OpenAI is reportedly using Google's home-developed TPU chips to power ChatGPT and its other products.

According to The Information, and citing a person involved in the arrangement, this is OpenAI's first meaningful use of non-Nvidia chips.

Reports that OpenAI had signed an agreement to use Google Cloud's computing capacity first emerged last month, despite the AI company being in direct competition with Google's DeepMind AI unit, and ChatGPT posing a threat to Google's search business.

According to The Information, OpenAI hopes that, by using TPUs rented through Google Cloud, it will reduce the cost of inference. Google is not renting its most powerful TPUs to OpenAI, however, said a Google Cloud employee.

DCD has reached out to OpenAI and Google Cloud for comment.

OpenAI uses Nvidia chips for both training and inference workloads.

The contract with Google is part of a move from OpenAI to diversify its compute resources, with Microsoft previously being the company's exclusive cloud provider. OpenAI is also developing data centers through its Stargate project, and is a customer of Oracle and CoreWeave.

Google's TPUs - or Tensor Processing Units - are chips optimized for AI workloads. The latest generation, dubbed Ironwood, was unveiled in April of this year.

Ironwood also offers 192GB of 7.4Tbps of high-bandwidth memory (HBM) per chip, a 6x increase on Google’s sixth-generation Trillium TPU, in addition to 1.2Tbps bidirectional Inter-Chip Interconnect (ICI) bandwidth, a 1.5x improvement over Trillium.

Other known customers of Google's TPUs include Apple, Anthropic, and Safe Superintelligence.