Colocation sites for GB300 NVL72 deployments.

What arrives: one rack, already built

NVIDIA’s rack-scale user guide lists the parts: 18 compute trays of one rack unit each, every one holding two Grace CPUs and four Blackwell GPUs; 9 NVLink switch trays with two NVLink switches each; an NVLink backplane of passive copper cable cartridges that joins the trays up and down the back of the rack; eight power shelves; a bus bar; and liquid cooling manifolds. NVIDIA’s NVL72 reference architecture adds two in-rack out-of-band management switches, powered from the bus bar. The rear of the rack carries the cable management, the manifold inlets and outlets, the cable cartridges and the bus bar; the network ports sit at the front, on the cold-aisle side.

None of this is assembled on site. The hall has to accept the rack as one unit, which drives everything that follows.

On names: GB300 NVL72 is the rack design. NVIDIA sells its own version as DGX GB300, and its NVL72 reference architecture delivers the rack as a pre-configured system through an OEM; server makers build their versions on NVIDIA’s MGX rack design. The figures in this guide are NVIDIA’s own.

Cooling: liquid to the chips, air for the rest

NVIDIA calls the rack fully liquid-cooled, and its DGX GB300 reference architecture describes the detail as hybrid: the most power-intensive components, the GPUs and CPUs, are liquid-cooled, and the other components are air-cooled. Coolant runs up and down the rack through manifolds, then through cold plates on the CPUs and GPUs; each compute tray has a cold-water inlet and a hot-water outlet. Networking, storage and the air-cooled power supplies are cooled by fans. Leak detection is built in at tray level and rack level.

Air cooling cannot host this rack at all. NVIDIA’s explanation of why liquid is used applies directly to the site: direct-to-chip cooling moves heat into a technology cooling loop and on to facility water, and because that transfer is efficient the facility can run warmer water, which reduces or removes the need for mechanical chillers in many climates. The site therefore needs a coolant distribution unit (CDU) between the rack loop and the facility loop, and a facility loop that is installed, not planned.

Liquid

A CDU and facility water loop sized for the racks you will install, with the supply temperature and flow rate guaranteed in writing. Warmer supply water saves chiller energy but needs more flow, so the plant must be sized for it.

Air

The hall does not become air-free: fans still push air through the trays and power supplies. Ask the vendor how much heat leaves by air and what inlet air it needs, and keep hot-aisle containment.

Leaks

The rack detects leaks; the site responds to them. Ask how a leak is isolated, who is on site to do it, how fast, and how coolant is stored, handled and disposed of.

Redundancy

A cooling loss at this density leaves very little time. CDU and pump redundancy, and the liquid loop on backed-up power, matter more than generator count.

Rack density: 72 GPUs, one rack, eight racks a pod

The density is fixed by the product. One rack holds 72 GPUs and 36 CPUs in 18 compute trays, with 9 switch trays and 8 power shelves, all one rack unit high. NVIDIA’s DGX SuperPOD reference architecture builds in scalable units of eight racks, 576 GPUs, and shows the design scaling to 128 racks and 9,216 GPUs. Its enterprise NVL72 reference architecture treats a single rack as a unit and tests up to eight of them.

• There is no "fewer GPUs per rack" option: the rack is the unit, and a pod is eight of them side by side.
• Rack-to-rack cable limits set how far apart the racks of a pod can stand, so a pod wants one contiguous row with its network racks next to it.
• Each DGX SuperPOD unit also brings management nodes, network racks and storage, which need their own rack positions, power and cooling.

Power: up to 142 kW per rack

NVIDIA’s NVL72 reference architecture states a full GB300 NVL72 rack requires up to 142 kW. It is fed by eight power shelves of 33 kW, each holding six air-cooled 5.5 kW power supplies; NVIDIA’s rack-scale guide describes the shelves as N+N redundant. For a DGX SuperPOD, NVIDIA budgets a thermal design power of 1.2 MW per eight-rack unit, which covers the racks and the networking, management and storage that go with them.

NVIDIA’s DGX GB300 power shelf carries larger energy storage than the previous generation, specifically to reduce the peak load the rack puts on the data-centre supply. That smooths swings inside the rack; it does not lower the 142 kW the feed must be able to deliver.

What the site commits to

Up to 142 kW per rack delivered to the rack, with the busway and panel sized for the peak. Ten racks is up to 1.42 MW of IT load before cooling plant and networking.

Power quality

A synchronised cluster swings its load as one. Ask how much load swing the site’s interconnection tolerates and what energy storage or ramp limits sit upstream of the rack.

Billing

Ask whether power is billed on committed, peak or metered draw. A contract priced on peak penalises a rack that NVIDIA itself describes as smoothing its peaks.

Infrastructure: distribution, networking, plant and facility

Power distribution

Power whips from a remote power panel energise the eight power shelves, which convert AC to a nominal 50 to 51 V DC and distribute it through the bus bar to every tray and switch. A site needs many high-amperage whips per rack, from panels that are themselves redundant and concurrently maintainable.

Networking

Three layers: NVLink inside the rack, so 72 GPUs act as one; InfiniBand or Spectrum-X Ethernet between racks as the compute fabric, with 800 Gb/s per GPU through ConnectX-8 SuperNICs (four dual-port adapters per tray); and Ethernet through a BlueField-3 DPU per tray for storage, in-band management and external traffic. NVIDIA’s enterprise design uses 64-port 800 Gb/s switches for the fabrics and a 48-port switch per rack for out-of-band management. Fibre pathways, meet-me-room density and management addressing are planned before delivery, not after.

Cooling plant

CDUs between the rack manifolds and the facility water, the facility loop with its pumps and heat rejection, leak detection and coolant handling, and the air side for the fans. All of it installed and commissioned before the rack arrives.

Facility

NVIDIA’s DGX SuperPOD reference architecture says the data centre should meet or exceed Uptime Institute Tier 3, or TIA-942-B Rated 3 or EN 50600 Availability Class 3, with concurrent maintainability and no single point of failure. NVIDIA’s guide also requires the rack to be anchored to an unmovable support and moved with mechanical assistance. Take the rack’s weight and dimensions from your vendor and have the floor checked for the point loads under it; slab floors are preferred.

Operations

NVIDIA’s DGX GB300 systems are run with NVIDIA Mission Control, and NVIDIA’s Infrastructure Specialist team offers data-centre planning, deployment and bring-up help. Ask the operator what its remote hands can do on a liquid-cooled rack, and where its responsibility ends and yours begins.

Deployment: a liquid-cooled hall, or a new one

An ordinary air-cooled hall cannot take this rack. It needs a facility water loop and CDUs, DC-fed power shelves on whips from a remote panel, up to 142 kW per rack position, a floor that carries a fully loaded rack delivered in one piece, and a Tier 3 facility around it. NVIDIA writes that the DGX SuperPOD design stays flexible so it can be tailored to integrate into existing data centres, but every item on that list has to exist before delivery.

In practice there are three paths. A hall that already runs liquid-cooled GPU racks, which NVIDIA’s DGX-Ready programme marks with its Liquid Cooled Ready designation. A hall being retrofitted with a liquid loop, where the schedule for the loop, the CDUs and the power upgrade decides your delivery date. Or a new build. The rack itself is delivered pre-configured through the OEM.

• Ask whether the liquid loop, CDUs and power are installed today, contracted, or planned, and get the dates in writing.
• Ask for the DGX-Ready certificate and its date; it tells you the operator has deployed liquid-cooled NVIDIA racks before.
• Ask who commissions the loop, who owns coolant quality, and who is on site when a leak alarm fires.

GB300 NVL72 or B300 servers?

The same Blackwell Ultra silicon also arrives as 8-GPU air-cooled servers, which far more existing halls can host. If your site shortlist is mostly air-cooled, read the B300 server guide before you commit to the rack.