GS-SYS-04Engineering datasheet

NVIDIA GB200 NVL72

Statusshipping
Verified2026-07-10

The GB200 NVL72 is a liquid cooled, 48U rack that wires 72 Blackwell B200 GPUs and 36 Grace CPUs into a single NVLink domain, so software addresses one accelerator with a 13.4 TB HBM3e pool instead of 36 separate servers stitched together over InfiniBand. It targets teams training or serving frontier scale models that need that pool at NVLink speed rather than sharded across PCIe or network hops. Hosting one draws 125 kW to 135 kW of continuous IT load through direct to chip liquid cooling with zero rack fans, a facility class that only purpose built AI data centers and a small number of retrofitted colocation halls can deliver today.

72x Blackwell B200
GPU count · one NVLink domain, no PCIe fabric between GPUs
13.4 TB HBM3e
NVLink domain memory · plus 17 TB LPDDR5X CPU memory
125 kW to 135 kW
Rack power · operating power, reported
direct to chip liquid cooling
Cooling · zero rack fans
130 TB/s aggregate
NVLink bandwidth · 1.8 TB/s per GPU
Compute01/05
GPU72x NVIDIA Blackwell B200 · one NVLink domain
CPU36x NVIDIA Grace · 72 core Arm Neoverse V2, 2,592 cores total
FP4 Tensor Core, sparse1,440 PFLOPS · rack aggregate
FP4 Tensor Core, dense720 PFLOPS · rack aggregate
FP8/FP6 Tensor Core720 PFLOPS · rack aggregate
Compute trays18x 1U trays · 4 GPU plus 2 Grace CPU per tray, 2 Superchips per tray
Memory02/05
GPU memory (HBM3e)13.4 TB total · roughly 186 GB per GPU, 372 GB per Superchip
GPU memory bandwidth576 TB/s aggregate · rack total
CPU memory (LPDDR5X)17 TB total · 480 GB per Grace CPU
CPU memory bandwidth14 TB/s aggregate · rack total
NVLink domain03/05
GPUs per domain72 · single non blocking NVLink domain
NVLink switch trays9x · 4 ports per compute tray
Aggregate NVLink bandwidth130 TB/s · GPU to GPU, rack total
Per GPU NVLink bandwidth1.8 TB/s · NVLink 5
External scale out fabricConnectX-7, up to 400 Gb/s per GPU · InfiniBand NDR or Spectrum-X Ethernet
Power and liquid cooling04/05
Operating power125 kW to 135 kW · Supermicro rack spec, reported
Peak/transient powerup to 192 kW · HPE EDPp figure, roughly 1.5x TDP, reported
Power shelves8x 1U, 132 kW total · 6x 5.5 kW PSUs per shelf, N+N redundant
Cooling architecture100 percent direct to chip liquid cooling · no air cooled components in the compute path
Facility coolant loopASHRAE W-series chilled water · reported supply range 18C to 25C, return up to 45C depending on CDU config
In-rack CDU capacityup to 250 kW · or 1.3 MW in-row CDU shared across multiple racks
Physical and weight05/05
Rack dimensions2,236 mm x 600 mm x 1,068 mm · 48U, reference chassis
Rack weightapproximately 1.36 metric tons · reported, varies by OEM chassis and options
Floor loadingconcentrated point load · reported around 1,875 kg per square meter, exceeds most 1,000 to 1,500 kg per square meter raised floor ratings
Rack footprintapproximately 1.34 sq m · 600 mm x 1,068 mm base
Field notes
  • 125 to 135 kW of continuous, liquid cooled load per rack is well outside what most enterprise data centers can deliver; a large majority of existing halls top out around 40 to 60 kW per rack on air cooling alone.
  • The rack's point load (reported near 1,875 kg per square meter) exceeds standard raised floor ratings, so slab on grade installation or floor reinforcement is a real line item, not a footnote, in any facility budget.
  • Zero rack fans means every watt of GPU, CPU, NVLink switch, and PSU heat lands on the facility water loop; a chiller or cooling tower undersized for this delta will throttle the whole rack, not just the hot node.
  • For teams that cannot host 120 kW plus liquid cooled racks, a fleet of air cooled 8-GPU HGX B200 servers remains the realistic near term alternative, at the cost of losing the single NVLink domain across all 72 GPUs.
  • Treat the 13.4 TB NVLink-domain memory figure as the actual reason to buy this system over a cluster of smaller nodes: it is what lets very large models live in one coherent address space instead of being sharded across a slower network fabric.
Questions we get on this part

How much does a GB200 NVL72 cost?

Reported figures cluster around $2.8M to $3.4M per rack, with an HSBC analyst estimate near $3M for a training class configuration. NVIDIA does not publish an official price; actual cost depends on the OEM, networking, storage, and support contract.

What power does a GB200 NVL72 rack need?

Operating power is reported at 125 kW to 135 kW continuous, with peak or transient draw reported as high as 192 kW by at least one OEM. That is delivered as 480V three phase into 8 power shelves feeding a fully liquid cooled compute path.

Can a normal data center host a GB200 NVL72?

Not without significant upgrades. Most colocation and enterprise facilities are built for 40 to 60 kW air cooled racks; this system needs a facility chilled water loop, reinforced or slab on grade flooring for its concentrated weight, and 480V distribution most legacy halls do not have.

Is GB200 NVL72 still current, or has GB300 replaced it?

Both are shipping. GB200 NVL72 remains in production and is deployed at CoreWeave, Oracle, Microsoft, and other hyperscalers; GB300 NVL72 is the newer Blackwell Ultra generation with more memory per GPU and higher FP4 throughput, sold alongside GB200 rather than as an immediate replacement.