Oracle and Google Cloud plan to begin deploying Google Tensor Processing Unit accelerators on Oracle Cloud Infrastructure in 2027.

Google Cloud and Oracle announced plans to deploy Tensor Processing Unit (TPU) capacity on OCI beginning in calendar 2027, making OCI the first hyperscaler outside Google to offer TPUs. A supplementary report describes multiyear scaling for model training, inference, and agentic applications, and identifies eighth-generation TPU 8t for training and TPU 8i for inference.

Google’s announcement describes eighth-generation TPUs as delivering up to 12.6 petaflops of native four-bit floating-point (FP4) computation per chip and up to twice the performance per watt of the previous generation. For historical context, it cites TPU v1 at 92 teraflops and 30 times the throughput per watt of contemporary processors; these are Google’s stated comparisons, not independent OCI benchmarks.

The architecture combines on-die Inter-Chip Interconnect (ICI) routing with Optical Circuit Switching (OCS) to reroute around degraded hardware or links without restarting jobs. Google reports more than 97% goodput, defined as the share of time spent on productive computation, across training runs involving thousands of chips. It also cites more on-chip static random-access memory, more HBM, a topology reducing network hops by over 50%, and up to 80% better performance per dollar for low-latency serving.

Google says its native PyTorch stack, TorchTPU, is now open source, with support for tools including vLLM, SGLang, llm-d, TorchTitan, Ray, TorchRec, Hugging Face, and LiteRT. The OCI announcement does not disclose committed chip quantities, sites, customer pricing, or procurement value, and the cited hardware performance figures should not be treated as measurements of an already deployed OCI service.