GB200 NVL72 Rack Power Requirements for Facility Planning

The GB200 NVL72 is a roughly 120 to 132 kW per-rack system that assumes direct liquid cooling and 480V three-phase power. Those numbers sit four to ten times above typical data center rack designs, so facility planning, not GPU selection, becomes the gating work. This page assembles the documented power envelope, the cooling and electrical assumptions behind it, the monthly energy arithmetic any planner can rerun, and the conditions that select between retrofitting, purpose-built space, and managed capacity.

The Documented Power Envelope: 120 to 132 kW per Rack

The documented envelope is approximately 120 kW of rack power consumption per the vendor user guide, with vendor collateral listing 132 kW nominal rack TDP and deployment specifications engineering toward roughly 150 kW per system including overhead at 480V three-phase. The drivers inside that number: 72 Blackwell GPUs and 36 Grace CPUs in one NVLink domain, NVLink switch trays, and power conversion overhead, which is why the per-GPU mental model of roughly 1.2 kW produces a 100 kW-plus surprise once switching and conversion losses land. Two planning rules follow. First, provision electrical capacity to the engineering number (150 kW class), not the nominal figure, because electrical design power sits above nominal TDP. Second, multiply rack draw by facility power usage effectiveness: at PUE 1.25, a 132 kW rack pulls roughly 165 kW from the grid before the chilled-water plant is counted, and facility contracts price that grid draw, not the IT load.

Figures are vendor-documented as of October 2026 and vary with integrator configuration; final engineering must use the specific system datasheet. The envelope is the reason NVL72 planning starts with electrical service rather than with the GPU roadmap.

Cooling and Electrical Assumptions Behind the Numbers

The power numbers assume direct liquid cooling, and the assumption is structural: deployment documentation specifies roughly 80 liters per minute of coolant flow per rack, with about 115 kW of heat removed by the liquid loop and only the remainder left to air. Facilities therefore need three inputs verified before the rack arrives: coolant distribution capacity (flow and supply temperature matched to the vendor loop), 480V three-phase power distribution at the engineered per-system rating, and condenser or heat-rejection capacity sized to the PUE-adjusted load. The cost arithmetic then becomes a formula any planner can rerun: monthly energy cost equals rack kW times 730 hours times tariff times PUE. At a 0.07 dollars per kWh industrial tariff, a 132 kW rack at PUE 1.25 costs roughly 8,400 dollars per month in energy alone (132 times 730 times 1.25 times 0.07), before distribution equipment, cooling capital, or maintenance. That monthly figure, annualized, is the number to compare against managed capacity pricing rather than the IT-load-only estimate most budgets start from.

The worked example uses stated assumptions, and readers substitute their own tariff and measured PUE. A Texas industrial tariff near 0.05 dollars per kWh moves the same rack near 6,000 dollars per month, while a 0.12 dollars per kWh coastal market moves it past 14,000 dollars; the formula, not the example, is the planning input.

Heat rejection deserves the same arithmetic honesty: 115 kW of liquid-side heat per rack at five racks is more than half a megawatt of condenser load, which is a chiller-plant conversation, not a rack conversation.

Retrofit, Purpose-Built, or Managed Capacity

Choose by three observable conditions. Retrofit your own room only if it already carries high-density power distribution (480V to the rack position), a liquid loop or a credible path to one, and floor loading for a liquid-cooled rack; retrofit projects otherwise become chiller-and-busway construction with multi-quarter timelines. Purpose-built or purpose-leased space fits programs with sustained multi-rack demand and an operations team ready to own facility engineering. Managed dedicated capacity fits teams whose product is models rather than facilities: power, cooling, and rack engineering arrive with the cluster under one contract, which converts a capital construction problem into an infrastructure service decision. The disqualifier to check first is density honesty: if a provider quotes NVL72-class hosting without stating per-rack power, coolant flow, and a liquid-ready facility, the quote prices a rack the facility cannot feed. Managed private infrastructure, delivered by providers like OneSource Cloud as single-tenant environments in U.S. data centers with flat-rate monthly pricing, exists precisely for teams that want the density without the construction project.

Whichever route wins, ask the density-honesty question first: per-rack power, coolant flow rate, and liquid-readiness evidence. OneSource Cloud delivers dedicated single-tenant GPU environments in secure U.S. data centers where power, cooling, and rack engineering are part of the managed service, which keeps the engineering burden inside one accountable contract.

FAQ

Can an air-cooled data center support a GB200 NVL72 rack?

No, not as the design point. The rack is engineered for direct liquid cooling, with roughly 115 kW of heat assigned to the liquid loop and about 80 liters per minute of coolant flow required, so an air-only room would need to dissipate a heat load far beyond typical hot-aisle design. Facilities without a liquid loop should treat liquid readiness as the entry condition for NVL72-class systems.

How does OneSource Cloud handle power and cooling for dense dedicated GPU clusters?

OneSource Cloud delivers dedicated single-tenant GPU clusters in secure U.S. data centers where facility power, cooling, and rack engineering are part of the managed environment under flat-rate monthly pricing with zero egress fees. Teams adopt dense configurations without owning the facility engineering behind them, and the density-honesty questions become contract terms rather than construction research.

Summary: Plan from the Engineering Number

The planning sequence that survives contact with an NVL72-class system is fixed: take the documented 120 to 132 kW envelope, provision electrical to the 150 kW class, verify coolant flow and 480V distribution, and compute monthly energy from rack kW times 730 times tariff times PUE. Compare that total cost against managed capacity before committing to construction. For teams that conclude the facility project is not their business, OneSource Cloud exists to deliver the density as a service.

To evaluate dedicated environments where power and cooling engineering is delivered with the cluster, see OneSource Cloud managed AI infrastructure.