their investors and founders are happy with. His actual claim is narrower and more useful: "these businesses may not make sense as independent businesses". If you are a founder, that reframes the question from whether you will survive to whether you are building something that only makes sense inside somebody else's company. Both can be good outcomes. They are not the same plan, and they justify very different amounts of capital.

Should a founder start a neolab or an application company?

Start with whether you can verify the output of the work you want to automate, because that is what decides how fast AI improves at it. Reyes calls verifiability "the single most important property of success with current AI systems", and NEA's own analysis of the category lands in the same place: in a reinforcement-learning-first paradigm, the question is not only who has the largest cluster but who has the best environments, curricula and verifiers. If the domain you care about already has clean verification, expect the frontier labs and the open models to arrive there quickly and price your advantage accordingly. If it does not, and you can build the verification because you have the data or the expert judgment nobody else has, that is a real position, and it is available to an application company as much as to a lab. Neither answer requires raising at a research-lab valuation.

Neolabs: What They Are and Which Ones Last

A neolab is an AI company funded on a research bet rather than a product. The category has absorbed an extraordinary amount of capital very quickly, mostly ahead of revenue, and the obvious founder question is which of these companies is still standing in two years.

Eno Reyes has a specific and unusually early answer. He is the co-founder and CTO of Factory, which sells an enterprise platform for autonomous software development and raised $200 million at a $5 billion valuation in September 2026. Asked on 20VC about the common prediction that 70 to 80% of today's neolabs die within three to five years, he went further:

"it could be 80 to 90% of Neolabs die in the next 18 months"

What makes the interview worth a founder's time is not the number. It is the test he uses to get there, which applies just as well to a company that is not a neolab at all.

What a neolab actually is

The definition is looser than the confidence with which people use the word. A neolab is a research-first AI startup: a small group of researchers, usually out of a frontier lab or a top university, raising on a technical thesis rather than a product roadmap, often before there is revenue, a product, or in some cases a model. The label spread through 2026 in press and venture coverage. Some people use it only for the very large generalist labs chasing superintelligence. Others include any research-first AI company, including ones pointed at one domain.

The scale is easier to pin down than the definition. A June 2026 analysis by four investors at NEA put AI research labs at roughly $1.68 trillion of aggregate private valuation, and found that 93.6% of that sits in a single bucket: frontier generalists, with OpenAI, Anthropic, xAI and Safe Superintelligence absorbing almost all of it. Everything else an investor might back in the category, biotech, robotics, scientific discovery, voice, edge and more, combines to barely 6%. Their conclusion is that the neolab market is not really a market at all, but one concentrated basket with four names.

That concentration is the thing to hold onto. When someone says neolabs are overfunded, they may be describing four companies. When someone says neolabs are the next wave, they are usually describing the other 6%, which is a different bet with different economics. The survival question below is mostly about the second group.

Why dying is the wrong word for it

Reyes is careful about the framing, and the care is the useful part:

"these businesses may not make sense as independent businesses"

For a lot of these teams the ending will be an acquisition that everyone involved is pleased with. He expects it. Asked in the same interview about Factory's stated plan for future hires to arrive by acquiring companies and bringing their founders and teams in, he describes the profile he looks for: people who quit a job to spend months building one thing, who are "sometimes companies of one", and who are willing to scrap what they built to work on something bigger.

So the prediction is not that 80 to 90% of this capital burns. It is that most of these companies end up as a component of somebody else's company rather than a standalone business. That is a genuinely different thing to plan for, and it is worth being honest with yourself about which one you are building, because the two justify very different amounts of capital and very different promises to the people you hire.

The three questions that decide whether a neolab lasts

Reyes's test is three questions, and none of them is about the research:

"is this business attached to a durable workflow"

Then: does that workflow change if new frontier models get better. Then: does the workflow survive an entirely new way of working, or would a business shown this workflow figure out something better to do instead.

His worked example is law. New models do not automatically get better at legal work without access to the data, the workflow is genuinely proprietary, and there will still be a legal system in ten or twenty years. All three questions come out the right way, so he expects neolabs pointed at legal work to do well. Compare that with the category he thinks is exposed:

"a lot of knowledge work that's related to intermediate tasks like people operating in Excel and Jira that's just not going to be differentiated"

The workflows are common, they are not proprietary to anybody, and there is a real chance we are not using tools of that shape in five to ten years. A company whose whole value is being good at an intermediate step in a workflow that everyone performs the same way has no answer when the model gets good at that step for free.

This is the same argument as what the model can't eat, arrived at from the lab side rather than the application side. The durable thing is the part of the work that is specific to an industry and expensive to learn, not the part that is technically impressive.

Verifiability is the property that decides the winners

The reason some domains move fast and others stall is not mystery, and Reyes is blunt about it:

"verifiability is ultimately the single most important property of success with current AI systems"

That explains the uneven progress founders keep running into. Stebbings puts the unevenness to him directly: coding and customer service are undeniable, legal is good but not at that level, and marketing copy and visuals are, in his words, so not there, with clipping this very podcast as the example. Reyes's explanation is the verifier. Clipping is still hard because the skill lives in the heads of people who, as he puts it, could not even describe how they know when to do the right clip, and nobody has yet done the work of writing it down.

Andrej Karpathy, who co-founded OpenAI and led AI at Tesla, framed the same boundary in his Sequoia interview by analogy to the previous era:

"traditional computers can easily automate what you can specify in code"

His point is that this generation of models automates what you can verify, and that because frontier training runs are enormous reinforcement learning environments rewarded on verification, the resulting models peak in verifiable domains like math and code and stay rough elsewhere. The NEA authors reach it from the investor side: in a reinforcement-learning-first paradigm, the question is not only who has the largest cluster but who has the best environments, curricula and verifiers.

Where Reyes goes further is in treating verification as something you build rather than something you find:

"I actually would argue the frontier right now of AI is AI systems that can build verification where there is none and thus they can progress into tasks that today humans consider to be too difficult for AI to resolve"

The mechanism he describes is the one every company already uses on people. A novice manager judges new hires by gut. A large firm cannot, so it writes down what good looks like and builds a framework to assess it. He extends the analogy with a warning founders should take seriously: writing down what good looks like changes behavior as well as measuring it, so if you specify the wrong three things, you will reliably get those three things. Anyone who has watched a team optimize a badly chosen metric knows the shape of this. It is the same failure mode as a security program that measures patches closed rather than exposure reduced.

For a founder, the practical read is a question to ask before you commit years: in the work I want to automate, does a good verifier exist, and if not, am I the one who can build it. If a clean verifier already exists, expect the frontier labs and the open models to arrive quickly, because that is exactly the surface reinforcement learning eats. If no verifier exists and you have the expert judgment or the proprietary data to create one, you have something that is hard to copy. That is a defensible position whether or not you call yourself a lab. The related problem of knowing when your system is actually working in production is LLM observability.

Why the research itself is not the moat

Reyes's second argument is the one that should unsettle a research-first founder most, and he grounds it in a deal rather than a theory. Stripe agreed to acquire OpenRouter, the model routing marketplace, in a deal announced on August 19, 2026. Neither company disclosed the price; reporting put it at roughly $7 billion to slightly more than $8 billion.

Reyes does not think Stripe bought the routing technology, and he is direct that routing itself is not especially differentiated. His reading is that Stripe bought position: tokens are intelligence, businesses have to allocate intelligence the way they allocate money, and Stripe already sits on the money flow. OpenRouter adds visibility into where the intelligence flow is going. The sentence that does the work:

"the technology is just no longer the moat"

His supporting observation is the one to write down. You would not pay that price for an identical company with better technology and no users. The technology was not the asset. The traffic and the information it generated were.

If you are running a neolab, that is the uncomfortable version of the question. Your pitch is that your research is ahead. Reyes's claim is that being ahead is a temporary state that does not, by itself, convert into a business, and that the durable position is a workflow other people depend on. He says as much when asked, in the quickfire round, whether he would buy or sell Salesforce:

"I believe that the businesses that are going to be most durable are the ones that have a workflow and a system of record that they've defined that everyone agrees is consensus"

Which is nearly the opposite of a research bet. We work through the general version of this in where a competitive moat comes from in AI.

Being model locked is a disadvantage when you sell outcomes

There is a structural trap here for any lab that also wants to sell an application, and it is worth understanding before you design your own company around owning a model.

Reyes's framing starts with pricing. He argues that founders think about price wrong:

"you should not be thinking about the inputs to the price you should be thinking about the outputs"

The unit that matters is the cost of the outcome, not the cost of the tokens. A code review done correctly first time by an expensive model can cost less than the same review done badly and repeatedly by a cheap one, which is why he expects that in many demanding tasks "I see a world where the smartest model is actually the cheapest". That is the same accounting we work through in AI inference cost.

Now apply it to a company that owns the model. If you sell outcomes but can only use your own weights, you cannot give the customer the best available answer, only your best available answer. Reyes calls the problem plainly:

"it's bad incentive alignment"

And the consequence:

"being model locked is actually a huge disadvantage if you're trying to sell outcomes"

He expects the shape of the market to make this worse rather than better, predicting that in three years "99% of workflows are going to be done on open models", with a small remainder of genuinely frontier tasks carrying a disproportionate share of the economic value. Whether or not the number lands, note that this is a company with a commercial interest in model independence making the argument, so weigh it as a position rather than a finding. The useful part is the structural point underneath, which does not depend on the percentage: if most work runs on commodity models, owning one of them is not where the margin sits. What to actually do with open weights is in where an open source LLM belongs.

He also thinks the place where advantage now accumulates is not the model at all:

"the harness is effectively the new sort of application"

His argument is that the problems people keep expecting the model to solve, longer context among them, were solved in the harness instead, with techniques like compaction. The harness holds the state and the logic, so it is where learning collects. That is the same conclusion we reach from a different direction in continual learning AI.

Who is the sovereign of your intelligence

The part of the interview with the longest shelf life is a governance question, and it is the one most founders selling to enterprises will meet within a year:

"Who is the sovereign of your intelligence? Is it you or is it some other company?"

What he means is concrete: who owns the record of which workflows succeeded, and the learnings that came out of them. His example is a law firm that outsources every case to ten different vendors and discovers five years later that those vendors know exactly how to run its business. He notes that two of the largest model providers have, in his account, explicitly said they intend to go after the industries they currently supply intelligence to, which makes the concern less abstract than it sounds.

This is not a fringe view among people selling to large enterprises. Glean founder and CEO Arvind Jain, asked on the same show whether the largest enterprises are more skeptical of frontier model providers than ever, said their leaders worry about their core IP, their data and their way of doing things, and about being subject to "too much of technology dependence on these model providers".

The security read is worth making explicit, because it is usually mistaken for a procurement preference. Reyes says the on-premise demand he sees is mostly not technical:

"on premise isn't even about the technology. It's just about the idea which is that if I need to I can take control and ownership over every dimension of this software and we get to stay in control"

He adds the detail that gives it away: many of the businesses he talks to still choose the hosted product once they know the on-premise option exists and understand how a switch would work, because what they wanted was the exit, not the deployment. That is a real requirement and founders dismiss it at their cost. Being able to answer what happens if we need to leave is a sales asset even for customers who never use it, and it is cheaper to build early than to retrofit under procurement pressure. The engineering version of that decision is in the self-hosted AI agent tradeoff.

If you are choosing between a lab and an application

Put the pieces together and the choice is less binary than it looks. What Reyes is describing is not research versus product. It is whether the thing you accumulate is something the frontier will hand out for free in eighteen months.

Research alone does not qualify, because the frontier moves and open models follow it. A verifier nobody else can build does qualify, because it takes data or expert judgment you have and others do not. A workflow and a system of record that an industry agrees on qualifies, because the agreement is the asset. Model ownership mostly does not, and actively works against you if you are selling outcomes.

None of that requires the label. A domain-focused company that builds the verification for a workflow nobody has automated is doing the defensible version of a neolab's job without raising at a research-lab valuation. That is the version of this bet most founders should be running. Where all of these decisions connect, across product, engineering, agents, go-to-market, pricing and team, is the AI for startups pillar.

What to do this week

Write down the verifier for your core task, in one paragraph. What does a correct output look like, and how would a machine check it without a human. If you can write it easily, assume the frontier gets there too and that your advantage is elsewhere. If you cannot write it, you have found either your moat or your dead end, and knowing which is the next quarter's work.
Answer Reyes's three questions about your own company in writing. Is the business attached to a durable workflow. Does that workflow change if the next frontier model is much better. Does it survive an entirely new way of working. Any answer of no that you cannot defend is your real risk, not competition.
Name the part of your product that a better model deletes. Every AI company has one. Write it down and check how much of your revenue depends on it, then decide deliberately whether you are defending it or replacing it.
Check whether you are model locked, and price it. If your outcome quality is capped by one provider's weights, estimate what the best available model would do for the same task. That gap is what a competitor can sell against you.
Decide who owns the learning loop, and write it into the contract. Who keeps the record of which runs succeeded, you or your vendor, and who keeps it when your customer asks the same question of you. This clause gets negotiated at renewal whether or not you planned for it.
Build the exit answer before a customer demands it. Document what leaving your product looks like: data out, models swapped, deployment moved. You are selling control, and customers buy the option far more often than they exercise it.

Most of these are governance decisions disguised as technical ones, which is exactly the category founders defer until an enterprise deal forces the issue. Making them on purpose, early, is what we build in AI Operating System for Startups.

Sources

Only 10% of Neo-labs Will Survive | Factory CTO, Eno Reyes (20VC with Harry Stebbings, August 2026), the interview this article distills. Reyes on the neolab survival rate and his three durability questions, verifiability as the deciding property, building verification where none exists, the technology no longer being the moat, model lock as an incentive problem, the harness as the new application, sovereign intelligence, and on-premise as optionality.
Andrej Karpathy: From Vibe Coding to Agentic Engineering (Sequoia Capital), for the verifiability boundary: what code could automate versus what verification can, and why reinforcement learning produces models that peak in verifiable domains. Karpathy's background is on Wikipedia.
Why OpenAI and Anthropic Won't Win the App Layer (20VC with Harry Stebbings), for Glean founder and CEO Arvind Jain on enterprise leaders' fear of technology dependence on model providers.
The AI Neolab Wild West (NEA, June 17, 2026), by Thomas Joshi, Madison Faulkner, Lila Tretikov and Andrew Schoen, for the $1.68 trillion aggregate private valuation figure, the 93.6% concentration in frontier generalists, and the reinforcement-learning-first argument that environments, curricula and verifiers now matter as much as cluster size. Radical Ventures is the other venture firm whose writing popularized the term.
Factory's September 2026 round of $200 million at a $5 billion valuation: Factory's own announcement and Reuters, which reports the valuation as more than triple the company's previous level. Background on Eno Reyes, including the machine learning role at Hugging Face and the earlier engineering role at Microsoft: his LinkedIn profile and his AI Engineer speaker bio.
Stripe's agreement to acquire OpenRouter, announced August 19, 2026: Stripe's newsroom post, which does not state a price, plus Reuters and Forbes for the reported range of roughly $7 billion to slightly more than $8 billion. The figure was never officially disclosed.