Google DeepMind's Leadership Shake-Up Tests Its Scientific Independence

The Leadership Reset Inside Google DeepMind

Google DeepMind, the research unit that created AlphaFold and Gemini, is undergoing a leadership reset that reaches far beyond a title change. Demis Hassabis, the cofounder and longtime CEO, said last week he would move into a chair role, handing day-to-day control to chief technology officer Koray Kavukcuoglu. Notably, Kavukcuoglu now reports directly to Alphabet CEO Sundar Pichai, rather than holding a standalone DeepMind chief title—a structural shift that pulls the unit closer to Google's Mountain View headquarters and away from its London roots.

The transition was followed almost immediately by news that chief scientist Jeff Dean is leaving. It also comes after Gemini 3.5 Pro, announced at Google I/O in May, missed three release dates. According to independent benchmarking firm Artificial Analysis, Google's best currently shipped model, Gemini 3.6 Flash, now trails Anthropic, OpenAI, xAI, Meta and at least one Chinese lab on raw intelligence—a reversal from a stretch last year when Google briefly topped the AI leaderboard.

The leadership and benchmark setbacks are not happening in isolation. In a single week in June, Google lost Gemini co-lead Noam Shazeer to OpenAI and AlphaFold co-inventor John Jumper to Anthropic; two more AlphaFold veterans, Jonas Adler and Alexander Pritzel, followed Jumper. Engineers cited aggressive poaching by cash-rich rivals, frustration over falling coding-benchmark performance and the lure of pre-IPO equity as reasons for the exodus.

A Google spokesperson pushed back on the idea that the firewall between DeepMind and parent Google is breaking down, saying London remains central and DeepMind keeps its research autonomy under the new structure. Still, several engineers told Fortune that the reshuffle compounds a slow-motion shift of power from London to Mountain View, leaving the unit to prove it can move fast enough without the two figures most associated with its scientific identity.

What DeepMind's Identity Crisis Means for Google's AI Race

Kavukcuoglu Inherits a More Integrated, Less Independent DeepMind

Kavukcuoglu's reporting line is the clearest sign of the structural change. Under the previous model, DeepMind's CEO ran the unit with substantial autonomy; under the new arrangement, the day-to-day leader is positioned as a senior executive inside Google rather than as the head of a semi-autonomous research lab. That may simplify coordination between DeepMind and Google's cloud and product teams, but it also removes a layer of buffer that researchers and engineers see as part of the unit's identity. Google disputes the firewall concern, but the organisational distance has undeniably shortened.

Google Still Has an Infrastructure Cushion, but the Model-Layer Gap Is Real

Artificial Analysis's finding that Gemini 3.6 Flash trails rivals on raw intelligence is not necessarily fatal for Google because the company's advantage has long been distribution and infrastructure: its chips, Tensor Processing Units, cloud platform and default access to products such as Search, Workspace and Android. In enterprise AI, raw benchmark scores matter, but so do cost, reliability, ecosystem integration and procurement path. The risk is that repeated delays and weaker benchmark showings erode the perception that Google is the default choice for frontier AI customers.

The Talent Exits Hit DeepMind's Scientific Core, Not Just Its Executive Ranks

Shazeer's move to OpenAI and Jumper's move to Anthropic are significant because both are associated with the technologies that gave DeepMind and Google their scientific credibility: large language models and AlphaFold. Losing Adler and Pritzel compounds the signal. The engineers' explanation—poaching, coding-benchmark frustration and pre-IPO equity—suggests the competition is now as much about compensation and equity upside as about research mission. That is a harder problem for Alphabet to solve with computing power alone.

What the DeepMind Reset Signals for Enterprise Buyers and AI Talent

The leadership reset and benchmark slide give enterprise AI buyers a concrete checklist when evaluating Google's frontier models, and give the broader market a clearer signal about where Alphabet's AI risk sits.

• Enterprise buyers evaluating Gemini should benchmark the actually shipped model, not the announced roadmap. The story names Gemini 3.5 Pro's three missed release dates and the Artificial Analysis finding that shipped Gemini 3.6 Flash trails Anthropic, OpenAI, xAI, Meta and a Chinese lab on raw intelligence. Procurement teams can ask Google to attach model-version performance metrics to contracts and include a right to re-test before renewal.
• Teams building on Gemini should pin workloads to specific model endpoints and avoid assumptions about feature cadence. The missed dates mean a migration or feature promised around Gemini 3.5 Pro could arrive later; technical teams can reduce rework by isolating model calls and maintaining a fallback to an alternative frontier model.
• Alphabet watchers and enterprise partners should tie their confidence to named signals, not general reassurances. The relevant signals from this story are whether Kavukcuoglu's reporting line improves coordination without further London departures, whether Gemini's next shipped model closes the benchmark gap identified by Artificial Analysis, and whether remaining DeepMind research leaders stay.