NVIDIA GB200 NVL72
72 × GB200 Blackwell per rack, with 72 compute rails at 400G. That is 28.8 Tb/s of scale-out network per rack.
- GPUs per rack72 × GB200 Blackwell
- Compute rails72 × 400G
- Scale-out bandwidth28.8 Tb/s
- Storage links36 × 400G
- In-band links36 × 100G
- BMC ports18
- Rack units48U
- Typical power120 kW
- Maximum power132 kW
Rack-scale. The indivisible unit is a whole rack of 72 GPUs, so GPU counts round to multiples of 72 instead of to a node boundary. The figures above are per rack. See GB200 and GB300 NVL72 networking for what the three internal networks do and which of them you actually buy.
What this means for the fabric
72 compute NICs means 72 rails. In a rail-optimized design each of those rails is sized on its own against the rack count, then multiplied back up, so the leaf tier is driven by the rail count and not by the total endpoint count.
Storage and in-band management are separate fabrics with their own switches and their own optics, and the 18 BMC ports per rack feed the out-of-band network, which also has to carry every switch and every PDU in the cluster.
Power and racks
At 120 kW typical draw, this rack fills a rack's power budget long before it fills the space. A 45 kW rack takes 1 of them on power alone, against 0 on rack units. Power is what actually sets the rack count.
Notes
Rack-scale: the indivisible unit is a whole rack of 72 GPUs, so GPU counts round to multiples of 72, not 8. Three networks, as on the GB300: per-GPU NVLink inside the rack over a copper backplane through 9 switch trays, ConnectX east/west, and BlueField-3 north/south. The tray itself is a different design in two ways that change the BOM. (1) East/west is 4 single-port 400G ConnectX-7 per tray, not dual-port 800G ConnectX-8, so 28.8 Tb/s of rack network against the GB300's 57.6. (2) North/south is TWO dual-port BlueField-3 per tray, not one, giving 36 storage and 36 management links per rack against the GB300's 18 and 18. The two management ports are bonded (LACP mode 4); the GB300 has no bond because it has only one card. GB200 also lacks the GB300's NIC-direct path: the GPU reaches the ConnectX-7 through the Grace CPU over PCIe. That does not change the port count. At roughly 120 kW per rack this needs liquid cooling and a very different power design from air-cooled DGX. One caveat: some secondary sources describe the tray front as 4 OSFP plus only 2 QSFP cages, which would imply 2 north/south ports instead of 4. NVIDIA's named-interface enumeration is taken as authoritative over those.
Source: NVIDIA DGX GB200 user guide (hardware and networking), and the Mission Control rack setup guide, which enumerates the per-tray interfaces by name. Verified 2026-08. Confirmed 2026-08-06