Google and Oracle announced on October 27, 2026, that Google Tensor Processing Units (TPUs) are planned to become available through Oracle Cloud Infrastructure (OCI) during calendar 2027, with capacity ramping over multiple years. Oracle says it will provide the power, cooling and data-center networking for the deployment. The announcement sets out a future cloud-access route, not a service that is already live.
Three ways to access Google TPUs
The planned OCI offering differs from buying TPU systems for a company’s own data center or renting TPU capacity through Google Cloud.
| Access route | Where the TPUs run | How customers access them | Infrastructure |
| Direct system sales | The customer’s own data center | Selected external customers can buy systems; deliveries and direct-sales revenue were reported for Q2 2026. | Customer’s own data center |
| Google Cloud rental | Google Cloud | Customers rent TPU capacity; Google lists rental prices per chip-hour. | Google Cloud |
| Planned OCI access | Oracle Cloud Infrastructure | Cloud access is planned for calendar 2027, with a multi-year capacity ramp. | Oracle says it will provide power, cooling and data-center networking. |
For organizations that want TPU systems on site, direct sales put the hardware in their own data centers. OCI is a planned cloud-hosted option instead. Google Cloud rentals are a separate route: Google’s pricing varies by TPU product, deployment model and region, and its per-chip-hour rates are not OCI prices. Network World reported on October 28, 2026, that OCI pricing, configurations and initial capacity had not been published.
TPU 8t is for training; TPU 8i is for inference
A TPU is a custom-built accelerator for AI and machine-learning workloads. Google’s eighth-generation lineup assigns TPU 8t to training, the process of developing a model, and TPU 8i to inference, when a trained model produces outputs. Google says OCI customers are planned to be able to access TPUs alongside GPUs and choose an accelerator by workload.
That choice does not establish a universal cost or performance winner. Stephanie Walter, AI stack practice leader at HyperFrame Research, told Network World that switching accelerators can involve adapting models and tools and maintaining the resulting software. Infrastructure price is only one part of that decision.
What has been reported about external customers
CNBC reported on October 28, 2026, that OpenAI planned to buy Google TPUs. Separately, DiarioBitcoin reported on October 28 that Meta was running production workloads on TPUs in its own data centers during October 2026. Those reports describe customer activity apart from Oracle’s future OCI route.
