What is a data center?

A data center is a building that supplies three things to computers: electricity, cooling and network. The computers sit in racks in rows, with a cold aisle on one side and a hot aisle on the other, and everything about the building exists to keep them running.

Traditional facilities are measured in floor area. AI facilities are measured in megawatts, because power is what runs out first.

Where the power goes

An AI accelerator draws 700 to 1,400 watts. Its server, networking and storage add roughly another 400 watts per chip. So a hundred thousand accelerators is about 140 megawatts of computing load.

On top of that comes the overhead for cooling and power conversion, expressed as PUE, power usage effectiveness. A PUE of 1.2 means the building draws 20 percent more than the computers consume. Modern sites reach 1.1; older ones run at 1.5 or worse.

That hundred thousand chip cluster therefore draws around 170 megawatts continuously, roughly the consumption of a city of several hundred thousand people. Our power calculator does this arithmetic with your own numbers.

Why cooling changed

Air cooling works up to roughly 30 to 50 kilowatts per rack. Modern accelerator racks exceed 100 kilowatts, which air physically cannot remove from that volume. So the industry moved to liquid: cold plates on the chips, or in some designs full immersion.

Liquid is more efficient and needs plumbing in the computer room, which is a habit the industry did not have. It also changes the water question completely: closed-loop liquid cooling consumes very little water, while evaporative cooling consumes a great deal. A water figure quoted without the cooling type tells you nothing.

Why they are built where they are

Not near users. Near power. A grid connection for a large site takes years to obtain, transformers have long lead times, and interconnection queues in mature markets run for years. So sites appear where electrical capacity already exists: retired power stations, former industrial sites, places with hydro or cheap gas.

For training workloads, latency to the customer does not matter at all, which is why the map of AI data centers is a map of available electricity.

Training and inference

Training is a large, bounded, visible expense that happens once per model. Inference is smaller per request and runs forever, scaling with the number of users. Inference now consumes more total compute than training, and it is the load that does not go away.