OpenAI Turns to Google's Chips, and Away From Nvidia

OpenAI is preparing to buy artificial-intelligence chips from Google, according to a report from The Information, a step that would loosen its heavy reliance on Nvidia and hand a rival one of the industry's most coveted customers. Reuters, Bloomberg and others picked up the story on 27 October. Neither OpenAI nor Google has confirmed it, and the terms, including how many chips and at what price, are not public.

The chips in question are Google's tensor processing units, or TPUs, custom silicon the company designed for its own AI work and has lately begun renting and selling to outsiders. OpenAI is no stranger to them. Reuters reported last year that the company had started renting Google's chips to help run ChatGPT. Buying TPUs outright, and deploying them at scale, would be a larger commitment, and a signal that OpenAI wants more than a single source for the hardware its models depend on.

For Nvidia, whose graphics processors have been the default engine of the AI boom, OpenAI is about as important a customer as exists. Any shift of orders toward Google dents that relationship, even modestly. It also fits a wider pattern. OpenAI's models recently started appearing on Google Cloud once Microsoft's period of exclusivity ended, and Oracle said on 27 October that it would become the first big cloud provider outside Google to offer TPUs to its own customers.

The economics help explain the manoeuvring. Amazon, Microsoft and Alphabet together plan to spend as much as 600 billion dollars on AI infrastructure this year, and the firms building frontier models are hunting for cheaper compute and insurance against any one supplier. Google, which has spent a decade refining its TPUs in-house, is suddenly placed to sell the shovels as well as dig.

Whether the reported deal closes, and on what scale, will say a good deal about how far the industry's dependence on Nvidia can be unwound. For now it is a report, not a contract.