OpenAI 据报计划向 Google 采购 TPU

OpenAI plans to buy AI chips from Google, potentially reducing its reliance on NVIDIA as it expands computing capacity to train and run its AI models, according to a report by The Information.

The planned purchase could strengthen Google’s Tensor Processing Unit (TPU) business, which competes with NVIDIA’s graphics processing units (GPUs) for AI workloads. The financial terms and other details of the proposed purchase remain unknown.

OpenAI has been one of NVIDIA’s largest customers, using its GPUs to train AI models and run inference workloads. Nvidia has also invested billions of dollars in OpenAI.

The reported plan comes amid growing commercial ties between OpenAI and Google, which compete in AI products and search. OpenAI recently announced plans to make its frontier AI models available through Google Cloud.

OpenAI has also been hiring an engineer to lead the optimisation of its frontier models on TPUs, suggesting that it is working to adapt its models to Google’s hardware.

Google developed TPUs as specialised processors for AI workloads and has rented them to customers through Google Cloud for years. The company has also begun selling the chips directly to selected customers.

OpenAI began renting Google’s TPUs through Google Cloud in 2025, according to The Information.

OpenAI previously began renting Google’s AI chips last year, according to The Information. A direct purchase would represent a further step in its use of hardware from a company that competes with it in the AI market.

Oracle and Google also recently signed a deal to bring Google’s Tensor Processing Units (TPUs) to OCI from 2027, with capacity expected to increase over several years.

The partnership will allow Oracle customers to use Google’s custom chips for training and running AI models without moving to Google’s cloud.