A decade after AlphaGo stunned the world in Seoul, one of the scientists whose name sits on the landmark 2016 Nature paper is betting that the machinery behind that win, not ever-larger language models, is the missing piece of modern AI. Thore Graepel, who left Alphabet's Google DeepMind this summer, is raising tens of millions of dollars for a new startup called Metis Reasoning, Bloomberg reported on Thursday, citing people familiar with the matter.
According to Bloomberg, Metis Reasoning is building AI that can respond to unfamiliar problems and decide what to do, with target applications in robotics, science and engineering. Graepel is seeking the initial capital from a small group of backers, and the company could later raise a further tranche of hundreds of millions of dollars at a higher valuation. The talks are described as preliminary, and the structure and size of the round could still change.
From Go Boards to the Physical World
Graepel's thesis has been hiding in plain sight. His personal website, updated after his departure, describes a new project focused on "bringing AlphaGo-style reasoning to frontier AI, so that machines can plan and act under uncertainty." The same page sums up the AlphaGo recipe as a policy that suggests moves, a value function that weighs them, and a search tree that looks ahead, and argues that the approach carries over naturally to embodied intelligence.
He has been blunter in interviews. "We need to go back to the architecture and fundamentally redesign it so that it does proper reasoning," Graepel told the European tech outlet Sifted when news of his exit broke earlier this month. The remark cuts against the prevailing industry strategy of stretching large language models with chain-of-thought training and ever-bigger compute budgets. Where an LLM predicts the next token, AlphaGo-style systems run explicit lookahead over possible futures before committing to an action, an approach that matters when a robot arm or a lab instrument cannot simply undo a mistake.
His resume gives investors plenty to underwrite. A physicist by training, Graepel spent 2003 to 2015 at Microsoft Research, where he co-created TrueSkill, the matchmaking system behind Xbox Live, and AdPredictor, the click-through model behind Bing's ads. He joined DeepMind in 2015 and worked on AlphaGo, which beat Lee Sedol 4-1 in March 2016, and on its successors AlphaGo Zero, AlphaZero and MuZero. From 2021 to 2025 he led machine learning at the longevity biotech Altos Labs before returning to Google DeepMind's post-AGI team. He also holds the Chair of Machine Learning at University College London.
A Crowded Field of DeepMind Neolabs
Metis Reasoning joins a fast-growing roster of so-called neolabs founded by DeepMind veterans. The most prominent is Ineffable Intelligence, launched by AlphaGo lead architect David Silver after he left DeepMind in January. Silver's company raised $1.1 billion at a $5.1 billion valuation in April in a round co-led by Sequoia and Lightspeed, with Nvidia and Google participating, according to TechCrunch and CNBC. Emulate, founded in August by three DeepMind veterans, is in talks to raise up to $700 million at a $3.7 billion valuation, PYMNTS reported. Outside the DeepMind orbit, Yann LeCun's AMI Labs raised $1.03 billion in March to build world models.
The exits have piled up quickly. Tech Times, citing Zeki Data, reported that DeepMind's share of AI research hiring in Europe, the Middle East and Africa fell from 49% in 2022-23 to 18.6% in 2025-26. Demis Hassabis has pushed back on the narrative. "We have by far the biggest and broadest research bench of any of the labs out there. We win our fair share of the top talent," he said at Cannes Lions in June, as reported by Semafor.
The financing structure is itself a sign of the times. Bloomberg notes that tranched rounds, in which capital arrives in stages at rising prices, are spreading among AI startups. The structure lets early backers buy in cheaply while allowing founders to signal a lofty valuation that helps with recruiting.
Why It Matters
Metis Reasoning is a clean test of one of the biggest open questions in AI: whether the next leap comes from scaling language models or from reviving explicit planning and search. Silver, Graepel and LeCun are all, in different ways, arguing that models which learn from their own actions and reason about consequences will outrun systems trained mainly on human text. Investors are now pricing that argument in the billions.
For Google, the departure stings twice. AlphaGo is arguably DeepMind's defining achievement, and two of its core architects are now building potential competitors. The company can still participate financially, as it did in Ineffable's round, but the research agenda increasingly lives outside its walls. For robotics and scientific-discovery startups, meanwhile, a well-funded lab focused on decision-making under uncertainty could become either a key supplier of foundation models or a formidable rival.
The technical risk is real. AlphaGo's search worked because Go has exact rules and cheap simulations. The physical world offers neither, and building a learned world model accurate enough to support reliable lookahead remains an unsolved research problem.
What to Watch
Look for the names of Metis Reasoning's first backers and whether Google, Nvidia or the Sequoia-Lightspeed pairing that funded Ineffable shows up again. The size and valuation of the promised larger tranche will show how much of a premium the market places on the AlphaGo pedigree. Graepel's scheduled November talk at the London School of Economics, billed as From AlphaGo to AGI, may offer the first detailed public look at the company's technical roadmap, and at whether it plans to publish research, ship models, or partner directly with robotics and lab-automation companies.
“We need to go back to the architecture and fundamentally redesign it so that it does proper reasoning.”— Thore Graepel, Founder, Metis Reasoning