Transformer Co-Inventor Noam Shazeer Left Google for OpenAI

On June 18, Noam Shazeer announced on X that he was leaving Google to join OpenAI. The post was polite, almost gentle: "I'm excited to share that I'll be joining OpenAI and look forward to working with the exceptional team there." Then the knife: "It was a difficult decision to move on. I'm incredibly proud of the amazing team at Google and everything we've built together." Less than two years earlier, Google had paid a reported $2.7 billion in a licensing deal to bring Shazeer back from Character.AI — along with co-founder Daniel De Freitas and a slice of their research team. The most expensive talent reunion in AI history just became the most expensive talent departure.

The hire is not a normal executive shuffle. Shazeer is a co-author of "Attention Is All You Need," the 2017 paper that introduced the transformer architecture now underneath ChatGPT, Gemini, Claude, and virtually every frontier system on the market. At Google he helped build LaMDA, contributed to mixture-of-experts research, and — after returning in August 2024 — became vice president of engineering and co-lead of Gemini at the moment Google was trying to close the gap with OpenAI. Google told Reuters it was "grateful for Noam's meaningful contributions to Google over the years." That is corporate understatement for losing the researcher most central to the technical foundation of your flagship product.

OpenAI CEO Sam Altman did not hide his satisfaction. "noam is one of the people i have most wanted to work with since the very beginning of openai," he posted on X the same morning. "only took 10 years. i think it will be worth the wait!" Trade press reports, citing OpenAI Chief Research Officer Mark Chen, say Shazeer will lead architecture research — the layer where capability per training dollar, inference cost, and the shape of whatever comes after today's transformer stack get decided. OpenAI has not published a formal announcement with his title or start date. What is on the record is the timing: the move landed ten days after OpenAI submitted a confidential S-1 to the SEC.

That IPO context is why this is a markets story, not just a résumé story. On June 8, OpenAI disclosed the filing with characteristic bluntness: "We recently submitted a confidential S-1. We expect it to leak so we're just announcing it." The company added it had "not decided on timing yet" and that some work is "likely easier as a private company" — but the option to go public sooner is now live. Private-market valuations have reportedly pushed above $850 billion. Anthropic filed its own confidential registration days earlier. Wall Street is preparing for a duel of trillion-dollar AI listings. In that environment, hiring the transformer co-inventor is not a research flex. It is a prospectus paragraph.

For Google, the loss cuts deeper than morale. Shazeer left Google the first time in 2021 because, as he has said publicly, the company would not move fast enough on conversational AI — frustration that produced Character.AI and, indirectly, the licensing deal that regulators are still examining. He came back to co-lead Gemini weeks before Google I/O 2026, where the company unveiled Gemini 3.5 Flash and its Spark agent. Now the architect walks out the door again, toward the rival racing to list. Google's playbook — pay billions to reacquire founders through licensing structures that avoid outright acquisition — bought less than two years of loyalty from the one person whose name is synonymous with the field's core invention.

The broader lesson is uncomfortable for anyone treating AI as a procurement problem. Compute can be rented. Models can be licensed. Capital can be raised. But architectural advantage still lives in a tiny pool of people who have already built what the industry is now scaling. Every major lab is in a talent war that looks increasingly like sports free agency with national-security stakes: Meta offering nine-figure packages, OpenAI filing for public markets, Google writing multibillion-dollar reunion checks. Shazeer's move signals that even that price was not enough — and that the next competitive edge may not be a better chatbot but a different architecture entirely. If transformers are the past everyone is monetizing, the person who invented them just bet his next decade on inventing what replaces them — at the company preparing to sell shares on the story.

Convina's view: Shazeer's departure is the clearest proof yet that AI moats are human, not technical. Google did not lose a VP. It lost the author of the blueprint its entire AI business is built on, less than two years after paying a fortune to get him back. OpenAI's hire is strategically brilliant and narratively perfect ahead of an IPO — but it also exposes how fragile "frontier" advantage really is when the frontier is people. Boards should stop asking which model vendor to standardize on and start asking which architectural bets their organization is structurally unable to make. The transformer era may be ending. The war for whoever builds the next one has already started — and the opening move was a resignation letter.