The tour
A big AI datacenter, from the wires in to the heat out.
The building on this page is a real design for a large AI datacenter, the size being built now. A megawatt is a million watts; this one takes a hundred of them. Six short steps, and every number links to its source.
Step 1
One building = about 83,000 homes.
A big AI datacenter takes about 100 MW from the grid, about what ≈82,800 average homes use. It arrives through about ≈60 electrical units along the building's sides, each the size of a shipping container and rated at 2 MW. Five of them share the work of four, so one can fail without anything noticing.
Bringing that much power to one place means new lines and sometimes new power plants, and your state's utility commission decides who pays for them. The story page answers "will it raise my bill?", with the figures and what to ask.
Step 2
One cabinet = about 120 homes, all of it becoming heat.
Inside are 552 racks, each holding 72 chips. A cabinet draws about 145 kW, as much as about 120 homes, and every bit of it becomes heat, the way a space heater turns electricity into warmth. That's the whole problem a datacenter has to solve: getting the heat out.
Step 3
A sealed loop = 93% of the heat carried away.
In this building, liquid runs right past the chips and carries the heat away, the way the coolant in a car carries heat from the engine to the radiator. It picks up about 93% of the heat that way. The liquid never leaves the pipes.
Step 4
Air instead of water = almost no water, a little more electricity.
The old way: evaporate water. A cooling tower works like sweat, and it uses water on site, sometimes a lot, especially somewhere hot: older buildings average 1.8 to 1.9 liters of water for every kilowatt-hour the computers use, about half a gallon. The other way: blow air over big radiators, like the fan on the back of a car. That uses almost no water at the building, but takes roughly a third to three-quarters more electricity to do the same job (0.30 to 0.74 kWh/kWh). Same company, same computers: one building in Jakarta uses 140× the water of one in Stockholm. The difference is this choice and the weather.
AWS, own disclosure 0.02 L/kWh in Stockholm, 2.85 L/kWh in Jakarta.
Step 5
Four drops in five = at the power plant, not the building.
Power plants evaporate water too, to make electricity for everything. For a typical modern AI building, 81% of the water it's responsible for is at the plant, not the building: about half a gallon for every kilowatt-hour on an average US grid (1.8 L/kWh). So the cooling choice on the last step decides the one drop in five. The electricity decides the other four.
The power plant. Its own cooling towers evaporate water to make electricity for everything, your TV included.
The wires. Often hundreds of miles of them.
The datacenter. One drop in five, for cooling, in a typical modern building.
Step 6
The whole staff = 24 to 30 people.
About 3.75 football fields under one roof, on ≈140 acres, run by 24 to 30 around the clock. Their sinks and bathrooms use about 528 gallons a day, far less than a drop each across the ≈1.19 billion people the building could serve. That's a datacenter. The drawings are below for anyone who wants them, and the sources page has every number.
The drawings
The drawings, for the engineers: section, plan and equipment list
| Tag | Component | Qty | Rating | Notes |
|---|---|---|---|---|
| IT-1 | GB300 NVL72 rack | 552 | 145 kW average | 72 GPUs per rack. 39,744 GPUs total. Direct liquid cooling, 93% capture. |
| CH-1 | Air-cooled chiller | ≈60 | 630 tons | N+1, six blocks of ten, integrated free cooling. Rejects heat to air. No evaporative makeup. |
| CDU-1 | Coolant distribution unit | per hall | facility loop, 32 °C supply | Isolates the chip loop from the facility loop. Free cooling available at 32 °C supply for most of the year. |
| E-1 | Modular electrical line-up | ≈60 | 2 MW | Distributed redundant, 5-to-make-4. Peak design PUE 1.35, annualized 1.20 PUE. |
| W-1 | Domestic water | 1 | ≈2,000 L/day | Whole building, 24 to 30 people. 0.7 µL per user-day. |
Peak design PUE 1.35 PUE, annualized 1.20 PUE. Liquid captures 93% of the heat at the chip. ≈60 line-ups at 2 MW, 5-to-make-4.