OpenAI is reported to be planning to buy AI chips from Google, a step that would loosen its dependence on Nvidia. The report comes from The Information, which cited one person familiar with the matter on 27 and 28 October 2026. The chips would be Google's tensor processing units, known as TPUs. Neither company has confirmed the plan, and OpenAI declined to comment to Reuters.
Buying would be a bigger commitment than what OpenAI does today. It began renting Google TPUs in June 2025, at a time described as early testing with no large-scale deployment planned, whereas a purchase would mean owning the hardware. Price, chip count and delivery dates have not been disclosed. Switching chips is also real engineering work, since compilers tie models to particular hardware, and OpenAI has advertised a San Francisco job for an engineer to tune its frontier models for TPUs.
The OpenAI logo and a keyboard in an illustration made on 19 October 2026
REUTERS/File · Image: The Financial Express (Reuters)
Google has been pushing TPUs outward. It unveiled eighth-generation versions in April, one for training and one for inference, and began delivering systems into customers' own data centres in the second quarter. Alphabet's finance chief has said revenue from TPU deals would be small by the end of 2026 and mostly arrive in 2027. Anthropic signed in April for multiple gigawatts of TPU capacity from 2027.
About OpenAI
OpenAI's search for chips beyond Nvidia's started well before this report. In 2024 it began working with Broadcom on a custom AI chip for training and inference, meant to cut its dependence on costly Nvidia GPUs. In January 2025 it announced the Stargate Project, a joint venture with Oracle, SoftBank and MGX to build AI infrastructure with the US government.
In March 2025 it struck a deal with CoreWeave that involved $350m of CoreWeave shares and access to AI infrastructure, in return for $11.9bn paid over five years. Each of these moves aimed at securing the computing power its models need, and the company has repeatedly sought it from more than one supplier.
11
Works credited
27
Models indexed
Models on AIPopDB
OpenAI is spreading its bets in other directions too. Nvidia's planned investment of up to $100bn ended up as $30bn in the latest funding round, AMD has given it a right to buy shares tied to deploying 6 GW of its chips, and its own Broadcom-built inference chip, Jalapeno, is due its first deployment by the end of 2026. The week before the report, OpenAI said its frontier models would be offered through Google Cloud. Alphabet reported third-quarter earnings after the US close on 28 October, with investors watching cloud growth.