TL;DR
On October 27, 2026, Oracle announced OCI as the first non-Google hyperscaler to offer Google TPUs, going live in calendar 2027: Oracle supplies power, cooling and datacenter networking, and customers can mix NVIDIA GPUs in the same cloud — versus Google's own TPU service, the difference lies in delivery timing, infrastructure ownership and same-cloud chip mix.
Comparison Overview
On October 27, 2026, at Oracle AI World, Oracle announced that OCI will be the first hyperscaler outside Google to offer Google TPUs, with the service launching in calendar 2027 and scaling gradually over multiple years. For the first time, Google TPUs will run on non-Google hyperscale infrastructure, with Oracle providing power, cooling and datacenter networking. Meanwhile, Google's own TPU service has already reached the eighth generation. This article compares OCI-hosted Google TPUs against Google Cloud GCP TPU on launch timing, scale, infrastructure ownership and workload positioning.
Comparison Table
| Dimension | OCI-hosted Google TPU | Google Cloud GCP TPU |
|---|---|---|
| Provider and launch timing | Oracle; launches in calendar 2027, scaling gradually over multiple years | Google's own cloud service; eighth-generation TPUs already on the GCP roadmap |
| First-wave generation | Not disclosed (Oracle release does not specify the first-wave generation) | TPU 8i (single pod interconnect of 1,152 chips) and TPU 8t (9,600-chip superpod) |
| Compute scale | Scaling gradually over multiple years; specific scale not disclosed | TPU 8i: 1,152 chips per pod; TPU 8t: 9,600 chips per superpod with 2 PB shared HBM |
| Who provides the facility and network | Oracle provides power, cooling and datacenter networking, deploying TPU clusters inside OCI infrastructure | Google supplies them within its own cloud infrastructure; official blog discloses TPU 8i / 8t hardware specs |
| Other chips available alongside | Customers can choose NVIDIA GPUs and Google TPUs in the same cloud | Not disclosed (research scope gives no same-cloud companion chips on the GCP side) |
| Target workloads | Training, post-training and production inference (inference, agents, long context) for dense and MoE models | Same: training, post-training and production inference for dense and MoE models |
| Software stack | Not disclosed (Oracle release does not detail the TPU software stack) | Not disclosed (Google's blog focuses on hardware specs; software stack not covered in this research) |
Evidence and Positions
Oracle's official line: OCI will be the first hyperscaler outside Google to offer Google TPUs, launching in calendar 2027 and scaling over multiple years; Oracle provides power, cooling and datacenter networking, deploying TPU clusters within OCI infrastructure, and customers can select NVIDIA GPUs and Google TPUs in the same cloud (Oracle press release, October 27, 2026, remarks by Mahesh Thiagarajan); workloads span training, post-training and production inference for dense and MoE models (inference, agents, long context).
Google's official line: TPU 8i interconnects 1,152 chips in a single pod with on-chip SRAM tripled versus the previous generation; TPU 8t superpods span 9,600 chips (Google blog, April 22, 2026).
Third-party reporting: per EET-China, Google's external TPU revenue began to be recognized in the fiscal quarter ended June 2026; Anthropic committed to scaling TPU usage to as many as 1 million chips with over 1 GW of capacity online in 2026.
Interpreting the Disagreement
First, launch timing differs: OCI-hosted TPUs go live only in calendar 2027, while Google's eighth-generation TPU specs (8i / 8t) are already on the GCP roadmap. There is no overlapping window in which customers can buy the same generation through both clouds.
Second, the scale figures are not comparable: OCI commits only to gradual multi-year scaling without disclosing numbers, while GCP publishes per-pod / per-superpod chip counts (1,152 and 9,600). One is a total-scale plan, the other a unit spec.
Third, the business signal: per EET-China, Google began recognizing external TPU revenue in the quarter ended June 2026, and Anthropic committed to up to 1 million chips. The OCI partnership adds another channel for external TPU sales, but Oracle's release discloses no purchase volumes, financial terms or pricing, so OCI's weight within Google's external TPU revenue cannot be judged.