Overview
- A joint announcement on Tuesday from Oracle and Google said Oracle Cloud Infrastructure will offer Google’s TPU 8t for training and TPU 8i for inference beginning in 2027 with a phased, multi‑year rollout.
- The TPU 8t is optimized for model training and the 8i for serving models, and Google markets the 8t as delivering substantially higher compute per pod than prior generations.
- Multiple outlets report that OpenAI plans to buy TPUs directly from Google as part of a broader capacity buildout, but neither company has confirmed final terms, volumes, or delivery timing.
- Oracle said it will continue support for Nvidia and add other accelerators such as Cerebras and SambaNova to OCI, and investors reacted positively to the TPU news with Oracle shares rising after the announcement.
- If executed with clear region coverage, pricing, and tooling, the move could speed a shift to multi‑vendor AI stacks by letting customers mix TPUs and GPUs without moving data between clouds and by reducing single‑vendor risk for large model builders.