Google’s chips leave Google’s buildings
Google Cloud and Oracle announced on 27 October that Google’s Tensor Processing Units will be installed in Oracle Cloud Infrastructure, making OCI, in the post’s words, “the first hyperscaler outside Google to offer TPUs”. Deployment of TPU capacity on OCI starts in calendar year 2027.
The post is signed by Karan Batta, senior vice president at Oracle Cloud Infrastructure, and Gurmeet Goindi, senior director of product management at Google. It carries no direct quotations and no commercial terms. It does not say which TPU generation Oracle will run, how many chips, in which regions, at what price, or what the arrangement is worth to either company.
What each side brings
The division of labour is legible from the benefits the post lists rather than stated outright. Google supplies the TPU systems — co-designed silicon, optical networking, liquid cooling and compiler software — along with TorchTPU, its PyTorch-native stack for TPUs, which it has now open-sourced on GitHub alongside support for vLLM, SGLang, llm-d, TorchTitan, Ray, TorchRec, Hugging Face and LiteRT.
Oracle supplies the buildings, the power and the cooling, described as AI-dense data centres built for direct liquid cooling at rack scale, and the networking that joins TPU capacity to everything else in OCI. That networking is Oracle Acceleron: a SmartNIC with hardware-level tenant isolation, a high-performance fabric and zero trust packet routing. TPU-to-TPU traffic stays on Google’s own interconnect and optical circuit switching; Acceleron handles the traffic between the TPU clusters and Oracle’s databases and storage.
The numbers that are in the post
The performance figures are a history lesson rather than a specification. The post notes that the first TPU delivered 92 teraflops and thirty times the throughput per watt of contemporary processors, and that the eighth generation reaches up to 12.6 petaflops of native FP4 compute per chip — more than a hundred times TPU v1 — at up to twice the previous generation’s performance per watt. It cites over 97% goodput across thousands of chips on large training runs, up to 80% better performance per dollar for low-latency serving on the latest generation, and an interconnect topology that cuts network hops by more than half.
None of those figures is attached to the hardware Oracle will actually deploy, because the post never names it. The eighth generation is described as Google’s latest, not as the one going to OCI.
Oracle is collecting accelerators
The TPU agreement is one of several. SiliconANGLE reported that Oracle also intends to deploy Cerebras’s wafer-scale accelerators and OpenAI’s Jalapeño inference processor, attributing those plans to CNBC and noting that neither is confirmed as final. Oracle’s pitch is that customers can pick an architecture without leaving the cloud their enterprise data already sits in — fine-tuning and inference running next to Oracle databases rather than across a cross-cloud transfer.
What to watch
The first thing to look for is a generation and a number. A multi-year arrangement that starts in 2027 has more than a year to acquire both, and until it does, the commercial shape of the deal is unknowable from the outside. The second is whether TorchTPU being open source actually lets teams move workloads between accelerators without rewriting model code, which is the portability argument the whole announcement leans on, and which can only be tested once TPUs on OCI are available.
