Here is the contradiction most people are papering over.
Google's own message to staff says Demis Hassabis is stepping up, not out — Chair of Google DeepMind, Chief Scientist of Alphabet, writing warmly to his team "from our awesome new London Platform 37 offices."
Pathfounders, citing industry sources, says something else entirely: that Hassabis originally planned to leave Google alongside Jeff Dean, and that management, terrified of what a simultaneous exodus would do to the share price, talked him into staying — and predicts he'll be fully gone within a year.
Both of these can't be a promotion. And once you hold them side by side, the whole reshuffle stops looking like succession and starts looking like something Google is managing rather than choosing.
Once you hold them side by side, the whole reshuffle stops looking like succession and starts looking like something Google is managing rather than choosing.
Start with the scene that diagnoses the disease. Before all this, Hassabis would finish dinner with his family in London, reopen his laptop, and work until 4 a.m. — not to think, but to overlap with the California team. Gemini's development was spread across offices spanning eight time zones, coordinated through hundreds of internal chat groups.
That is not a research problem. That is a map problem. And Google decided the map was the thing to fix.
Notice what the fix was not. It was not better async tooling, not a rethink of how a globally distributed lab collaborates. The fix was geography: bring the people back to Mountain View and put the decisions in one building.
Koray Kavukcuoglu, DeepMind's longtime CTO and now the SVP actually running Google DeepMind, already relocated from London to California — named Chief AI Architect back in 2025 with an explicit mandate to shorten the distance between research and product and make Gemini iterate faster.
Sebastian Borgeaud, who runs a key Google AI coding project, has moved from London to California too. Per The Information, the Gemini post-training team — the people doing the final tuning before a model ships — has been pulled from scattered offices into headquarters, with Sergey Brin personally watching the pace.
So when Hassabis writes that he'll advise "from our awesome new London Platform 37 offices," read it against where power actually sits. The operational chain has already crossed the Atlantic. The London office is now a place he waves from.
This is why the "promotion" and the spin-out are better understood as one event, not two.
Jeff Dean — employee number 30, 27 years in, foundational to Search infrastructure, Brain, TensorFlow, TPUs, and a Gemini technical co-lead — left to found Discovery Loop with Sanjay Ghemawat, Oriol Vinyals (DeepMind research VP and Gemini technical lead), and Quoc Le (a Google Brain co-founder). Alphabet is a founding investor and cloud partner, and the whole thing is framed as a friendly departure.
Friendly or not, that is four of the most decorated builders in the company walking out the same door on the same day. Dean's own pitch deck reads like Google's product catalog — Search, Ads, Gmail, Translate, Gemini, Cloud TPU, GFS, MapReduce, Spanner, Word2Vec, MoE, PaLM, AlphaFold — because these people, collectively, built the modern Google AI stack.
Now put Pathfounders' claim next to that: Hassabis meant to leave with them. Alphabet's stock dropped on the news as it was; a Dean-plus-Hassabis exit on the same morning would have read, externally, as the core of DeepMind collapsing at once.
So the story goes: Google absorbed the Dean cluster as a supportive spin-out, and retained Hassabis with a bigger title and a smaller job — enough of a soft landing that the market could digest it, enough of a transition that a later, quieter departure becomes the expected next step.
The tell is that the real handoff had already happened. Hassabis has spent the past year gradually offloading day-to-day management of Gemini and consumer products to Koray. Pichai's memo is almost explicit about it: the models are "in good hands with Koray and the leads, as they have been for a while."
Which is to say — by the time the reshuffle was announced, Hassabis was already advising a company that no longer ran the way he ran it.
Zoom out and the pattern is unmistakable: Google is simultaneously bleeding icons and recentralizing. The Financial Times reported the standalone AlphaFold team — the work that produced a Nobel — has been broken up, its researchers redistributed into Gemini, AI coding, and genomics. Other DeepMind people have simply left London.
And where Google can't recentralize fast enough, it buys. The company is reportedly in talks worth more than $1.5 billion for a non-exclusive license to Mechanize's technology plus a pull of its core staff — the San Francisco startup founded by Tamay Besiroglu, Matthew Barnett, and Ege Erdil — to close its gap in AI coding, where Claude Code, Codex, and Cursor took the lead Google thought its full stack would defend.
Relocate the decision-makers into one building; break up the autonomous science teams; buy a coding team off the shelf. These are three moves aimed at one target: velocity.
Because the uncomfortable backdrop is that Google, with planetary-scale data, custom TPUs, distribution to billions, and ~$200B in annual AI capex, is still mid-pack on the benchmarks that define the frontier — trailing Anthropic's Claude family and OpenAI's GPT-5.6, with Gemini strong on multimodal, long-context, and cost but not on pure reasoning and agentic coding.
Google has concluded the bottleneck is not research culture. It's shipping speed. And its remedy is command from Mountain View.
Which lands us on the irony that makes this more than an HR story.
Hassabis has said, in Fortune, that one reason he refused to move DeepMind to Silicon Valley in the first place was to prove a point: that you could build world-class AI outside it. DeepMind was born in London as an argument against the assumption that the frontier had to run on California time.
Google is now betting the opposite. It has decided the London-centered, freedom-to-explore model that made DeepMind is the thing slowing it down — and that the cure is to pull power, talent, and decisions back inside headquarters, and to buy velocity where it can't grow it.
That is a defensible bet in a race being lost on release cadence. But it is a very specific one, and worth naming for what it is.
Google is not choosing between its founders' vision and its rivals' urgency by keeping both. It is choosing urgency, and letting the vision walk — Dean to Discovery Loop, the AlphaFold team scattered, Hassabis retained on a countdown that his own gravitational pull toward Isomorphic Labs and AI-for-science makes hard to reverse.
The founding generation is cashing out of an organization that no longer wants what they built. The remarkable part is that Google seems to have decided it can afford to let them.