Noam Shazeer, Google DeepMind’s vice president of engineering and co-lead of the Gemini model family, announced Wednesday that he is leaving Google to join OpenAI. The departure strips Google of one of its most consequential AI researchers and hands OpenAI the co-inventor of the transformer architecture that underpins every modern large language model and autonomous agent system.
“I’m excited to share that I’ll be joining OpenAI and look forward to working with the exceptional team there,” Shazeer wrote on X. “It was a difficult decision to move on. I’m incredibly proud of the amazing team at Google and everything we’ve built together.”
The $2.7 Billion Return That Lasted 22 Months
Shazeer’s departure is particularly costly for Google because of the price it paid to get him back. In August 2024, Google brought Shazeer and fellow researcher Daniel De Freitas back to its DeepMind unit as part of a $2.7 billion partnership with Character.AI, the chatbot startup the pair had co-founded after leaving Google in 2021. They had originally left after Google declined to aggressively pursue a chatbot project they championed.
Less than two years later, Shazeer is leaving again, this time for Google’s primary competitor in frontier model development.
Why the Timing Matters for Agent Development
Shazeer is not just any researcher. He is a co-author of “Attention Is All You Need,” the 2017 paper that introduced the transformer architecture. That paper is the technical foundation for GPT, Gemini, Claude, and every agent-capable large language model shipping today. His expertise spans model scaling, inference optimization, and the architectural decisions that determine whether a model can reliably execute multi-step agentic workflows.
The move comes weeks after Google unveiled Gemini 3.5 Flash and the Gemini Spark agent at its I/O developer conference, and days after OpenAI confidentially filed for its IPO, setting up what could be the most closely watched technology listing in years.
The Agent Infrastructure Talent Race
For OpenAI, the acquisition is a statement about where it sees its competitive advantage: not just in model capabilities, but in the architectural talent required to build models that function as reliable autonomous agents. Shazeer’s background in scaling models and optimizing inference pipelines is directly relevant to the challenge every agent platform faces, getting models to execute complex, multi-step tasks without degrading in quality or speed.
For Google, losing Shazeer for a second time raises questions about DeepMind’s ability to retain the researchers who built its core model infrastructure, particularly as the agent development race accelerates across the industry.