Google is dismantling DeepMind's independence, putting one man in charge of Gemini

On August 6, Google DeepMind held an all-hands meeting and announced a structural change: the lab's non-technical teams, including business, communications, and operations, would be moved out of DeepMind and into Google's central corporate reporting line. Research teams stay. The move, first reported by Reuters on August 12, is the largest unwinding of the independence buffer Google built when it acquired DeepMind in 2014.

The 2014 deal gave DeepMind operational distance from Google's main business, on the theory that a frontier research lab needs protection from quarterly product pressure. Twelve years later, that buffer has become the bottleneck, and Google is collapsing it on purpose.

Google's Gemini flagship model is roughly two months behind schedule, internal testing has it trailing on coding benchmarks, and the company has cancelled the planned Gemini 3.5 Pro release, according to Reuters and Ars Technica citing people familiar with the matter. DeepMind product lead Logan has said Gemini 4 pre-training has begun, but the company has not named a release date.

Three failure modes explain the slippage.

First, compute contention. Gemini is trained on TPUs, Google's in-house AI chips, and the same chips serve Search ranking and Google Cloud customers. When supply tightens, training loses.

Second, internal project-lead conflict. Reuters reports that multiple Gemini teams are running parallel training runs without unified prioritization. Koray Kavukcuoglu, who took over model-development oversight in 2025, has been the main liaison between DeepMind and Google Cloud. A Google spokesperson confirmed last week that Kavukcuoglu now has final-decision authority on major DeepMind decisions.

Third, release friction. OpenAI and Anthropic ship model updates on a roughly two-month cadence; Google's public release timeline has slipped behind both. The reorg moves shipping coordination into the same corporate function that ships Search and Cloud features, on the theory that the friction is structural, not technical.

Demis Hassabis, the DeepMind co-founder who has run the lab since the acquisition, remains as its head but no longer runs day-to-day operations. Kavukcuoglu, a long-time DeepMind principal who joined the company in 2015, is the central figure shaping Gemini's direction. The change is not a departure; it is a reporting-line shift designed to put a single accountable owner between the lab and the product surface.

Sergey Brin has also re-entered the picture. The 52-year-old Google co-founder holds no formal management title but is pushing resource allocation toward "recursive self-improvement," a research direction in which AI models help train and refine the next version of themselves, reducing per-iteration dependency on human researchers. Brin's previous re-engagement came after ChatGPT's late-2022 launch. This one was triggered in part by Anthropic's April 2026 Claude Mythos demonstration, which internal Google reviewers treated as a frontier shift.

The structural question is whether centralizing model development under Kavukcuoglu actually cuts shipping time, or simply relocates the friction. Three things to watch: whether the next Gemini release ships inside two months; whether TPUs allocated to Gemini training increase at the expense of Cloud customer supply; and whether Brin's "recursive self-improvement" program shows up as a public capability, or stays inside the lab.

Google's bet is that the 2014 buffer was the right structure for a research lab and the wrong structure for a model that ships into Search, Cloud, Workspace, and a coding benchmark war. The reorganization is the test of that bet.