Why AI data centers use so much power and water
AI data centers turn almost all the electricity they draw into heat, and many use evaporating water to get rid of it. Here is how the power and water add up, and the trade-offs between them.
Kiln, covering chips, data centers and power
, 3 min read
~1,600 literswater evaporated to carry away 1 MWh of heat
What to know
- Nearly every watt a data center's chips use ends up as heat that has to be removed.
- AI racks pack many times more power into the same floor space than ordinary server racks, which pushes operators toward liquid cooling.
- Cooling by evaporation saves electricity but consumes water; closed-loop systems save water but use more electricity.
Fights over new data centers keep coming back to two numbers: electricity and water. Both come from the same physical fact. Computer chips turn electricity into heat, and the heat has to go somewhere.
Where the power goes
A data center's electricity has two jobs. Most of it runs the IT equipment: processors, memory, storage and network switches. The rest runs everything that keeps that equipment alive, above all cooling, plus power conversion losses and lighting.
The industry measures the split with a ratio called power usage effectiveness, or PUE: total facility power divided by the power used by the IT equipment. A PUE of 1.0 would mean no overhead at all. Older server rooms often ran above 1.5; large modern cloud facilities commonly report figures closer to 1.1.
Almost none of the electricity used by the chips leaves the building as anything but heat. So a data center that draws 100 megawatts is, in effect, a 100-megawatt heater that has to be cooled around the clock.
Why AI made it worse
AI accelerators are far more power-hungry than the processors in a typical server. A single data center GPU from Nvidia or AMD can be rated at around 700 watts to more than 1,000 watts, and servers hold eight or more of them, alongside the high-bandwidth memory described in our HBM explainer.
Packed into racks, that adds up quickly. A conventional server rack might draw somewhere in the range of 5 to 15 kilowatts. Nvidia's rack-scale systems built around its Blackwell chips are designed to draw well over 100 kilowatts per rack. That is the difference between cooling a room and cooling something closer to an industrial furnace.
Air cannot carry heat away fast enough at those densities, which is why new AI halls pipe liquid directly to cold plates on the chips, and why cooling-equipment makers have become acquisition targets, as in Samsung's reported talks to buy Modine.
Scale matters too. An average US home uses roughly 1.2 kilowatts on average over a year. A single 100-megawatt campus therefore draws about as much as 80,000 homes, and several proposed AI campuses are planned in the gigawatts. Utilities often cannot add generation and transmission that fast, which is why developers sign long-term power deals with utilities or build their own gas generation.
Where the water goes
Water use depends on how the heat is finally released outdoors.
Evaporative cooling. Many large data centers send warm water from the building to cooling towers, where some of it evaporates. Evaporation is very effective at moving heat: turning one liter of water into vapor absorbs about 2.26 megajoules, or roughly 0.63 kilowatt-hours of heat. Carrying away one megawatt-hour of heat by evaporation alone therefore takes on the order of 1,600 liters of water.
Closed-loop and dry cooling. Other designs keep the water sealed in a loop and dump heat to the air through large radiators, sometimes helped by refrigeration chillers. They use little water after the initial fill, but they need more electricity, especially on hot days, which raises the PUE.
Indirect water. Most thermal power plants, whether gas, coal or nuclear, are themselves cooled with water, so a data center's total footprint includes the plants generating its power.
The industry tracks on-site water with a second ratio, water usage effectiveness (WUE): liters of water used per kilowatt-hour of IT energy.
The trade-off
There is no option that saves both. Evaporative cooling is efficient with electricity but draws on local water, which matters most in dry regions where land and sunshine also make sites attractive. Closed-loop cooling protects water supplies but adds electricity demand, and therefore more power plants and more indirect water.
Why it matters
The power and water figures are what local governments and regulators are actually voting on, in measures like Minneapolis's new data center ordinance and in fights over projects such as the Kraft Group's proposal in Sutton, Massachusetts. Knowing that heat is the root of both helps make sense of a developer's claims: a site promising little water use is usually trading it for more electricity, and the reverse.
AI chips and data centersAI law and policy
Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.