Google's TPUs are coming to Oracle Cloud in 2027: the architecture goes plural, but the metrics still need checking

At Oracle AI World in Las Vegas on 27 October 2026, Oracle and Google announced in a joint press release that Google's Tensor Processing Units (TPUs) will be available on Oracle Cloud Infrastructure (OCI). The deal breaks new ground: it is the first time these chips will be offered on a hyperscale cloud that competes with Google Cloud. Availability is planned for calendar year 2027, with capacity ramping up gradually over several years. Under the partnership, OCI will supply the data center infrastructure, including power, cooling and networking, so the units can be deployed at scale for training and inference workloads.

The offering covers the eighth generation of accelerators, split into two variants: TPU 8t for training and TPU 8i for inference. According to the two companies' release, TPU 8t would deliver nearly three times the compute performance per pod of the previous generation; this is, however, a vendor claim that has not been independently verified. The release also leaves out operational details such as initial capacity, cluster size and first customers, and neither pricing nor regions have been disclosed. As Unite.AI points out, Google's own product page still lists TPU 8t and 8i as 'coming soon', while the seventh-generation Ironwood is already available.

Bringing in Google's accelerators is part of a broader multi-vendor infrastructure strategy. According to Reuters, as cited by Finimize, Oracle said at the same event that it will scale AI servers built on NVIDIA Vera Rubin systems to 'multi-gigawatt' levels and add chips from Cerebras and SambaNova. The same report says Oracle confirmed capital expenditure of roughly 90-95 billion dollars for fiscal year 2027. Following the announcements, some secondary sources noted Oracle's stock rising by almost 5%.

Proprietary chips moving onto rival infrastructure point to a shift toward AI infrastructure with multiple accelerator suppliers, especially for inference, where moving workloads from one chip to another is easier. Until the real 2027 volumes are known, the deal's actual reach will stay within the bounds of a statement of intent. — Olya