Google Loses a Transformer Pioneer to OpenAI

Noam Shazeer has spent his career building the machinery that everyone else now races to catch. On June 18 he told colleagues he was leaving Google DeepMind, where he co-led the Gemini models, to join OpenAI as its Lead for Architecture Research. The move costs Google one of the few people whose fingerprints sit on nearly every modern language model, and it lands in the middle of a talent war that has turned unusually personal.

Shazeer is a co-author of the 2017 paper "Attention Is All You Need," which introduced the Transformer architecture. That design underpins GPT, Gemini and Claude alike. His new brief at OpenAI is to work on what comes after the current generation of models, drawing on his earlier research into mixture-of-experts systems and attention mechanisms. Sam Altman, in comments reported by CNBC, called the hire "only 10 years in the making" and described Shazeer as someone he had wanted to work with "since the very beginning."

The departure stings partly because of what Google paid to secure him in the first place. Shazeer left the company in 2021 to co-found Character.AI. Google brought him back in August 2024 through a deal for Character.AI valued at roughly $2.7 billion, an arrangement widely read as an expensive way to reinstall him at the head of Gemini. Less than two years later he is walking to a direct rival.

He is not the only one heading for the door. Days before Shazeer's announcement, the Nobel laureate John Jumper, who led the AlphaFold work that won him a share of the 2024 chemistry prize, confirmed he was leaving DeepMind for Anthropic. Surveying the run of exits on June 23, Axios described a lab losing its star power at an awkward time, with Gemini 3.5 Pro still stuck in limited preview while rivals ship aggressively.

What does one researcher's move actually buy OpenAI? Possibly a great deal, possibly very little. Frontier labs have spent two years wringing gains out of sheer scale, and there is a growing sense that the next jump may need a different architecture rather than a bigger one. If anyone is placed to find it, it is the person who helped write the blueprint the whole field has been building on. Whether that turns into a shipped model is a separate question, and the honest answer is that nobody knows yet.

For Google, the loss is as much about signal as substance. Talent follows talent, and a lab that cannot keep the people it spent billions to recruit will find the next recruitment harder. For OpenAI, preparing for a reported public listing later this year, a name of Shazeer's weight is a useful headline as well as a research bet. The part worth watching is what he builds next, and that will take a lot longer than a news cycle to judge.