Google DeepMind has produced more startup founders than OpenAI, Anthropic or Meta, more than 40 of them by Bloomberg's count. The newest crop skips chatbots entirely1. Nando de Freitas wants diffusion models to replace transformers. David Silver wants reinforcement learning to replace human text as a model's teacher. Jack Parker-Holder and Thore Graepel each want a piece of the same bet. Each raised money separately, and each aimed at one target: the architecture DeepMind still ships to billions of Gemini users1. Four ventures, read together, are asking investors to price the end of the paradigm that built their founders' careers. That paradigm still runs DeepMind's own flagship product.
Fifteen People, One Breakfast, in London
Numbers alone miss the texture behind this exodus. Fifteen Google DeepMind employees and alumni sat down for breakfast in central London this year. They swapped notes on one subject: which Silicon Valley investors write checks for ventures that skip large language models entirely1. Researchers who once competed for the same internal promotion now compare notes on outside term sheets instead. A shared breakfast table reveals something a funding chart misses on its own. Leaving DeepMind has become routine enough to plan for over coffee, and common enough that the people doing it already know who else is doing it too.
The Pressure Behind the Departures
Janusz Marecki spent years as a DeepMind lead scientist. Now he works as an AI partner at Ahren Innovation Capital, reading pitch decks from researchers he once sat beside. He named the reason they left in plain terms. "There's a top-down pressure to drop everything and focus on large language models," he told Bloomberg1. That pressure reads as strategy from inside DeepMind's leadership: focus the lab on the product already paying its bills, Gemini. From a researcher's chair, the same focus reads as a ceiling. Marecki's framing carries weight for a specific reason. He is paid now to find and fund the researchers who hit that ceiling and left, which gives his read on DeepMind its own bias, the same way Hassabis's read carries his.
Four Bets Against the Transformer
Revolution Labs, de Freitas's venture, is raising more than $100 million. Its bet is diffusion models, the technique that turns noise into images and video, applied instead to the reasoning tasks transformers now dominate1. Silver's Ineffable Intelligence went further and faster. It closed a $1.1 billion seed round in April, the largest in Europe, at a $5.1 billion valuation, co-led by Sequoia Capital and Lightspeed Venture Partners3. Silver frames the ambition as a research mission first and a business second. "Any money that I make from Ineffable will go to high-impact charities that save as many lives as possible," he said, pledging away the bulk of his personal equity before the company had shipped a product3. Parker-Holder's Emulate is seeking a $700 million round of its own. Graepel's Metis Reasoning is chasing tens of millions more, and his resume gives the pitch its edge. Graepel co-authored the Nature paper introducing AlphaGo. He later led machine learning work at Altos Labs, rejoined Google's research arm, then departed again this summer6. His thesis borrows straight from that history. Capability's next leap, in his telling, comes from machines that plan and search through uncertainty, the way AlphaGo once beat the world's best Go player. Scaling a language model further stalls out well short of that goal, he argues6. All four ventures sit in early stages, still years from anything a paying customer could buy. Investors have already priced the wait, betting the science arrives on schedule.
Google's Money Follows Its Own Researchers Out the Door
Follow the money. The stranger detail sits in the cap tables: Google itself appears among Ineffable Intelligence's seed investors, alongside Nvidia, Sequoia, Lightspeed, Index Ventures and the U.K.'s Sovereign AI Fund3. Alphabet backed Discovery Loop too, the science automation startup Jeff Dean founded after 27 years at Google, alongside Radical Ventures and Khosla Ventures2. Dean's own description of Discovery Loop's ambition explains what Google chose to fund. "Imagine a future where a handful of people can conduct scientific research and engineering tasks much more rapidly, and with higher quality, than massive teams of scientists and engineers do today," he said2. Google CEO Sundar Pichai sent Dean off on friendly terms. "After an incredible 27-year run, Jeff Dean is at a moment where he wants to try something new, and we're excited to support him in that," Pichai said2. A company that funds the researchers betting against its own core product has made a choice. It would rather own a piece of the bet than let a rival place it alone.
Hassabis Calls It Business as Usual
Demis Hassabis stepped back from running DeepMind day to day in August, taking the title of Alphabet chief scientist. Koray Kavukcuoglu, DeepMind's former technology chief, took the wheel in his place4. Asked whether the alumni exodus signaled trouble, Hassabis pushed back on the premise. "There's a lot of talent movement between all the leading labs and we win our fair share of the top talent," he said, adding that DeepMind still has "by far the biggest and broadest research bench of any of the labs"1. Both claims can be true and still miss the point his own alumni are making. Talent moving between labs describes a different pattern entirely from talent leaving to build a rival kind of lab. A deep bench answers a question about DeepMind's present roster. It leaves the harder question standing: whether the architecture that bench works on survives the decade unchanged.
What Each Side Needs to Be True
Marecki and the founders he now funds need the transformer's dominance to be a plateau rather than a peak. They need a pause to arrive before diffusion models, or reinforcement learning trained on synthetic experience, take over the hardest reasoning tasks. Their case gained a data point outside language models entirely. DeepMind alumni have also begun building tools aimed at speeding up fusion power. That is a sign the departures track a broad appetite for problems current architectures handle poorly, rather than a single grudge against one product line5. Hassabis needs the opposite to hold. Scale and engineering depth inside one lab need to beat four smaller, separate bets on unproven designs. DeepMind's own roadmap, Gemini included, needs to keep growing faster than any startup working alone. Both sides are pricing the same uncertainty. The founders price it in equity they gave up a paycheck to hold. Hassabis prices it in headcount he still commands.
By the numbers
- Startup founders: more than 40, the count Bloomberg attributes to DeepMind and its alumni, ahead of OpenAI, Anthropic or Meta1.
- Ineffable Intelligence's April seed round closed at $1.1 billion, the largest ever raised in Europe3.
- That round valued the company at $5.1 billion, before it had shipped a product3.
- Emulate, Jack Parker-Holder's venture, is seeking a $700 million round of its own1.
- Revolution Labs opened its pitch to investors seeking more than $100 million1.
- 27 years: Jeff Dean's tenure at Google before he left to found Discovery Loop2.
- Hassabis stepped back from DeepMind's daily operations in August 2026, becoming Alphabet's chief scientist4.
What to watch
Watch whether any of the four ventures publishes an independent benchmark result that beats a transformer model on a task researchers actually care about. That result would turn a funding story into a technical one. Gemini's own release cadence under Kavukcuoglu matters just as much. A lab that keeps shipping faster than its alumni can raise money makes Hassabis's case for him. One more number is worth tracking too: how many Google researchers leave for ventures Google itself ends up funding. A rising count would suggest the company treats the departures as a hedge, rather than a loss.