AI Data Centres and Water: Myth vs Reality

Servers generate heat, and most data centres remove it with water, not just electricity. The dominant method, evaporative cooling, sprays or circulates water over hot surfaces so it evaporates and carries heat away — efficient, but the water is gone for good once it turns to vapour.

Not every facility works this way. Closed-loop systems recirculate the same water indefinitely, and air-cooled designs skip water entirely in favour of fans and chillers. Equinix's San Jose facility uses this air-cooled approach, consuming a fraction of the water an evaporative-cooled facility of the same size would use.

The claim has a kernel of truth. Research published in Environmental Science & Technology mapped 9,347 data centres worldwide against monthly gaps between renewable water availability and consumption, and found that roughly one in four sit in areas where demand exceeds supply for at least one month a year — 2,315 facilities in water-stressed zones.

But the national picture tells a different story: Lawrence Berkeley National Laboratory's study of US data centre water consumption found that direct water use is a small fraction of total US freshwater withdrawal, nowhere near the scale needed to cause a national crisis. The myth survives on local examples; it collapses under national data.

Golf courses make the sharpest comparison. A typical 18-hole course uses 90 to 130 million gallons of water a year (Golf Course Superintendents Association of America), and desert courses use far more — in Palm Springs, a single course can consume over 1 million gallons a day, roughly what a family of four uses in four years.

US data centres used about 17.4 billion gallons directly in 2023 (Lawrence Berkeley National Laboratory). The US has more than 16,000 golf courses; at the low end of the range, golf irrigation alone approaches or exceeds total data centre consumption nationally.

The 'data centres use more water than golf courses' claim does not hold up.

A gallon is not a gallon. The source of the water, its quality, and whether it returns to the system all change its real-world impact. Freshwater drawn from a stressed aquifer is a very different proposition from reclaimed municipal wastewater running through a closed cooling loop.

Some operators now use reclaimed water — a genuine mitigation. Others draw potable water in dry regions, which is exactly the scenario that generates local opposition. Consumption is also seasonal: the same facility can be harmless in spring and damaging during a summer drought.

Water researchers split usage into categories, and understanding them clarifies most of the debate. Blue water is surface and groundwater — rivers, lakes, aquifers — the category data centre cooling and irrigation compete for directly, and the one behind nearly all local concern.

Grey water measures the freshwater needed to dilute pollutants to safe standards — a pollution indicator, not a withdrawal. Green water is moisture stored in soil, relevant to agriculture but not data centres. Virtual water is the water embedded in a product or service — including the water used to generate the electricity a data centre draws, which is often larger than its direct consumption and almost never counted in headline figures.

At national scale, data centre water use is real but small beside agriculture, and comparable to or below golf irrigation. The reason it gets attention is growth: direct US consumption rose from about 5.6 billion gallons in 2014 to 17.4 billion in 2023 — roughly a threefold increase in nine years (Lawrence Berkeley National Laboratory).

Two caveats matter. First, this figure excludes the water used to generate the electricity data centres draw, which a University of California study estimated can double a facility's true water footprint once power-plant cooling is included. Second, national averages hide local extremes — a facility in water-rich Virginia and one in drought-stricken Arizona are not the same story, even if the national total looks modest.

The objections that hold up are specific, not general. Four stand out: drought regions, where facilities drawing potable water in already-stressed basins genuinely compete with households and farms; evaporative cooling in dry climates, which turns a manageable load into a significant permanent loss; opaque disclosure, since companies report inconsistently and most off-site water use (like power-plant cooling) isn't captured at all; and rapid, concentrated growth, where multiple large facilities cluster in the same basin faster than water authorities can plan for them.

Each of these is a local planning failure, not a national emergency.

The concern is largely overblown as a national story and largely justified as a local one. Data centres use real water, and that use is growing fast — but measured against agriculture and golf courses, the national total is modest, and the claim that AI is draining the country's water supply does not survive the numbers.

Where the concern holds is specific: dry basins, evaporative cooling in hot climates, and facilities sited without regard to local water balance. A single poorly sited facility in Arizona does more local damage than a thousand well-sited ones in Virginia — which is exactly why national statistics miss the real risk.