NON-SATIRICAL REPORT · Field Notes

A Dry Data Center Can Still Depend on Water Elsewhere

A low onsite-water number can describe a real design choice while leaving a separate question unanswered: what water is associated with the electricity that keeps the servers running?

By Field Notes Editorial Desk · October 27, 2026

Ceres's 3.4 trillion gallons is a modeled 2024 estimate of freshwater withdrawals associated with data-center electricity in seven states. LBNL's nearly 800 billion liters is a separate estimate of U.S. 2023 indirect water consumption. The figures use different years, geographies, water metrics, and methods, so they are neither equivalent nor addable.

A dry data center is not necessarily a water-free data center. The question is where the measurement stops.

The boundary behind a low-water claim A data center can reduce the water used inside its property line and still rely on electricity whose generation involves water elsewhere. The Los Angeles Times reported that the water question around AI infrastructure includes both on-site cooling and the power system that serves the facility. A low onsite figure can therefore be accurate without describing the whole relationship between a campus and water. [1]

That is not a semantic objection to a cooling design. It is a request to name the boundary of the claim. Some data-center cooling systems use water at the site, while designs intended to use less onsite water can require more electricity. The relevant comparison is not simply a building's water meter. It is the onsite design together with the water associated with the power needed to run it. [1]

Withdrawal is not consumption The terms matter because they answer different questions. The U.S. Geological Survey defines a withdrawal as water removed from a source. Consumptive use is the share that is not returned locally, often because it evaporates. Thermoelectric generation can withdraw water for its steam and cooling processes while consuming only part of that withdrawal. A large withdrawal number and a large consumption number do not describe the same burden. [4]

That distinction also prevents a shortcut in public debate. An onsite water figure may describe water delivered to or used at a campus. A power-sector figure may describe water withdrawn at a generating plant, or water consumed there, depending on the study. Before comparing any of them, readers need to ask what was counted, where it was counted, and whether the water was returned to the source system. [4]

Two estimates that should not be merged Ceres estimated that data centers in Virginia, Texas, California, Illinois, Georgia, Ohio, and Arizona were associated with about 3.4 trillion gallons of freshwater withdrawals through electricity generation in 2024. The analysis covers seven states, not the nation, and allocates power-plant withdrawals to data centers through a model. It is not an estimate of national water consumption, a measurement of water used up, or a meter reading for any named facility. [2]

Lawrence Berkeley National Laboratory, or LBNL, made a different estimate. It associated U.S. data-center electricity use in 2023 with nearly 800 billion liters of indirect water consumption. LBNL derived grid water-intensity factors from balancing-authority and plant data. Its analysis does not incorporate every individual facility's power-purchase agreement or behind-the-meter generation, so it is a grid-based estimate rather than a facility-by-facility ledger. [3]

The numbers should not be converted into a single total or used as rival scorecards. Ceres reports 2024 freshwater withdrawals for a seven-state allocation. LBNL reports 2023 indirect consumption for U.S. data centers. The years, geographies, metrics, and methods differ. Treating them as equivalent, or adding them together, would erase the information each number is meant to convey. [2] [3]

Questions the records can answer For a proposed or operating campus, the first useful record is an annual load forecast in a permit, interconnection filing, utility-service document, or integrated-resource plan. Readers can then ask what onsite cooling system is proposed, what water metric the operator reports, which balancing authority or power resources are expected to serve the added load, and whether the documents separately state electricity-related withdrawals and consumption. These questions ask for definitions and evidence rather than presuming that a low-water claim is deceptive. [1] [4]

A renewable-procurement statement also needs a boundary. Does it describe an annual contract, a claim about hourly physical supply, or another accounting method? Does the record identify its geography and time period? The answer may not yield a precise water ledger, but it can show whether the claim concerns the facility, the grid, or both. LBNL's stated limits illustrate why a grid-average estimate cannot automatically resolve an individual facility's supply arrangement. [3]

Uncertainty is part of the result Ceres describes its work as a directional estimate and notes that its state-based model excludes several factors, including saltwater and brackish-water cooling, behind-the-meter generation, water-source origins, and seasonal variation. Electricity mixes, generating technologies, cooling configurations, and the timing of load can all change the relevant water picture. Those limits are reasons to read local records carefully, not grounds to fill the gaps with allegations. [2]

The available reporting and analyses do not establish that any named company concealed a water footprint, violated a rule, or harmed a particular river or community. They also do not support a fixed water footprint for an AI prompt. What they support is a narrower and more useful conclusion: a data center can be dry at the fence line while remaining dependent on water elsewhere, and the public deserves claims that say which water, where, and by what measure. [1] [2] [3] [4]

Source: Los Angeles Times