NVIDIA Server Rack Power Consumption Calculator

NVIDIA rack power is a different question from ordinary rack power, because on the rack-scale systems the rack is the product. A GB200 NVL72 is not seventy-two servers you install in a cabinet. It is one machine, 48U tall, delivered as a rack, with no per-server power supplies at all: eight power shelves convert facility AC onto a shared 50 V DC busbar, and 18 compute trays plus 9 NVLink switch trays draw from it.

So this calculator opens on a real GB200 NVL72, built tray by tray. Switch the preset for a GB300 NVL72, a rack of DGX B200s, a rack of DGX H100s, or air-cooled HGX nodes, and every line stays editable if your configuration differs.

The numbers are the story. A well-filled traditional rack of 2U servers runs around 12 kW. A GB200 NVL72 draws about 120 kW in one cabinet, and peaks well above that. That is ten conventional racks of power in a footprint of roughly 0.64 square metres, which is why these systems are mandatory direct-to-chip liquid cooled, sit on slab floors rather than raised tiles, and arrive with their own coolant manifolds.

Three things make NVIDIA racks catch people out, and the calculator surfaces all three:

• Peak is a long way above nominal. NVIDIA quotes the GB200 NVL72 near 120 kW and HPE at 132 kW nominal, but the electrical design peak runs to roughly 155 to 192 kW. Feeds, breakers and upstream capacity get sized on the peak, not the average.
• Weight. A loaded NVL72 is around 1.4 tonnes. A standard raised floor tile is rated near 570 kg. These racks go on slab, and the delivery route needs checking before the crate arrives.
• The heat has nowhere to go without liquid. 120 kW is about 410,000 BTU/hr, or 34 tons of cooling, from one cabinet. No amount of air handles that.

Rack-scale systems are summed at the tray level, which is how they are actually built:

rack watts = (compute trays × W each) + (switch trays × W each) + (power shelves × standby)

For a GB200 NVL72 that is 18 compute trays, 9 NVLink switch trays and 8 power shelves in a 48U frame. Each compute tray carries two Grace CPUs and four Blackwell GPUs, so 18 trays give the 36 Grace and 72 Blackwell the name refers to.

The tray figures here are AC-side: the power shelves' conversion loss is already inside them, so the shelves themselves contribute only standby draw rather than being counted twice. That matters because unlike a conventional server there is no per-node PSU efficiency to apply, the conversion happens once at the shelf.

Facility side

amps = watts ÷ (√3 × volts × PF) for the three-phase feeds these racks require
usable circuit watts = √3 × volts × amps × 0.8 × PF
circuits = ceil(peak watts ÷ usable circuit watts), doubled for A+B

At 415 V three-phase a 63 A circuit delivers about 35 kW usable after the 80% continuous-load derate. A 120 kW NVL72 therefore needs four to five such circuits, or eight to ten once you provision redundant A and B feeds, each sized to carry the whole rack alone.

Heat

BTU/hr = watts × 3.412 and tons = BTU/hr ÷ 12,000

On an NVL72 the great majority of that leaves through the coolant loop rather than the air. HPE's figures for their GB200 NVL72 split it roughly 115 kW liquid to 17 kW air, so the room still needs air handling for the residue even though the rack is liquid cooled. A CDU, in-row or facility-provided, sits between the rack loop and the building water.

Where the DGX figures differ

DGX systems are conventional servers with their own power supplies, so they go through the same component model as everything else on this site and report a realistic sustained peak. NVIDIA separately publishes a maximum system power rating, which is the nameplate ceiling: 6.5 kW for the DGX A100, 10.2 kW for the DGX H100 and H200, and 14.3 kW for the DGX B200. Size the electrical feed on NVIDIA's published maximum; use the modelled figure for energy, cost and heat, which is what the machine actually spends its life drawing.