Oracle and Google say TPUs are coming to Oracle Cloud in 2027

The figure is Oracle and Google's, not an independent benchmark. In their October 27 announcement, the companies said Google's eighth-generation TPU 8t, built for training, delivers nearly three times the compute performance per pod over the previous generation, and cuts frontier-model development cycles from months to weeks. Google published the same figure on its own blog when it introduced the eighth generation in April 2026, and its technical material compares against the seventh-generation Ironwood TPU. The Oracle release gives no benchmark name, workload, or outside test behind the figure.

The figure is stated per pod, not per chip or per job, and neither company's announcement says how it was measured on any given workload. When checked for this briefing, Google's TPU page still listed TPU 8t and TPU 8i as “coming soon,” so the chips are not yet generally available even on Google Cloud.

The second architecture, TPU 8i, is built for inference, the work of running a trained model to answer requests. Oracle and Google describe it as designed for the iterative work of many specialized agents, the back-and-forth of agentic applications rather than one big training run. Oracle's release says Google TPUs can support large-model training and post-training, including dense and mixture-of-experts models, and production inference for reasoning, agentic, and long-context applications.

Oracle will provide the power, cooling, and datacenter networking to run the TPUs at scale inside its own infrastructure. Customers will be able to choose Google TPUs within Oracle Cloud Infrastructure based on workload, performance requirements, and cost objectives, the companies said.

Oracle Cloud Infrastructure will be the first hyperscale cloud platform outside Google to offer Google TPUs, the companies said, announcing the deal at Oracle AI World in Las Vegas on October 27. Google has built custom TPU silicon for more than a decade, and the companies pitch the portfolio as purpose-built compute for the full AI lifecycle, from frontier training to large-scale inference. The qualifier is “hyperscale”: DCD, citing The Information, reported that Google has also arranged TPU deployments with the smaller provider Fluidstack, so the claim is specific to hyperscale clouds.

Nothing is available yet. Google TPUs on OCI will arrive in calendar year 2027 and ramp over a multi-year period, the release says, with no month, region, or price named. Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, said different workloads have different compute needs and that bringing TPUs to OCI gives customers access to multiple leading AI compute architectures in one cloud. Amin Vahdat, Google's senior vice president and chief technologist for AI and infrastructure, said the next generation of AI will require a diverse range of compute.

Reuters reported the same day that Oracle and Nvidia said Oracle will expand to “multi-gigawatt scale” with Nvidia's newest Vera Rubin AI servers, with no deal value disclosed, and that Oracle will use its own Acceleron networking chips in place of some Nvidia networking chips. Oracle also plans to offer inference chips from Cerebras Systems and SambaNova Systems, Reuters reported; Cerebras chief executive Andrew Feldman told Reuters the Oracle deal is “a massive expansion of our customer base of tier-one customers.”

What the announcement leaves open: pricing for TPUs on OCI, the exact 2027 arrival date, which regions get them first, and which TPU generations will be offered. The 3x figure is the companies' own claim about the hardware, not a result measured on Oracle's cloud.