Noam Shazeer Leaves Google DeepMind for OpenAI

The AI talent war just produced its highest-profile transfer yet. Noam Shazeer — one of the most influential researchers in modern machine learning — is leaving Google DeepMind to join OpenAI, the two companies confirmed in June 2026. For an industry where a single researcher can shape the direction of an entire field, this is not a routine hire.

To understand why this move matters, you have to understand what Shazeer helped build.

The man behind the Transformer

Shazeer is one of the eight co-authors of the 2017 paper “Attention Is All You Need.” That paper introduced the Transformer, the neural network architecture that underpins essentially every major large language model today — ChatGPT, Claude, Gemini, and the open-weight models you can run yourself. If you want the full picture of how that architecture works, our explainer on what a Transformer is walks through the mechanics.

Before the Transformer, sequence models processed text one step at a time, which made them slow to train and bad at holding long-range context. The Transformer’s attention mechanism let a model weigh every word against every other word in parallel. That single idea unlocked the scale-up that produced the modern large language model. Almost a decade later, the architecture still sits at the foundation of the field — which is exactly why the people who invented it are so sought after.

The research record beyond attention

Reducing Shazeer to “Transformer co-author” actually undersells him. Across two decades at Google he kept producing ideas that became industry defaults:

• Mixture-of-Experts (2017). Months before the Transformer paper, Shazeer was lead author on the “sparsely-gated” Mixture-of-Experts work that showed how to build enormous networks where only a small expert subset activates per input. That design was ahead of its time; today it is the standard architecture for frontier-scale models at nearly every major lab.
• Multi-query attention (2019). A modification to the Transformer’s attention that dramatically cuts the memory traffic needed to generate text. Variants of it now sit inside most production models, quietly making inference affordable.
• Scaling infrastructure and training tricks. From the Mesh-TensorFlow system for splitting giant models across chips to the Adafactor optimizer and contributions to Google’s T5, much of the unglamorous machinery of large-scale training carries his fingerprints.
• Conversational AI. He co-led Meena, Google’s 2020 research chatbot, and worked on the LaMDA line that followed — the direct ancestors of today’s assistant products.

Few researchers have one contribution that defines the field’s direction. Shazeer has had several, spanning architecture, efficiency, and product — the exact combination frontier labs now compete on.

The Character.AI detour — and the $2.7B return

Shazeer’s relationship with Google has broken once before. He left in 2021, reportedly frustrated that the company would not ship the chatbot technology he had helped build, and co-founded Character.AI with fellow Google researcher Daniel De Freitas. The consumer roleplay-chat service became one of the earliest breakout products of the LLM era, reaching a billion-dollar valuation by 2023 on the strength of engagement numbers that rivaled major social apps among young users.

In August 2024, Google brought him back in one of the defining deals of the talent war: a reported $2.7 billion licensing arrangement with Character.AI whose practical effect was returning Shazeer — and a core group of his colleagues — to Google, where he was named a technical co-lead of the Gemini program. The structure of that deal became a template other giants copied: acquire the people and the technology license, leave the company standing.

That return made him a central figure in Google’s push to compete at the top of the model race, including the lineage behind its latest flagship, Gemini 3. His departure for OpenAI is therefore a loss felt at the core of a direct rival — and a striking coda to a deal designed, less than two years ago, specifically to secure him.

Why the talent war keeps escalating

The competition for frontier AI talent among OpenAI, Google, and Anthropic has grown fierce because the work is unusually leverage-heavy. A handful of researchers can determine whether a lab ships a leading model or falls a generation behind. When the supply of people who have actually built foundational systems is this thin, each one becomes strategically valuable in a way that few other industries experience.

The stakes are compounded by the sheer cost of staying competitive. Frontier labs are committing to enormous compute buildouts — OpenAI’s multi-gigawatt partnership with Nvidia is one example of the infrastructure arms race running in parallel with the hiring one. Capital buys the hardware; people decide how to use it. Securing both is now the price of admission at the frontier.

It’s worth being precise about what we don’t know. Neither company has disclosed Shazeer’s exact role or title at OpenAI, and the terms of the move haven’t been made public. What’s clear is the signal: OpenAI has pulled a foundational researcher out of a top rival, and Google has lost one.

What researcher moves have signaled before

If history is a guide, marquee researcher moves tend to precede shifts in lab momentum rather than merely reflect them. OpenAI itself was founded in 2015 around a concentrated hiring push that pulled leading researchers away from academia and Google. The 2021 departure of a senior OpenAI research group produced Anthropic, now one of the three frontier labs. Google’s 2023 merger of Brain and DeepMind was in part a consolidation play to stop internal talent fragmentation. In each case, where the researchers went, frontier results followed within a couple of years.

That pattern is why labs treat individual hires as strategic events, and why a move like this one draws the kind of attention normally reserved for major product launches. The bet OpenAI is making is straightforward: the person who helped invent both the Transformer and the Mixture-of-Experts paradigm may have another architecture-defining idea left, and it is worth a great deal to be the lab where he has it.

The takeaway

Noam Shazeer helped invent the architecture the entire industry is built on — twice over, counting the Mixture-of-Experts designs that now power frontier-scale models — and where he chooses to work is a meaningful indicator of where momentum is gathering. His move to OpenAI underscores how the contest among the leading labs is now fought as much over individual researchers as over chips and capital. In a field this concentrated, talent is infrastructure — and the biggest names move the market.