The Vera Rubin NVL72 is NVIDIA’s rack-scale AI platform housing 72 Rubin GPUs and 36 Vera CPUs in a single fully liquid-cooled rack, connected by NVLink 6 delivering 260TB/s of aggregate fabric bandwidth. It is the direct successor to GB200 NVL72, and the generational jump is largest in exactly the place that currently constrains AI: memory.
Key benefits at a glance
| Benefit | Why it matters |
| 288GB HBM4 per GPU | Substantially more accelerator memory, which is what limits model size and concurrency |
| Around 22TB/s memory bandwidth per GPU | Roughly a 2.75x jump over Blackwell, directly improving inference throughput |
| 50 PFLOPS NVFP4 inference per GPU | Around five times Blackwell for inference workloads |
| NVLink 6 at 260TB/s aggregate | Double the fabric bandwidth of the previous generation across the rack |
| 72 GPUs in one coherent rack | Very large models fit within a single NVLink domain rather than spanning racks |
| Vera CPU with 88 custom Arm cores | Host processing designed alongside the GPUs rather than adapted to them |
| Fully liquid cooled | Necessary at this density and considerably more efficient than air at the same load |
Generation on generation
| Specification | GB200 NVL72 | Vera Rubin NVL72 |
| GPUs per rack | 72 Blackwell | 72 Rubin |
| NVLink fabric | NVLink 5, around 130TB/s | NVLink 6, around 260TB/s |
| Per-GPU chip-to-chip | 1.8TB/s | 3.6TB/s |
| Memory per GPU | HBM3e | 288GB HBM4 |
| Memory bandwidth | Around 8TB/s | Up to around 22TB/s |
| Rack power | Around 120 to 130kW | Around 190 to 230kW |
The power figure deserves attention
This is the practical planning issue and it is easily glossed over in a specification comparison.
A Vera Rubin NVL72 rack draws roughly 190 to 230 kilowatts depending on configuration, against around 120 to 130 for GB200 NVL72 and roughly 40 for a Hopper-era rack. That is not an incremental increase, and most existing facilities were not designed for it.
If Rubin is on your roadmap, the electrical and cooling conversation with your facilities team needs to start well before the hardware conversation. Power delivery, not silicon availability, is increasingly the thing that determines when you can actually deploy.
Why memory is the headline
Compute figures get the attention, and for inference the memory numbers matter more.
Generating tokens requires repeatedly reading model weights from memory, so bandwidth sets throughput far more directly than FLOPS do. Rubin’s jump from around 8TB/s to around 22TB/s per GPU is therefore the specification most likely to change what your infrastructure can serve.
Capacity matters equally. 288GB per GPU means larger models fit without splitting across devices, and splitting is where latency and complexity both increase.
Who should be planning for it
Being straightforward about positioning: this is not a general enterprise upgrade.
- Worth planning for if you are serving very large models where per-token cost is the primary constraint
- Worth planning for at 200 billion parameter scale and above, where the economics shift decisively
- Probably not yet for inference workloads below roughly 70 billion parameters, where Blackwell remains the sensible choice
- Definitely worth planning the facility work now, whichever generation you eventually buy
Frequently asked questions
NVIDIA’s rack-scale AI platform with 72 Rubin GPUs and 36 Vera CPUs in a single liquid-cooled rack connected by NVLink 6.
It entered full production in 2026 with partner availability in the second half of the year. Allocation queues are expected for buyers without hyperscaler-scale commitments.
It is built for very large models. All 72 GPUs share one NVLink domain, so very large models fit within a single rack, and the economics favour it from around 200 billion parameters upwards. Below roughly 70 billion parameters, Blackwell remains the sensible choice for inference.
288GB of HBM4 with memory bandwidth up to around 22TB/s.
Roughly 190 to 230 kilowatts depending on configuration, considerably more than the previous generation.
It depends on model scale. Rubin makes the clearest economic sense at very large model sizes where per-token cost dominates.
Talk to our team about planning for Vera Rubin, including the facility work it requires.