A GB300 NVL72 rack holds 72 Blackwell Ultra GPUs and 36 Grace CPUs in 18 compute trays, joined by 9 NVLink switch trays into one 130 TB/s NVLink domain, with 20 TB of HBM3e (NVIDIA). Server makers rate the rack at about 132 to 135 kW, up to 155 kW at peak, with about 90% of the heat removed by liquid.
GB300 NVL72 is NVIDIA's Blackwell Ultra rack: one 48U rack that works as a single 72-GPU system. It is bought and rented as a whole rack. Every figure below links to its source.
| Spec | GB300 NVL72 | Source |
|---|---|---|
| GPUs and CPUs | 72 Blackwell Ultra GPUs, 36 Grace CPUs (2,592 Arm Neoverse V2 cores) | NVIDIA GB300 NVL72 |
| Trays | 18 compute trays of 2 Grace CPUs and 4 GPUs; 9 NVLink switch trays | DGX GB200/GB300 user guide |
| GPU memory | 20 TB HBM3e, up to 576 TB/s; 21 TB in Supermicro's sheet | NVIDIA GB300 NVL72, Supermicro |
| CPU memory | 17 TB LPDDR5X, 14 TB/s | NVIDIA GB300 NVL72 |
| NVLink | 130 TB/s | NVIDIA GB300 NVL72 |
| FP4 Tensor | 1,080 dense, 1,440 sparse (petaFLOPS) | NVIDIA GB300 NVL72, NVIDIA DGX GB300 |
| FP8 / FP6 Tensor | 720 sparse (petaFLOPS) | NVIDIA GB300 NVL72 |
| FP64 | 100 TFLOPS | NVIDIA GB300 NVL72 |
| Networking | 800 Gb/s per GPU through ConnectX-8, Quantum-X800 InfiniBand or Spectrum-X Ethernet | NVIDIA GB300 NVL72 |
| Spec | Blackwell Ultra in GB300 | Source |
|---|---|---|
| Memory | 279 GB HBM3e at 8 TB/s in NVIDIA's datasheet; 288 GB in NVIDIA's technical blog and OEM sheets | Blackwell Ultra datasheet, NVIDIA technical blog, Supermicro |
| Power | Configurable up to 1,400 W (Lenovo lists 1,100 W) | Blackwell Ultra datasheet, Lenovo Press |
| FP4 Tensor | 15 dense, 20 sparse (petaFLOPS) | Blackwell Ultra datasheet |
| NVLink | 1.8 TB/s | Blackwell Ultra datasheet |
| Source | Rack power |
|---|---|
| Supermicro | Eight 1U power shelves of 33 kW (six 5.5 kW supplies each), 132 kW of total power |
| Lenovo Press | 135 kW TDP, up to 155 kW at peak; six or eight 33 kW power shelves |
| DGX GB200/GB300 user guide | About 120 kW for a DGX GB rack; eight power shelves with N+N redundancy turn AC into 50 to 51 V DC on the busbar |
NVIDIA added energy storage to the GB300 NVL72 power shelves, 65 joules per GPU, which cut the peak power the grid sees by 30% in its Megatron training test (NVIDIA power blog).
In our offers GB300 NVL72 racks can start from Q1 2027, usually on terms of 3 to 5 years. See GB300 NVL72 rental, what a rack costs to buy and B300 vs GB300 NVL72.
About 132 to 135 kW by server makers' ratings: Supermicro 132 kW, Lenovo 135 kW and up to 155 kW at peak.
Eight 1U shelves of 33 kW, each with six 5.5 kW supplies, in Supermicro's and NVIDIA's descriptions; NVIDIA describes them as N+N redundant, and Lenovo offers six or eight.
20 TB of HBM3e for the 72 GPUs and 17 TB of LPDDR5X for the 36 Grace CPUs in NVIDIA's figures.
Yes: NVIDIA calls it fully liquid-cooled; its SuperPOD document says GPUs and CPUs are liquid-cooled and other parts air-cooled, and about 90% of the rack's heat goes to liquid in Lenovo's figures.
We found no full-rack weight published by NVIDIA, Supermicro or Lenovo; Lenovo gives 185 kg for the empty rack, 29 kg per compute tray and 17 kg per switch tray.
