Ask how much electricity a data centre uses and the answer is already complicated. Ask how much water it uses and things get messier still. There are plenty of numbers available. Some measure water used inside the facility. Others include the water needed to generate its electricity.
Some measure what’s taken from a water supply, while others count only what isn’t returned. Then there are estimates that try to turn all of this into the water used by a single AI query. The problem isn’t that these numbers are necessarily wrong. It’s that they often describe different things.
As AI drives more high-density computing into data centres, data center water use is becoming harder for enterprises to ignore. Understanding it starts with something much more basic: why computing infrastructure needs water at all.
Why Do Data Centers Use Water?
Servers produce heat while they work. The harder the hardware works, the more heat a data centre has to remove to keep it operating safely and efficiently. Water can be one of the most effective ways to do that. Some facilities use evaporative cooling, where heat is removed as water evaporates.
Others use closed-loop liquid systems, dry cooling or combinations of different technologies. The exact design changes how much water the facility needs, and how much of that water is actually consumed rather than returned. AI adds another layer because high-density computing can pack much more processing power into the same physical space.
That increases the amount of heat that needs to be managed, even as newer cooling technologies try to do it more efficiently. But cooling is only the most visible part of the story. A 2026 Water Research review identified three parts of AI's wider water footprint: on-site cooling, water associated with electricity generation and water used to manufacture semiconductors.
So even a data centre that uses very little water on-site can still have a wider water footprint through the infrastructure supporting it.
How Much Water Does A Data Center Actually Use?
There isn't one reliable figure for how much water a data centre uses. It depends on the facility, the computing workload, the cooling system, the climate, the electricity supply and several other factors. The scale of that variation is surprisingly large.
A 2025 study published in Resources, Conservation and Recycling found that water use at workload level can vary by more than 10,000 times. The biggest factors included server efficiency, the water used to generate electricity, server utilisation, cooling technology, infrastructure efficiency and climate.
That makes apparently simple per-query estimates difficult to interpret. OpenAI CEO Sam Altman, for example, has said an average ChatGPT query consumes around 0.32 millilitres of water. However, OpenAI hasn't published the methodology behind that figure, so it's unclear which parts of the service's wider water footprint are included.
This doesn't mean per-query estimates are useless. They can help make an abstract resource easier to understand. But an average shouldn't be mistaken for a measurement that applies equally to every query, model or facility. Where the workload runs can change the calculation.
So can when it runs, what hardware processes it and how the data centre keeps that hardware cool. Before comparing the numbers, then, we need to know exactly what they're counting.
Water Use Isn't One Measurement
One of the easiest ways to get confused about data center water consumption is to treat withdrawal and consumption as interchangeable. Water withdrawal is the water taken from a source. Water consumption is the portion that isn't returned to the immediate water system, often because it has evaporated.
A facility can therefore withdraw more water than it ultimately consumes. Then there's the difference between direct and indirect use. Direct water use happens at the data centre itself, particularly through cooling. Indirect use happens elsewhere, such as at the power plants generating its electricity.
A broader water-footprint calculation can extend further into semiconductor manufacturing and the technology supply chain. This creates a fairly obvious problem when companies report different combinations of those numbers. Researchers from the University of Illinois warned in AGU Advances in February 2026 that facility-level disclosure remains limited and reporting isn't standardised.
Published figures also rarely separate direct and indirect water use, making comparisons between facilities difficult. A water figure without its reporting boundaries can tell you much less than it appears to.
Where A Data Center Uses Water Changes The Risk
Quantity is only half of the question. Location can change what the same quantity means. A data centre using water in an area with plentiful supply faces a very different situation from one drawing on a system already serving households, agriculture, manufacturing and power generation during periods of scarcity.
Research published in Environmental Science & Technology in September 2026 mapped 9,347 identified data centres worldwide. It found that 25 per cent would have experienced at least one month of annual "water gap", where regional demand exceeds available supply, under the study's historical hydrological conditions.
Under its future conditions, that rises to 33 per cent. Importantly, the researchers don't claim those facilities caused the shortages. They explicitly say the effect on data-centre operations and other water users remains an open question. There's another issue hiding behind annual consumption figures too: the water system has to cope with demand when it peaks.
A March 2026 preprint estimates that US data centres could require between 697 million and 1.451 billion gallons of additional daily water-system capacity by 2030 if 2024 water-use intensity continues. That's a modelled scenario, not a forecast of measured consumption. The distinction is useful.
A facility's annual consumption might look manageable while its hottest days place much greater demands on local pipes, pumps and treatment capacity. South Africa is already having this conversation. Civil society groups have called for a temporary halt to further data-centre development while resource requirements are investigated, with water, electricity, land and disclosure all part of the debate.
The South African Human Rights Commission told AP it had received more than 250 submissions after requesting public input. Water is becoming a siting and infrastructure question, not simply an environmental metric.
What Does Water Usage Effectiveness Actually Tell You?
Water Usage Effectiveness (WUE) measures how much water a data centre uses relative to the energy consumed by its IT equipment, usually in litres per kilowatt-hour. Lower figures generally indicate greater water efficiency. But WUE still needs context. Microsoft reports a global WUE of 0.27 L/kWh for FY25 across qualifying owned and controlled data centres.
Its calculation uses water for cooling and humidification relative to IT electricity consumption. AWS reports 0.12 L/kWh for 2025, defining its metric as water withdrawn per kWh of IT load. Those figures can help track efficiency, but they don't answer every question.
WUE alone won't tell an enterprise whether water is potable or reclaimed, whether the local watershed is stressed, how much water electricity generation requires or what happens to demand during the hottest part of the year. It's better treated as one part of the picture than a sustainability scorecard on its own.
What Should Enterprises Ask Their Data Center Providers?
Most enterprises don't need to calculate the litres consumed by individual workloads. They do need enough information to understand the water dependency attached to the infrastructure they're buying. That starts with asking more specific questions:
- Is water use disclosed globally, regionally or for the actual facility hosting the workload?
- Does the provider report withdrawal, consumption or both?
- What type of water does the facility use?
- How does its cooling system behave during hot weather, and is the surrounding area already water-stressed?
WUE deserves the same scrutiny. Enterprises should understand how it's calculated before comparing one provider's number with another. Claims around water replenishment or becoming "water positive" also need to be separated from the water physically consumed by the facilities delivering the service. There are signs that reporting is improving.
Uptime Institute's 2026 Global Data Center Survey found that more than half of operators now track water consumption. But tracking water and giving customers enough facility-level information to assess their exposure aren't quite the same thing.
For cloud and colocation buyers, the useful question is increasingly whether they can see enough of that dependency to make an informed infrastructure decision.
Final Thoughts: Water Use Makes More Sense When You Know What You're Measuring
So, how much water do data centres use? There still isn't one number that answers the question properly. The useful answer depends on what water is being measured, how the facility is cooled, what workload is running, where the electricity comes from and where the data centre is located.
Once those differences are visible, wildly different water-use figures start to make considerably more sense. For enterprises, this turns water from a distant sustainability concern into another part of understanding the infrastructure they depend on.
As AI drives more compute into larger and denser facilities, knowing where that infrastructure gets its resources, and how resilient those resources are, becomes part of knowing what you're actually buying. EM360Tech will continue following how those physical dependencies are changing as AI, data and infrastructure grow increasingly connected.
Comments ( 0 )