The Number Behind the Exodus
Google and DeepMind have produced more than 40 founders of what the industry calls neolabs, independent AI research companies started by alumni of a bigger lab, according to a June 2026 analysis by venture firm Radical Ventures. That is more than the combined founder output of OpenAI, Anthropic and Meta, reported by Bloomberg on September 25. Earlier that same month, about 15 current and former DeepMind researchers met for breakfast at a five-star hotel across the street from Google's UK headquarters in London to compare notes on which investors were writing checks.
Four Spinouts, One Missing Ingredient
David Silver's Ineffable Intelligence, focused on AI that learns without human-generated data, raised roughly 1 billion dollars earlier in 2026. Jack Parker-Holder's Emulate, which simulates and predicts real-world systems, is in talks for a 700 million dollar round. Nando de Freitas, a decade-long DeepMind veteran, is seeking at least 100 million dollars for Revolution Labs, built around diffusion models as an alternative to transformers. Thore Graepel, an AlphaGo co-creator who left again this summer, is raising tens of millions for Metis Reasoning, a bet on structured search and probabilistic planning over LLM scaling.
| Startup | Founder | Approach | Funding status |
|---|---|---|---|
| Ineffable Intelligence | David Silver | Learning without human data | Raised about 1 billion dollars |
| Emulate | Jack Parker-Holder | Simulation of real-world systems | In talks for 700 million dollars |
| Revolution Labs | Nando de Freitas | Diffusion models, not transformers | Seeking at least 100 million dollars |
| Metis Reasoning | Thore Graepel | Structured search and planning | Seeking tens of millions |
Every one of the four is explicitly built around an approach that is not a large language model. That is not a coincidence, according to Janusz Marecki, a former DeepMind lead scientist now an AI partner at Ahren Innovation Capital. He told Bloomberg there is top-down pressure inside Google to drop other research and focus on large language models, and that researchers whose work does not fit that mandate see no point in staying.
Google Is Funding Its Own Departures
This is not a story of a rival poaching Google's talent. Alphabet was a founding investor in Jeff Dean's Discovery Loop, the spinout announced August 5 alongside Demis Hassabis's move from DeepMind chief executive to Alphabet chief scientist, and it supplies that company with Google Cloud credits. Discovery Loop's own valuation talks jumped from roughly 10 billion to about 50 billion dollars within weeks, before it had shipped a product. Google, in effect, is both the source of the researchers leaving and a shareholder in what they build next.
Marecki also pointed to the price investors are now willing to pay for that talent: valuations that used to run at ten times revenue are now landing at 200 to 300 times, on companies with no shipped product and, in Metis Reasoning's case, not yet a working demo.
What This Means If You Rely on DeepMind's Research
None of this changes what ships in Gemini this quarter; Koray Kavukcuoglu, who already ran production models day to day, remains in operational control regardless of who leaves for a startup. What it does change is where the next generation of non-LLM AI research happens. If diffusion-based alternatives to transformers, or reasoning systems built on structured search, turn out to matter, they are now more likely to ship from a two-year-old startup with a single-digit headcount than from inside Google. An enterprise betting its roadmap on whatever Google's research lab does next is betting on a lab that keeps exporting the researchers most likely to try something else.
Read next: Your AI Agents Can Catch Fraud, Not Stop It | Half of Google's Weather Claims Are Independently Checked
