How AI data centres use water for cooling

A liquid-cooled AI server can keep recirculating the same coolant while the facility around it consumes water. A sealed circuit may carry heat away from processors, while a separate cooling tower releases that heat outdoors by evaporation. To understand the water requirement, trace the complete system from the chip to the outside environment.[1]

An October 2026 review preprint by Eric Masanet separates heat collection, transfer and outdoor rejection. Different stages can use different fluids, so each circuit needs its own explanation.[1]

Water taken in and water consumed answer different questions

The U.S. Geological Survey distinguishes withdrawal from consumption. Withdrawal describes water removed from a surface-water or groundwater source. Consumption includes water that evaporates or is otherwise removed from the immediate water environment. Water returned after use is a separate flow. Evaporation moves water into the atmosphere; it does not destroy the water molecules.[2]

For an invented example, suppose a facility directly withdraws 1,000 litres, returns 800 litres to the same water system and evaporates 200, with no other flows or storage change. Withdrawal is 1,000 litres and consumption is 200. Both numbers describe the facility correctly. Counting the 800 litres again each time they circulate inside a pipe would measure circulation, rather than additional withdrawal.

Design and water source can both change

Microsoft’s December 2024 design announcement describes water continuously moving between servers and chillers, avoiding evaporation. It still allows water for kitchens and restrooms. The announced programme scheduled pilots in Phoenix and Mount Pleasant for 2026 and new sites from late 2027; a design and schedule are different evidence from a completed fleet conversion.[3]

The company’s June 2026 update also reports operational use of alternative water sources. It gives recycled, reused or non-potable shares of 74% at Quincy, 99% in Singapore and 79% at San Antonio. Changing the source can reduce reliance on drinking water even where a cooling system still needs replacement water.[4]

Hamina illustrates how location opens another route. Google converted a former paper-mill site into a data centre that opened in 2011 and uses seawater for cooling. Its current site description also identifies an offsite heat-recovery facility with Haminan Energia. Heat can become a resource for neighbouring buildings as well as something to remove from servers; that does not supply a universal water-saving percentage for other sites.[5]

Some of the water footprint is at the power station

Electricity generation creates a further pathway. A September 2026 University of Michigan preprint connects data centres with the generators responding to their demand. Its hypothetical Oakland County facility had a fixed 200 MW computing load. Switching from water-cooled to air-cooled chillers cut total modelled water consumption by 44% while raising electricity demand and indirect consumption by 32%. Those scenarios used worst-case efficiency assumptions for both cooling types; they do not measure Microsoft’s design.[6]

A cooling choice can therefore reduce water consumed on site while changing demand elsewhere. Following the heat, the water supply and the electricity connections explains why two AI data centres can require different arrangements—and how a specific engineering change can make one less demanding.[6]

Report an error or suggest a correction