Google’s Gemini co-lead Noam Shazeer is leaving for OpenAI

Noam Shazeer is leaving Google for OpenAI. The man widely credited as a principal architect of Google’s Gemini models and a co-author of the 2017 transformer paper that underpins virtually every modern large language model announced the move himself, on X: “I’m excited to share that I’ll be joining OpenAI and look forward to working with the exceptional team there.”

The departure is striking for its speed. Shazeer returned to Google less than two years ago in a deal that was itself remarkable: the company reportedly paid around $2.7bn to bring him back, along with a team of his researchers, from Character.AI, the startup he had co-founded after leaving Google the first time.

To lose him again, this quickly, to the company Google most directly competes with, is the kind of talent loss that money was supposed to prevent.

Shazeer is a vice president of engineering at Google and a co-leader of the Gemini effort. As one of the authors of “Attention Is All You Need”, the paper that introduced the transformer architecture, he is among a small group of researchers whose work sits underneath ChatGPT, Claude, Gemini, and most of what the industry now calls AI. His decision to switch sides carries weight beyond a single hire.

OpenAI gains him at a particular moment. The company is widely reported to be heading towards an initial public offering, and adding a researcher of Shazeer’s standing to work on new model architectures is both a research move and a signal to the market about who wants to be there.

For a company whose pitch rests heavily on staying at the frontier of model design, recruiting one of the transformer’s co-inventors is about as direct a statement as it can make.

For Google, the loss lands amid an intensifying contest with OpenAI and Anthropic across the enterprise AI stack. At Cloud Next this year the company rebuilt its AI platform around agents, consolidating products and lining up partners in an explicit bid to out-package its rivals.

The Gemini models are central to that bet, which makes the exit of one of their co-leads more than a routine personnel matter.

The competitive churn around Gemini has been visible elsewhere too. Google has been pushing its models into high-profile deployments, including a Pentagon rollout that scaled from tens of thousands of users to more than a million in six months.

Holding the technical talent that builds those models is part of sustaining that momentum, and Shazeer’s move is a reminder of how mobile the most senior researchers remain even after enormous sums are spent to retain them.

The departure also reframes a debate inside the industry about whether enormous retention packages actually work. Google’s reported $2.7bn move to reacquire Shazeer and his Character.AI team was held up as evidence that the giants would pay almost anything to hold the people who build frontier models.

His exit, so soon after, suggests money buys presence but not permanence, and that the pull of a company seen as the field’s frontier can outweigh a nine-figure reason to stay.

It also feeds a recruitment dynamic that OpenAI has been pressing hard. The company has spent the past year hiring aggressively from rivals while expanding its commercial footprint, including distribution deals that put its models in front of far more enterprise customers. A researcher of Shazeer’s standing is both a capability gain and a morale signal, the kind of hire that tells the rest of the field where the gravity is.

Neither company has detailed the terms of Shazeer’s move, his start date, or precisely what he will work on at OpenAI beyond new architectures. What is on the record is the announcement itself, made by Shazeer, and what it represents: less than two years after a multibillion-dollar effort to bring him home, Google’s Gemini co-lead is walking across the street.