AI

Google merged its two AI labs in 2023 and kept two capitals. This week it picked one.

The reset that knocked $190 billion off Alphabet on Wednesday was also a map change. Koray Kavukcuoglu moved from London to Mountain View in the past year and now runs Google DeepMind; Sebastian Borgeaud, who leads a key coding effort, relocated from the UK too. Former employees describe split compute, one-way document access, and the “America-fication” of a lab Demis Hassabis spent eleven years keeping British.

N Noah · The Sharp Brief · August 6, 2026 · 5 min read
An empty desk and chair in a glass-walled AI research office at dusk, with colleagues still working in the background

The market read Wednesday’s memos as an exit story: Jeff Dean gone after 27 years, three senior researchers with him, and roughly $190 billion briefly off Alphabet’s valuation. That is the loud version.

The quieter version, reported Thursday by Bloomberg and by Forbes, is geographic. Google’s AI leadership is being pulled to Mountain View. Kavukcuoglu, until this week the division’s chief technology officer, relocated from London to Google headquarters in the past year when he was made the company’s chief AI architect. Borgeaud moved from the UK to California as well. Hassabis — who is based in London, who founded DeepMind there in 2010, and who made staying in London a condition of the roughly £400 million sale to Google in 2014 — becomes chairman.

Note the title Kavukcuoglu actually got. Not chief executive. Senior vice president of Google DeepMind, reporting to Sundar Pichai. The CEO seat is not being filled. It is being dissolved into the org chart at headquarters.

Our take: The 2023 merger of Google Brain and DeepMind was sold as one company and one AI army. It produced one org chart and two capitals — Hassabis running the combined division from London, Dean in Silicon Valley, both reporting to Pichai. Three years later the tie-break finally happened, and London lost. That is a real answer to a real problem: you cannot ship a frontier model when two continents each think they own it. It is also the most expensive kind of fix, because the people you lose in the process are the ones with the most options. Google did not choose between speed and talent this week. It chose speed and is now finding out what talent costs.

What the split actually looked like inside

Former employees told Forbes the fault lines predate Wednesday by years, and they were mundane in the way real organizational rot usually is.

Document access ran one way. In some cases London-based DeepMind researchers could see projects and documents from Mountain View researchers, but not the reverse — a detail two former employees described as a source of lasting mistrust. Where projects overlapped between the offices, one team sometimes got favored on the basis of which building it sat in. “Things are quite politicized based on who you report to,” one said.

Then compute, which is the only currency that matters in an AI lab. Shortly before Noam Shazeer left Google in June, computing resources dedicated to one of his projects were reassigned to a London-based team, according to Bloomberg’s earlier reporting. Shazeer is the lead author of the 2017 transformer paper and a researcher Google paid roughly $2.7 billion to acqui-hire back in 2025. He went to OpenAI. Two days later John Jumper — who shared the 2024 Nobel Prize in Chemistry with Hassabis for AlphaFold — joined Anthropic.

One long-tenured London employee called the direction of travel the “America-fication” of DeepMind. “It’s the loudest person in the room,” that person said. “There’s just a sharper edge.”

Google’s position, given to Forbes: it is not aware of any rivalries between employees in the two cities, the merger went very smoothly, it remains committed to London where it opened a new office earlier this year, and Google DeepMind is “a global team with a global footprint.”

Why this is the fix Google reached for

The scoreboard explains the urgency. Google has not shipped a major Gemini update in months. Gemini 3.5 Pro was due in June and still is not out; Axios reports the delay is partly a morale problem. Meanwhile Alphabet has guided to roughly $205 billion of capital spending this year and ran negative quarterly free cash flow to fund it.

That combination is the whole argument. When you are outspending the field and still late, the bottleneck is not silicon. It is the number of people who have to agree before a decision sticks — and a five-to-eight-hour time zone gap between two rival power centers is a structural tax on exactly that. Bloomberg frames the consolidation as an attempt to reverse it.

Kavukcuoglu is a reasonable bet to run it. Turkish-born, he studied under Yann LeCun at NYU, joined DeepMind in London in 2012 — two years after founding, two years before the Google sale — and spent 13 years there before moving to Mountain View. He is fluent in both buildings, which is the actual job description. A Google executive put it bluntly to Forbes: “Koray was already running probably the majority of DeepMind for quite some time.”

What to watch

The through-line here is that Google’s AI problem was never a compute problem — a point that also explains why rivals are building their own silicon while Google already has TPUs and is still behind. Alphabet spent three years running the most valuable research organization in the industry as a federation, discovered federations are slow, and dissolved it. The bill for that arrives in headcount, not capex.

Advertisement

Get the day, decoded — at 7 PM ET

The Sharp Brief: AI, money, business & performance in five sharp minutes. Free.

Free bonus: subscribe today and The 2026 AI Playbook (PDF) lands with your welcome email.

Recommended by 5+ newsletters across AI, markets & business.