Google and Oracle plan to bring Google Tensor Processing Units, or TPUs, to Oracle Cloud Infrastructure starting in calendar year 2027. Google says OCI will be the first hyperscale cloud provider outside Google to offer its AI chips.
The partnership will let OCI customers use TPUs for AI training and inference alongside GPUs. Google says placing TPU capacity within OCI will also let organizations fine-tune models and run inference near data they already keep in Oracle databases.
Google outlines the TPU hardware
Google says its eighth-generation TPUs deliver up to 12.6 PFLOPS of native FP4 compute per chip and up to twice the performance per watt of the previous generation. It says its chip and optical networking design can route around hardware or link faults without restarting a job, sustaining more than 97% productive computation time across thousands of chips in large training runs.
For low-latency serving, Google says changes to on-chip memory, high-bandwidth memory capacity and the interconnect allow the latest TPUs to deliver up to 80% better performance per dollar. These are Google’s performance claims; the announcement does not give a price for TPU capacity on OCI.
Developers can use an open-source PyTorch stack
Google says TorchTPU, its native PyTorch stack for TPUs, is now open source on GitHub. It also names tools including vLLM, SGLang, TorchTitan, Ray and Hugging Face as options for training, serving and deploying models on TPUs.
Within OCI, Oracle Acceleron will connect TPU clusters to Oracle databases and storage, while Google’s interconnect handles traffic between chips. Google and Oracle say they will begin deploying TPU capacity on OCI in calendar year 2027.
