Google Cloud is booming by selling the AI chips its own researchers can’t get

Google is having the best and worst month of its AI life at once. Its cloud arm just grew 82% in a quarter, and the researchers who made the company an AI pioneer keep leaving. The two facts are connected, and the link is compute.

Some Google researchers cannot get the chips they need for ambitious projects, and they have grown frustrated, CNBC reported. Google Cloud, meanwhile, sells those same chips to outside customers, including Anthropic, whose models compete directly with Gemini. The chips, Google’s home-grown TPUs, are the scarce resource everyone is fighting over.

Selling the picks and shovels

You can watch the strategy in the open. This week Google Cloud announced that Mirendil, a frontier AI lab, will train its models on Google’s TPUs and Nvidia chips. Nearly every big AI lab now rents Google’s infrastructure. It is a lucrative business, and a far surer one than building the best model yourself.

That is the choice facing a $4 trillion company. Frontier models cost a fortune to train, with no guaranteed payoff. Selling the compute is highly profitable, and growing faster than Amazon’s or Microsoft’s clouds. So the question, four years into the AI boom, is blunt. Does Google need to build the best model, or is it better off letting rivals foot that bill?

Plenty of insiders lean towards the second. If most enterprise work runs fine on cheaper, good-enough models, the frontier looks less like a prize than a cost. “The models are good enough for 90% of what needs to get done,” said Dan Niles, a Google shareholder. “You don’t need a Ferrari for this stuff. A Ford will work.”

The people who built it are leaving

The researchers do not see it that way. Chief scientist Jeff Dean is leaving after 27 years, taking three senior colleagues to found a startup, Discovery Loop. With them goes the last of the eight authors of Google’s 2017 transformer paper, the work that started the whole generative-AI era. Every one has now left.

“They’re not interested in commercialising AI,” said Gil Luria of D.A. Davidson. They are interested in being part of history, he added, and they see that history being made at Anthropic and OpenAI. Google’s layers of approval, the ones that turn research into products, look like a cage to people who would rather do the research.

Google disputes the idea that it starves its own labs. Sundar Pichai says securing frontier compute is the company’s “first priority”, the foundation for everything it does. Both things can be true. Google can put the frontier first and still find that, when the cloud is this profitable, the researcher chasing the next breakthrough is not always first in the queue.

So that researcher leaves, sometimes to a lab that then rents Google’s chips. Even Mirendil is chasing the self-improving research loop that Dean’s new company wants to build. Google may end up selling the shovels to the very people it could not keep.