Google paid 2.7 billion to bring back a Transformer author. He left for OpenAI 22 months later.
Noam Shazeer posted on X shortly after midnight Pacific on June 18, 2026. He was leaving Google. He was joining OpenAI.
“I’m excited to share that I’ll be joining OpenAI and look forward to working with the exceptional team there,” Shazeer wrote. “It was a difficult decision to move on. I’m incredibly proud of the amazing team at Google and everything we’ve built together.”
The post gathered 16,000 likes within hours. Sam Altman responded publicly the same night.
“Noam is one of the people I have most wanted to work with since the very beginning of OpenAI,” Altman wrote. “Only took 10 years. I think it will be worth the wait.”
If you have used ChatGPT, Gemini, Claude, Grok, or Llama, you have used Shazeer’s work. He co-authored “Attention Is All You Need,” the 2017 paper that introduced the Transformer architecture. Every major large language model is built on it. He authored the Sparsely-Gated Mixture of Experts paper in 2016, the technique behind GPT-4, Gemini 1.5, and models like Kimi K2.7 Code. He designed Multi-Query Attention, the inference optimization that made running large models at scale economically viable. These are not academic citations. They are the engineering primitives the industry runs on.
Shazeer first joined Google around 2000. He stayed 21 years. In 2021, he left to co-found Character.AI. In August 2024, Google paid roughly $2.7 billion in a licensing arrangement to bring him and co-founder Daniel De Freitas back into DeepMind. The deal was structured to avoid the regulatory friction of a full acquisition. The commercial reality was simpler: Google was paying $2.7 billion primarily for Shazeer.
He stayed 22 months.
The math is brutal. Google received roughly 22 months of Shazeer’s time for $2.7 billion in licensing fees. That works out to about $122 million per month. Whether the Gemini improvements delivered during that window justify the cost is a question Google’s leadership and investors are now confronting.
During those 22 months, Shazeer served as VP of engineering and technical co-lead for Gemini. Google closed the gap with OpenAI on multiple benchmarks. Gemini 3 Flash shipped. The Antigravity agentic platform launched. Gemini 3.5 Flash became the default model across all Gemini products. The next major release, Gemini 3.5 Pro, remains unshipped as of his departure.
Google told Reuters it was “grateful for Noam’s meaningful contributions to Google over the years.” A source speaking to Windows News described the internal reaction as “shell-shocked.”
The competitive risk is specific and immediate. Shazeer’s deep knowledge of where Gemini’s architecture is strong and where it faces fundamental limitations now sits inside OpenAI. An architecture lead at Google would know the weaknesses of their own model the same way they know the strengths. OpenAI does not need to copy Gemini. It can design its next generation to be strong exactly where Gemini is weak.
For Google, the timing is painful. Gemini 3.5 Pro, the model Google has been building toward since its I/O announcement in May, is still not released. A next generation codenamed Gemini Nova exists on the roadmap with no public timeline. The architectural contributor who understood the tradeoff space behind both is now across town.
For OpenAI, the hire is a strategic win at a moment when the company has confidentially filed for an IPO. Shazeer’s role, per ThePlanetTools.ai, will be Lead for Architecture Research. His sparse MoE expertise, attention mechanism innovations, and inference efficiency work will shape the generation of models beyond GPT-5.5. The GPT-6 family that determines whether OpenAI enters the public market with a technical lead or faces credible competition from Gemini, Claude, and Grok will carry Shazeer’s fingerprints.
OpenAI has been trying to recruit him for a decade. Altman said so himself. Google outbid them in 2024 with $2.7 billion. Altman waited. Shazeer came anyway.
Sources: CNBC, Neowin, 9to5Google, ThePlanetTools.ai, BuildFastWithAI