Why AI data centers need water and power planning together

A low-WUE data center can still carry a large water footprint

A data center can sharply reduce water use on site while still depending on large amounts of water elsewhere in the infrastructure that keeps it running. Operators are using closed-loop liquid cooling and dry cooling to cut operational cooling-water requirements, and those choices can materially improve water usage effectiveness (WUE), which measures water use at the data center. But the WUE facility boundary excludes water consumed or withdrawn to generate electricity, so facility efficiency captures only part of the water requirement.

For AI infrastructure, that boundary matters because electricity demand can move much of the water burden away from the campus. Shaolei Ren, professor of electrical and computer engineering at the University of California, Riverside, points to indirect, or Scope 2, water associated with electricity as a major part of the calculation. “Yes, the indirect (Scope 2) water use for electricity generation is significant and often outnumbers by far the direct water use,” Ren said.

That upstream burden changes what a low WUE establishes. WUE remains useful for understanding and improving direct facility water use, particularly as operators change cooling systems. A broader assessment also has to follow the electricity supplying the data center because generation technologies and power-plant cooling systems can have radically different water requirements for the same electrical load.

AI’s electricity demand moves the water question upstream

Following the electricity supply starts with how much power data centers consume. Lawrence Berkeley National Laboratory (LBNL) estimated that U.S. data centers consumed about 176 terawatt hours (TWh) of electricity in 2023. Using location-based electricity-water factors, LBNL estimated that producing this electricity consumed nearly 800 billion liters of water, equivalent to about 211 billion gallons.

That 2023 load is the baseline for a much larger projected requirement in 2030. LBNL’s 2025 update gives a reference case of 649 TWh, with sensitivity scenarios ranging from 521 TWh to 843 TWh. Electricity demand is one variable in future upstream water requirements; the generation and cooling systems supplying that electricity are the other major variables.

Holding water intensity constant shows the potential scale of those electricity scenarios. The 2023 national-average indirect water-consumption factor was 4.52 liters per kilowatt-hour. Applying the same factor across LBNL’s 2030 range produces roughly 622 billion to more than 1 trillion gallons of indirect water consumption per year. These are scenario calculations rather than LBNL water forecasts because they assume the 2023 water-intensity factor remains unchanged.

That assumption matters because generation technologies and their cooling systems have very different operational water profiles. National Renewable Energy Laboratory (NREL) median water-consumption factors show large differences even within one fuel type. For natural-gas combined-cycle generation, changing the cooling configuration shifts median operational consumption by roughly two orders of magnitude, and a constant 1 GW data-center load makes the scale of that difference visible.

Because the 1 GW calculation holds data-center load steady, the differences among the natural-gas cases come from generation-side cooling. Electricity sourcing is consequently a water decision even when the data center uses a closed-loop liquid system or dry cooling. Facility engineering can reduce direct consumption while the power supply carries substantial water requirements beyond the campus.

Those generation differences also change what an electricity forecast means for infrastructure planning. A TWh figure establishes how much energy is required, while generation technology, plant cooling design and location determine much of its operational water footprint. As AI pushes data-center electricity requirements higher, power procurement and water planning have to account for where the resulting water demand occurs.

Consumption and withdrawal reveal different kinds of water pressure

Once the boundary extends into the power system, the type of water use matters too. Consumption is water that is not returned to its original source after use, while withdrawal is the volume taken from a river, lake or other source, including water that may later be returned. A planning decision can produce very different numbers depending on which measure it tracks.

Once-through generation cooling shows the practical difference. These systems can consume less water than recirculating systems because much of the withdrawn water returns to the source. But they can withdraw much larger volumes, creating a different kind of pressure on the source and surrounding region.

That wider pressure appears in a 2026 Ceres analysis covering Virginia, Texas, California, Illinois, Georgia, Ohio and Arizona. Ceres estimated that data-center electricity across the seven states was associated with approximately 3.4 trillion gallons of freshwater withdrawals each year for power generation. Selected state estimates show how unevenly that regional exposure can be distributed.

Because Ceres measures withdrawals and LBNL’s 211 billion-gallon figure measures consumption, the 3.4 trillion gallons represents a different kind of water pressure. Keeping the definitions separate matters when a decision depends on how much water leaves a source, how much returns and how much is ultimately consumed. Both figures can inform the same planning process while answering different questions.

Once the type of water use is clear, location determines how consequential that use can be. Ceres found that 66% of electricity generated by water-dependent plants in the states it studied came from facilities facing medium-high to extremely high water stress. Electricity’s water footprint therefore depends on local water conditions as well as the number of gallons involved.

Those local conditions also make state boundaries an imperfect way to assign power-system effects to individual campuses. Electricity moves across state lines, so Ceres describes its state-level figures as approximations of regional water risk rather than volumes physically supplied to individual data centers. For decision-makers, the relevant boundary can extend across the power system and the water resources supporting it, far beyond the jurisdiction where the servers operate.

Low annual water use can still create a peak-capacity problem

Geographic boundaries are one limit of annual water accounting; time is another. Power usage effectiveness (PUE), which tracks data-center energy efficiency, and WUE are commonly expressed as annual measures, but infrastructure must operate through peak conditions as well as average ones. Ren says annual measures can obscure water-and-electricity tradeoffs during extreme summer conditions: “That’s exactly the time that data center capacity planning is designed for.”

Those summer tradeoffs depend directly on cooling technology. Evaporative and adiabatic cooling use water to improve cooling performance during hot conditions, which can reduce the electricity needed to keep equipment within its operating limits. Dry cooling can reduce operational water consumption, but during the same extreme heat it can require more electricity and add demand when the grid is already under pressure.

Because the hottest periods can be short, a facility can have modest annual water use alongside a demanding peak requirement. “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. The engineering requirement follows from that peak: “Water systems must be sized to accommodate the peak.”

Peak conditions also connect facility cooling to the upstream electricity boundary. Cooling choices directly affect site water demand and can change electricity demand, especially during extreme heat. The changed electricity requirement then affects generation-side water use, coupling site water with upstream water through the cooling and power systems.

That coupling creates a capacity tradeoff during severe heat. A facility designed around dry cooling may have low annual operational water consumption while putting more demand on electricity infrastructure during the hottest periods. An adiabatic design may intentionally use more water during those periods because doing so reduces its electrical requirement, so capacity planning has to account for both resources alongside annual WUE and PUE.

Companies are beginning to account beyond the fence line

The connection between facility and power-system water is also beginning to appear in operator accounting. Ceres highlights CyrusOne’s WUE Source, which incorporates electricity-related water use into its water accounting. Ceres reports that CyrusOne’s WUE Source has declined 67% since 2018, primarily because of renewable-energy procurement, showing how electricity sourcing can change the measured water footprint independently of a facility cooling intervention.

Meta’s reporting offers another view of the difference between direct and electricity-related water. Its 2025 Environmental Data Index, covering 2024, reported 72,207 megaliters of water embedded in purchased electricity, equivalent to about 19.1 billion gallons. Meta separately reported direct water consumption of 2,974 megaliters, or roughly 786 million gallons, at its data centers.

Those reported figures put water embedded in purchased electricity at about 24 times Meta’s reported direct data-center consumption, but their boundaries differ. The purchased-electricity figure is companywide, whereas the direct-water figure covers Meta’s data centers, so the ratio does not represent the upstream-to-direct relationship for a particular Meta campus. CyrusOne and Meta are themselves data-center operators, which gives both companies a commercial interest in how their water performance is measured and presented.

That operator interest makes clear boundaries especially important while leaving a direct role for facility engineering. Meta has pursued closed-loop liquid-cooling designs using dry coolers to reduce operational cooling-water requirements. Electricity-related accounting addresses water associated with the power supply, while direct-water engineering addresses requirements at the facility, so operators have to manage both parts of the infrastructure decision.

Water and power planning have to become one capacity decision

The peak tradeoff turns water and power into linked capacity decisions. Generation mix determines part of electricity’s water footprint; cooling technology changes direct water requirements and facility power demand; local water availability determines how consequential those requirements are. Extreme heat can tighten several of those constraints at once, making separate capacity plans vulnerable to the same event.

“We need to plan the water and power infrastructures together,” Ren said. His argument includes deliberate water use where it reduces a larger power constraint: “If used responsibly, water can be the most efficient way to lower the peak power demand of data centers.” Ren extends that relationship to infrastructure investment itself: “In other words, investing in water infrastructure is essentially investing in power grids.”

Infrastructure investment then carries the capacity decision beyond the operator and into surrounding communities. Julie Bolthouse, director of land use at the Piedmont Environmental Council, identifies “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” among the regional effects. Data-center growth contributes to the infrastructure demand behind these effects even when they occur outside the facility.

Those regional effects lead Bolthouse to argue that planning should begin before attention narrows to an individual development: “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.” The Piedmont Environmental Council has gone further by calling for a statewide moratorium on data-center development. Its position makes the linked water-and-power capacity decision a regional policy question because power infrastructure, land, air pollution and shared water supplies can all be affected together.

Key takeaways for leaders

• Track water beyond facility WUE: Data center operators need electricity-related water use in their planning because power generation can consume far more water than facility cooling. Generation mix and power-plant cooling technology materially change that upstream footprint.
• Treat electricity sourcing as a water decision: AI growth is increasing data center power demand, while generation technologies vary widely in water consumption. Procurement and infrastructure teams can evaluate electricity supply by water intensity, cooling technology and location alongside cost and carbon.
• Separate water consumption from withdrawals: Power generation can create large freshwater withdrawals even when much of that water returns to its source. Planners need both measures, combined with local water-stress data, to understand regional capacity and resource pressure.
• Plan for peak water and power demand together: Annual WUE and PUE can obscure infrastructure requirements during extreme heat. Data center designers and utilities need to model peak conditions because dry and adiabatic cooling create different tradeoffs between water use and electricity demand.
• Extend water accounting into purchased electricity: Operator reporting shows that electricity-related water can substantially exceed direct facility consumption. Sustainability and procurement teams can incorporate upstream water into performance metrics and use energy sourcing to reduce the broader footprint.
• Make water and power one capacity decision: Data center expansion can affect grids, water systems, land and surrounding communities through shared infrastructure. Operators, utilities and regional planners need coordinated capacity planning before individual projects lock in power and water requirements.