NVL72

NVL72 is NVIDIA's rack-scale system design in which 72 GPUs share one NVLink domain and act as a single large GPU. The number counts the GPUs in that domain. The GB200 NVL72 pairs 72 Blackwell GPUs with 36 Grace CPUs in a liquid-cooled rack, and the GB300 NVL72 swaps in 72 Blackwell Ultra GPUs. NVIDIA has since rebranded its Vera Rubin rack from NVL144 to NVL72.

A GB200 NVL72 rack holds 18 compute trays and nine NVLink switch trays. Four NVLink cartridges with over 5,000 copper cables let every GPU reach every other at 1.8 TB/s, for 130 TB/s across the rack. An HGX H200 baseboard, by contrast, caps the NVLink domain at eight GPUs and 900 GB/s per GPU. The rack carries 13.4 TB of HBM3E and 17 TB of LPDDR5X memory. NVIDIA deepened the rack busbar to carry 1,400 A and sized direct liquid cooling for 120 kW per rack.

Choose an NVL72 rack when one model must span more than eight GPUs at full NVLink speed, such as real-time trillion-parameter inference. Confirm first that the site can deliver 120 kW of liquid cooling to one rack. Where it cannot, an eight-GPU HGX server keeps full NVLink speed inside one chassis.

Sources

Last verified October 7, 2026.