Why 37 AI-lab alumni founded startups in 2026

The frontier-lab talent drain finally has a number on it: an Amplemarket tally now circulating on LinkedIn and Reddit counts 37 OpenAI and Anthropic alumni who left to found companies this year. The headline number is less interesting than what the founders are building — because the sharpest bets all point the same direction: automating AI research itself, and selling the tools the labs keep to themselves.

The tally

Per the Amplemarket tally now doing the rounds, 37 people left OpenAI or Anthropic in 2026 to start companies. The list spans automated-research labs, personal-AI plays, and agent-infrastructure bets. A few of the names: Jerry Tworek, ex-OpenAI VP of research behind o1, o3, and Codex; Rohan Anil, who left Anthropic after Tworek, in his words, "nerdsniped" him; Behnam Neyshabur, who co-led Anthropic's Discovery team; and Joanne Jang, who led OpenAI's model-behaviour work. The exits span safety, infrastructure, and research — but the through-line is consistent: these people believe the next big thing can be built outside the labs.

The flagship bets: automating the lab

Two companies concentrate the trend, and both are chasing the same prize from slightly different angles.

Core Automation was founded in late March by Tworek, who spent nearly seven years at OpenAI leading reinforcement learning and reasoning work. The company's stated objective is to build "the world's most automated AI lab" — systems that optimize and automate work, starting with research itself. It closed a roughly $100 million round at a $1 billion valuation backed by Nvidia, Spark Capital, and Accel within weeks of launch, and by May was reported to be seeking $300–500 million at a $4 billion valuation. Technically, it is not chasing ever-larger models: the pitch is continual learning — models that keep learning from real-world experience after training — plus architectures designed to outscale transformers, combined training stages, and far less training data than rivals use.

Mirendil was founded by Neyshabur and Harsh Mehta, who left Anthropic in December after barely a year, along with Shayan Salehian and Tara Rezaei. It raised $200 million at a $1 billion valuation — one of the largest seed rounds in AI history — co-led by Andreessen Horowitz and Kleiner Perkins with Nvidia participating, and assembled a founding team of 20 researchers from Anthropic, xAI, DeepMind, and OpenAI. Mirendil's pitch is a platform that does the work of an AI researcher: designing experiments, searching hyperparameters, evaluating models, and running the next round of training — packaged so that outside organizations can point it at their own problems. A university biology lab could use it to build and refine a drug-target model without a dedicated ML engineering team.

The two approaches differ in a revealing way. Core Automation is building the automated lab for itself — the moat is the lab. Mirendil is building the automation as a product — the moat is the platform everyone else can rent. Same observation, opposite business models.

Why it matters: the moat thesis

The pattern across the 37 is not random. The most crowded new category attacks the labs' own moat: startups building "automate the AI lab" systems staffed by the people who used to run those labs. If Tworek and Neyshabur are right, the frontier's internal research tooling — experimental design, evaluation, iterative training — becomes a commodity anyone can buy. That is a direct challenge to the assumption that frontier AI capability is permanently gated by a handful of deep-pocketed labs.

There is a democratization argument at the center of it, and it is not just marketing. Mirendil explicitly frames itself as a counterweight to capability concentration: give more labs, businesses, and scientists the ability to own their AI infrastructure rather than depend on a few giants. Neyshabur has said the goal is "AI-accelerated science," pursued through recursive self-improvement under human supervision.

The skeptical case

Three objections, in order of strength.

First, the funding math is running ahead of the science. Core Automation went from zero to a reported $4 billion valuation in six weeks with no product; Mirendil raised $200 million at a $1 billion valuation with no product. Investors are paying for team and thesis, not demonstrated capability — and "automate AI research" is precisely the kind of claim that sounds inevitable until it meets an actual experiment. The history of "AI will automate AI" bets is littered with demos that worked on the problem they chose and nothing else.

Second, recursive self-improvement is the single most contested idea in the field right now. Anthropic and OpenAI have both publicly called for a global oversight committee to monitor recursively self-improving AI — and the people building it are the labs' own alumni. The regulatory conversation has not caught up with the fact that the capability is being productized outside the labs' control.

Third, the talent these startups are buying is expensive and shallow. A 20-person founding team of elite researchers is a real asset; it is also a rounding error next to the thousands of researchers at the labs they left. The labs are not standing still — they are hiring and retaining aggressively, and the defection wave has already produced countermeasures (OpenAI's retention packages were widely reported in early 2026).

What to watch

Three things. Whether Core Automation's reported $4 billion raise closes — that would make the automated-lab thesis the most valuable unfunded product in AI. Whether Mirendil ships a platform that an outside lab can actually use this year, which would separate the product bet from the research bet. And whether the labs respond by productizing their own internal research tooling, which would turn the alumni startups' market into a competition against their former employers' best engineers. The 37 founders are betting that automation beats scale. The labs are betting they can build it themselves. 2026's second half should tell us who is right.

Is 'automating AI research' a real business — or a funding bubble? Tell us in the comments.