OpenAI plans to purchase Google's tensor processing units, or TPUs, according to a person familiar with the matter cited by The Information. Neither company has confirmed the plan, and the price, chip count and delivery schedule remain undisclosed. The purchase would reduce OpenAI's dependence on Nvidia, which supplies most of the hardware supporting the AI boom. OpenAI ranks among Nvidia's largest customers.
Google's TPUs specialize in the mathematical operations used by AI models. Google introduced TPUs in 2016 and said Google had already used the processors in data centers for over a year. Reports published in June 2025 said OpenAI had started renting Google's TPUs. OpenAI characterized the activity as preliminary testing and said OpenAI had no plans for large-scale deployment. Buying TPUs would give OpenAI ownership of the hardware rather than access paid for by the hour.
OpenAI's San Francisco job advertisement seeks an engineer to oversee optimization of frontier models for TPUs, lending support to the reported plan. The position involves compilers, which convert AI models into instructions compatible with particular chips and make changing hardware more complicated than replacing components. Richard Ho, the head of OpenAI's hardware program, worked at Google for nearly nine years and helped lead Google's Cloud TPU project.
Nvidia signed a letter of intent in September 2025 to invest up to $100 billion in OpenAI alongside OpenAI's deployment of at least 10 gigawatts of Nvidia systems. A gigawatt equals a billion watts and serves as a measure of project size for data center builders. AMD granted OpenAI the right to purchase up to 160 million AMD shares, approximately 10% of AMD, two weeks later. The purchase right depends on OpenAI deploying 6 gigawatts of AMD chips. Nvidia ultimately invested $30 billion in OpenAI's funding round, substantially below the letter of intent's $100 billion. Nvidia CEO Jensen Huang said in March that the investment might be Nvidia's last such investment before OpenAI becomes publicly traded.
OpenAI introduced Jalapeño in June as OpenAI's first custom chip, developed with Broadcom. Jalapeño targets inference, the process through which trained models produce answers, and OpenAI plans initial deployment by the end of 2026. Google introduced two eighth-generation TPUs in April, with the 8t intended for model training and the 8i for inference. Google's performance comparisons use the previous-generation Ironwood chip as the benchmark rather than Nvidia hardware.
Anthropic, OpenAI's principal competitor, reached an agreement with Google and Broadcom in April for multiple gigawatts of next-generation TPU capacity. Anthropic expects that capacity to begin operating in 2027. Alphabet's CFO said Alphabet started supplying TPU systems for installation in customers' own data centers during the second quarter. Alphabet's CFO expects 2027 to account for most revenue from TPU system sales. Alphabet will release third-quarter earnings after U.S. markets close on Wednesday. Google Cloud started giving selected customers private-preview access to OpenAI's GPT-6 models last week.