Data centre water use: which figures can you actually compare?

Two figures about a data centre’s water use can both be accurate yet incomparable. One may measure water withdrawal, the other the portion consumed; one may stop at the building, while the other includes electricity generation. Before comparing litres, identify the quantity, boundary, period and location.

This guide helps you read an environmental report or project dossier. It does not calculate the footprint of an AI query or provide a global average to apply to a particular site.

Withdrawal and consumption describe different things

Withdrawal is water taken from a resource. Consumption refers, among other things, to the portion evaporated, incorporated into products or otherwise removed from the immediate water environment. These distinctions are set out in the US Geological Survey glossary.

Some withdrawn water may therefore be returned. This does not make its use impact-free: the location, timing, quality or temperature of the return can matter in a local analysis. Conversely, calling every withdrawal “consumption” can suggest that the entire volume disappears from the resource.

Do not automatically infer consumption from a simple difference between two meters without checking the boundaries. A volume may be stored, transferred or accounted for in another period. For a real installation, the balance must explain relevant inputs, outputs and changes in storage.

Unpack the boundaries of the published figure

A balance can cover cooling in one building, the whole site or a broader chain. Producing the electricity consumed by the centre is a different boundary from water used directly inside it.

Published figure	Information needed to interpret it
Litres withdrawn on site	Source, period and destination after use
Litres consumed on site	Definition used and measurement or estimation method
Water associated with electricity	Attribution method and electricity generation considered
Total across several sites	Sites included and local differences concealed
Litres per unit of activity	Definition of the denominator and allocation method

Berkeley Lab’s 2024 report on US data centres illustrates another important boundary: it contains national estimates and scenarios, not a universal measurement of each installation. A projection must remain a projection, even when it comes from a public laboratory.

Avoid adding figures with overlapping boundaries. A company total may already incorporate an estimate associated with electricity. Adding a second estimate without checking the definitions would count some water twice.

What WUE relates to IT activity

Water Usage Effectiveness, or WUE, relates a volume of water to the energy used by IT equipment. The DOE/NREL guide of July 2024, section 8.3, distinguishes a site-level version, in litres per IT kilowatt-hour, from a boundary that includes water associated with energy production.

Seeing “WUE” therefore does not remove the need to read the method. Ask which water appears in the numerator, which energy appears in the denominator and whether both cover the same period. A value without these details remains difficult to compare.

This indicator relates water to an energy-based activity measure. It does not directly express the usefulness of the computation or pressure on the local resource. A ratio can be useful for monitoring a site while remaining insufficient to assess a project as a whole.

A falling ratio can accompany a rising volume

Take an entirely fictional example, with an identical definition across two years. In the first year, a site records 1 million litres for 2 million IT kilowatt-hours: the ratio is 0.5 litres per kilowatt-hour. In the second, it records 1.5 million litres for 5 million kilowatt-hours: the ratio is 0.3.

The ratio falls by 40%, but water volume rises by 50%. Both results are compatible. A presentation limited to the ratio would describe a relative improvement; it would not show the absolute increase in volume.

Fictional example	Year 1	Year 2
Volume within the same boundary	1,000,000 L	1,500,000 L
IT energy	2,000,000 kWh	5,000,000 kWh
Ratio	0.5 L/kWh	0.3 L/kWh

This calculation describes no operator. It shows why both lines, volume and ratio, should appear together. The same precaution applies when a lower average results from adding a very different site to a company total.

Location and season change the question

For a local project, an annual average is insufficient to describe periods when water is most constrained. A useful dossier must connect the site’s needs to the relevant resource and seasonal variations. If only annual data are available, the absence of a seasonal view must remain visible.

Also ask what kind of water is used: drinking water, surface water, groundwater or reused water, depending on the actual case. These categories are not certificates of harmlessness. They help frame the right question about the resource and associated treatment, rather than interpreting every litre as equivalent.

This checklist is a reading recommendation by Pnebbula. It does not replace a hydrological study or local regulatory dossier. It helps identify which evidence is missing before concluding that a use is compatible with its surrounding area.

Why water per AI question requires more evidence

Moving from a site balance to a query requires an allocation rule. You need to know the service executed, its attributable consumption, the location and period, then the water boundary selected. A data centre average divided by a number of queries is not automatically a measurement of each query.

Equipment can serve several uses, and queries do not all require the same amount of work. Without data supporting attribution, announcing a precise figure per question conveys a certainty the original balance does not provide. Explaining the missing data is more informative than turning an average into a universal constant.

The guide to data centre infrastructure (in French) describes the systems involved. To extend the analysis to other impacts, the guide to lifecycle assessment (in French) explains why boundaries must remain consistent.

Sources and scope

Consulted on 22 October 2026: USGS, Water Science Glossary, hydrological definitions; DOE/NREL, Best Practices Guide for Energy-Efficient Data Center Design, July 2024, section 8.3; Berkeley Lab, 2024 report page, scope of the study. The table values are fictional; no French site’s water balance is assessed.