Their’s a lot of talk around data centres and water at the moment, particularly as AI drives demand for more computing infrastructure.
Some of the numbers being shared are huge, and the conversation can quickly become either "AI is draining our water supplies" or "there's nothing to worry about".
As usual, the reality sits somewhere in the middle.
Data centres do use water. Some use a lot of it. But how much they use depends heavily on where they are, how they're designed, how they're cooled and where their electricity comes from.
So here are the facts.
Why do data centres need water?
Servers generate heat, and that heat needs to be removed. Some data centres use evaporative cooling, where water helps remove heat from the facility. Others rely more heavily on air cooling, chillers or increasingly liquid cooling systems that take heat directly away from high-powered processors.
This distinction matters because not every cooling system actually consumes water in the same way.
A closed-loop system, for example, can circulate the same cooling liquid repeatedly. Water may still be required elsewhere in the facility, but that is very different from a system continually losing water through evaporation.
There is also indirect water consumption to consider. Electricity generation itself can require water, particularly with some thermal power stations. So when you see estimates for the "water footprint" of a data centre or AI, check whether they are talking about water used at the facility or water associated with generating its electricity.
They're not the same thing.
How much water are we actually talking about?
In the US, Lawrence Berkeley National Laboratory estimated that data centres directly consumed around 66 billion litres of water in 2023. That's clearly a substantial amount, but large national numbers don't necessarily tell us much about an individual data centre, particularly when we try to apply US figures to the UK.
A 2026 House of Commons Library briefing gives us some useful UK context. Research based on anonymised water-meter data from 257 data centres estimated that they consumed almost 1.9 billion litres of potable water per year in England.
Again, that sounds like an enormous number. Put into context, though, it represented around 0.2% of non-household potable water consumption. The distribution is perhaps even more interesting, more than two-thirds of the data centres studied used less than one million litres each year. Just six facilities accounted for 65% of the measured consumption, and the top 15 accounted for 80%.
In other words, talking about the water consumption of "data centres" as though every facility behaves in roughly the same way isn't particularly helpful. They don't.
Where does AI fit into this?
AI has brought much more attention to the subject because the infrastructure behind it can be extremely power dense. Modern GPUs use significant amounts of electricity and generate significant amounts of heat, put thousands of them together and cooling becomes a serious engineering challenge. That means continued growth in AI infrastructure could increase demand for both electricity and water.
But there is another side to it. Cooling technology is changing precisely because operators need to deal with these higher-density workloads. Microsoft, for example, has introduced a new data-centre design using closed-loop chip-level cooling, the cooling liquid circulates between the servers and chillers without requiring a fresh supply of water to replace evaporation. Microsoft estimates that the design can avoid more than 125 million litres of water per facility each year compared with the evaporative cooling systems it replaces.
That doesn't mean the whole data centre suddenly uses no water, there are still normal building requirements and there is still the wider water footprint associated with energy generation. But it does show why simply taking today's water consumption and multiplying it by expected AI growth doesn't give us a particularly reliable picture of the future. The infrastructure is changing too.
Location matters
This is probably the part of the discussion that matters most. A litre of potable water used in a region experiencing water stress is not the same practical issue as a litre used somewhere with plentiful water resources. Likewise, a data centre drawing drinking water from the public network is different from one using reclaimed wastewater, rainwater or another non-potable source.
This is why a national percentage can sometimes hide a local problem. Data centres might represent a relatively small proportion of total water consumption nationally, while still putting significant additional pressure on a particular water network.
So when a new facility is proposed, the sensible question isn't just how much water the data-centre industry uses nationally, it's how much this facility will use, where that water will come from and whether the local supply can comfortably support it.
There are trade-offs
There is another point that tends to disappear from headlines: reducing water use can have consequences elsewhere. Evaporative cooling is popular partly because it can be very energy efficient, switching to a cooling technology that consumes less water may require more electricity, depending on how that electricity is generated, electricity production itself may have a water footprint.
So you can't really look at water, energy and carbon as completely separate problems. The best answer for a data centre in one part of the world may not be the best answer somewhere else. Climate, local water availability, electricity generation, the type of computing being carried out and the design of the facility all matter.
What should we actually be asking?
I don't think the right conclusion is that concerns about data-centre water use are overblown, the growth of cloud computing and AI means we're building more physical infrastructure, and that infrastructure has a real environmental footprint.
Water absolutely needs to be part of that discussion, but broad claims about how much water "AI" or "data centres" consume can hide more than they explain. If we want a useful discussion, we need to ask better questions:
- Where is the data centre being built?
- Is the area already under water stress?
- Is it using potable or reclaimed water?
- What cooling technology is being used?
- Is water being consumed through evaporation or circulated in a closed loop?
- How much does the facility actually consume?
- And what is the water footprint of the electricity supplying it? Those answers tell us far more than one global headline number.
The cloud has always been a slightly misleading name, behind every cloud service, AI model and online application sits real infrastructure: buildings, servers, networks, electricity and cooling. That infrastructure has an impact and we should understand it properly, but understanding it properly also means recognising that not every data centre is the same.
Water consumption matters. The context behind that consumption matters just as much.