Two star AI exits have turned Google’s Gemini race into a talent-retention test.
Noam Shazeer, a Gemini co-lead and co-author of the paper that introduced the Transformer architecture, has joined OpenAI. John Jumper, a Nobel laureate known for his work on AlphaFold, has left Google DeepMind for Anthropic.
The moves raise fresh questions about whether Google can hold on to the researchers behind its AI ambitions, even as Gemini usage, infrastructure, and enterprise adoption metrics show no visible stall.
Google’s talent risk hits the Gemini race
Shazeer’s move to OpenAI and Jumper's move to Anthropic made Google’s AI talent losses unusually visible in a market where elite researchers can shape recruiting momentum and investor sentiment.
The Shazeer exit carries extra weight because Google previously spent more than $2 billion to license Character.AI technology and bring him and part of his team back. Other reporting put the deal at $2.7 billion.
So far, the departures look more like a roadmap-confidence risk than a measurable Gemini slowdown. Google is fighting OpenAI and Anthropic for a small pool of researchers whose reputations can shape recruiting and market confidence.
Alphabet’s stock reaction showed how quickly investors connected the departures to questions about Google’s AI leadership, with shares extending losses Tuesday after falling 5% Monday.
Talent movement alone is not a reason to abandon Gemini, Vertex AI, or Google Cloud roadmaps. For enterprise teams, the stronger question is whether future Gemini releases keep pace in reasoning, coding, and agentic workflows that could affect vendor selection and long-term platform bets.
Gemini metrics still point to platform strength
Gemini already runs inside Google products with large built-in audiences, including Search, Android, and Workspace.
In Alphabet’s Q1 2025 earnings call, CEO Sundar Pichai said active users in AI Studio and the Gemini API had grown more than 200% since the start of 2025. He also said AI Overviews had more than 1.5 billion monthly users and that all 15 Google products with at least 500 million users used Gemini models.
AI Mode queries were twice as long as traditional Search queries on average, while Workspace delivered more than 2 billion AI assists a month. That distribution strategy is already showing up in productivity tools, including Google’s rollout of Gmail AI summaries to more than 1 billion users.
In Alphabet’s Q4 2024 earnings call, Pichai said Trillium, Google’s sixth-generation TPU, delivered four times the training performance and three times the inference throughput of the previous generation. Vertex AI usage increased 20 times during 2024.
Google said in its Gemini 2.0 launch post that TPUs powered all Gemini 2.0 training and inference, and that Trillium was generally available to customers in December 2024. Its Project Mariner prototype scored 83.5% on the WebVoyager benchmark, but that figure came from Google’s own evaluation and should be read as vendor-reported.
As Gemini moves further into agentic AI, Google DeepMind is also working on security controls for AI agents that could shape how enterprises evaluate autonomous tools.
Google’s advantage is structural, even if the competitive gap is not independently measured: its AI stack spans chips, models, cloud infrastructure, and developer tooling. The next test is whether Gemini gains traction in third-party enterprise systems and keeps pace on independent reasoning and agentic AI benchmarks.
Also read: Berkshire Hathaway’s $10 billion Alphabet bet shows how investors are weighing Google’s AI infrastructure, Gemini strategy, and long-term platform moat.