Summary
On October 27, 2026 at Oracle AI World in Las Vegas, Oracle announced it will offer Google's eighth-generation TPU on OCI from 2027, becoming the first hyperscaler outside Google to provide TPU. This card compares TPU 8t (training, 216GB HBM, 12.6 PFLOPs FP4) and TPU 8i (inference, 288GB HBM, 10.1 PFLOPs FP4, ICI 19.2 Tb/s).
Background and Data Scope
- Event: Oracle announced on October 27, 2026 (Tuesday, Las Vegas, Oracle AI World) that it will offer Google TPU to customers for the first time, becoming the first hyperscaler outside Google to provide TPU; the service is expected to launch in 2027 and scale over multiple years, supporting training, inference and agentic workloads, per IT Home.
- Sources: Chip specs and supercluster figures come from Google's official blog (The Keyword, 2026-10-28) and the Google Cloud technical deep-dive blog; the OCI partnership was reported by IT Home.
- Scope: Peak FP4 figures are per-chip native FP4 peaks; HBM capacity is per-chip configuration; supercluster size follows the TPU 8t single-pod basis.
Core Data Table
| model | positioning | HBM capacity | on-chip SRAM | peak FP4 compute | interconnect bandwidth | supercluster scale |
|---|---|---|---|---|---|---|
| TPU 8t | training | 216GB | 128MB | 12.6 PFLOPs | cluster-to-cluster bandwidth 2x previous generation | single supercluster 9,600 chips, 2PB shared HBM, 121 ExaFlops |
| TPU 8i | inference | 288GB | 384MB (3x previous generation) | 10.1 PFLOPs | ICI 19.2 Tb/s (2x previous generation) | — |
| Ironwood (previous-gen reference) | prior flagship baseline | — | — | — | — | TPU 8t single-pod compute nearly 3x higher |
| OCI listing deal | cloud partnership | — | — | — | — | launches 2027; first hyperscaler outside Google |
Supplementary Data
- TPU 8t: Google says single-pod compute is nearly 3x the previous Ironwood generation, targeting goodput above 97%; the Virgo network supports near-linear scaling of a single logical cluster to one million chips.
- TPU 8i: Google says on-chip latency is reduced by up to 5x, performance-per-dollar improves 80% over the previous generation, and energy efficiency (performance per watt) is up to 2x.
- Google's official blog (2026-10-28) confirms the eighth-generation TPU delivers up to 12.6 pflops of native FP4 compute and formally announces TPUs coming to OCI.