Google is reporting record cloud momentum while losing some of the people who helped build modern AI. CNBC said the company is facing one of the clearest contradictions of the current AI cycle: its commercial business is expanding, but research and scientific talent continues to walk out the door.
A wave of departures has hit Google AI
Google’s AI organization has gone through a fresh round of exits over the past two weeks. Chief scientist Jeff Dean said he was leaving after 27 years at the company. He later co-founded AI startup Discovery Loop with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. The company is structured as a public-benefit corporation and says it aims to use automated machine learning to speed up breakthroughs in science and engineering.
Shortly after that, Google DeepMind co-founder Demis Hassabis stepped down as chief executive and moved into the role of chairman. He also became Alphabet’s newly created chief scientist, where his focus will be long-term research and the social impact of AGI, or artificial general intelligence.
The report also highlighted the 2017 paper Attention Is All You Need, widely seen as a foundation of modern AI. All eight co-authors have now left Google. Noam Shazeer joined OpenAI in June, less than two years after Google spent nearly $3 billion to bring him back. John Jumper, the Nobel Prize winner and a central figure behind AlphaFold, moved to Anthropic. Taken together, the author list now reads like a map of Google’s AI talent drain.
TPU access has become a core point of tension
CNBC traced much of the frustration to one issue: compute allocation.
Google’s Tensor Processing Units, or TPUs, are a key resource for training and inference on large AI models, and they compete directly with Nvidia GPUs. But those chips are limited, and Google has to divide them across internal research teams, internal products such as Search and Gemini, and outside cloud customers.
That last category includes Anthropic, which competes directly with Gemini. According to people cited in the report, some researchers were frustrated to see Google Cloud selling TPU capacity to Anthropic while their own research requests were stuck in internal approval channels. The gap between those priorities became a source of resentment that compensation alone could not fix.
Alphabet Chief Executive Sundar Pichai said on an earnings call that AGI research is the “number one priority” for compute allocation. The report said accounts from inside the company suggest a noticeable gap between that statement and what some researchers are experiencing.
Dan Niles, founder of Niles Investment Management and a Google shareholder, put it bluntly: “Google has too many businesses to manage. You have to decide who gets the resources, and in that situation someone is always unhappy.”
Bureaucracy has made startups more attractive
Compute is only part of the story. The report said Google’s internal bureaucracy is another major reason top researchers are willing to leave.
Inside Google, turning research into product often means passing through several layers of review and approval. In a fast-moving AI race, that kind of process can become a disadvantage by itself. OpenAI, Anthropic, and newer AI startups offer a different setup. Research can move directly into products, and lab work is under less pressure to match quarterly financial targets.
Gil Luria, an analyst at D.A. Davidson, said: “They’re not interested in commercialization. They want to be part of history. In their eyes, Anthropic and OpenAI are the places where history will be written.”
For a company valued at $4 trillion, that is a difficult problem to solve. Google’s scale makes it more bureaucratic, more cautious, and more tied to measurable commercial returns than younger rivals.
Strong results show Google’s market thesis is still working
Even with the talent losses, Google Cloud posted strong numbers. Second-quarter revenue rose 82% year over year, outpacing Amazon Web Services and Microsoft Azure. The company also said 90% of the Fortune 100 are using Gemini Enterprise.
Alphabet’s stock has climbed 16% so far this year after rising 65% last year, giving it one of the strongest performances among major technology companies.
Those results also point to Google’s reading of customer demand. Theory Ventures founder Tomasz Tunguz said that for most white-collar work, a model only needs to be good enough. Niles made the same point more sharply: “Ninety percent of use cases don’t need a Ferrari. A Ford is enough.”
Google’s Gemini Flash model reflects that approach. The company positions it as a model built for top efficiency and speed, using performance-per-cost and responsiveness to win broader adoption.
A strategic choice with no easy answer
Google now faces a broader decision. It can keep spending heavily to compete for the frontier of AI research, or it can lean into a different identity as one of the biggest infrastructure and enterprise-service platforms of the AI era.
The departure of all eight Attention Is All You Need authors has come to symbolize deeper dissatisfaction with Google’s commercial priorities and internal culture. The company is now at a point where balancing talent retention with corporate positioning has become a central challenge.