Google DeepMind is navigating a turbulent period marked by high-profile leadership departures, stalled product releases, and an exodus of key talent to rival AI labs.

Last week, Demis Hassabis, longtime CEO and cofounder of Google DeepMind, announced he was stepping back from the chief executive role to assume a new position as chair. Hassabis handed day-to-day operational control to DeepMind chief technology officer Koray Kavukcuoglu, who now reports directly to Google CEO Sundar Pichai rather than retaining a standalone DeepMind leadership title. Within minutes of that announcement, news emerged that chief scientist Jeff Dean was also departing.

These leadership changes cap a difficult stretch for the AI division. Gemini 3.5 Pro, which was unveiled at Google I/O in May, has now missed three scheduled release dates. According to independent benchmarking firm Artificial Analysis, Google's best currently shipped model, Gemini 3.6 Flash, trails Anthropic, OpenAI, xAI, Meta, and at least one Chinese laboratory in raw intelligence—a notable reversal from last year, when Google briefly held the top position on the AI leaderboard. The benchmark gap matters beyond bragging rights: enterprises increasingly select cloud and productivity platforms based on which provider offers the strongest underlying models, meaning model-tier slippage can ripple into Google Cloud and Workspace revenue.

Several engineers told Fortune that the organizational reshuffle accelerates a gradual shift of influence from London, where DeepMind was founded, to Mountain View, the home of Google's headquarters. They expressed concern that Hassabis's departure from the CEO role signals the erosion of the longstanding firewall between DeepMind and its parent company—a separation that was culturally significant because DeepMind originally operated as a semi-autonomous research lab after Google acquired it in 2014, with its own ethics board and independent research agenda. A Google spokesperson disputed that characterization, stating that London remains central to operations and that DeepMind retains its research autonomy under the new structure.

Talent retention has also become a pressing challenge. During a single week in June, Google lost Gemini co-lead Noam Shazeer to OpenAI and AlphaFold co-inventor John Jumper to Anthropic. Two additional AlphaFold veterans, Jonas Adler and Alexander Pritzel, subsequently joined Jumper in leaving. Engineers attributed the departures to aggressive recruitment by well-funded competitors, frustration over declining performance on coding benchmarks, and employees seeking pre-IPO equity at rival firms before those companies go public. The losses reflect a broader industry pattern in which top AI researchers command compensation packages reportedly reaching into the tens of millions of dollars annually, as labs backed by multibillion-dollar war chests compete for a shallow pool of experienced practitioners.

Despite these setbacks, Google retains significant structural advantages. The company's proprietary chip infrastructure—its Tensor Processing Units—cloud computing capabilities, and global distribution network remain unmatched by any AI-native competitor—assets that could compensate for weaker performance at the model layer. However, four years after Google's first internal AI "code red"—declared in response to ChatGPT's viral launch in late 2022—DeepMind faces renewed pressure to demonstrate it can iterate quickly enough to stay competitive. The coming months are likely to be judged on whether the next major Gemini release ships on schedule and whether the leadership transition stabilizes or accelerates the talent drain. This time, DeepMind must navigate that test without the two individuals—Hassabis and Dean—most closely identified with the organization's scientific identity.

This story was originally featured on Fortune.com.