Google, which is seen as falling behind OpenAI and Anthropic in the artificial intelligence (AI) race, is speeding up a reorganization to scale back the independence of its in-house AI lab DeepMind. It is also reportedly pursuing a plan to carve out parts of DeepMind and move them to Google's headquarters.
Reuters reported on the 12th (local time), citing sources familiar with the matter, that Google held an all-hands meeting on the 6th and decided to transfer nontechnical teams under DeepMind to Google's headquarters. The announcement came right after Demis Hassabis, CEO of Google DeepMind, stepped back from day-to-day management to become chair, and Jeff Dean, a senior fellow known as Google's "legendary engineer," left the company.
According to the sources, Google's latest restructuring is part of Alphabet's "AI all-in" strategy to gain an edge in the AI model race. DeepMind had long operated with a strong research-lab identity that guaranteed autonomy and independence, but Google now appears set to emphasize commercialization and focus on strengthening the AI business's revenue. Sergey Brin, Google's co-founder, has in recent months urged the company to devote itself to improving the performance of its Gemini model.
After acquiring DeepMind in 2014, even though the company is headquartered in London, Google did not move it to the United States and encouraged long-term scientific research and various projects over commercial results. But as Gemini's performance has lagged behind OpenAI's "GPT" and Anthropic's "Claude," the company appears to be changing that policy.
Unlike Hassabis, who co-founded DeepMind and has remained in London, the newly appointed head of DeepMind, Koray Kavukcuoglu, has been working at Google's headquarters in Mountain View, California, since last year. Reuters said Kavukcuoglu is known within Google as someone with a strong grasp of commercialization.
Google briefly drew reviews that it had overtaken OpenAI when it unveiled "Gemini 3" in Nov., but this year it has fallen behind "Claude Mythos" and "GPT-5.6 Sol," and recent analyses say it now even trails open-source models from China.
According to AI model analysis firm Artificial Analysis, Google's highest-scoring model in performance tests, "Gemini 3.6 Flash," is tied for 11th alongside OpenAI's low-cost model "GPT-5.6 Luna" and China's "DeepSeek V4 Flash 0731."
Google's next-generation AI model "Gemini 3.5 Pro," which was slated for release in June, has had its debut delayed.
As this situation persisted, even the founders who had stepped back from front-line management stepped in and began pressing DeepMind. Co-founder Brin also stressed the need to concentrate resources on areas such as "recursive self-improvement (RSI)," which would allow models to advance on their own without human intervention.
Reuters said, "Google has not disclosed how it plans to run teams or research projects from which key executives have departed," adding, "Some employees worry that DeepMind had served as a shield preventing Google from being swayed by excessive commercial interests, and that this function will weaken."