TL;DR
On October 27, 2026, Oracle AI World announced Google TPUs on OCI and Vera Rubin NVL72 GA on the same day: Google's TPU 8t fields a 9,600-chip superpod with 216 GB of HBM per chip, against NVIDIA's 288 GB HBM4 Rubin GPU and its 40,000-GPU Vera Rubin AI factory — the eighth-generation training compute roadmaps of custom ASIC vs GPU.
Comparison Overview
On October 27, 2026, at Oracle AI World in Las Vegas, Oracle made three announcements on a single day: Google TPUs coming to OCI, Cerebras inference coming to OCI, and NVIDIA Vera Rubin NVL72 reaching GA on OCI. This article compares the two training compute routes among them: Google's eighth-generation training chip TPU 8t (codename Sunfish) and NVIDIA's Vera Rubin (Rubin GPU). The goal is to lay out the publicly disclosed specs — per-chip HBM capacity, HBM bandwidth, peak FP4 compute and cluster scale — so the magnitude of the two training routes can be judged.
Comparison Table
| Dimension | Google TPU 8t | NVIDIA Vera Rubin (Rubin GPU) |
|---|---|---|
| HBM per chip | 216 GB | 288 GB HBM4 (Vera Rubin superchip totals 576 GB HBM4) |
| HBM bandwidth | 6.52 TB/s | 22 TB/s (superchip total 44 TB/s) |
| On-chip SRAM | 128 MB | Not disclosed |
| Peak FP4 compute | 12.6 PFLOPS per chip (FP4) | Per chip not disclosed; Vera Rubin AI factory: 1.4 ZFLOPS training (NVFP4), 2 ZFLOPS inference (NVFP4) |
| Maximum cluster scale | Single superpod: 9,600 chips, 2 PB shared HBM, 121 ExaFlops total | Vera Rubin AI factory: 40,000 GPUs |
| Chip interconnect | Inter-chip interconnect bandwidth doubled vs previous generation | NVL72: 72 GPUs per rack interconnected via NVLink |
| Current availability | Codename Sunfish; preview H2 2026, full GA around end of 2027 | Vera Rubin NVL72 and Vera CPU reached GA on OCI on October 27, 2026, scaling to gigawatt scale |
| Software and system complement | Oracle states TPU 8t compresses frontier-model development from months to weeks, with ~3x more compute per pod vs previous generation | Planned HGX Rubin NVL8: 8 Rubin GPUs, dual-socket 64-core Intel Xeon 6, 4 TB DDR5, 6.4 Tbps RDMA |
Evidence and Positions
Google's official line: TPU 8t superpods pack 9,600 chips with 2 PB of shared HBM and 121 ExaFlops of total compute, with inter-chip interconnect bandwidth doubled versus the previous generation (Google blog, April 22, 2026); per-chip specs are 216 GB HBM at 6.52 TB/s, 128 MB on-chip SRAM and 12.6 PFLOPS peak (FP4), versus the previous-generation TPU 7x at 192 GB at 7.37 TB/s and 4.6 PFLOPS (FP8) (Google Cloud technical deep dive).
Oracle's official line: TPU 8t compresses frontier-model development cycles from months to weeks, with per-pod compute nearly 3x higher than the previous generation (Oracle press release, October 27, 2026).
NVIDIA's official line: the Rubin GPU delivers 288 GB of HBM4 at 22 TB/s per chip, and the Vera Rubin superchip totals 576 GB HBM4 at 44 TB/s; a Vera Rubin AI factory (40,000 GPUs) delivers 1.4 ZFLOPS training and 2 ZFLOPS inference (NVFP4); NVL72 interconnects 72 GPUs per rack via NVLink (NVIDIA Vera Rubin NVL72 spec page / NVIDIA Newsroom, October 27, 2026).
Third-party research: according to Gyre Research's June 2026 competitive landscape report, TPU 8t is codenamed Sunfish, with a H2 2026 preview and full GA around the end of 2027.
Interpreting the Disagreement
First, total-compute definitions differ: TPU 8t's 121 ExaFlops is the aggregate of a 9,600-chip superpod (FP4), while the Vera Rubin AI factory's 1.4 ZFLOPS training / 2 ZFLOPS inference figures grade totals across 40,000 GPUs (NVFP4). Cluster scale and precision formats both differ, so per-chip efficiency cannot be derived by simple division.
Second, disclosure is asymmetric: Google publishes per-chip FP4 peak (12.6 PFLOPS) and on-chip SRAM (128 MB) for TPU 8t, whereas NVIDIA does not disclose Rubin GPU's per-chip FP4 peak or on-chip SRAM in the published specs; those cells are marked not disclosed.
Third, delivery stages differ: Vera Rubin NVL72 already reached GA on OCI on October 27, 2026 and scales to gigawatt levels, while Gyre Research pegs TPU 8t's full GA at around end of 2027 after a H2 2026 preview. The two are at different delivery stages, and no independent end-to-end benchmark puts them head to head yet.
Data Sources
- Oracle press release - AI World: Oracle brings Google TPUs to Oracle Cloud Infrastructure
- NVIDIA Newsroom - NVIDIA and Oracle expand NVIDIA Vera Rubin to multi-gigawatt scale and bring NVIDIA Vera CPU to OCI
- Google blog - Eighth-generation TPU for the agentic era
- Google Cloud technical deep dive - TPU 8t and TPU 8i technical deep dive
- NVIDIA Vera Rubin NVL72 spec page