Oracle to Offer Google TPUs on OCI

Oracle announced on the second day of Oracle AI World 2026 that it plans to offer Google’s Tensor Processing Units (TPUs) through Oracle Cloud Infrastructure (OCI). It would be the first time a hyperscaler other than Google offers Google TPUs as a service. OCI’s TPU service is scheduled to launch in 2027 and expand gradually over several years to meet customer demand.

Google TPUs have so far been available only through Google Cloud. Google said this year that it would supply TPUs to outside business customers for the first time, with shipments set to begin in earnest next year. Oracle plans to integrate TPUs into OCI’s AI computing portfolio and provide the power, cooling and data center networking needed for large-scale deployments. Google describes TPUs as custom accelerators designed alongside open-source software to support the AI lifecycle, including training, large-scale inference and the multistep reasoning used by agentic AI.

Google said that as models grow and demand for inference increases, the performance and economics of training runs and inference requests become more important. The TPU architecture is designed to access model data quickly and enable fast communication between chips, maximizing power use and improving AI processing efficiency. OCI customers will be able to choose TPUs based on their workloads, performance needs and cost goals. TPUs support training large models, including dense models and Mixture of Experts (MoE), as well as post-training, inference, agents and production inference for long-context applications.

Google has developed custom AI silicon for more than 10 years. Its eighth-generation TPU lineup consists of TPU 8t for training and TPU 8i for inference. Google says TPU 8t can shorten development cycles for state-of-the-art models from months to weeks and deliver nearly 3 times the computing performance per pod compared with the previous generation. TPU 8i is designed to handle repeated tasks by multiple specialized agents.

Amin Vahdat, Google’s senior vice president and chief technology officer for AI and infrastructure, said next-generation AI requires a range of computing capabilities and that TPUs offer an option optimized for demanding training and inference workloads. Mahesh Thiagarajan, president of OCI, said the addition of TPUs will let customers use several major AI computing architectures in one cloud and choose infrastructure suited to each workload’s performance, throughput and economics.

Oracle announced on the second day of Oracle AI World 2026 that it plans to offer Google’s Tensor Processing Units (TPUs) through Oracle Cloud Infrastructure (OCI). It would be the first time a hyperscaler other than Google offers Google TPUs as a service. OCI’s TPU service is scheduled to launch in 2027 and expand gradually over several years to meet customer demand.

Google TPUs have so far been available only through Google Cloud. Google said this year that it would supply TPUs to outside business customers for the first time, with shipments set to begin in earnest next year. Oracle plans to integrate TPUs into OCI’s AI computing portfolio and provide the power, cooling and data center networking needed for large-scale deployments. Google describes TPUs as custom accelerators designed alongside open-source software to support the AI lifecycle, including training, large-scale inference and the multistep reasoning used by agentic AI.

Google said that as models grow and demand for inference increases, the performance and economics of training runs and inference requests become more important. The TPU architecture is designed to access model data quickly and enable fast communication between chips, maximizing power use and improving AI processing efficiency. OCI customers will be able to choose TPUs based on their workloads, performance needs and cost goals. TPUs support training large models, including dense models and Mixture of Experts (MoE), as well as post-training, inference, agents and production inference for long-context applications.

Google has developed custom AI silicon for more than 10 years. Its eighth-generation TPU lineup consists of TPU 8t for training and TPU 8i for inference. Google says TPU 8t can shorten development cycles for state-of-the-art models from months to weeks and deliver nearly 3 times the computing performance per pod compared with the previous generation. TPU 8i is designed to handle repeated tasks by multiple specialized agents.

Amin Vahdat, Google’s senior vice president and chief technology officer for AI and infrastructure, said next-generation AI requires a range of computing capabilities and that TPUs offer an option optimized for demanding training and inference workloads. Mahesh Thiagarajan, president of OCI, said the addition of TPUs will let customers use several major AI computing architectures in one cloud and choose infrastructure suited to each workload’s performance, throughput and economics.