How much power does AI need?

It depends on the facility, but the ranges commonly discussed in the industry run like this: traditional enterprise data centers draw single-digit megawatts; hyperscale cloud facilities run in the tens to low hundreds of megawatts; and the AI campuses now being planned sit in the hundreds of megawatts, with the largest heading toward a gigawatt and beyond. A gigawatt-class campus consumes power on the scale of a major city — delivered to a single customer, at a single site, around the clock.

The orders of magnitude

The tiers are less about square footage than about what the racks are doing. An enterprise facility serves one company's IT. A hyperscale facility serves the cloud. An AI campus serves training and inference for frontier models — and it is the training clusters, which want as many accelerators as possible in one electrically contiguous place, that pushed site sizes from large building to industrial complex.

These campuses also arrive in phases. A site announced at gigawatt ambition typically energizes its first block of tens or hundreds of megawatts early, then grows as halls fill — which means the power supply has to scale with it, on the same ground, without going back to the end of any line.

Why AI bent the curve

Legacy CPU racks draw power in the single-digit-kilowatt class. GPU racks built for AI training draw an order of magnitude more — tens of kilowatts and climbing with each hardware generation. Multiply that density across thousands of racks, add the cooling required to remove the same energy as heat, and facility-level demand moves up in step. The computer got denser, so the building became a power project.

The shape of the load matters as much as the size

AI load is firm and flat. Training runs continuously for weeks at a stretch, and the capital cost of idle accelerators makes downtime punishingly expensive, so operators hold utilization — and therefore load — as high and as steady as they can. That shape wants dispatchable, always-on generation behind it, a very different profile from loads that can flex around whatever supply happens to be available. Sizing a supply for AI means sizing for the flat top of that curve, every hour of the year.

Where the megawatts come from

Blocks of this size are arriving faster than grid interconnection can absorb them — the ERCOT queue, explained covers why — which is pushing developers toward dedicated generation built together with the load. What is a power foundry? describes that model, the cost math compares it with grid service, and why is Waha gas so cheap? explains where the fuel comes from.

About Corley Energy

Corley Energy is a behind-the-meter independent power producer, founded in 2024 by Jake Corley, Tim Bozeman, and Mark Meyer. We convert stranded Permian Basin natural gas into firm, contracted electricity for AI data centers at Power Foundry, our ~1,000-acre development in Upton County, Texas. Start with what a power foundry is, see the company facts, or check current capacity on the Sites page.

Keep reading

Behind-the-meter power generation, explained · Can flared gas power data centers? · Can AI data centers be flexible loads? · Browse the full library