Google's AI empire is growing. The researchers who built it keep leaving.

Key points

• Jeff Dean, Google's chief scientist for AI, announced his departure on Wednesday after 27 years at the company, to co-found a new AI startup called Discovery Loop.
• Demis Hassabis, who led Google DeepMind, is stepping down as its CEO to become chairman and take a new Alphabet-wide chief scientist role focused on long-term research.
• All eight authors of the landmark 2017 paper that laid the groundwork for today's AI chatbots have now left Google.
• Google's cloud division posted 82% revenue growth in its most recent quarter, showing the commercial side of its AI bet is paying off.
• Internal frustration centres on access to TPUs, Google's custom AI chips, which researchers say are being sold to outside customers, including rival lab Anthropic.

Google is, by almost any financial measure, winning at AI. Its cloud division, the part of the business that sells AI tools and computing power to other companies, grew revenue by 82% in its latest quarter. Alphabet, Google's parent company, is worth roughly $4 trillion. Sundar Pichai, the CEO, said last month that 90% of Fortune 100 companies are using Gemini Enterprise, Google's AI assistant product aimed at large businesses.

And yet the people who built the intellectual foundation for all of it keep walking out the door.

What actually changed this week?

Two significant leadership moves landed on Wednesday. Jeff Dean, one of the engineers most responsible for Google's early dominance in AI, announced he is leaving after 27 years. Demis Hassabis, who co-founded DeepMind in 2010, sold it to Google in 2014, and ran it ever since, is stepping back from day-to-day management to take a broader research and policy role across all of Alphabet.

Dean is departing alongside three other prominent Google researchers: Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. The four are starting Discovery Loop, a startup backed by Google and structured as a public benefit corporation. Its stated aim is to automate machine learning and scientific research to speed up discovery.

Koray Kavukcuoglu, DeepMind's technology chief, will take over daily management of the division and lead development of the next version of Gemini, Google's flagship AI model.

Why are top researchers leaving?

The exits follow a pattern that goes beyond this week. Noam Shazeer, one of the eight co-authors of the 2017 paper "Attention Is All You Need" (the research that gave modern AI its core architecture, the technology underlying chatbots like ChatGPT and Gemini), left for OpenAI in June 2025. Google had paid close to $3 billion to bring him back just two years earlier. Nobel laureate John Jumper left DeepMind for Anthropic shortly after. All eight authors of that 2017 paper have now left Google.

Analysts and insiders point to two forces. First, internal bureaucracy: getting research into a product at Google requires layers of approval that smaller labs do not have. Second, and more concretely, access to computing power.

Google builds its own AI chips, called TPUs (Tensor Processing Units), specialised processors designed specifically to train and run AI models. Researchers inside the company have grown frustrated watching those chips get allocated to Google Cloud customers, including Anthropic, whose Gemini-competing models benefit from the same hardware Google's own teams are waiting to use, according to people familiar with the matter cited by CNBC Tech.

"Somebody's always going to be unhappy in that situation," said Dan Niles, a Google shareholder and founder of Niles Investment Management.

Pichai has said publicly that DeepMind's research gets first priority on compute, but the internal picture is more complicated. Google models demand years in advance how much chip capacity different teams will need, and priorities shift when a consumer product grows faster than expected.

What does this mean for people who use Google's AI tools?

For most users, nothing changes immediately. Gemini, Google's AI assistant built into Search, Gmail and other products, will continue to be developed under Kavukcuoglu, who has already been directing model releases for the past year.

The longer-term question is whether Google can keep attracting the researchers needed to stay competitive. Investors seem cautiously comfortable: Alphabet shares are up 16% this year after rising 65% in 2024. But the stock did dip after this week's announcements.

One venture capital investor, Tomasz Tunguz of Theory Ventures, summed up the commercial reality bluntly. For most business uses, he said, good-enough AI models are already good enough. The next leap will matter most in highly specialised fields that need serious computing power.

That framing may explain why Google's cloud business is thriving even as its research bench thins. The harder question is whether those two things can stay true at the same time.

Common questions

What is Google DeepMind?

Google DeepMind is the research division of Alphabet that builds Google's most advanced AI systems, including the Gemini family of models. It was formed in 2023 by merging two earlier AI teams, and until this week was led by Demis Hassabis.

Does this affect Google Search or Gmail?

Not directly. The Gemini models powering Google's consumer products will still be developed and updated. Leadership changes at the research level typically take months or years to affect the tools ordinary users see.

What is Discovery Loop, the startup Jeff Dean is starting?

Discovery Loop is a new company, backed by Google, that wants to use AI to automate scientific research and engineering. It is structured as a public benefit corporation, meaning its founding documents prioritise a public mission alongside profit.