Five frontier-lab veterans have raised -- or are raising -- more than $3 billion in 2026 betting the next AI leap isn't a bigger LLM

Thore Graepel, a co-creator of AlphaGo who left Alphabet's research lab this past summer, is raising tens of millions of dollars from a small initial group of backers for a new venture called Metis Reasoning, Bloomberg reported Sept. 24. The pitch, per people familiar with the fundraising, is to bring "AlphaGo-style reasoning to frontier AI, so machines can plan and act under uncertainty" -- combining learned judgment with search and planning, aimed at robotics, science and engineering, rather than another round of scaling up a chatbot. Graepel could raise a further tranche of hundreds of millions later at a higher valuation; no product, name, or launch date has been disclosed.

Metis Reasoning is the smallest and newest of a pattern that's now run through most of 2026: a researcher who helped build one of the last decade's defining AI systems leaves a frontier lab, and raises an amount with a comma and a period in it to chase a bet that the next leap in capability doesn't come from a bigger language model. Four comparable bets have already closed or are close to it this year, from three different labs, for a combined total north of $3 billion:

The shared thesis: search and self-play, not more pretraining

What unites five founders who otherwise disagree on almost everything technical is a common target, not a common method: every one of them is betting that whatever comes after today's transformer-based language models will look more like AlphaGo or AlphaZero -- systems that got better by playing against themselves and searching ahead, not by reading more text -- than like GPT-6 with a bigger context window. Silver's Ineffable, valued at $5.1 billion, wants a "superlearner" that discovers skills through trial and error rather than studying human-generated examples. LeCun's AMI Labs is building on JEPA, the world-model architecture he's been proposing since 2022, on the argument that a system that has never observed physical reality can't reason reliably about it -- AMI's own CEO, Alexandre LeBrun, points to hallucination risks that "could have life-threatening repercussions" in a healthcare context as the reason language-only models aren't enough. Socher and Tian's Recursive Superintelligence wants a loop that "autonomously discover[s] knowledge, continuously optimise[s] itself, and evolve[s] in an open-ended loop." And Emulate -- founded in August by the three researchers who built DeepMind's Genie world-model demos, and reportedly in talks at a $3.7 billion valuation -- wants to simulate how objects break and robots move well enough that a robotics company can test in the simulation before it ever touches hardware.

Emulate is not even the only world-model startup racing to close a round this size: techfundingnews.com's own reporting on the deal notes it now competes for talent and capital with World Labs (reportedly $1.23 billion raised, roughly a $5 billion valuation) and Generalist (reportedly valued around $3 billion) -- meaning "world models" alone is already a crowded sub-category with at least three well-capitalized entrants competing for a thesis none of them has yet shipped a public product to prove.

One investor shows up on the cap table of three of the four closed rounds regardless of which thesis it's betting on: Nvidia put money into Ineffable Intelligence, AMI Labs and Recursive Superintelligence alike, alongside its more familiar role selling all three the chips their research will run on. That is not evidence any particular thesis is right -- it's evidence Nvidia is hedging across every credible non-scaling bet in the field at once, which is a rational thing for a chip vendor with no stake in which paradigm wins to do, and a useful tell that no outside investor, including the one with the best view into compute demand across the whole industry, actually knows yet which of these bets pays off.

What's proven, and what's pedigree

Graded the way this desk grades any capability claim -- against what baseline, under what conditions, with whose scoring -- none of the five clears the bar. Every dollar of the roughly $3.5 billion already closed, plus whatever Metis Reasoning and Emulate eventually raise, is priced against a resume rather than a released result: AlphaGo and AlphaZero for Silver and Graepel, Vision Transformer co-authorship and the JEPA papers for LeCun's team, DeepMind's Genie demos for Emulate's founders, a self-play research pedigree spanning DeepMind, Meta FAIR and OpenAI for Recursive Superintelligence's roster. That is a real signal -- these are, credibly, some of the researchers most likely to find whatever comes next -- but it is a different kind of evidence than a benchmark result, and investors are pricing it as if the two were interchangeable. Recursive Superintelligence's roster makes the pedigree-over-product pattern most explicit: a team of fewer than 30 people, drawn from DeepMind, Meta FAIR, OpenAI and Google DeepMind's reinforcement-learning group, with a Turing-Award-adjacent advisor in Peter Norvig, priced at $4.65 billion four months after incorporation and before a single released system.

The case against the pedigree bet

The strongest objection to all five deals isn't specific to any one of them -- it's a pattern this industry has lived through before, and the round sizes themselves make it worth stating plainly rather than treating five separate raises as five separate stories:

“In six months, every company will call itself a world model to raise funding.” -- Alexandre LeBrun, CEO, AMI Labs

That line is worth sitting with, because AMI Labs' own chief executive said it about the exact category his company is raising nine figures inside. It is the rare moment in this wave where an insider states the skeptical case better than an outside critic could -- and it's a reasonable prediction: once a label works as a fundraising word, it stops reliably describing a method. Watching whether the next round of "world model" or "reasoning" startups can actually name a baseline, a benchmark, and a scoring method -- the same three questions this desk asks of every capability claim -- is the only way to tell a real bet on a new paradigm from a pattern-matched pitch deck wearing this year's most fundable words.