Google dismantles DeepMind as Hassabis heads for the exit
Google DeepMind appears to be losing its independence and turning into an ordinary Google subsidiary, and founder Demis Hassabis may leave the company for good in the coming months. According to a Guardian report cited by The Decoder, a former Google manager believes the "era of DeepMind as an independent actor is over," and a current DeepMind employee now calls the lab "just another subdivision" of Google. AI researcher Koray Kavukcuoglu is taking over day-to-day operations from the U.S., but without the CEO title, meaning DeepMind no longer has its own chief executive.
On Hassabis himself, a Pathfounders report says he wanted to leave at the same time as Jeff Dean, but Google worried a second departure would hit the stock even harder than the drop that followed the news of Dean's exit and Hassabis's changed role. Instead, Hassabis was "promoted" to chairman, a move meant to smooth the way toward an orderly exit. He plans to spend more time on his scientific work and on the pharmaceutical startup Isomorphic Labs, and could leave DeepMind for good in the coming months. With Hassabis sidelined, all Gemini-related AI development is shifting to the Bay Area, and co-founder Sergey Brin is expected to take on even more control. Brin had already come out of retirement during Google's first "red alert" over ChatGPT; shortly after, Google Brain and DeepMind merged into "Google DeepMind," which is also where the Gemini name came from, since the "twin" teams from both organizations were meant to work jointly. Tensions over that restructuring, including frustration from Hassabis, reportedly go back years.
Two outside interpretations of the shakeup are circulating. SemiAnalysis, in a newsletter piece titled "Gemini is Cooked but GCP is Cooking," argues through Dylan Patel and colleagues that DeepMind is no longer a frontier lab, pointing to years of overly cautious compute allocation and a bureaucratic, risk-averse culture that left Google unable to get its AI systems to users fast enough despite having built them early. SemiAnalysis compares the moment to IBM's retreat from PCs into mainframe consulting or Intel abandoning bold bets to protect its legacy business, framing both as mistakes. It also argues Google is prioritizing short-term cloud profits, selling compute to third parties rather than resourcing its own teams to lead in AI applications: Gemini's annualized revenue was $12 billion in the second quarter of 2026, while Google Cloud is projected to generate over $73 billion in AI infrastructure revenue plus an additional $120 billion in TPU sales by the end of 2027.
The second reading comes from Tim O'Reilly, in his Substack newsletter "Asimov's Addendum." O'Reilly argues Google is deliberately running a different race, prioritizing infrastructure over model leadership, on the reasoning that in major tech shifts it is not the inventors who win but those who spread the technology. He points to Thomas Edison, who invented the light bulb and the first power plant but was overtaken by George Westinghouse, whose alternating-current grid brought electricity to all of society; Google's efficiency-focused Flash models, in this reading, fit the Westinghouse playbook. A third, less flattering account from the rumor mill holds that Google did try to train frontier models but the team could not catch the leaders, especially on coding performance: Gemini 3.1 Pro remains in preview only, and Gemini 3.5 Pro, announced in May for a June release, has been largely shelved.
Key facts
- Operational control of DeepMind is shifting to Koray Kavukcuoglu in the U.S., who will run day-to-day operations without holding the CEO title.
- Demis Hassabis wanted to leave alongside Jeff Dean but was instead "promoted" to chairman to allow an orderly exit; he could leave DeepMind for good within months to focus on his scientific work and Isomorphic Labs.
- Sergey Brin is expected to take on more control as Gemini development centralizes in the Bay Area.
- SemiAnalysis's Dylan Patel argues the restructuring shows Google losing the race for frontier AI models, citing Gemini's $12 billion annualized revenue in Q2 2026 against Google Cloud's projected $73 billion-plus in AI infrastructure revenue and $120 billion in TPU sales by end of 2027.
- Tim O'Reilly counters that Google is deliberately choosing infrastructure over model leadership, while a third account says Google tried and failed to keep pace on frontier models, with Gemini 3.1 Pro still in preview and Gemini 3.5 Pro largely shelved.
Why it matters
DeepMind has operated for years as Google's semi-autonomous AI research arm, distinct in culture and leadership from the rest of the company. Folding it into a centralized Bay Area operation under an executive who does not hold the CEO title, while its founder moves toward the exit, marks a structural shift in how Google runs its AI effort at a moment when it is under direct competitive pressure from OpenAI and Anthropic. It also reframes the public story around Gemini: whether Google's AI push succeeds or falls behind now traces to a single, more centralized command structure rather than DeepMind's semi-independent one.
Who it affects
Demis Hassabis, moving from chief executive to chairman and reportedly toward a full exit; Koray Kavukcuoglu, who inherits day-to-day operational control without the CEO title; Sergey Brin, expected to take on more direct influence over AI development; and DeepMind's staff, now described internally as working inside "just another subdivision" of Google rather than an independent lab. Investors and rivals watching Google's AI trajectory are also affected, since the reshuffle touches both Gemini's model roadmap and Google Cloud's infrastructure business.
How to use it
Readers tracking Google's AI strategy have two competing frameworks on offer here. SemiAnalysis's reading treats the restructuring as a warning sign, worth weighing against Gemini's actual model performance and release cadence. Tim O'Reilly's reading treats it as a deliberate bet that infrastructure and distribution, not frontier-model bragging rights, decide who wins the AI race long term, the way Westinghouse's power grid beat Edison's light bulb. Neither is confirmed as Google's official reasoning, so the two are best used as competing lenses rather than a settled explanation.
How solid is it
The account rests on a Guardian report, a Pathfounders report, and quotes from an unnamed former Google manager and a current DeepMind employee, none of it confirmed by an official Google statement on the restructuring or its reasons. The financial figures, Gemini's $12 billion annualized revenue and Google Cloud's $73 billion and $120 billion projections, are attributed but not independently sourced in this piece. The two big-picture interpretations, from SemiAnalysis and from Tim O'Reilly, are explicitly framed as competing outside commentary rather than facts, and the piece itself does not endorse one over the other.
Risks and caveats
No confirmed departure date exists for Hassabis: the reporting says only that he "could leave DeepMind for good in the coming months" and is "potentially leaving the company in the near future." There is no official Google statement confirming the restructuring or the reasons behind it, and Kavukcuoglu is explicitly reported as not holding the CEO title, so any read of him as a de facto chief executive overstates what is known. The 2027 revenue and TPU-sales figures are forward projections, not results, and the two competing explanations for the shakeup, a strategic pivot to infrastructure versus a failure to keep pace on frontier models, are not reconciled by the source.
“era of DeepMind as an independent actor is over”
— a former Google manager, per a Guardian report