Nabugu - stock.adobe.com

News

A shadow water problem: AI's power demand shifts consumption upstream

Data center operators are cutting cooling-water use inside the fence, but the electricity powering AI campuses drives water consumption and withdrawals at plants serving the grid.

Shane Snider
By
Published: 17 Sep 2026

A data center can use very little water on site while relying on an electricity system that consumes or withdraws enormous quantities of water elsewhere.

That upstream footprint is becoming harder to ignore as AI drives data center electricity demand higher. Operators are deploying closed-loop liquid cooling and dry cooling to reduce or eliminate operational cooling-water use at the facility. But the power feeding those facilities still has a water footprint, depending on how and where it is generated.

"Yes, the indirect (Scope 2) water use for electricity generation is significant and often outnumbers by far the direct water use," said Shaolei Ren, professor of electrical and computer engineering at the University of California, Riverside.

Water usage effectiveness (WUE) measures water use at the data center, so a low site WUE does not necessarily mean the broader infrastructure supporting the facility has a low water footprint.

Lawrence Berkeley National Laboratory (LBNL) estimated that U.S. data centers collectively consumed about 176 terawatt hours (TWh) of electricity in 2023. Using location-based electricity-water factors, LBNL estimated that generating that electricity consumed nearly 800 billion liters, or about 211 billion gallons, of water.

LBNL's 2025 update projects U.S. data center electricity consumption could reach 649 TWh in 2030 in its reference case, with sensitivity scenarios ranging from 521 TWh to 843 TWh.

Applying the 2023 national average indirect water-consumption factor of 4.52 liters per kilowatt-hour to those scenarios would imply roughly 622 billion to more than 1 trillion gallons of annual indirect water consumption. Those are scenario calculations, not LBNL water forecasts, and assume the 2023 water-intensity factor remains constant.

The power source matters

Power-generation technologies have very different operational water requirements, particularly because of their cooling systems.

National Renewable Energy Laboratory (NREL) estimates put median operational water consumption for natural gas combined-cycle generation at about 2 gallons per megawatt hour (MWh)with dry cooling, 100 gallons per MWh with once-through cooling and roughly 200 gallons per MWh with recirculating cooling. NREL's median factors are 250 gallons per MWh for once-through coal, 687 gallons for recirculating coal, 269 gallons for once-through nuclear and 672 gallons for recirculating nuclear.

If a constant 1 GW data center load were supplied for a year by one of those configurations, the factors would translate to roughly 18 million gallons of operational water consumption for gas with dry cooling, 876 million gallons with once-through cooling and about 1.8 billion gallons with recirculating cooling.

Consumption also differs from withdrawal. Once-through systems can consume less water than recirculating systems while withdrawing much larger volumes from rivers, lakes or other sources and returning much of that water after use.

The regional footprint can be much larger

A 2026 analysis from Ceres estimated that data centers in Virginia, Texas, California, Illinois, Georgia, Ohio and Arizona combined were associated with roughly 3.4 trillion gallons of freshwater withdrawals annually for electricity generation.

Ceres measures withdrawals rather than consumption, so its estimate is not directly comparable with LBNL's 211 billion-gallon consumption figure. Ceres estimated roughly 753 billion gallons of annual power-generation water withdrawals associated with data center electricity in Virginia, 520 billion gallons in Arizona, 1.4 trillion gallons in California and 25 billion gallons in Ohio.

Ceres also found that 66% of electricity generated by water-dependent power plants in the states studied came from facilities facing medium-high to extremely high water stress.

Ceres notes that power-system boundaries cross state lines, making the state-level figures approximations of regional water risk rather than water physically delivered to individual data centers.

Julie Bolthouse, director of land use at the Piedmont Environmental Council, said data center development can affect communities through the infrastructure required to support the facilities.

"Issues like escalating cost of energy infrastructure, use of eminent domain to take private property for right of way for transmission lines, regional air pollution from backup generators and power plants, and water consumption that creates water stress on shared water sources during periods of drought" are among the regional impacts, she said.

"These regional impacts must be addressed first and in addition to these sorts of local projects and protections for the community where the data center is located," Bolthouse said.

Piedmont Environmental Council is calling for a statewide moratorium on data center development.

WUE misses the peak

Annual water consumption can also obscure what happens during extreme heat.

Ren said annual PUE and WUE can obscure the water and electricity tradeoffs that emerge during summer peaks. "That’s exactly the time that data center capacity planning is designed for."

Evaporative and adiabatic cooling can use water during extreme heat to improve cooling performance and reduce electricity demand. Dry cooling can reduce operational water use while potentially requiring more electricity during those same peak conditions.

"Many data centers only need water for adiabatic assistance on the hottest days of the year and have a high peak demand but low annual total," Ren said. "Water systems must be sized to accommodate the peak."

Ceres highlighted CyrusOne as an operator that incorporates electricity-related water use into its accounting through its WUE Source metric. The report said CyrusOne's WUE Source declined 67% since 2018, primarily because of renewable-energy procurement.

Meta provides another concrete example. In its 2025 Environmental Data Index, covering 2024, Meta reported 72,207 megaliters of water embedded in purchased electricity, equivalent to about 19.1 billion gallons. It separately reported 2,974 megaliters, or about 786 million gallons, of direct water consumption at its data centers.

On those reported figures, embedded water associated with purchased electricity was roughly 24 times direct water consumption reported for Meta's data centers. The figures have different boundaries -- Meta's embedded-electricity figure is companywide, while the direct figure covers its data centers -- so they are not a facility-level accounting of one campus.

Meta has also pursued closed-loop liquid-cooling designs with dry coolers that reduce operational cooling-water requirements.

Water becomes a power-planning issue

"We need to plan the water and power infrastructures together," Ren said. "If used responsibly, water can be the most efficient way to lower the peak power demand of data centers."

For AI data centers, the generation mix, cooling technology, local water availability and peak electricity demand are increasingly intertwined.

"In other words, investing in water infrastructure is essentially investing in power grids," Ren said.

 

Dig Deeper on Data Center