People are very worried about the water use of AI, with good reason. There is a constant stream of stories on both social and traditional media which give the impression that data centres are consuming huge quantities of water in cooling.
But most of the stories and statistics come from the US, so I wanted to understand the situation in the UK.
I’ll explain what I found, why I used the data I did and then I’ll look at what I think the biggest issue is and how it might be resolved.
A note: this piece is about direct mains water used by data centres in England, not the water consumed by the electricity they use.1
Data centres don’t currently use that much water
This graph shows the picture according to MOSL, the market operator for the water retail market in England. You can see that in 2024 data centres took 1,879,000 m³ a year from the public water supply, or 1.879 billion litres.
That might sound like a lot, but in comparison to other industries it’s not that much. Data centres use as much water as the postal and courier sector and significantly less than food or beverage manufacturing or sports and recreation.
(As an aside, one of the reasons I think people are cross about data centres is that they don’t realise that everything we do comes with a cost in water. You probably don’t think that about the water involved in every package you get delivered, but Royal Mail says it used 1.3 million m³ of water in 2021-22, driven mainly by “washroom facilities and vehicle and equipment cleaning”, and that’s before you get to the water used in the packaging and the product. If you’re really worried about your personal water use, buy your next pair of jeans on Vinted. Yes, there’ll be water involved in cooling Vinted’s servers, but in it’ll be more than worth it.)
Anyway, back to the graph. If anything, it overstates the extent of the water usage, because the percentages on top of each bar represent the proportion of non-household potable water used by each industry. Factor in household water use and leaks (which are HUGE by the way – I had no idea) and those percentages shrink down even further.
The Environment Agency say that in 2024 to 2025 13.95 billion litres of water was put into the network each day, an annual total of 5.09 trillion litres or 5.09bn m³. (This doesn’t include the energy and agriculture sectors, which are massive users of water, but generally not from the mains.) Rework the percentages on that basis and you find that data centres use around 0.04 percent of England’s total potable water supply.
1.879 billion litres/year ÷ 5.09 trillion litres/year x 100 = 0.037%
And, let’s remember, this is an industry which for better or worse consumes our entire lives. The average UK adult spends 4 hours 30 minutes online each day just in their personal time. I have been looking at national water usage, but when you start thinking about usage per person per hour the sums become infinitesimal.
But I’ve heard figures that are way higher
Me too! A lot of time it’s just water usage presented in a scary way, like the BBC story from last year with the headline: “Scottish data centres powering AI already using enough water to fill 27 million bottles a year”. Only further down did the article note that 27 million bottles a year is about 0.005% of the water supply.
One story that drove me mad when I was researching this piece was an article in The Times with the headline: “Thirsty data centres are sucking up Britain’s scarce water supplies.” It was based on a set of Freedom of Information requests by campaign group Foxglove and a 2022 report for Thames Water by consultants Jacobs. The Thames Water report was the crucial element because almost all of the water apparently consumed by data centres came from the Thames Water – a total of almost 10 billion.
Foxglove and Adam Vaughan at The Times kindly shared their information with me. But even so I was still not much clearer. The Jacobs report showed enormous water use for data centres in London. Its “best initial value” for the Thames Water region was 47.8 million m³ or 47.8 billion litres a year, 25 times larger than the MOSL figure for the whole of England. Even the “low” value was around four times larger.
Then I realised: that 47.8 billion litres a year wasn’t based on measured consumption. It was a “potential demand” forecast, estimating what Thames Water might have to supply if a large share of data centres in its region used the most water-intensive kind of data centre cooling, where the water is evaporated in a constant loop to take away the heat.
In reality, very few data centres in the UK use this kind of cooling – why would you when it’s cold enough to get free cooling for roughly 250 days of the year? So I decided not to use this data and focus instead on measured consumption.
I still don’t know exactly how The Times came up with their figure of 10 billion a year and given the uncertainty around this area it’s possible that it’s correct. But if it is then data centre water usage would be about 1.1% of non-household water usage and 0.2% of total mains usage. More – but not, it seems to be, dangerously so.
Okay, but what’s the catch?
The catch is location and timing.
Most English data centres don’t use much water at all, so the majority of the water use for the industry comes from a few big data centres. According to the Water Research Centre, the top six data centres used an average of 115,800 m3/year in 2024 “and accounted for 65% of the sector’s water use.”2
You can see the problem in the graph below, which shows the metered water use of two anonymous data centres. As you’d expect, their usage spikes in summer. The trouble is, so does everyone else’s.
This is why the water companies are worried: not because the overall amount of water is so large, but because it might increase the demand on peak days.
The nightmare scenario probably goes something like: a cluster of five or so new AI sites is approved in an area with a scarce water supply like Cambridge. There is talk of new reservoirs but they are held up by local politics and the planning system. For a year or two there are no problems – then comes a hot, dry fortnight. Domestic demand rises, river flows are low and reservoirs are being conserved.
At that moment the AI sites’ cooling systems switch into water-assisted mode and try to pull hundreds of thousands of litres from the mains.
The total draw is small nationally but locally it becomes a crisis. Then it hits the media: “AI data centres sucking Cambridge taps dry”. It emerges that because the data centres were classed as Critical National Infrastructure some operators took to this to mean they were entitled to a guaranteed supply. The cry goes up: it’s the government’s fault! Then the local hospital is left without water for a day. It later emerges that this was actually caused by a plumbing error, but by that time it is too late. Spying an opportunity, the opposition parties commit to a two-year moratorium on data centres in the UK. Among nervous government MPs, the mood begins to shift: maybe we need to get ahead of this…
This is the situation we have to prevent – and the good news is that we can. Data centre operators can make their operations more efficient and have water storage tanks on site. The water companies could build more supply; they could even charge the data centre operators to do so. Two large data centre operators told the Water Research Centre they’d enquired about paying for more resilient water connections, the way they do with electricity – but apparently the water companies weren’t interested. Why not?
What happens next?
AI chips run hotter than regular ones so data centre water usage is bound to go up as more AI facilities are built. But that doesn’t mean it will explode. First, because it’s proving really hard to actually build data centres in the first place. Second, because these days a lot of data centres use direct-to-chip liquid cooling rather than evaporative cooling, which can lower water use by as much as half.
The new Isembard-AI supercomputer in Bristol uses this method. The heat from the chips is transferred to a set of fans a bit like the ones in a fridge and most of the time they vent into the air. “Wet mode” turns on when it’s above 19°C in Bristol and if it’s running at full tilt it will use 3.67 million litres of water, equivalent to the annual water use of a few dozen homes.
This isn’t an idle hope. Remember that Jacobs report? They assumed that data centres would use evaporative cooling because it’s the most efficient option. They were wrong.
But still – Isembard might be an exceptional case. So if we want to build data centres without causing problems for the water supply, there will need to be some coordination.
Unfortunately…
The data is bad
The best source on data centre water use in the UK is a report by the Water Research Centre using anonymised data from the market operator, but even this only had information on 257 of the 453 English data centres identified on Data Center Map. The situation doesn’t get clearer if you ask anyone else:
Anglian Water and the Environment Agency also stated that it was difficult to get a clear picture of data centre water use due to the lack of data available for analysis. When asked about this, the data centre operators stated that they are being billed for the water they use, so the water wholesalers/retailers have the data.
This is mad. First, because it allows people to make claims which aren’t related to the facts. The BBC regularly run stories about data centres being built “amid concerns”. Wouldn’t it be helpful to have real data to help put those concerns into context?
Second and more importantly, because decisions are being made on the basis of poor information. In their report for Thames Water, Jacobs explain that the largest water provider in the country did not have metered data from most data centres and the data that it did have was not sufficient to identify or verify cooling-water demand. One of their key recommendations was:
Undertake a specific data collection exercise to clearly identify all the major DC users of water within the TW region…
Did this happen? When I asked Thames Water if they could help with data they said I would have to file FOI requests to find out.
What should we do?
If you’ve made it this far I assume you are interested in technocratic fixes. So let me offer you some more!
Step one: don’t panic. Don’t create another planning veto by drawing water companies into data centre planning applications. Britain is already very good at making it hard to build things and if data centres are going to be part of the country’s critical infrastructure then it doesn’t help to create further blockages.
Step two: don’t do nothing. Data centre figures are a mess and a big part of the problem is that everyone is talking about different things. To give you just one example: according to WRc, water companies think some data centre engineers are being highly conservative, requesting supply capacity for an extreme hot day rather than a typical day. Multiply that by 365 and the annual demand looks enormous. But the site may only need that level of water once or twice a year. This is often the source of worrying new stories about data centres, but more importantly, it’s a problem from the water companies, because if the capacity has been promised, it has to be planned for.
So: make new data centres disclose their likely peak-day demand, cooling method, use of potable water, backup requirements and plans for on-site storage or non-potable sources. Standardise the categories so everyone knows whether they are talking about average use, peak demand, requested capacity, or actual consumption. And encourage projects to build or help build resilience in the system.
Perhaps I am naive – but can this really be that hard?
As far as I can tell most power stations in the UK use very little drinking water and get their water from the sea or freshwater sources, but I haven’t had time to dig into this
Just to go back to my point about everything we do using water: did you know that hospitals use a TON of water? Hand-washing, toilets, kitchens, cleaning, sterilisation, surgical scrubbing, laundry, cooling, dialysis – it all adds up. One of those big data centres uses as much water as a medium-to-large specialist hospital; or, if you must, the daily household water use of a town of about 1,200 people
A very good, nuanced piece. Thank you
I'm not sure many people in the UK will be interested in a pro data center lobby piece. First thought is "who are you like buying on behalf of" second thought is "why"? If it's for the benefit of the world at large there are plenty of far more important issues. Inequality? The rise of fascism? Poverty? Declining values? The impacts of technology on humanity?