Summary
On October 27, 2026, Oracle and Google jointly announced at Oracle AI World in Las Vegas that Google TPUs will be offered through Oracle Cloud Infrastructure (OCI), making OCI the first hyperscaler outside Google to provide TPUs; availability is planned for calendar year 2027.
Background and Positioning
Why is Oracle bringing Google TPUs into OCI?
Per Oracle's official press release, OCI will be the first hyperscaler outside Google to offer TPUs. Oracle and Google made the joint announcement at Oracle AI World in Las Vegas on October 27, 2026. OCI President Mahesh Thiagarajan officially said the move gives customers multiple mainstream AI compute architectures within the same cloud, rather than forcing fragmented deployments across chip ecosystems.
Does this conflict with Oracle's simultaneous Nvidia expansion?
No — Oracle officially positions OCI as a cloud where multiple compute architectures coexist. On the same day, Oracle also announced expanding Nvidia Vera Rubin NVL72 OCI Superclusters to multi-gigawatt scale. Per Oracle's press release, TPUs and Nvidia GPUs target different preferred workloads on OCI, and customers can pick the architecture per model and use case instead of choosing one over the other.
Availability and Specifications
When will Google TPUs be available on OCI?
Per Oracle's press release, Google TPU on OCI will be available in calendar year 2027 and ramp up over multiple years. The exact launch month, initial regions and supply volume have not been disclosed.
Which TPU generation is offered, and how do training and inference variants differ?
Per Oracle's press release, OCI will offer Google's eighth-generation TPUs in two architectures: TPU 8t for training and TPU 8i for inference. The two are optimized for large-model training and online inference respectively, so customers can choose the matching variant by task type.
How much faster is TPU 8t than the previous generation?
Per Oracle's press release citing Google TPU specifications, TPU 8t delivers nearly 3x the compute per pod versus the previous generation. Oracle says this generation can compress frontier-model development cycles from months to weeks.
Customers and Workloads
Which customers and workloads fit TPU on OCI?
Per Oracle's press release, supported workloads include training and post-training of dense and MoE large models, inference, and long-context and agentic applications. OCI will provide power, cooling and data-center networking to support large-scale TPU deployment; AI labs and enterprises that need multi-architecture options while keeping compute in one cloud are the primary target.
How does OCI secure power, cooling and networking for large-scale TPU deployment?
Per Oracle's press release, OCI handles the overall supply of power, cooling and data-center networking. Google provides the TPU hardware and software stack, while OCI provides data-center-level infrastructure; specific power capacity and facility configurations have not been disclosed.