Oracle Cloud Infrastructure is adding Google TPUs to its offering. It is the first time Google has offered its own AI chips on a hyperscale cloud platform outside the company itself.
Oracle and Google have announced that Google’s Tensor Processing Units (TPUs) will become available through Oracle Cloud Infrastructure (OCI). The announcement was made at the Oracle AI World trade show in Las Vegas. Oracle is adding the TPUs to its existing range of compute for AI. This lets customers choose the hardware that best suits each workload.
Role of both companies
Google develops the TPUs as its own accelerators, together with open software. Oracle provides the power supply, cooling and networking in its data centres to run the chips at scale. Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, says different AI workloads have different needs. According to him, customers get access to multiple architectures in a single cloud and can choose based on performance, throughput and cost.
Amin Vahdat, senior vice president and chief technologist AI & Infrastructure at Google, expects the next generation of AI to need a wide range of compute. With OCI, more organisations gain access to the technology.
What the TPUs can do
According to Google, the TPUs are designed to get as much AI processing as possible out of every watt of power. The chips support the training of large models, including dense models and mixture-of-experts (MoE) models, and inference for reasoning AI, AI agents and long-context applications.
The eighth generation consists of two architectures: the TPU 8t for training and the TPU 8i for inference. Google says the TPU 8t cuts the development time of frontier models from months to weeks. Per pod, it delivers almost three times more compute than the previous generation. The TPU 8i is intended for the iterative work of many specialised AI agents.
The TPUs will become available on OCI in calendar year 2027. After that, Oracle will expand the offering over several years, depending on customer demand.