From Compute to Heat to Cooling
Essentially all electrical energy drawn by servers leaves the building as heat. Removing that heat is the entire reason a data center touches water at all.
The physical chain
Electricity in, computation done, heat out. AI training and inference run processors at high sustained utilisation, so the heat load per rack is far higher than in a traditional server hall. Higher heat load means a larger cooling system, and the choice of cooling system is what determines whether water is evaporated on site.
Cooling designs that consume water
Evaporative cooling towers and adiabatic (water-assisted) systems deliberately evaporate water to shed heat, because evaporation is thermodynamically cheap compared with running mechanical chillers. Water leaves the site as vapour and is not returned to the local system. Chilled-water plants also carry water losses through the tower side of the loop.
Cooling designs that do not
Air-cooled and closed-loop designs circulate a fixed charge of fluid that is filled once and then recirculated, rejecting heat to the air rather than by evaporation. Direct-to-chip liquid cooling moves coolant to the processor itself, allowing higher operating temperatures and, in a sealed loop, effectively no ongoing water consumption for cooling. The trade is usually more electricity for fans and chillers.
The trade-off to keep in view
Evaporative cooling saves energy and spends water. Closed-loop cooling saves water and can spend energy. Neither is free, and which is preferable depends on the local grid and the local watershed — which is exactly why siting matters more than any single national number.