Google paid approximately $2.7 billion in August 2024 to bring Noam Shazeer back to the company. Shazeer had left Google in 2021 to found Character.AI; the deal returned him as Vice President of Engineering and co-lead of the Gemini project. Less than two years later, on June 18, 2026, Shazeer announced he was leaving Google again, this time for OpenAI. He's not just any executive making a career move: Shazeer is co-author of the Transformer paper, on which nearly every major language model today is built.
The man behind the architecture
In June 2017, a team of eight Google researchers published the paper "Attention Is All You Need." It described the Transformer architecture, which has since become the foundation of all major language models: GPT, Gemini, Claude, LLaMA and their successors. Among the authors: Noam Shazeer. Without this paper, generative AI as we know it today would not exist.
Shazeer also worked on two other core technologies of modern AI systems. Mixture-of-Experts architectures (MoE) activate only specialized sub-networks within a model rather than the entire system when responding to a query. This saves computational resources and makes very large models economically feasible. Multi-Query Attention accelerates inference, the speed at which a model generates responses. Both concepts are embedded in Gemini and GPT-4o today.
After leaving in 2021, Shazeer built Character.AI, a chatbot service for roleplay-based conversations. During its growth phase, the platform attracted over 20 million monthly active users. Google viewed this as both a threat and an opportunity. The acquisition of Character.AI in August 2024 for approximately $2.7 billion is considered one of the most expensive talent deals in AI history. Shazeer is believed to have personally received between $750 million and $1 billion from the sale.
Shazeer's role at OpenAI
At OpenAI, Shazeer will lead architecture research. This sounds like a specialized staff position, but it's central to AI competition. A model's architecture determines how efficiently it learns, how scalable it is, and what fundamental capabilities it can develop. Whoever makes the next breakthrough here can achieve advantages that can't be overcome by computing power alone.
OpenAI is reportedly already working on a new generation of architecture beyond classical Transformer design. Shazeer will explore how the underlying structure can be evolved or supplemented by newer approaches like State-Space models. Reuters called the move one of the most significant talent departures in the AI industry in 2026, describing it as "surprising" given that Gemini had just made significant progress under Shazeer's leadership.
For Google, the departure is a reputation problem. The company spent heavily to bring him back, and now he's leaving for a competitor. Gemini's Transformer core is not at risk: hundreds of thousands of hours of training work and a large team remain. But the architect's departure signals discontent at a time when Google is gaining ground in the AI market.
Gemini's rise and the Apple deal
Yet Gemini is performing better than ever. According to analysis from data service Vertu and market research firm First Page Sage, ChatGPT's global market share of AI chatbot traffic has fallen from around 87 percent to 68 percent; Gemini rose from five to approximately 18 percent in the same period. ChatGPT fell below the 50 percent mark for the first time in May 2026 among monthly users.
In July 2026, Google Cloud confirmed that a custom Gemini model with 1.2 trillion parameters will power Apple's Siri function and Apple Foundation Models beginning with iOS 27. This deal, which would establish Gemini as the default AI for approximately 1.4 billion iPhone users, is independent of Shazeer's departure. It was negotiated over several years and relies on the existing product team. Shortly after Shazeer's departure, Google also released Frozen v2, a version in which Gemini is integrated directly into specialized chips.
Before the IPO: Who benefits from the move
Why would Shazeer leave Google after less than two years? The answer lies partly in OpenAI's capital plans. The company is pursuing an IPO that analysts project at a valuation around one trillion dollars. Anthropic plans its own IPO for October 2026; OpenAI is waiting for an even higher valuation. Anyone given company shares before an IPO at such valuations can profit significantly.
For Shazeer, already wealthy from the Character.AI sale, financial considerations likely play a role, but not the only one. As an architect wanting to work on the next generation of models, OpenAI at this moment is a different platform than Google: freer in research, with clear focus on frontier models and without the internal coordination constraints of a technology conglomerate with dozens of product divisions. For a researcher building architectures, that can be the decisive difference.