NVIDIA GB200 NVL72

One rack as one GPU

GB200 NVL72 connects 72 Blackwell GPUs and 36 Grace CPUs into a single NVLink domain, so a trillion-parameter model sees the whole rack as one accelerator. NVIDIA's own figure is roughly 30 times the real-time inference throughput of an equivalent H100 cluster on the largest models, at a fraction of the energy per token. It ships only as a liquid-cooled rack, built by Supermicro and other partners to NVIDIA's reference design.

Who it is for

• Frontier-scale training and inference on models above the 400B class.
• Providers selling Blackwell capacity by the hour.
• Nobody running 70B models: an H200 node does that for far less.

The facility question comes first

• Around 120kW per rack means most existing European colocation halls cannot host it without an upgrade. We check this before quoting.
• Facility water or CDU capacity is mandatory.
• Allocation is still tight; lead times are quoted case by case.

Cloud pricing for B200 by the hour is on our price comparison if you want the capacity without the building work.