The chips OpenAI is reportedly after are Google’s tensor processing units, or TPUs, processors designed for one job: the math that AI models run on. Google revealed the chip in 2016, saying it had already run TPUs in its data centers for more than a year.
Reports in June 2025 said OpenAI had begun renting Google’s TPUs, and the company described it as early testing with no plans to deploy them at scale. A purchase would mean owning the hardware instead of paying Google by the hour. Other details of the plan have not been made public.
A job listing from OpenAI in San Francisco asks for an engineer to lead the optimization of its frontier models on TPUs so the rumors have weight. The work involves compilers, the software that translates an AI model into instructions a specific chip can run, which is why switching chips is harder than swapping parts.
Richard Ho, who leads OpenAI’s hardware program, spent nearly nine years at Google and was one of the leaders of its Cloud TPU project before joining OpenAI.
In September 2025, Nvidia signed a letter of intent to invest up to $100 billion in OpenAI as it deployed at least 10 gigawatts of Nvidia systems. A gigawatt is a billion watts, the unit data center builders now use to size their projects.
Two weeks later, AMD gave OpenAI the right to buy up to 160 million AMD shares, roughly 10% of the company. The right is tied to OpenAI deploying 6 gigawatts of AMD chips.
Nvidia ended up putting $30 billion into OpenAI’s funding round, well below the $100 billion in the letter of intent. CEO Jensen Huang said in March that it might be the last such investment before OpenAI goes public.
In June, OpenAI unveiled Jalapeño, its first custom chip, built with Broadcom. It is designed for inference, the step where a trained model generates answers, with initial deployment planned for the end of 2026.
Google unveiled two eighth-generation TPUs in April: the 8t for training AI models and the 8i for inference (running them). Its performance figures are measured against Google’s previous-generation chip, Ironwood, not against Nvidia’s hardware.
Anthropic, OpenAI’s main rival, signed an agreement with Google and Broadcom in April for multiple gigawatts of next-generation TPU capacity that it expects to come online starting in 2027.
Alphabet began delivering TPU systems to customers’ own data centers in the second quarter, according to its CFO. Google Cloud also began offering OpenAI’s GPT-6 models to select customers in private preview last week.
