AI Network Designer Design a network →

NVIDIA GB300 NVL72 (Spectrum-X Ethernet)

72 × GB300 (Blackwell Ultra) per rack, with 144 compute rails at 400G. That is 57.6 Tb/s of scale-out network per rack.

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

144 compute NICs means 144 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

The Ethernet face of the GB300 NVL72. Identical hardware to gb300-nvl72 — same rack, same 72 ConnectX-8, same per-GPU NVLink domain internally over the copper backplane, same 18 BlueField-3 for north/south. The one difference is that ConnectX-8 cannot run 1x800G in Ethernet mode, only 2x400G, so the scale-out fabric terminates 144 links per rack instead of 72. Set planes to 2 to match the twin-planar Spectrum-X design, giving 72 rails to each plane; leaving it at 1 sizes a single-plane 144-rail fabric, which is a different topology.

Source: NVIDIA NVL72 AI Factory enterprise reference architecture and the ConnectX-8 SuperNIC datasheet (1x800G InfiniBand or 2x400G Ethernet per adapter), verified 2026-08. Confirmed 2026-08-06

Related