GB200 & GB300 NVL72 Power and Cooling Requirements

NVIDIA and its integrators commonly describe the GB200 NVL72 as an approximately 120 kW rack. That's the right order of magnitude to plan a facility around, and it's the single fact that reframes everything else: a conventional enterprise server rack runs in the single-digit-to-low-tens of kilowatts, so an NVL72 concentrates roughly ten racks' worth of power into one cabinet.

The arithmetic starts at the GPU. Each Blackwell GPU draws roughly 1,200 W, and there are 72 of them in the rack — before the 36 Grace CPUs, the NVLink switch trays, and power-conversion losses. Add those and the ~120 kW rack figure follows.

Treat 120 kW as a planning figure, not a spec. Reported draw on deployed racks varies with the integrator build and configuration, and the exact site power and rack weight are confirmed at quote. What doesn't vary is the class of the problem: this is a high-density rack that most existing data-center halls were never designed to feed or cool. For the full datasheet, see the GB200 NVL72 specs.

There is a rough industry line — around 30 kW per rack — above which air cooling stops being practical. A GB200 NVL72 sits at roughly four times that. You cannot move enough air through a cabinet to carry ~120 kW of heat away from the chips, so direct-to-chip liquid cooling is mandatory, not an optimization.

That's a structural change, not a tuning knob. Where an air-cooled hall relies on hot-aisle/cold-aisle airflow and CRAC/CRAH units, a liquid-cooled NVL72 rack moves the heat in coolant piped directly across cold plates on the Grace CPUs, Blackwell GPUs, and NVLink switches. Liquid cooling is what makes the density thermally viable at all — it is the enabling technology for the whole rack-scale design, which is why the NVL72 ships factory-integrated with its cooling built in rather than assembled on site.

The rack is served by an in-rack coolant distribution unit (CDU) running direct-to-chip liquid cooling. The CDU is the heart of the thermal design and it separates two loops:

• The technology (secondary) loop circulates coolant through the cold plates on the chips inside the rack.
• The facility (primary) loop is your building's water, which carries the heat away to the outside world through the CDU's heat exchanger.

The practical consequence is that an NVL72 deployment is a facility-water project, not just a power project. The site has to supply and return water at the right temperature and flow to the CDU, which means plumbing, heat rejection (dry coolers, cooling towers, or chillers), and the floor loading and leak management that liquid in the white space implies. Some integrator builds specify a CDU rated on the order of 250 kW of cooling capacity — sized with headroom above a single rack's heat load. This is the part of a Blackwell buildout that most often gates the schedule, because it touches the building, not just the IT.

Inside the rack, power is delivered differently from a conventional server cabinet. Rather than each server carrying its own PSUs fed from rack PDUs, the NVL72 uses rack-level power shelves that convert facility input and feed the compute and switch trays over a common busbar running the height of the rack.

That design is what lets a single cabinet distribute ~120 kW cleanly to 18 compute trays and 9 NVLink switch trays. For the facility side, it means the rack needs high-density power distribution — the upstream feed, breakers, and redundancy have to be sized for a load that behaves more like a small power system than a row of servers. As with cooling, the exact input configuration depends on the integrator build and is confirmed at quote. The rack topology that these power shelves feed is broken down in the GB200 NVL72 specs.

The Blackwell Ultra GB300 NVL72 keeps the same rack-scale shape — 72 GPUs, 36 Grace CPUs, one NVLink domain, liquid cooling, busbar power — but raises the thermal and power envelope.

Each Blackwell Ultra GPU draws roughly 1,400 W, up from the GB200's ~1,200 W, so the rack is meaningfully denser than the GB200's ~120 kW. Some GB300 integrator builds specify an in-rack CDU rated around 250 kW of cooling capacity to carry it. Everything about the GB200's requirements applies here, only more so: liquid cooling is still mandatory, facility water is still the gating item, and power distribution has to be sized higher. If you're planning a new rack-scale buildout, size against the GB300 envelope rather than the GB200's. Full 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.