Power, cooling and the building in front of you

The constraint on serious GPU infrastructure has moved. Floor space still matters, and so does the network, but the conversation that stops a project is usually power and heat. A design that fits the white space and misses the switchboard is not a design. It is a delay.

This briefing collects published figures for current high-density systems and turns them into questions for the building you already have. The figures differ by vendor and by configuration. That spread is the point. Use the number that belongs to the bill of materials in front of you, and confirm it with the facility and the utility. Do not average a brochure into a load schedule.

What the vendors have published

Start with a single accelerator, because the rack number is just the consequence.

NVIDIA’s H100 product page (opens in a new tab) lists a maximum thermal design power of up to 700 W, configurable, for the H100 SXM, and 350 to 400 W, configurable, for the H100 NVL card. The same table shows why the form factor belongs in the brief: memory, NVLink bandwidth and server options change with it. An air-cooled PCIe card and an SXM module in a dense tray are different facility problems, even when the marketing name is the same generation.

Rack-scale Blackwell systems move the decimal point. These are published figures for specific products, not a universal constant:

Source	Published figure	What it refers to
NVIDIA DGX GB200 user guide (opens in a new tab)	About 120 kW	Rack power consumption stated for the documented DGX GB rack
HPE, NVIDIA GB200 NVL72 (opens in a new tab)	132 kW	HPE’s figure for that rack, of which 115 kW is liquid cooled and 17 kW is air cooled
Supermicro GB200 NVL72 datasheet (opens in a new tab)	125 kW to 135 kW operating	Supermicro’s published operating range, with power shelves listed at 132 kW total

Read the three rows together. A planner who carries “120 kW” into a room, and a planner who carries “132 kW”, are both quoting a real document. They are not quoting the same configuration. Redundancy makes the gap wider: the NVIDIA guide describes multiple power shelves arranged for N+N redundancy, so the nameplate of the shelves is not the same number as the expected draw of the compute. Ask the vendor for the expected operating load, the peak, and the power-shelf arrangement, in writing, for the exact BOM.

Liquid cooling is part of these designs because the heat flux is the design. NVIDIA’s NVL72 material describes a liquid-cooled rack. HPE splits its 132 kW figure into a liquid-cooled share and an air-cooled share. Supermicro’s datasheet discusses coolant distribution and liquid-to-air options for sites without facility water. The practical translation: “we will air-cool it like the last generation of enterprise racks” is not a plan for this class of machine. Some auxiliary load may remain on air. The GPU domain, at this density, is a liquid problem unless a vendor’s specific configuration says otherwise.

The global number, and why it will not size your room

The International Energy Agency’s Energy and AI (opens in a new tab) report estimates that data centres used around 415 terawatt-hours in 2024, about 1.5% of global electricity. In the IEA’s base case, that demand roughly doubles to around 945 TWh by 2030. The same report notes that the local effect is sharper than the global share, because the load concentrates.

A companion section on energy demand from AI (opens in a new tab) describes the share of cooling in a data centre’s electricity. It ranges from about 7% in efficient hyperscale sites to over 30% in less efficient enterprise facilities. Cooling is not a rounding error, and it is highly sensitive to how the room was built.

Use those figures as context for a board paper. Do not use them as a diversity factor for a switchboard in an Australian suburb. Your constraint is the feeder, the transformer, the generator or the contract with the utility, the heat-rejection plant, and the planning conditions on the site. A global terawatt-hour does not tell you whether this room can accept one 120 kW rack, let alone a row of them.

What “site readiness” actually means

Readiness is a joint review by facilities and IT. It is a set of measured facts.

Utility capacity. What power is contracted, what is spare, and what a request for more would take in months and in capital. Include redundancy. A rack with N+N power shelves wants two paths that each survive the loss of the other. If the building has a single path, the architecture has to say so, and the availability claim has to follow.

From the street to the rack. Switchboards, distribution, the busbar or the whip, and the connector the rack expects. A 50 V DC busbar, which the NVIDIA guide describes for the DGX GB rack, is a different installation from a room of conventional power distribution units. The electricians need the vendor’s installation manual, not a summary.

Heat rejection. Where the heat goes after it leaves the rack. Facility water, a coolant distribution unit, or a liquid-to-air path when the building has no suitable water loop. Each has a capacity, a temperature range and a failure mode. Leaks, condensation and water quality are design items, not commissioning surprises.

Weight, access and the route from the dock. These racks are heavy and deep. The lift, the corridor and the floor rating are part of the schedule. A machine that cannot reach the tile is a storage problem.

Containment and failure. What happens when a pump stops, a hose leaks, or a power shelf drops. The room needs a response that facilities staff can actually perform, and a monitoring path that reaches the people on call.

Network entry. The fabric inside the pod is only half the drawing. The pathways into the room, the demarcation, and the latency to the systems the factory must reach belong on the same page.

Commissioning tests. Agree them before delivery. Power draw against the design, coolant temperatures and flow, leak checks, fabric health, storage throughput, and one failure case (loss of a power path, loss of a pump) that you are willing to perform. Write the results down. That record is the handover.

Questions to put in the same meeting

Facilities and the infrastructure team should be in the room together. Split meetings produce split numbers.

• What is the spare capacity today, in kilowatts, on each path, after the loads you have already promised to other projects?
• What is the lead time for more capacity, and who pays for the upstream work?
• Is there a heat-rejection path with a stated capacity, or only a hope that the existing computer-room air conditioning will cope?
• If the design is liquid-cooled, who owns the secondary loop, the water chemistry and the leak response?
• What floor loading and access route has a structural engineer already accepted?
• Which monitoring system will see both the IT telemetry and the facility telemetry, and who is paged?
• What will commissioning prove, and who signs?

If the answers are “we will find out”, the next spend is a survey, not a GPU order.

How this changes the business case

A business case that prices servers and omits the facility will understate both cost and time. The facility work can dominate the schedule: utility upgrades, structural checks, and heat-rejection plant do not arrive on a server lead time. Put them on the critical path in the first version of the plan.

It also changes the operating cost. The IEA’s range for cooling share is a reminder that two rooms with the same IT load can have very different electricity bills. Power usage effectiveness is a useful conversation if you measure it on this site. A number copied from a hyperscale case study is not a forecast.

For the architecture that sits on top of this, see What an AI factory is, and how to plan one. Dynacore’s work on the facility interface is part of the AI Factory practice: the brief, the readiness note, and commissioning against the design. If you have a building and a workload and you want those two facts in one conversation, book a briefing.