OpenAI plans to buy Google’s AI chips
OpenAI wants to buy Google's AI chips to rely less on Nvidia as competition over computer hardware grows
According to sources, OpenAI plans to buy AI chips from Google. The goal is to depend less on Nvidia, its main supplier of hardware. The price, number of chips and delivery dates have not been shared. OpenAI is a major customer of Nvidia.
The chips OpenAI wants to buy are Google's tensor processing units, known as TPUs. They are made especially for AI tasks. Google first introduced TPUs in 2016, after using them successfully in its data centres for more than a year.
Reports from June 2025 said OpenAI had started testing Google's TPUs but did not plan to use them on a large scale. If OpenAI buys the chips, it would move from renting hardware to owning the hardware it needs.
OpenAI's planned purchase
An OpenAI job listing for an engineer to make AI models work better on TPUs adds support to reports that the company may buy these chips.
The job focuses on compilers, which help AI models run on specific chips. This shows that changing to different hardware can be complicated.
Richard Ho, who leads OpenAI's hardware work, previously spent nearly 10 years at Google. While there, he played an important role in the Cloud TPU project.
In September 2025, Nvidia said it planned to invest up to $100 billion in OpenAI. It aimed to provide computer systems with a total capacity of at least 10 gigawatts.
Two weeks later, AMD gave OpenAI the option to buy up to 160 million shares of AMD. This deal depended on OpenAI using AMD chips with a total capacity of 6 gigawatts.
In the end, Nvidia invested $30 billion in OpenAI, which was less than it had first planned to invest.
In June, OpenAI announced Jalapeño, its first custom chip made with Broadcom for AI inference, the process of using a trained AI model to produce results. The chip is expected to be put into use by the end of 2026.
Google also introduced two new TPUs in April. They are designed for training AI models and running them to produce results. Google did not directly compare how well they perform with Nvidia's hardware.