Thore Graepel, one of the co-creators of AlphaGo, has left Google DeepMind to found an artificial intelligence reasoning startup, according to an exclusive report by Sifted published September 7, 2026. The venture is betting on a thesis that runs against the industry's dominant current: that structured search, rather than ever-larger language models, holds the key to reliable machine reasoning.
A DeepMind Veteran Steps Away
Graepel is best known for his work on AlphaGo, the DeepMind system that defeated Go world champion Lee Sedol in 2016 — a milestone widely credited with igniting the modern AI boom. The move ends a long tenure at Google DeepMind, where he was one of the research organization's most senior and visible figures.
Sifted reports that Graepel quit to pursue the reasoning venture, and trade outlets including Tech Times framed the new company's core bet explicitly: structured search over LLM scaling. Details about funding, staffing, and product plans have not been disclosed, and neither Graepel nor DeepMind has published a formal announcement — but the move lands in a year of aggressive AI startup formation around alternatives to pure scaling.
The Thesis: Search Over Scale
The bet matters because it challenges the strategy that has defined frontier AI since 2020 — train bigger transformer models on more data and compute, then teach them to reason through chain-of-thought prompts and reinforcement learning.
The AlphaGo lineage represents a different tradition. AlphaGo and its successors combined neural networks with tree search and, in later iterations, learned from self-play rather than human games. Systems in this family could verify their own moves against fixed rules, which made their reasoning exact in a way that language-model outputs, generated one plausible token at a time, are not.
Advocates of the search-first approach argue that as LLMs are pushed into domains like mathematics, formal verification, and planning, their failure modes — fluent but wrong answers, unreliable multi-step logic — stem from architecture, not just scale. Search-based systems that can check each step against formal rules offer a different path to dependability, and recent progress on Lean-based theorem proving by both startups and major labs has kept that debate alive.
Part of a Broader Exodus
Graepel is not the first AlphaGo veteran to leave DeepMind in pursuit of this thesis. David Silver, who led the AlphaGo project, left Google DeepMind in late 2025 after more than a decade building AlphaGo, AlphaZero, and AlphaStar, and founded Ineffable Intelligence, which was incorporated in November 2025. As reported this week, Sequoia and Nvidia have backed the startup at a $5.1 billion valuation.
The departures add to a stretch of high-profile turnover at Google DeepMind. Earlier this year, the lab saw leadership changes including the Hassabis chair transition and the exits of senior figures such as Koray Kavukcuoglu and Jeff Dean, coverage that AI Buzz Wire tracked at the time. When the people who built a lab's most famous systems leave to chase a different architecture, the signal is less about internal politics and more about genuine disagreement over where the next breakthrough comes from.
Why It Matters for the Field
For the startup ecosystem, Graepel's move is another marker that the post-transformer argument is attracting serious capital and serious resumes. Investors spent 2024 and 2025 almost entirely on LLM applications and infrastructure; 2026 has seen a steady stream of funding for companies pursuing world models, neuro-symbolic systems, formal reasoning, and search-based planning.
For Google, the loss is symbolic as much as technical. AlphaGo remains DeepMind's most famous creation, and its alumni now headline at least two well-funded ventures pitching alternatives to the scaling playbook. Whether structured search can actually out-reason frontier language models on open-ended problems — rather than narrow, formally verifiable ones — is the question these startups will have to answer with products, not manifestos.
There is also a middle path the industry is already converging on: hybrid systems that use LLMs for intuition and search for verification. If Graepel's venture lands there, it will be competing less against the scaling narrative than alongside it.
Neither Graepel nor Google DeepMind had responded publicly with further details at the time of writing. AI Buzz Wire will update this story as more information about the startup emerges.
Stay Ahead of AI
Keep up with the people and companies shaping the next wave of AI — read more AI news on AI Buzz Wire.
Read more AI news →
