Nvidia says Vera Rubin NVL72 trains large models with 75% fewer GPUs
In brief
- Vera Rubin NVL72 combines 72 Rubin GPUs and 36 Vera CPUs in one NVLink domain.
- Nvidia claims the system needs about 75% fewer GPUs than Grace Blackwell NVL72 for a large training run.
- Nvidia projects inference at one-tenth the price per million tokens versus Blackwell racks.
- MLPerf Inference v6.1 recorded up to 3.7x the throughput of GB300 NVL72 on Qwen3-VL.
- Nvidia began ramping production shipments in fall 2026 with more than 80 MGX partners, Crypto Briefing reported.
What's inside the system
Vera Rubin NVL72 is the direct successor to Nvidia's Blackwell NVL72 racks, which shipped in GB200 and GB300 variants. Each GPU carries 288 GB of HBM4 memory. The architecture also brings new interconnect designs and updated precision formats (both aimed at better performance per watt).
That's the hardware. The cost claims are the bigger story.
Nvidia projects that inference on the Vera Rubin NVL72 will cost one-tenth the price per million tokens compared with Blackwell racks. It's a projection, not a measured result. The 75% GPU reduction on the training side is Nvidia's own figure too, and it hasn't been presented as an independent test.
What the benchmarks show
The third-party benchmarks cited in the report cover inference, not training. MLPerf Inference v6.1, released in September 2026, recorded throughput up to 3.7 times higher than the GB300 NVL72 on the Qwen3-VL task, Crypto Briefing reported.
Artificial Analysis benchmarks showed approximately 5.7 times higher throughput per megawatt than the GB300 NVL72 on agentic AI workloads, according to the same report. That's the metric that lines up most closely with what Nvidia says it designed for, since the interconnect and precision changes were pitched at performance per watt rather than raw speed alone.
Those are two different tests measuring two different things.
Shipments and cloud partners
Production shipments began ramping in fall 2026 with support from more than 80 MGX partners, per Nvidia's account as reported by Crypto Briefing. Google Cloud and Microsoft Azure are among the partners supporting the system.
Crypto Briefing also reported that Google Cloud anticipates scaling to as many as 960,000 Rubin GPUs by November. Neither cloud provider was quoted confirming specific deployments in that report, so the 960,000 figure should be read as Crypto Briefing's account rather than a Google statement.
Frequently asked questions
What is the Nvidia Vera Rubin NVL72?
The Vera Rubin NVL72 is an Nvidia system that combines 72 Rubin GPUs and 36 Vera CPUs in a single NVLink domain. It is the direct successor to the Blackwell NVL72 racks, which shipped in GB200 and GB300 variants. Each GPU carries 288 GB of HBM4 memory.
How much more efficient is Vera Rubin NVL72 than GB300 NVL72 in benchmarks?
MLPerf Inference v6.1, released in September 2026, recorded throughput up to 3.7 times higher than the GB300 NVL72 on the Qwen3-VL task, as reported by Crypto Briefing. Artificial Analysis benchmarks showed approximately 5.7 times higher throughput per megawatt on agentic AI workloads.
Is the 75% fewer GPUs figure independently verified?
No. Nvidia itself says the Vera Rubin NVL72 can train a 10-trillion-parameter mixture-of-experts model on 100 trillion tokens in one month with about 75% fewer GPUs than Grace Blackwell NVL72, as reported by Crypto Briefing. The independent benchmarks cited cover inference throughput, not this training figure.
