A grounded look at the actual power numbers behind an AI data center — from single-rack draw through hall-scale and campus-scale totals — and why total IT load, not building square footage, is the number that actually drives design.
Power, Not Floor Space, Is the Real Design Constraint
For a traditional data center, usable capacity was often described in square footage or the number of racks a building could hold. For an AI data center, the binding constraint is almost always electrical power — a facility can have plenty of physical floor space left and still be completely full from a power-delivery standpoint, because GPU racks draw so much more power per square foot than traditional servers ever did. This is why AI data center capacity is described in megawatts of IT load, not square feet, and why utility interconnection capacity (not land or building size) has become the real bottleneck on how much a given site can host.
The Numbers, from Rack to Campus
At the individual rack level, a modern GPU server rack can draw anywhere from roughly 40 kW to well over 130 kW depending on the GPU generation and how densely servers are packed — see How Much Power Does an AI GPU Rack Use? for the detail behind that range. Scale that up: a single data hall housing a few hundred racks can draw tens of megawatts, and a full hyperscale AI campus with multiple halls can reach several hundred megawatts to over a gigawatt of total IT load. For comparison, a large traditional enterprise data center of similar physical footprint might have drawn a small fraction of that — the increase is driven almost entirely by GPU power density, not by the facility being larger.
Total Facility Power vs. IT Load
The power a facility actually draws from the utility is higher than its IT load alone, because cooling, power conversion losses, and other overhead all consume additional power on top of what reaches the GPUs. This relationship — how much extra power the supporting infrastructure requires per watt of IT load — is captured by the PUE (Power Usage Effectiveness) metric, covered in the studio's PUE and cooling efficiency article, with a hands-on PUE Calculator to work through real numbers. A facility with 100 MW of IT load and a PUE of 1.2 draws 120 MW total from the utility — the extra 20 MW is cooling, power conversion losses, and other overhead, not compute.
Why the Numbers Keep Climbing
Successive generations of AI accelerator hardware have consistently drawn more power per chip than their predecessors, even as they've also become more computationally efficient per watt for a given workload — the net effect on facility-level power demand has been upward, not flat, because the industry has generally chosen to deploy more, denser compute rather than hold total facility power constant. This trajectory is a major reason data center electrical design increasingly plans for headroom well beyond a facility's initial GPU generation, and why utility interconnection requests for new AI facilities have grown so large that they're now a meaningful line item in regional grid planning discussions.
Putting a Number on "How Much"
There's no single correct answer to how much power an AI data center needs — it depends entirely on the number of racks, the GPU generation deployed, the cooling architecture chosen, and the facility's redundancy design. But as a working frame: a single rack today can plausibly draw 40-130+ kW, a mid-size AI data hall can draw tens of megawatts, and the largest hyperscale AI campuses under construction are being designed for several hundred megawatts to over a gigawatt of total capacity — numbers that would have been almost unthinkable for a data center a decade ago.