Noam Shazeer, Google vice president of engineering and co-lead of the Gemini model, has left the company and announced on June 18 that he is joining OpenAI. In a post on X, Shazeer said he was excited to join OpenAI and work with its team, adding that leaving Google was a difficult decision and that he was deeply proud of what had been built there.
A key architect behind modern attention models
Shazeer is widely known in AI for proposing Multi-Head Attention in 2014, a technique that later became a core building block of the Transformer architecture. That architecture underpins leading large language models including GPT, Gemini, and Claude. His departure stands out because his work sits close to the technical foundations of today’s generative AI systems.
He joined Google Brain in 2015 and went on to contribute to Google’s first-generation Transformer model, the PaLI multimodal model, and the Gemini family. Within Google, he had been seen as one of the people central to the company’s AI competitiveness, especially as Gemini became the group’s flagship model effort.
OpenAI keeps adding senior talent
The move comes as competition for AI staff keeps intensifying. According to a June 17 report by The Information, OpenAI and Anthropic have hired more than 85 employees away from Salesforce this year, with about 40 joining OpenAI and nearly 45 going to Anthropic. The figures point to an aggressive hiring cycle across the top tier of AI firms.
The source material also says OpenAI’s audited financial data for last year showed $34 billion in total spending, including roughly $19 billion devoted to research and development. At the same time, the company is pushing ahead with listing plans, with its overall valuation said to have the potential to exceed $1 trillion. In that setting, high-level research hires carry unusual weight.
Gemini’s multimodal direction now passes to a new team
The report notes that Shazeer also played a major part inside Google in advancing a “multimodal prior” strategy. Gemini was designed from the start as a natively multimodal model rather than a system assembled by stitching together separate single-stream models. With his exit, that approach will now be carried forward by a new team.
For Google, the change affects more than leadership structure; it also touches how resources are assigned for the next stage of model development. For OpenAI, bringing in a researcher tied to the early Transformer path shows how hard it is still pushing on top-tier talent. In AI, compute and capital matter, but the movement of researchers is shaping the field in real time.