Power Is the Bottleneck: A Back-of-Envelope Model for AI Data Center Capacity

When people talk about AI capacity, they usually count GPUs. The people building it count megawatts. GPUs can be ordered; utility power, substations, and cooling capacity take years. So the useful question isn't "how many GPUs can we buy?" It's "how many GPUs can this site actually run?"

Here's a simple model I use to get from a site's power to a GPU count. It's deliberately rough. The goal is to get within reach of the right answer and to make every assumption visible so you can change it.

Step 1: from utility power to IT power

A site's headline number, say 100 MW, is usually the power it can draw from the grid. Not all of that reaches the servers. Cooling, power conversion losses, lighting, and everything else eat into it.

The ratio is PUE (power usage effectiveness): total facility power divided by IT power.

A modern liquid-cooled AI facility might run around 1.1 to 1.3.
Older air-cooled facilities often sit around 1.4 to 1.6.

At a PUE of 1.25:

100 MW ÷ 1.25 = 80 MW of IT power

One caution: PUE varies with weather and load. Size for the hot summer day, not the annual average.

Step 2: carve out non-GPU IT load

Not all of the IT power goes to GPU racks. Storage, scale-out networking (switches and optics), management, and CPU-only services take a share. For a training-focused cluster, I usually start by assuming 10 to 15% goes to that supporting infrastructure.

At 12%:

80 MW × 0.88 ≈ 70 MW for GPU racks

Step 3: decide how much headroom to keep

Racks are rated by their maximum power, and you can't plan to run the facility at 100% of its electrical capacity all the time. Redundancy schemes (N+1, 2N, or distributed redundancy) and operating margin take a share. Some operators plan very close to the limit and rely on power capping; others keep a real buffer.

For a first pass, I hold back about 10%:

70 MW × 0.9 = 63 MW of usable GPU-rack power

Step 4: convert to racks and GPUs

Now pick a platform and use its per-rack or per-node power.

GB200 NVL72, at roughly 130 kW per rack:

63,000 kW ÷ 130 kW ≈ 485 racks → 485 × 72 ≈ 35,000 GPUs

HGX H100, at roughly 10.2 kW per 8-GPU server:

63,000 kW ÷ 10.2 kW ≈ 6,175 servers → ≈ 49,000 GPUs

The H100 build gets more GPUs out of the same power, but that's the wrong comparison. Each Blackwell GPU does considerably more work, and the NVL72 rack puts 72 of them in one NVLink domain. The fairer way to compare is useful work per megawatt for your actual workload, measured as tokens per second for inference or time to train for training. The power model tells you how much you can deploy. Benchmarks tell you what it's worth.

Step 5: check the floor and the cooling

Power isn't the only constraint. Two others tend to show up right after you do this math:

Rack density versus floor space. 485 NVL72 racks at 130 kW is a much smaller footprint than 6,000 air-cooled servers, but only if the building can deliver that much power per row. The busway, PDUs, and switchgear have to be sized for density, not just total megawatts.
Heat rejection. Nearly all of that 63 MW turns into heat. With direct-to-chip liquid cooling, most of it goes into the liquid loop and out through dry coolers or chillers; the rest goes to air. The facility's heat-rejection capacity, at its worst-case outdoor temperature, has to match the IT load you just planned.

The whole model on one line

GPUs ≈ (site MW ÷ PUE) × (1 − non-GPU share) × (1 − headroom) ÷ (kW per rack) × (GPUs per rack)

For 100 MW, PUE 1.25, 12% non-GPU, 10% headroom, and NVL72 racks at 130 kW, that comes out to roughly 35,000 GPUs. Change any input and the answer moves a lot, which is the point: the assumptions are where the real conversation is.

Why this matters for the economics

This is why AI infrastructure deals are increasingly priced and financed by the megawatt, not by the GPU. The GPUs depreciate in a few years and get refreshed. The powered, cooled building and its grid connection are the scarce, long-lived asset. A site with power available in the next 12 months is worth more than a site with more power available in four years, because GPU generations turn over faster than substations get built.

If you're evaluating a site, a provider, or a deal, start with the power: how much is really there, when it arrives, and what PUE and density it can support. The GPU count falls out of that.