NVIDIA GB300 NVL72
72 × GB300 (Blackwell Ultra) per rack, with 72 compute rails at 800G. That is 57.6 Tb/s of scale-out network per rack.
- GPUs per rack72 × GB300 (Blackwell Ultra)
- Compute rails72 × 800G
- Scale-out bandwidth57.6 Tb/s
- Storage links18 × 400G
- In-band links18 × 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 Blackwell Ultra. An NVL72 rack carries THREE distinct networks, and only two of them are sized here. (1) NVLink, intra-rack: each GPU has 18 fifth-generation NVLink links, one to each in-rack NVSwitch over a copper backplane, through 9 switch trays. It is per-GPU, it is what makes the rack behave as one accelerator, and it ships with the rack: no optics, no purchasing decision, not in this BOM. (2) East/west scale-out: four dual-port ConnectX-8 SuperNICs per tray at a stated 1:1 GPU-to-NIC ratio, so 72 adapters giving the 57.6 Tb/s of rack network NVIDIA quotes. This is what the compute fabric below sizes. (3) North/south: one dual-port BlueField-3 B3240 DPU per tray, 18 per rack, for storage acceleration, out-of-band provisioning and secure infrastructure. Those are the storage and in-band figures here. NOTE on Ethernet: those ConnectX-8 are dual-port, and ConnectX-8 runs 1x800G in InfiniBand mode but 2x400G in Ethernet. A Spectrum-X build is therefore 144 rails at 400G instead of 72 at 800G: the same 57.6 Tb/s, arranged differently. 100% liquid-cooled at 120 kW continuous, 132 kW nominal TDP, ~155 kW peak. The indivisible unit is a whole 72-GPU rack, so GPU counts round to multiples of 72. Raise the facility rack budget above 132 kW or this correctly reports a power overrun. For an Ethernet build use gb300-nvl72-spectrumx, which has the same rack but a different rail count.
Source: NVIDIA GB300 NVL72 product page and the NVL72 AI Factory enterprise reference architecture (docs.nvidia.com/enterprise-reference-architectures/nvl72-ai-factory), verified 2026-08. Confirmed 2026-08-06