The Kansas Health Institute put a range on the water question. The range is wide enough to be a planning problem. US data centers consumed an estimated 17 billion gallons of water directly in 2023. That draw could double or quadruple by 2028, according to Utility Dive's coverage of the report, which Robert Walton wrote up on July 21. Emma Uridge and Jasmin Kamruddin published the underlying work on July 15. The electricity figure runs beside it: 183 TWh in 2024, more than 4% of total US consumption, projected to grow 133% to 426 TWh by 2030. The water range comes out of scenarios modeled at Lawrence Berkeley National Laboratory. KHI carries the full spread and declines to pick a midpoint.
Quadrupling means 68 billion gallons. Doubling means 34. That is a 34 billion gallon uncertainty band on a forecast resolving inside thirty months. Almost none of it turns on technology that has yet to ship. The buildings producing the 2028 number are permitted. Their heat rejection method is already drawn. Long-lead equipment for 2027 and 2028 commissioning sits in the queue right now: cooling towers, centrifugal chillers, dry coolers, coolant distribution units, plate heat exchangers, and the switchgear that has to be sized around whatever rejection method the design team locked when the site plan went in. Nothing invented this quarter moves the number. What moves it is the ratio of evaporative towers to closed-loop and liquid-cooled plant across projects breaking ground through next spring. Permit officers and capex committees are setting that ratio.
KHI puts cooling at as much as 40% of a data center's total electricity draw, citing NREL. The report splits the load three ways: computing at 40%, cooling at 40%, other IT equipment at 20%. It expects the cooling share to climb as racks densify. Average density runs from 36 kW in 2023 to 50 kW by 2027, with some AI rack designs reaching 600 kW. That is why the 17 billion gallon figure and the 426 TWh figure cannot be argued separately. Closed-loop and dry cooling cut onsite consumption hard, then hand the difference back as fan and compressor power. Fossil generation supplies 56% of the electricity feeding these sites, with 22% renewable and 21% nuclear. Thermoelectric plants consume water to make that power. Uridge and Kamruddin say it plainly. Water consumed offsite at power plants is the largest share of a data center's total footprint, and Water Usage Effectiveness hides all of it. That is the mechanism behind our earlier finding that 72% of data center water consumption happens off the property. One in five US data centers already sat in a water-stressed area in 2021, before the current buildout landed.
KHI is a health policy institute in Topeka, and the report reads like one. It synthesizes published modeling for county commissioners and public health staff. Kamruddin is credited as an intern at the time of the work. The 17 billion gallon baseline has been circulating for a while. KHI's footnotes point back to Mytton's 2021 paper in npj Clean Water and the 2024 LBNL energy usage report. No new metering here. The fresh material is the framing and the policy table. Uridge and Kamruddin push practitioners toward cumulative exposure tracking, air quality monitoring near diesel generators, water availability screening in stressed regions, and electricity cost burden on low-income households. Their local inventory names Loudoun County, which ended by-right zoning for data centers in 2025, and Marana, Arizona, which prohibited potable water for cooling in 2024. Walton adds New York's one-year statewide moratorium. That one does not appear in the report itself.
Marana is the item operators should read twice. A potable water prohibition removes open evaporative cooling from the design menu before anyone runs the capex comparison. KHI notes that some facilities currently pull more than half their water from potable supplies. What survives is closed-loop with a dry cooler and a summer power penalty, reclaimed water with a municipal treatment contract and a chemistry problem, or a different county. Vendors have spent two years selling warm-water loops that cut onsite draw to almost nothing as the answer to exactly this. On the water meter they are right. The power draw and the upstream generation mix are where that claim gets tested. KHI flags the offsite water and never puts a number on the swap.
Anyone specifying heat rejection for 2028 commissioning in the Southwest or the Great Lakes should price the dry-cooling summer power penalty into the pro forma at concept design. The rejection method is still cheap to change there. Marana's ordinance is a template, and Tucson and Kansas City have already written water conditions of their own. Redesigning a rejection plant after the tower pads are poured costs a schedule year.