How much power does an AI data center need?

Total facility power = (number of racks × kW per rack) × PUE.

Everything else is refinement. The two terms that matter are rack density, which has moved by an order of magnitude in five years, and PUE, which converts the IT load into what the site actually draws from a generator or a utility.

Two habits separate a usable number from a wrong one. First, size at the rack, not at the GPU — a GPU is 60–75% of a node’s power, and the rest is real. Second, size for the peak and the rate of change, not the average, because AI clusters do not behave like enterprise workloads.

Rack density is the unit of account for a data hall, and AI has broken the historical bands.

• Conventional enterprise racks: 3–10 kW. Air-cooled, standard hot-aisle containment.
• General high density: 10–30 kW. This is where the industry average has been slowly migrating.
• Air-cooled AI nodes: an 8-GPU HGX-class server with 1,000 W-class GPUs is roughly 8 kW of GPU silicon and 11–12 kW at the node. Four such nodes is already a 45 kW rack, near the practical ceiling for air.
• Rack-scale AI systems: a GB200 NVL72 rack draws on the order of 120 kW. Schneider Electric’s reference design for GB300 NVL72 supports rack densities up to 142 kW. These are liquid-cooled by necessity, not preference.

The inverse — racks per megawatt of IT load — is the number that sizes a room. At 10 kW you fit 100 racks in a megawatt. At 120 kW you fit eight. The floor area per megawatt collapses; the power and cooling density per square foot explodes.

A GPU’s TDP is a thermal design point, not a meter reading. Across the platforms in the GPU catalog, per-GPU TDP runs from 700 W on H100 and H200 SXM parts, to 1,000 W on B200, to about 1,100 W on the 8-GPU HGX B300 baseboard, to 1,400 W on the liquid-cooled AMD MI355X and on the rack-scale GB300 NVL72 superchip. Note the last two: "B300" alone is ambiguous, because Blackwell Ultra ships in two bins and the package decides which — the HGX B300 node runs ~1,100 W per GPU while the GB300 NVL72 rack runs ~1,400 W. Multiplying any of these by GPU count gives you a number, but not the right one, in two different directions.

Upward: GPU silicon is only part of a node. Host CPUs, system memory, NVLink switches, NICs, local storage, fans, and power-supply conversion losses add materially — commonly 20–30% on top of the GPU figure. An 8×700 W node is 5.6 kW of GPU and roughly 10 kW of server.

Downward: real workloads do not pin every GPU at TDP continuously. Dense pretraining comes close for long stretches; inference, data loading, checkpointing, and mixed serving run well below. Measured utilization across a fleet is usually far under nameplate.

But the electrical system is not sized on the average. It is sized on the peak — and on how fast the load gets there.

This is the part that surprises people who have designed conventional data centers.

In a traditional facility, thousands of uncorrelated workloads average out. The aggregate load is smooth, and the electrical infrastructure sees a well-behaved curve.

AI training does the opposite. As NVIDIA describes it, thousands of GPUs operate in lockstep performing the same computation on different data, so the cluster transitions between idle and high-power states together — abrupt, synchronized swings rather than an averaged draw. The grid, meanwhile, responds to sudden demand changes over one minute to ninety minutes, because generation is physical (NVIDIA, 2025).

GB300 NVL72 attacks this at the rack: about 65 joules of capacitive energy storage per GPU to absorb fast fluctuation, ramp-rate-limited power capping on startup so increases match what the grid tolerates, and power-burn hardware so job shutdown is a taper rather than a cliff. NVIDIA reports roughly a 30% reduction in peak power demand on a Megatron LLM training run with these features engaged.

The design consequences are concrete: distribution and generation must be sized for the step, protective coordination must not trip on a synchronized ramp, and a multi-unit generation plant that can add and shed units follows this profile far better than a single large machine. That trade is covered in gas turbine vs reciprocating engine.

PUE — power usage effectiveness — is total facility power divided by IT power. A PUE of 1.5 means that for every watt reaching a server, half a watt goes to cooling, distribution losses, lighting, and everything else.

The industry has not improved much. Uptime Institute’s 2025 global survey put the weighted average annual PUE at 1.54, effectively flat for six consecutive years (Uptime Institute, 2025). Purpose-built liquid-cooled AI halls do considerably better — 1.1 to 1.25 is a realistic modern target, because direct-to-chip liquid removes most of the fan and chilled-air energy.

The difference is not a rounding error. A campus with 100 MW of IT load draws about 154 MW at the industry average and about 115 MW at 1.15. That ~39 MW gap is roughly a whole aeroderivative turbine package, or a large block of engines, or the difference between a grid connection that clears and one that does not.

When someone quotes a data center in megawatts, always ask whether they mean IT load or facility load. Developers usually quote facility (it is the utility ask); operators usually quote IT (it is what they sell). The two differ by 15–55%.

Essentially all electrical energy delivered to compute leaves as heat. That makes the cooling load a direct function of the IT load, and it is the dominant term in PUE.

The conversions worth memorizing: 1 kW is 3,412 Btu/h, and 12,000 Btu/h is one ton of refrigeration. So 1 MW of IT load is about 284 tons of heat rejection. A 60 MW hall needs roughly 17,000 tons.

Air cooling runs out somewhere between 30 and 50 kW per rack — beyond that you cannot move enough air through the cabinet at a tolerable temperature rise and fan power. Direct-to-chip liquid cooling with in-rack coolant distribution units is what makes 120 kW-plus racks thermally possible at all, which is why every rack-scale AI platform assumes it.

Liquid does not reduce the heat; it changes where the energy goes to move it. Fan and chilled-air energy drops sharply, which is the main reason liquid-cooled halls hit lower PUE. In exchange you take on a facility water system, CDU redundancy, water quality and treatment, and — often the binding constraint on a site — a water permit. Where water is scarce, closed-loop and dry-cooler designs trade some efficiency back for consumption.

Take a hall of 500 rack-scale AI racks at 120 kW.

• IT load: 500 × 120 kW = 60 MW.
• At a well-run liquid-cooled PUE of 1.2: 72 MW facility load.
• At the industry-average 1.54: 92 MW facility load.
• Heat rejection at 60 MW IT: roughly 17,000 tons.

Now a smaller air-cooled comparison. 1,000 H100-class GPUs is 125 eight-GPU nodes: 700 kW of GPU silicon, roughly 1.25 MW at the node once hosts and losses are counted, and about 1.5–1.8 MW of facility load at a PUE of 1.2–1.4. A thousand GPUs is not a rounding error on a campus; it is a substation feeder.

What the facility number then determines:

• The interconnection request. You file for facility load, and that number sets your position and study scope in the interconnection queue.
• The generation block count if you are going behind the meter — how many engines or turbines, plus N+1.
• The distribution voltage. Past roughly 10 MW you are distributing at medium voltage, not 480 V, which means MV switchgear line-ups and a unit substation per pod.
• The standby plant. The critical fraction of the IT load, plus the mechanical plant that has to restart with it, sizes the generator sets.

The useful discipline is to carry all four numbers together — GPU count, IT load, facility load, and heat rejection — because a plan that is coherent in one and incoherent in another is the normal failure mode.