Oracle to offer Google's AI chips, and OpenAI reportedly wants them too

Google's chips leave home

Oracle and Google announced that Google's TPUs, its custom AI chips, will be added to Oracle Cloud Infrastructure (OCI). Service is expected to begin in 2027. As TechNews reported, this is the first time TPUs will be placed in a large cloud platform that Google does not run itself.

Oracle will supply the power, cooling and data-center networking needed to run the chips at scale. Google will supply the accelerators, which are chips built specifically for AI work.

Two chips for two jobs

Google's eighth-generation TPU comes in two designs. TPU 8t is built for training, the stage where a model learns. TPU 8i is built for inference, meaning running a finished model to answer requests, and it targets AI agents that repeat many such tasks.

Google says a single TPU 8t pod (a large group of linked chips) delivers nearly 3 times the performance of the previous generation. It also says this could cut development of top-tier models from months to weeks. Oracle customers are meant to pick the chip type that fits their needs for performance, throughput and cost.

OpenAI's reported interest

The Information reported on October 28, citing people familiar with the matter, that OpenAI plans to buy Google TPUs to depend less on Nvidia. OpenAI is also recruiting engineers to adapt its most advanced models to run on TPUs.

This has come up before. Last year the same outlet reported that OpenAI had begun using TPUs, and an OpenAI spokesperson said the company had no plans to use them exclusively. The new report is unconfirmed, and no deal terms have been given.