GB200 & GB300 NVL72 Power and Cooling Requirements

NVIDIA's DGX GB200 NVL72 rack guide lists approximately 120 kW for the reference rack. NVIDIA's current GB300 NVL72 reference architecture lists up to 142 kW. Those are the cleanest generation-level planning numbers; an OEM may publish a different nominal load, peak-design figure, or facility provisioning requirement for its build. See the NVIDIA DGX GB200 rack guide and NVIDIA GB300 reference architecture.

The density is the important facility fact. These are 72-GPU, 36-Grace-CPU systems with nine NVLink switch trays, rack-level power conversion, and direct liquid cooling. Do not derive the facility feed by multiplying a quoted per-GPU value: the rack total also includes CPUs, switches, fans or pumps, conversion losses, and the OEM's design margin.

Use the published figure that matches the actual build and the actual engineering question. The table below compares like-for-like platform characteristics; the later power table keeps unlike electrical definitions separate. For the compute datasheet, see the GB200 NVL72 specs.

The GPUs and CPUs in an NVL72 are cooled through direct-to-chip liquid loops; this is part of the platform design, not an optional rack upgrade. Coolant runs through cold plates on the high-heat components and carries most of that heat to a CDU and then to the facility side.

That does not mean the rack has no airflow. OEM technical documentation can still specify fans and a residual air-cooling requirement for components outside the liquid loop. Lenovo, for example, documents two liquid loops and eight system fans for its GB300 rack. The accurate facility statement is therefore: air alone cannot cool the NVL72, but liquid cooling does not necessarily remove every air-side load. See Lenovo's GB300 technical specifications.

A coolant distribution unit (CDU) separates the technology loop through the rack from the facility-water loop that rejects the heat. Its placement is deployment-specific: it can be integrated with a rack, installed in-row, or provided as facility equipment. Treating every NVL72 as an in-rack-CDU design is too broad.

For one published example, HPE specifies an in-row 1.3 MW CDU supporting up to eight racks for its GB200 and GB300 systems—not one CDU inside every compute rack. The facility design still has to resolve supply temperature, flow, pressure drop, water quality, redundancy, leak detection, and heat rejection. See the HPE GB200 QuickSpecs.

Coolant flow is also temperature-dependent, not a single universal number. The values below are Lenovo's published GB300 requirements and should be used for that implementation only; warmer facility water demands substantially more flow. Source: Lenovo GB300 environmental specifications.

The NVIDIA reference design uses eight 33 kW power shelves. Each shelf contains six 5.5 kW power supplies that convert AC input to roughly 50–51 V DC, then feed the compute and switch trays through a common rack busbar. This is the current published architecture—not an 800 V DC rack input. The reference design supports redundant power-shelf groups; the exact upstream feed and redundancy scheme remain OEM- and site-specific.

The numbers below are deliberately labeled because they are not interchangeable. Reference power describes an architecture, nominal power a stated operating point, EDPp a peak design value, and busway provision the facility capacity reserved for the rack. Do not turn these four columns into a single ranking chart. Sources: NVIDIA DGX GB200 rack guide, NVIDIA GB300 reference architecture, HPE GB200 QuickSpecs, and HPE GB300 system specifications.

The Blackwell Ultra GB300 NVL72 keeps the 72-GPU, 36-Grace-CPU rack shape but increases GPU memory and the NVIDIA reference power ceiling to up to 142 kW. That is higher than NVIDIA's approximately 120 kW DGX GB200 reference figure, but it is not correct to claim that every GB300 OEM rack has a higher nominal draw than every GB200 rack: HPE publishes 132 kW nominal for both, with different peak-design values.

For a new facility, size from the selected OEM's maximum or peak design requirement plus the agreed redundancy and growth margin—not from a generic per-GPU estimate. Confirm the CDU topology and coolant envelope for the same bill of materials. Full compute datasheet: GB300 NVL72 specs.

A single NVL72 rack is a planning unit, but a cluster is many of them plus the network fabric, storage, and support systems around them — and the facility has to be sized for the whole thing, not one cabinet. The end-to-end arithmetic (IT load, power-usage effectiveness, and what it means for the site) is walked through in how much power an AI data center needs.

Two adjacent decisions interact with the power-and-cooling plan. First, rack-scale vs discrete nodes: an NVL72 concentrates the load and the cooling into one integrated cabinet, versus a fleet of air- or liquid-cooled HGX nodes you rack yourself — the trade-off is laid out in GB300 NVL72 rack vs HGX nodes. Second, the network fabric adds its own load and its own sizing questions (InfiniBand vs Ethernet for AI). To browse the integrator builds themselves, see the GB200 and GB300 catalog.