The Frontier Lab Exodus

In June 2024, three people incorporated a Delaware C-Corp called Safe Superintelligence Inc. Their website was a single page of text. They had no product, no model, no roadmap. By the end of the year, investors had valued the company at $32 billion.

SSI is the most extreme example of a pattern that has hardened over the last two years: senior researchers walking out of OpenAI, Anthropic, DeepMind, and Meta AI, and capital chasing them at increasing speed. The running total since mid-2024: fifteen new labs, roughly $50 billion in pre-product valuations, three commercial products.

The talent flow is the headline. The more useful story is that these labs disagree about what to build. I sort the cohort into five groups.

Sutskever

The cleanest break from the incumbents is Ilya Sutskever's. SSI, co-founded with Daniel Gross and Daniel Levy in June 2024, exists because the founders concluded that no commercial lab could do responsible work toward superintelligence. SSI has published no roadmap, no technical paper, no model. The plan is to stay pre-product until something it considers safe is ready to ship.

Investors put $1 billion into the seed at a $5 billion valuation, then added capital at $32 billion later that year. They aren't buying a product. They're buying Sutskever, and the conviction that he is right about what AGI requires.

This is the only lab in the cohort whose valuation defends itself. Nobody else has a comparable reputation. SSI's pitch is also the most honest one in the cohort: trust the founders. It pays off if the alignment thesis is right and the resulting product turns out to be commercially viable. The downside is a decade of silence and a $32 billion crater.

The product companies

Thinking Machines Lab is the opposite of SSI. Mira Murati raised her $2 billion seed in mid-2025 to build a product company at frontier scale, with no pretense of stealth. Her team is the largest cluster of OpenAI alumni anywhere outside OpenAI: Barret Zoph, John Schulman, Lilian Weng, Jonathan Lachman, and a long list of senior researchers from the company. The lab shipped its first product, Tinker, an API for fine-tuning open models, in early 2026.

Murati's claim is that the frontier-lab playbook (massive compute, massive team, opaque models, API-only access) is no longer the only way to be at the frontier. Their public research emphasizes customization, interpretability, and open-weight base models. If it works, Thinking Machines becomes the lab that built at frontier scale without becoming OpenAI 2.0.

Eureka Labs and Cognition AI are in the same cluster without the OpenAI lineage. Andrej Karpathy left OpenAI in early 2024 to build an AI-native education company. Cognition (founded by Scott Wu, no frontier-lab background) built Devin, the agentic coding assistant, then acquired Codeium/Windsurf in late 2025. Both are product-first companies that happen to need frontier-lab engineering. Neither looks anything like SSI.

Against the consensus

A third group thinks the transformer-everywhere consensus is wrong. AMI is the most public entrant: Yann LeCun has spent a decade arguing that autoregressive next-token prediction leads nowhere, and that JEPA-style world models are the path forward. AMI raised $1 billion at $3.5 billion pre-money inside four months of LeCun's late-2025 departure from Meta. The lab exists to test the JEPA thesis at scale. It rests on questions the field has argued for years without resolution.

Liquid AI, an MIT CSAIL spinout, takes a different architectural family: Liquid Foundation Models, derived from continuous-time neural networks. Reflection AI, founded by ex-DeepMind researchers Misha Laskin and Ioannis Antonoglou, is open-weight rather than open-architecture. Their pitch: closed labs are pulling away on capability but losing the ecosystem fight, and the third position belongs to whoever can do open-weights at frontier scale.

Three different rejections of the consensus. If any of them lands, the upside is enormous. If none does, the result is three expensive academic projects.

World Labs alone

Fei-Fei Li's World Labs sits in a category by itself. The pitch, "spatial intelligence," holds that 3D space is the next platform for AI, and that mastering generation, reasoning, and interaction in 3D is a frontier-scale problem on its own. World Labs released Marble, a research preview of generated explorable environments, in late 2025. The company has raised $230 million at a $1.3 billion valuation.

The open question: is spatial intelligence a separable platform, or is it one capability of a general-enough model? The World Labs answer is that 3D warrants its own frontier lab. That answer looks more defensible the longer pure language scaling shows diminishing returns. It looks less defensible if Google and OpenAI absorb 3D into their multimodal models, which is the obvious risk.

Autonomous science

A fourth group thinks the highest-impact use of frontier AI is autonomous science. Periodic Labs, founded by ex-DeepMind and Meta researchers in 2024, builds robots that run closed-loop hypothesis testing in real labs. FutureHouse, funded by Eric Schmidt, builds agentic AI for scientific literature and experiment design. Lila Sciences, incubated at Flagship Pioneering, does the same with biology at the center.

This is the quietest group in the cohort and arguably the most ambitious. None of these labs is competing for chatbot mindshare. The thesis is that you can compress the discovery cycle in materials, biology, and chemistry by an order of magnitude, and entire new industries appear behind that compression. The downside is that all three are pre-product and the science is hard.

Goodfire, founded by ex-Anthropic interpretability researchers, sits near this group. Its product, Ember, opens AI internals to scientific inspection, and went commercial in late 2024. Goodfire is the lab in the cohort whose work is most likely to get used inside the labs the founders left.

The shippers

A fifth group has no grand thesis. They want to ship product. Magic has worked on its long-term-memory architecture for years; it raised $400M at $1.5B in early 2025 on the strength of LTM-2, and has yet to convert that into a product that beats Cursor or Cognition's Devin on any benchmark anyone respects. Decart, an Israeli startup, ships real-time generated video (Oasis) and has actual users. Prime Intellect runs decentralized training experiments (INTELLECT-1 was the largest decentralized run to date) and competes on infrastructure novelty.

These are the easiest labs to assess. They will succeed or fail on whether anyone wants what they ship. Magic is the one I watch most closely, because its valuation and its release cadence have decoupled uncomfortably.

Patterns

A few things hold across the cohort.

Geography. Almost every one of these labs is in the Bay Area or New York. AMI is the rare exception, in Paris. The exodus has produced no meaningful European or Asian counterpart, despite years of EU policy aimed at producing one. Mistral, Kyutai, and DeepSeek live outside the cohort; they were founded earlier and on different premises.

Funding velocity. The time between announcement and unicorn status keeps collapsing. Anthropic took two years. SSI did it in under a month. Thinking Machines did it before it had a product name. Reflection raised at $500M within a year of founding. The market is underwriting reputation at a velocity that would have been unthinkable in 2022.

Team composition. None of these are solo-founder companies. Almost every lab in the cohort has a co-founder duo plus 5 to 15 named hires from the founder's previous employer. The pitch to investors is that you are buying a team that already shipped together. That's the most defensible thing a pre-product lab can sell.

Public communication. The cohort splits into talkers (Thinking Machines, Goodfire, World Labs, Reflection AI) and silent labs (SSI, AMI, Periodic Labs, FutureHouse, Lila Sciences). The talkers ship more. The silent labs assume product-market fit before disclosure. Either approach has historically been made to work. But silent-and-pre-product has a high mortality rate, and right now it holds a lot of capital.

The compute tax. None of these labs can compete with OpenAI or Anthropic on raw scale. The cohort's largest training runs are an order of magnitude smaller than the incumbents'. This is part of why so many have chosen architectural or vertical theses. Scale isn't a viable axis for the cohort. Each lab has had to find a different one.

Why now

The exodus isn't random.

The second wave of AI capital arrived. The funds that had missed OpenAI's 2019 round and Anthropic's 2021 round had spent two years hunting for the next generation of frontier-lab founders. When researchers began leaving in mid-2024, the capital was waiting.

The alignment debate inside the big labs went public. Sutskever's role in the November 2023 OpenAI board crisis, Jan Leike's resignation in May 2024, and the dissolution of OpenAI's superalignment team told senior researchers that their arguments were losing the internal politics. Some left. The most prominent of those started labs.

And compute supply outside the hyperscalers got real. Cerebras, Groq, NVIDIA-on-CoreWeave, and a handful of other providers made it possible to train at frontier scale without an Azure or GCP alignment. The economics still favor the hyperscalers. But the floor under "you can train a real model outside Big Tech" has dropped.

Who's actually shipping

The scorecard, as of April 2026:

Of fifteen labs valued near $50 billion, three have shipped commercially: Thinking Machines, Goodfire, Cognition. Six are pre-product. The remaining six are research demos or limited releases. This isn't damning for a cohort this young, with most labs 12 to 24 months old. But it is the ratio I watch.

What to watch

Five concrete signals over the next twelve months.

SSI's first publication. If Sutskever's lab publishes anything technical, it will be the most-read AI paper of the year. The longer it doesn't, the more pressure builds on the valuation.

Thinking Machines' research output. Murati's team has the people to publish at OpenAI's pre-2023 cadence. If they do, it will be the most concrete signal that "open frontier lab" is a viable model. If they don't, the lab risks looking like a high-priced product company with frontier overhead.

Magic's next round. A flat or down round at Magic is a signal about the architecture cluster. The next round will tell you whether investors still believe in long-term-memory architectures or whether they have moved on.

AMI's first model. LeCun has been arguing for JEPA for years. AMI is the empirical test. The first model that comes out of the lab, even a small one, will say whether the alternative-architecture thesis is real or rhetoric.

Acquisitions. The most likely outcome for several labs in the cohort is acquisition within 24 months. Watch Magic, Decart, Liquid AI, and possibly World Labs. The acquisition price will reveal what the market values these companies at, stripped of pre-product valuation theater.

The frontier-lab exodus is the biggest AI talent story since the original founding of OpenAI. The ending isn't yet visible. I expect the verdict around 2027 or 2028. By then I'll know which of these founders were right, and which were just expensive.

Sources used in this piece:

• Crunchbase and PitchBook funding data, cross-referenced with TechCrunch and The Information coverage of each lab's most recent round.
• LMArena and Artificial Analysis benchmark standings (where applicable).
• Public statements from founders on Twitter/X, podcast interviews, and lab blog posts.
• The Nextomoro lab catalog (April 2026 snapshot) for organizational details and team composition.

Last updated: April 28, 2026. Funding figures and valuations are as of the most recent disclosed round; private-company numbers are necessarily approximate. Send corrections.