Blog · Water and Infrastructure · October 2, 2026

The Water Cost of AI Has a Location and a Season

A liter count describes a volume. To understand its consequences, follow the water to its source, distinguish what returns from what evaporates, and ask when competing demands are greatest.

The same total, a different season

Imagine two proposed data centers reporting identical annual water consumption. In this hypothetical comparison, one draws steadily from a source with spare capacity; the other concentrates demand in months when that source struggles to meet existing needs. The annual numbers match. The decisions facing their neighbors do not.

Our argument is that a useful account of AI's water cost needs a map and a calendar alongside the volume. This is a narrower question than the public bargain explored in The Data Center Becomes a Civic Machine. Before judging whether a particular bargain is reasonable, readers need to know what its water figures actually describe.

A national total can establish scale. A company average can describe its portfolio. Neither, by itself, tells a resident what happens to a particular source during a dry month. Treating these numbers as interchangeable obscures both real problems and real improvements.

Name the water being counted

David Mytton's peer-reviewed 2021 analysis distinguishes water taken from a source from consumptive losses, commonly evaporation. Withdrawal describes the intake; consumption concerns the portion unavailable for immediate reuse after the process. Cooling water may circulate repeatedly while some evaporates and some is discharged. Repeated circulation therefore does not establish zero consumption. The paper also separates water used directly in cooling from water associated with electricity generation.

Our practical inference is to resist comparing a withdrawal figure with a consumption figure, even when both are expressed in liters. They answer different questions. Intake matters when assessing the capacity needed to supply a facility. Consumptive loss matters when assessing what remains available after use. A meaningful comparison should retain both, with the source and return destination identified.

Berkeley Lab's 2024 report estimates that US data centers directly consumed 66 billion liters in 2023. That is a historical estimate for data centers collectively, not a measurement of AI alone or a current-year total. The report defines site water usage effectiveness, or WUE, as water consumption divided by IT electricity use, expressed in liters per kilowatt-hour. Its modeling accounts for cooling design, operating practices and climate.

The denominator matters. Our reading is that a lower WUE can represent useful engineering progress while leaving the question of absolute demand open. If computing activity expands faster than water intensity falls, total consumption can rise. Conversely, a larger absolute figure does not prove that a facility became less efficient. A reader needs the intensity and the volume to distinguish those possibilities.

Follow the electricity upstream

Siddik, Shehabi and Marston's 2021 study connects data centers to the power plants, water suppliers and wastewater facilities supporting their operation. It uses 2018 data and excludes non-operational stages such as server manufacturing. Its estimates depend in part on how electricity generation is attributed to consumers; the authors test that sensitivity. This is an operational footprint study, not a census of today's AI systems.

We draw a simple consequence: a facility's address is only the beginning of its water map. An assessment that counts electricity-related water should identify the geography where that consumption occurs. Assigning every upstream liter the scarcity conditions at the data center would erase the distinction the accounting was meant to reveal.

Boundary discipline also prevents exaggerated totals. Adding an electricity estimate to a figure that already includes electricity counts part of the footprint twice. Adding manufacturing to an operational estimate can be legitimate, but changes the question and requires compatible periods and methods. A larger number assembled from overlapping categories is less informative than a smaller number with a clear boundary.

What the monthly view changes

The 2026 ACM paper Balancing Bits and Drops applies monthly county-level water-scarcity factors to direct consumption and electricity-related consumption. Its electricity mapping uses regional generation portfolios as an approximation. The authors' scheduling analysis is an offline upper bound, and alternative hydrological models could change the absolute stress estimates. The work supports examining seasonal differences; it does not establish that an individual facility caused a particular shortage.

Our interpretation is that monthly accounting changes the comparison itself. Ask whether high consumption coincides with constrained supply, rather than merely whether an annual total went down. In a hypothetical facility, savings achieved mainly during abundant months could coexist with unchanged demand during the difficult season. Another facility might deliver a smaller annual reduction but relieve demand precisely when the local system needs it.

This does not make an annual reduction meaningless. It makes the timing relevant to what can be claimed about that reduction. Preserve the ordinary volume figures alongside any scarcity-weighted indicator. The indicator is a model of relative pressure, not extra physical water flowing through a pipe.

A calendar also has limits. Monthly averages can hide brief peaks, while modeled scarcity does not describe every household's access or a utility's distribution constraints. We would use the monthly view to identify questions for closer investigation, then seek local operating and supply records before making claims about specific harm.

Evaluate the cooling trade

The IEA's 2025 Energy and AI report illustrates how much facility designs differ: cooling and environmental control account for about 7% of electricity in efficient hyperscale centers and over 30% in less-efficient enterprise centers. These are background comparisons from that report, not a measured energy penalty for replacing a particular cooling system.

The 2026 seasonal-water study examines dry cooling as a way to reduce direct water demand that can increase electricity needs. Whether that trade improves the modeled water outcome depends on the energy penalty and the water conditions associated with electricity supply.

Our recommendation is therefore to evaluate the proposed replacement as a complete change. What direct consumption does it avoid? What additional electricity does it require, under the weather conditions that matter? Where does the associated water demand fall? A genuine improvement may survive all these questions. The upstream inquiry should help identify it, rather than become an excuse to dismiss local savings.

The same reasoning makes broad labels insufficient. Calling a design water-efficient should invite an explanation of the comparison and operating conditions. It should not end the inquiry. A site may have compelling reasons to prioritize local water availability while accepting an energy trade; readers should be able to see that choice clearly.

How to read a water claim

We propose a short reading sequence. First identify whether the figure measures withdrawal or consumption. Next locate its boundary: the facility, electricity supply, or a wider life cycle. Then ask for the source geography and reporting period. Finally compare the proposed improvement with the previous arrangement under matching assumptions.

For an AI-specific claim, add the allocation question: how was a shared facility's footprint assigned to this workload? A universal water-per-prompt number cannot answer the location, boundary and timing questions posed here. A carefully scoped estimate may still be useful if those conditions travel with it.

The useful endpoint is a statement a reader can check: what water demand changes, where it changes, and during which conditions. That makes room for both scrutiny and credit. It allows a real reduction in a constrained season to count for what it accomplishes, and keeps an attractive annual average from settling a question it never measured.

Sources

Production: commissioned by the site operator; researched and drafted by GPT-6 Astra; editorial and source review by the coordinating AI assistant.

Return to Blog