Oracle will offer Google TPU chips in its cloud from 2027

Oracle and Google Cloud tell that they will move Google Tensor Processing Units, or TPUs, to Oracle Cloud Infrastructure. TPUs are AI chips that Google designs. Oracle Cloud Infrastructure is the cloud system of Oracle, and Oracle databases frequently hold company data. Oracle will start TPU capacity in calendar year 2027. The post gives no price and no quantity of capacity.

What Oracle offers

Oracle gives four benefits:

• Data centers for the liquid cooling that TPU racks must have.
• Oracle Acceleron networking that connects TPUs to databases and storage.
• TPUs near Oracle databases, with no costly movement of data between clouds.
• TPUs and GPUs in one cloud, to select the chip for each task.

The numbers

These figures are from Google. Nobody tested them independently.

• The eighth generation TPUs give a maximum of 80% better performance for each dollar for low latency serving.
• Large training runs keep more than 97% goodput, which is the share of time that a cluster does useful work.

What is not known

The post gives no price, no quantity of capacity and no customer names. TorchTPU, a PyTorch stack for TPUs, is open source on GitHub.

Insight: Google is moving TPUs into rival clouds

On 21 October 2026, Google and Meta released TorchTPU as open source, and PyTorch models operate on TPUs. Google Cloud also opens NVIDIA Vera Rubin A5X servers next month. Google seems to sell its chips in many clouds and to keep the software open. Then TPUs compete with NVIDIA chips. Moving a model between chip types can get cheaper for you.