How much power does a GB300 NVL72 rack need?

The short answer

An NVIDIA GB300 NVL72 rack, as configured by HPE, has a nominal draw of 132 kW and an estimated peak (EDPp) of about 155 kW. HPE recommends provisioning busway for 192 kW per rack. Each rack holds 72 Blackwell Ultra GPUs and 36 Grace CPUs across 18 compute trays, fed by eight power shelves of up to 33 kW each.

Cooling: liquid, direct to chip

The GPUs and CPUs are cooled with cold plates, which is direct-to-chip liquid cooling. A coolant distribution unit (CDU) moves coolant through the rack and transfers the heat into the building's facility water loop. HPE's reference design uses in-row CDUs rated for 1.3 MW of cooling, each supporting up to eight racks. Rear-door heat exchangers are listed as an option for heat that still goes to air.

What it means in megawatts

At 132 kW per rack, 1 MW of IT load is about seven racks (7 × 132 kW = 924 kW). A 10 MW deployment is roughly 75 racks before networking, storage and the cooling plant. Plan electrical capacity around the 192 kW busway figure, not the nominal draw.

What to check before you sign

• Busway and power distribution rated for 192 kW per rack, with 400 V or 480 V, 60 A whips.
• Facility water to the CDUs: available capacity, supply temperature, and who owns the secondary loop.
• Floor loading for about 3,300 lb per rack, plus the delivery path from the dock to the hall.
• A rack footprint of 600 mm by 1,068 mm, with room for in-row CDUs.
• A confirmed energization date for the specific hall, in writing.
• Who operates and maintains the liquid cooling, and the response time if it fails.

Many existing halls were designed for air-cooled racks at a fraction of this density. That's why AI-ready capacity is scarce even in large markets. Need a site that's ready for GB300? Send us your requirement.

Source: HPE QuickSpecs, NVIDIA GB300 NVL72 by HPE. Related: Colocation vs powered shell vs GPU cloud, US data center markets for AI.