On Wednesday, Noam Shazeer — co-author of the “Attention Is All You Need” paper, VP of engineering at Google, and co-lead of the Gemini project — announced he was leaving the company to join OpenAI. The post on X was gracious. He was proud of the team. It was a difficult decision. He was excited to work with the exceptional people at OpenAI.

The problem, for Google, is not that a star researcher left for a rival. Talent moves. The problem is the price tag.

In 2024, Google paid $2.7 billion in a licensing deal with Character.AI, the chatbot startup Shazeer co-founded. The deal’s structure was curious from the start: it wasn’t an acquisition, technically. Character.AI would continue to exist as an independent entity. Google would license its technology. But the real object of the transaction, according to The Wall Street Journal, was Shazeer himself. His return to Google was considered the primary rationale for the deal. More than half the valuation, in other words, was a signing bonus.

Less than eighteen months later, he’s gone. That works out to roughly $150 million per month of Shazeer’s second tour at Google — an extraordinary figure even by the standards of an industry that has lost its mind over compensation.

The Deal Wasn’t a Deal; It Was a Regulatory Workaround

Why structure a talent acquisition as a licensing agreement in the first place? The answer, as Bloomberg reported last year, is that the U.S. Department of Justice was already investigating whether the arrangement was designed to circumvent antitrust oversight. A straightforward acquisition of Character.AI would have triggered regulatory review. A licensing deal with some talent attached? That’s murkier territory.

The irony is that the workaround created its own vulnerability. When you don’t actually acquire a company — when you license IP from an entity that continues to exist and pay a sum that is functionally a retention bonus for a handful of key people — you have not bought loyalty. You have rented it. And in a market where OpenAI is reportedly headed toward an IPO and the competition for frontier model-builders is a genuine arms race, the renter has almost no leverage when the lease comes up.

A lawyer who has structured several AI talent deals told me, on a phone call from outside the Palo Alto courthouse where a separate noncompete case was being heard, “The problem isn’t that Shazeer left. The problem is that everyone structured these deals to avoid calling them acquisitions, and now they’re discovering that calling something a license means the asset can reprice itself every twelve months.”

The Real Cost of Regulatory Dodgeball

There is a familiar story about Big Tech and regulation, and it goes like this: antitrust scrutiny prevents companies from making sensible acquisitions, which entrenches incumbents and hurts innovation. It’s a story with some truth to it. But the Character.AI episode suggests something messier.

Google didn’t fail to acquire Character.AI because regulators blocked it. Google chose a workaround — one that allowed it to avoid the scrutiny of a formal merger review while still getting what it wanted, which was Shazeer back inside the building. The workaround worked, for a while. Shazeer reportedly contributed meaningfully to Gemini. But the structure meant Google never actually locked in the asset. When OpenAI came calling, there was no golden handcuff clause in a licensing agreement that could stop him.

This is not a victimless miscalculation. Alphabet shareholders effectively funded an eighteen-month sabbatical for one of the world’s most in-demand AI researchers, after which he took his expertise to the company’s most direct competitor. The $2.7 billion wasn’t wasted — Character.AI still holds the license fee, and Google retains whatever IP it extracted — but the primary rationale for the price is now working on ChatGPT’s successor.

What the Market Actually Looks Like in 2026

If there is a broader lesson here, it’s that the AI talent market has evolved past the structures big companies use to contain it. The old model — acquire the startup, vest the founders over four years, hope the cultural antibodies don’t kill them — assumed that the acquired party needed the acquirer more than the acquirer needed any individual. That assumption no longer holds. The people who can build frontier models have more options, more leverage, and less patience for organizational friction than any cohort of technical talent in modern history.

Google is not a dysfunctional company. But it is a large one, and large companies operate through processes that brilliant, idiosyncratic researchers tend to find suffocating. Shazeer left Google once before, in 2021, reportedly frustrated by the company’s reluctance to release the chatbot technology he had helped develop. He came back because Google made it worth his while — literally, to the tune of a multibillion-dollar deal. But money that arrives in a lump sum does not create ongoing reasons to stay.

The $2.7 billion figure will be cited for years as a cautionary tale. Not because Google shouldn’t have tried to bring Shazeer back — he is, by all accounts, one of the most consequential engineers in the field — but because the structure of the attempt guaranteed its expiration date. If you want to keep the talent, you eventually have to acquire the company for real. And if antitrust law won’t let you do that, you have to accept that you’re renting, not buying.

Sources