Vera Rubin NVL72 Request capacity

Vera Rubin NVL72 / Vera Rubin vs GB300

NVIDIA Vera Rubin NVL72 vs GB300 NVL72

One generation apart. What's different in the rack, what it means for your cost per token, and when Blackwell is still the right call.

Updated 10 Nov 2026 · Sources linked inline

Side by side

Per rackGB300 NVL72Vera Rubin NVL72
GPU72 Blackwell Ultra72 Rubin
CPU36 Grace36 Vera
GPU memory20 TB HBM3e20.7 TB HBM4
GPU memory bandwidthUp to 576 TB/s1,580 TB/s
CPU memory17 TB54 TB
NVLink in the rack130 TB/s260 TB/s
FP4 inferenceUp to 1,440 PFLOPS3,600 PFLOPS

Vera Rubin figures are NVIDIA's preliminary numbers (NVIDIA, launch release). GB300 figures and their sources are on gb300nvl72.com.

What it means in practice

More tokens per megawatt. NVIDIA claims up to 10x against GB200 NVL72 and one-tenth the cost per million tokens (NVIDIA). If power is your constraint, this is the number that matters.

Faster memory. Capacity per rack is similar, but HBM4 moves data almost three times as fast. Long-context and reasoning inference is bound by memory bandwidth, so it benefits most.

A real CPU upgrade. Vera and 54 TB of CPU memory give agent orchestration and data preparation more room in the same rack.

Measured, not just claimed. Cognition reported 4.8x total token throughput against a GB200 NVL72 baseline in production (CoreWeave).

When GB300 is still the right call

  • You need racks this quarter. GB300 is in volume; Vera Rubin is mostly reserved.
  • Your software is tuned for Blackwell and you can't spend a quarter requalifying.
  • You want a shorter commitment. Vera Rubin deals lean multi-year.

Do both

A common structure is Blackwell now, Vera Rubin as it lands, with one operator or two. Tell us the timeline and we'll bring options for each leg. For Blackwell detail see gb300nvl72.com and gb200nvl72.com.

Questions

Is Vera Rubin faster than GB300?
Yes. NVIDIA's figures put FP4 inference at 3,600 PFLOPS per rack against up to 1,440 for GB300, with twice the NVLink bandwidth and much faster memory.
Do Vera Rubin and GB300 use the same rack?
Both are 72-GPU, liquid-cooled NVL72 racks, but Vera Rubin's reported power draw is higher, so a hall built for GB300 may not take it without changes. Ask the operator.

RackShore runs this site and may be paid a referral fee by the provider you choose. Buyers pay us nothing. Spot a mistake? Tell us.

Keep reading

Vera Rubin NVL72 specsWhat's in the rack and what each number means, with sources.Vera Rubin availabilityWhich clouds have Vera Rubin NVL72 running, and how to get in.Vera Rubin NVL72 priceWhat drives the cost of Vera Rubin NVL72 capacity, and how to get a firm quote.