Water, AI, and the Ledger That Actually Matters

Buckhannon, West Virginia; December 26th, 2025

When people ask whether water used by artificial intelligence and data centers is “returned,” the answer depends on numbers, categories, and destinations, not slogans. In U.S. water accounting, the first measurable quantity is withdrawal, the total volume drawn from a river, reservoir, aquifer, or other source so that it can move through pipes and cooling systems. The second quantity is consumptive use, the portion of that withdrawn water that does not return to the immediate local system because it evaporates, is incorporated into products, or is otherwise removed from near term reuse. Everything else is return flow, water that reenters surface or groundwater systems after use, often slightly warmer, sometimes at a different discharge point, but still available downstream.

Across modern data centers, confirmed engineering ranges place cooling related water withdrawal at roughly 1 to 5 liters per kilowatt hour of electricity consumed, depending on facility design and climate. What matters next is the fraction consumed. In evaporative cooling systems, the consumed portion is commonly 0 to 1.5 liters per kilowatt hour, while the remainder is returned as discharge or return flow. Facilities using air cooling, closed loop liquid cooling, seawater cooling, or reclaimed wastewater can consume little to no freshwater at all. This is not theory, it is how federal water agencies define and measure use, and it is why withdrawal and consumption are treated as separate columns in water ledgers.

The distinction is not academic. Withdrawn water that is returned remains part of the local hydrologic system. Consumed water leaves the local system, usually as vapor, before eventually returning elsewhere through the global water cycle. Both matter, but they are not interchangeable. Treating them as the same turns infrastructure math into environmental theater.

Cooling choice drives the outcome because cooling is a heat problem before it is a water problem. Evaporative cooling removes heat by letting water carry it away as vapor. Air cooled and closed loop systems remove heat without turning large volumes of liquid water into vapor. That is why the U.S. Department of Energy has long identified cooling towers as a major efficiency opportunity in data centers, and why operators focus so intensely on cooling design when building new facilities.

Private operators have begun attaching concrete figures to this difference. Microsoft has publicly described a newer data center design that consumes zero water for cooling, while acknowledging that water is still used for ordinary building needs such as restrooms and kitchens. The company has stated that avoiding evaporative cooling can prevent the use of more than 125,000,000 liters of water per data center per year, a figure that would otherwise be lost to evaporation. The claim is not that a building becomes waterless, but that the cooling method stops converting large volumes into vapor.

Google, which publicly reports water metrics for its owned and operated data centers, separates withdrawals, consumption, and discharge in its environmental reporting, and also identifies where water is sourced relative to local water stress. That reporting framework mirrors federal definitions, and it exists precisely because the difference between what is withdrawn, what is consumed, and what is returned determines local impact.

Placed next to other sectors, the scale becomes clearer. Agriculture accounts for roughly 70 percent of global freshwater withdrawals, and consumes vast amounts of water permanently through crop growth and evapotranspiration. Food and beverage production removes water from aquifers and redistributes it elsewhere, meaning it does not return to the source basin. Bottled water, often invoked in comparisons, is by definition consumptive. By contrast, most data center water is temporary use, not permanent removal.

Day to day AI use, such as routine text generation or search, adds only a marginal load to existing data center operations, requiring fractions of a watt hour of electricity per interaction and often resulting in zero direct freshwater consumption at the facility level. Large scale AI model training does require more energy, but it occurs infrequently, and its water footprint depends entirely on location, cooling design, and power source.

The confirmed numbers do not support claims that AI is draining rivers or rivaling entire cities in water consumption. What they show instead is a regional infrastructure question. In water scarce areas paired with evaporative cooling, consumption matters and must be managed. In many other regions, most withdrawn water is returned, regulated, and accounted for. The difference is not ideology, it is engineering and accounting.

The honest way to talk about water and AI is therefore a ledger, not a meme. Withdrawals by source. Consumption by cooling method. Return flow by discharge point and timing. When those numbers are kept intact, the conversation shifts from panic to planning, which is where water policy actually belongs.

Sources

Primary First Hand Sources

• U.S. GEOLOGICAL SURVEY, official definitions and water use terminology, including withdrawal, consumptive use, and return flow
• U.S. DEPARTMENT OF ENERGY, Federal Energy Management Program guidance on data center cooling water efficiency
• MICROSOFT, official environmental sustainability reporting and public statements on zero water cooling data center design
• GOOGLE, official environmental reporting on data center water withdrawals, consumption, and discharge