AI's hidden water footprint may have reached 765 billion litres in 2025 — consumed in data-centre cooling and the power plants supplying them. That's more than twice the volume of bottled water the entire world buys in a year.

AI feels almost weightless when it arrives through a browser. A prompt goes in, an answer comes back, and the machinery doing the work remains out of sight.

A 2025 analysis tried to make part of that machinery visible. It estimated that AI systems could have consumed between 312.5 billion and 764.6 billion litres of water during the year, once water used inside data centres and at the power plants supplying their electricity was counted.

The top of that range is enormous. It is also not a meter reading. It is the upper end of a model built from incomplete corporate disclosures, average data-centre performance and an estimate of AI hardware power demand.

I think that distinction matters because a dramatic number can be directionally useful while still being much less precise than it sounds.

Where the 765 billion-litre figure comes from

The estimate comes from a peer-reviewed paper by Alex de Vries-Gao in Patterns. He began with an earlier estimate that AI systems were drawing about 9.4 gigawatts at the end of 2024 and could reach 23 gigawatts through 2025.

He then applied a water intensity of 3.40 litres per kilowatt-hour. That figure was derived from environmental reporting by Google, Meta and Apple, paired with grid data for their US data-centre locations.

Holding that intensity constant produced the range of 312.5 to 764.6 billion litres. The lower number corresponds to the lower power-demand estimate and the upper number to the higher one.

This is one study, not a settled measurement of AI’s global water footprint. De Vries-Gao repeatedly says the uncertainty is significant. The analysis was published online in December 2025 and appeared in the January 2026 issue of Patterns.

Most of the estimate is not water poured over servers

Water consumption has two operational parts in this analysis. Direct consumption occurs at the data centre, often when cooling systems evaporate water to carry away server heat.

Indirect consumption occurs when power stations consume water while generating the electricity used by the facility. Thermal plants, including coal, gas and nuclear stations, may use water to make steam and remove waste heat.

The paper cites an International Energy Agency estimate of 560 billion litres for total data-centre water consumption in 2023. About 140 billion litres were direct, while 373 billion litres were tied to electricity generation. A further 47 billion litres came from hardware manufacturing.

That split is important. The AI figure covers operational water from cooling and electricity. It is not a full lifecycle assessment, so water used to manufacture chips and construct facilities sits outside its boundary.

The IEA separately reports that data centres used roughly 415 terawatt-hours of electricity in 2024, about 1.5 per cent of the global total. It expects data-centre demand to rise to around 945 terawatt-hours by 2030 in its base case, although the pace depends heavily on AI adoption and efficiency.

The missing data create a very wide range

Operators generally report company-wide or fleet-wide environmental totals. They do not consistently separate AI from search, storage, office software and other cloud workloads. That leaves researchers trying to infer AI’s share from hardware shipments and broader data-centre averages.

Location matters as well. The paper found water-intensity estimates ranging from 0.68 to 11.98 litres per kilowatt-hour across the relevant US power grids. The same computation can therefore carry a very different footprint depending on where and when it runs.

Lawrence Berkeley National Laboratory estimated that US data centres consumed 66 billion litres directly in 2023 but nearly 800 billion litres indirectly through electricity use. Those numbers cover all data-centre workloads, not AI alone, but they show why off-site water can dominate the total.

They also show why corporate cooling figures tell only part of the story. A facility can reduce the water evaporated on its own site while remaining connected to a water-intensive electricity supply.

The bottled-water comparison depends on the boundary

A United Nations University review estimated annual sales of three major bottled-water types at about 350 billion litres. Against that figure, 764.6 billion litres is more than twice as large.

The same review said the total rises to around 470 billion litres when every bottled-water type is included. The Patterns paper used another market estimate of 446 billion litres and described AI’s possible water footprint as being in the same range as global bottled-water consumption.

The comparison is useful for scale, but only if its definitions stay attached. The AI range is modelled water consumption, meaning water that becomes unavailable for immediate reuse. Bottled-water data measure product volume sold. They are not measurements of the same process.

Global totals can also hide the question that communities care about most: where the water comes from. A litre consumed in a water-rich region does not impose the same local pressure as a litre taken during drought or from an already stressed basin.

Efficiency can move water rather than remove it

Cooling technology can cut water consumed on site. Microsoft said in its 2025 environmental report that a new design for AI workloads consumes no water for cooling and could avoid 125,000 cubic metres per facility each year.

That is a meaningful engineering change. It does not automatically make the system water-free. The electricity supply can still have a water footprint, and dry cooling may require more energy in some conditions.

Other levers include locating flexible workloads where water stress and grid water intensity are lower, using reclaimed rather than potable water, improving chip and software efficiency, and scheduling computation when cleaner electricity is available.

No single metric captures all those trade-offs. Site-level direct consumption, indirect power-related consumption and local water scarcity each answer a different question.

Transparency is the finding that matters most

I recently wrote about why the bottle-per-10-to-50-responses claim is not a universal measurement. The 765 billion-litre figure deserves the same care.

Companies could narrow the uncertainty by reporting where AI hardware operates, how much electricity it uses, and site-level values for power usage effectiveness and water usage effectiveness. They should distinguish withdrawal from consumption, direct from electricity-related water, and AI from other cloud workloads.

Until then, 765 billion litres is best read as a warning about the possible scale of an opaque system, not as a globally measured total accurate to the nearest litre.