Google co-founder Sergey Brin has in recent months strongly urged key AI employees to focus entirely on the company's Gemini model as it strives to catch up with competitors, according to two people familiar with his remarks.
When Anthropic was advancing with a preview of its powerful Claude Mythos model, Brin addressed hundreds of employees at a company-wide meeting in April, calling on Google DeepMind's AI lab to work faster, one anonymous source reported.
Google's Gemini model briefly surpassed competitors last November, but new updates from Anthropic and OpenAI have once again placed it in a catching-up position. According to one source and another person knowledgeable about the situation, Google postponed the release of a new flagship Gemini version by two months in August after internal testing showed that its performance continued to lag behind rivals in areas such as coding.
Restructuring Leadership at Google DeepMind
On August 5, the Mountain View, California tech giant announced a major reorganization of the leadership of its AI division, Google DeepMind. As part of this restructuring, leader Demis Hassabis stepped down in favor of his deputy, Karan Kavukchuoğlu. Hassabis took the role of chairman, taking on fewer management responsibilities. Simultaneously, two original Gemini technical leads left to found a startup.
Although some industry observers question whether Google is falling out of the model race, interviews with seven people familiar with Gemini development efforts indicate that Google's commercialization strategy and the recent staff reorganization are linked to its drive for model superiority.
These changes mark a further erosion of DeepMind's autonomy, which Google has gradually weakened since acquiring the London lab in 2014 as part of a continuous effort to use its AI across most of its software and increase profits. The restructuring included information shared at the general meeting on August 6 that some teams are moving from DeepMind to corporate Google, according to two sources.
Commercial Interests
Google declined to answer questions regarding Brin's recent involvement, the structural changes in DeepMind, and employee concerns.
Under Hassabis's leadership, Google DeepMind undertook a broader range of AI research projects while simultaneously creating Gemini models vital to corporate commercial interests. However, in previous years, Hassabis opposed some initiatives that could have brought new revenue to Google or helped strengthen its position in the AI race. His allies at Google previously claimed that he was not neglecting business or focusing too much on long-term science.
In 2025, Google CEO Sundar Pichai appointed Kavukchuoğlu, who was then DeepMind's technical director, to an additional senior role as Google's Chief AI Architect. His task was to integrate Gemini into all company products. Four sources reported that in the following months, the influence of some leaders at DeepMind diminished, while Kavukchuoğlu became the central figure determining Gemini's direction.
One source noted that Kavukchuoğlu had Brin's support. The news of the restructuring caused Google's stock to drop by 4% on the day of the announcement.
At the company-wide meeting last Thursday, Hassabis and Kavukchuoğlu attempted to ease employee concerns—most of whom learned about the change when it became public—about the future of the AI lab. Google leaders stated that daily operations in the lab would remain the same, according to three sources with direct knowledge of the meeting. Sources reported that Kavukchuoğlu would remain in charge of Gemini; he took over oversight of the model's development more than a year ago, according to five sources.
Hassabis delivered a presentation on how AI can advance science and change society—long-term issues that will be the focus of his new role as chairman, according to three sources familiar with the meeting. Despite this message, new concerns arose at the general meeting that DeepMind's autonomy is being absorbed by Google. Hassabis and Kavukchuoğlu informed employees that several non-technical teams are being moved from DeepMind to the corporate reporting structure, according to two sources. For an organization that once encouraged employees to pursue auxiliary projects, the message of regaining dominance in models was also evident.
Leaders, including Pichai, have stated in recent months that Google considers itself at the forefront in certain areas, such as video generation, and remains committed to catching up in other fields despite setbacks.
Meanwhile, senior leaders in the cloud division expect Kavukchuoğlu's appointment to alleviate tension within DeepMind regarding the allocation of Google's limited supply of AI chips, known as Tensor Processing Units (TPUs), another source reported.
Boundaries of Progress
Google's restructuring occurs against the backdrop of Brin's efforts to mobilize personnel, something he also did after OpenAI launched ChatGPT in 2022, which shook Google.
Although he no longer holds an executive position, the 52-year-old Brin unofficially influences model training efforts. Over several months, he has told key personnel that they need to reach cutting-edge milestones, according to one source familiar with his efforts. Brin uses the hidden power he possesses as a Google co-founder to guide resource allocation in specific areas, such as recursive self-improvement—a point where technology can improve without human intervention, this source reported. Google declined to provide Brin for an interview.
Meanwhile, Hassabis reduced his involvement in managing Gemini and handed over this responsibility to his deputies shortly after initially launching the AI model in 2023, according to two sources.
Although Google caught up to the cutting edge with the release of Gemini 3 last November, it quickly fell behind again. Four sources attributed its lag to persistent structural issues that initially led to it. For example, computational power limitations and disagreements among many Gemini project leaders resulted in fewer resources for critical areas like coding, according to two sources. Google's bureaucratic nature also led to it releasing models slower than more agile competitors, according to four sources.
Google declined to comment on these claims, but Pichai told Wall Street analysts last month that the company will continue to build out its infrastructure to remove constraints around TPUs, which it needs for training models, as well as for consumer and enterprise products.
Against this backdrop, Kavukchuoğlu is consolidating his power. He moved from London to Mountain View last year upon promotion to Chief AI Architect, reporting to Pichai. Some leaders from London found their influence on Gemini's technical direction diminished after Hassabis reduced his involvement in Gemini, and Kavukchuoğlu moved to Google headquarters, according to a source with direct knowledge.
Kavukchuoğlu has also begun building a reputation in other parts of Google as a person more focused on commercialization. Even before his move, Kavukchuoğlu, not Hassabis, was the main contact between DeepMind and Google Cloud, the company's primary AI revenue source, according to a source with direct knowledge. He will now make final decisions on major matters for DeepMind, a Google representative stated last week.
Some of the most prominent leaders, besides Kavukchuoğlu, including original Gemini technical co-leads Jeff Dean and Oriol Vinyals, left the company during last week's reorganization. Employees who reported to some of these leaders had a rare degree of autonomy to conduct research unrelated directly to Gemini training, according to three sources. After their departure, Google did not immediately share plans regarding the fate of these teams or projects, causing concern among some about the further erosion of the lab's bastion against commercial interests.
