NVIDIA Vera Rubin started shipping in Q3 2026 with 10× perf/watt over Grace Blackwell, 30× performance in NVL72 racks, and 35× lower token cost (NVIDIA newsroom, 2026). The platform uses HBM4 memory with 2.7 TB/s per-GPU bandwidth and targets hyperscalers, sovereign AI clouds, and enterprise DGX customers.
Data last verified September 9, 2026 from NVIDIA newsroom and the GTC 2026 keynote.
What is Vera Rubin
Vera Rubin is NVIDIA's next-generation AI accelerator platform, named after the American astronomer who provided the first evidence of dark matter. The platform succeeds Grace Blackwell (GB200/GB300) and consists of:
- Vera Rubin GPU — Rubin-based accelerator with HBM4, 2.7 TB/s memory bandwidth.
- Vera CPU — Grace-2 successor with 144 ARM Neoverse V3 cores per socket.
- NVLink Switch 6 — 6th-generation NVLink with 3.6 TB/s bisection bandwidth.
- NVL72 rack — 72 GPUs + 36 CPUs + liquid cooling.
Performance benchmarks
| Metric | GB300 NVL72 | Vera Rubin NVL72 | Improvement |
|---|---|---|---|
| FP4 inference (per-watt) | 1× | 10× | 10× |
| Aggregate FP8 training | 1× | 30× | 30× |
| Token cost (per million tokens) | 1× | 1/35× | 35× lower |
| Memory bandwidth (per GPU) | 1.8 TB/s | 2.7 TB/s | 1.5× |
| NVLink bisection | 1.8 TB/s | 3.6 TB/s | 2× |
Source: NVIDIA GTC 2026 keynote (2026).
What's new vs Blackwell
Vera Rubin's headline change is the move to a rack-scale system. Rather than optimising a single GPU, NVIDIA designed Rubin as a 72-GPU tightly-coupled compute fabric. This mirrors the hyperscaler trajectory — AWS Trainium 2, Google TPU v7, and Microsoft Maia 100 all target rack-scale rather than per-GPU performance.






