Cooling and water
How much water does a data center use? It depends on one choice.
Almost every watt of electricity entering a data center turns into heat that must be moved outside, around the clock. How the building sheds that heat — by evaporating water, or by fans and sealed liquid loops — determines whether it drinks millions of gallons a day or roughly what an office park does. For the same 30 MW facility, the difference between cooling designs is about fifty-fold.
The cooling systems, plainly
- Open cooling towers (evaporative) — warm water is sprayed through outdoor towers; evaporation carries the heat away. Energy-efficient, but the thirstiest design: roughly 6.75 million gallons per year per megawatt, per Uptime Institute.99Some water is also regularly discharged to the sewer as “blowdown” to flush out concentrated minerals.100
- Air-cooled (dry) systems — giant radiator-and-fan units reject heat straight to the air. Near zero water, but more electricity (evaporative assist can cut peak-summer cooling electricity 10–35%) and more fan noise.113
- Direct evaporative (“swamp cooler”) — outside air is cooled through wetted media. Uses far less water than open towers; Meta’s version runs about a tenth of traditional consumption.99
- Adiabatic / hybrid — dry coolers that mist water only on the hottest days. Low, seasonal water use.99
- Closed-loop, direct-to-chip liquid cooling — the new AI-era standard. A sealed water loop is filled once at construction and recirculated indefinitely; Microsoft says its zero-water-for-cooling design avoids more than 125 million liters (~33 million gallons) per facility per year, leaving only restroom-scale water use.101
- Immersion cooling — servers submerged in non-conductive fluid; essentially no on-site water. Still niche (crypto, HPC).99
What that means for a 25–30 MW facility
Using published water-efficiency figures (WUE — liters of water per kilowatt-hour of computing), here is the computed range for a 30 MW facility running at full load. These are estimates from cited inputs, not the developer’s numbers — the actual figure depends entirely on the cooling design the end user chooses.
Math: 30 MW × 8,760 hours = 262.8 million kWh/yr; multiply by the cited WUE and convert liters to gallons. WUE inputs: Uptime Institute,99 industry average,99 Microsoft,101 Meta.102The closed-loop row uses Microsoft’s real Wisconsin facility: 2.8M gallons/yr in Phase 1, ~8.4M gallons/yr at later phases.107
For comparison: Google disclosed that its average data center campus used about 450,000 gallons per day in 2021103 — so a worst-case evaporative 30 MW build would behave like a typical Google campus, while a closed-loop build would use about as much water as a large office park.107
What Pikeville has said about water
The trade-off, honestly
Water and electricity trade against each other. Evaporating water is the cheapest way (in energy) to shed heat; going dry costs more electricity.113 Google has argued water-cooling cuts its energy use about 10% versus air cooling.103There’s also a hidden ledger: peer-reviewed research found that roughly 75% of data centers’ total water footprint is indirect — consumed by the power plants generating their electricity, not at the site itself.104A “zero-water” data center still drives water use somewhere on the grid. Nationally, data centers consumed about 66 billion liters (~17 billion gallons) directly in 2023, a figure the federal LBNL report projects could double or quadruple by 2028.11 The industry trend helps here: AI chips run too hot for air alone, so new builds increasingly use sealed direct-to-chip loops — which happen to nearly eliminate on-site water use.101
Can the waste heat be put to use? Yes — with honest caveats
Nearly all electricity entering a data center leaves as heat, so a 25–30 MW facility is also, in effect, a 24/7 heat source of roughly the same size. In Northern Europe, where cities have district heating networks, that heat genuinely warms homes: Meta’s Odense, Denmark plant recovers heat for the city network (about 45 MW of heat production serving thousands of homes and a hospital),116Microsoft’s data centers near Helsinki now supply up to 180 MW of district heat — on track to cover ~40% of heating demand for an area of 250,000 people,117Google’s Hamina facility will cover up to 80% of its town’s heating demand free of charge,118and Stockholm’s utility literally pays data centers for their heat.119 Germany now requires new data centers to reuse 10–20% of their energy as heat.125
The honest catch: data center exhaust is low-grade heat — roughly 100°F air from air-cooled halls (hotter, more usable 50–60°C water from modern liquid-cooled designs)124 — and using it requires heat pumps plus, crucially, a heat customer nearby. The Nordic projects work because city-wide hot-water networks already existed; almost no U.S. town has one,123 and we found no documented example of a small-city American data center heating homes.
What isprecedented in the U.S. is single-neighbor reuse: Amazon’s Seattle towers are warmed by ~5 MW of waste heat piped across the street from a data-dense building,120Notre Dame’s server racks heat South Bend’s municipal greenhouse, saving the city about $70,000 a year,121and in the UK a 28 kW micro data center cut a public swimming pool’s gas bill 62%.122 A regional policy group, ReImagine Appalachia, argues for exactly this model here — using data center heat for schools, public buildings, or greenhouses on former mine land.126 For Pikeville, heating the town is not realistic; heating an adjacent greenhouse complex, pool, or campus building is — if heat-recovery plumbing is designed in from day one, which is a reasonable thing to raise while a Development Agreement is still being negotiated.
Where water became a problem — and where it didn't
- The Dalles, Oregon:Google’s data centers used 355.1 million gallons in 2021 — 29% of the entire city’s water — a fact disclosed only after a 13-month public-records fight with the local newspaper.105 Google also funded $28.5 million of city water infrastructure.106Lesson: the problem wasn’t only the volume — it was the secrecy.
- Newton County, Georgia:residents 1,000 feet from Meta’s construction site reported wells running dry and sediment; Meta’s commissioned study found no connection, the county did no pre-construction well survey, and causation remains unresolved.85
- Mesa & Goodyear, Arizona: desert cities negotiated tiered water caps (1 up to 4 million gallons/day in Mesa)108 and in Goodyear, Microsoft agreed to switch its design to air cooling and contributed $36M toward a wastewater plant.109
- Memphis: xAI pledged an $80M wastewater recycling plant (13 million gallons/day design) to stop drawing on the drinking-water aquifer — groundbreaking happened, though the project was later paused, a reminder that pledges need enforceable timelines.110
- Mount Pleasant, Wisconsin (the quiet success): Microsoft’s closed-loop campus uses 2.8M gallons a year in Phase 1 — less than many ordinary industrial customers on the same utility.107
- Loudoun County, Virginia: the water utility delivers reclaimed (treated wastewater) through purple pipe for data center cooling — over 750 million gallons in 2025 — sparing the same volume of drinking water.111Google’s Georgia site similarly cools with recycled municipal effluent.112