Expansion of datacentre capacity – lately driven by demand for artificial intelligence (AI) processing – has sparked growing concern about the industry’s impact on water supplies. With datacentres set to multiply, some are warning that the shift towards high-density computing could trigger unsustainable demands on water.
But the debate is characterised by an information gap. On one hand, we can find narratives that rely on worst-case “do-nothing” scenarios and cooling models that may not accurately reflect the UK climate or the evolution of modern hardware. On the other hand, the industry argues it has largely transitioned to cooling systems that radically reduce water consumption, but finding out from planning applications what they actually propose to deploy is a tough job.
This article attempts to get under the skin of these looming projections of a datacentre-driven water crisis. It looks at water company data about the supply situation 20-plus years hence, the nuances of current cooling design, and the glaring information gaps in the planning system that mean we lack transparent industry-wide data and hamper the ability to accurately forecast the true impact of digital expansion.
Push for mandated reporting
Last year’s report from the Government Digital Sustainability Alliance (GDSA) looked in detail at water consumption associated with AI and datacentres.
The report makes some excellent recommendations, but some of the supporting arguments amount to over-egging the pudding. Here we refer to its use of future water deficit predictions and the extent of datacentre water use, which we will deal with later.
The report, Water use in AI and data centres, makes its case around a few key factors, which include:
- Datacentres consume water through cooling systems, electricity generation, and the manufacturing of semiconductor chips needed to equip them.
- England is projected to face a daily water supply deficit of nearly five billion litres by 2050, an amount equivalent to more than one-third of the current public water supply.
- Many datacentres use “evaporative cooling”, which relies on continuous replenishment with fresh, potable water.
- The demands of AI processing, particularly during heatwaves, create the need for water-intensive cooling at the time supplies are most constrained.
- There is a geographical overlap between water-stressed regions and areas where datacentres are clustered.
- Current national water infrastructure plans don’t account for the water requirements of datacentres.
- There is a “transparency deficit” in the industry, in which a minority of datacentre operators track their water usage.
To combat these, the report’s appendix recommends mandating all UK datacentres above 1MW to report location-specific, granular data on water, energy and carbon usage; integrating datacentre water demand into national water resource planning; mandating high-efficiency cooling technologies in water-stressed regions and providing incentives for using non-potable and recycled water sources; restricting new datacentre developments in water stressed areas; and introducing mandatory water efficiency targets and disclosure requirements for cloud and datacentre suppliers.
The report’s recommendations are clear and would make a difference. Water use reporting by commercial facilities is not mandatory in the UK, so no unified set of data exists. Also, anyone who has ever tried to divine projected water use in proposed datacentres via planning application documents will know it is, at best, often obscured, and at worst, and in most cases, completely absent.
Getting an accurate ongoing picture of datacentre water use is vital with regard to planning and consents for datacentre location, as well as to build future water supply infrastructure.
But there are reasons to believe the report paints a slightly grimmer picture than necessary.
Why trust one and not the other?
The GDSA report says England will face a water deficit of five billion litres a day in 2050. This number appears to come from the Environment Agency’s Water Resource Management Plan (WRMP).
The WRMP is the formal, legally mandated blueprint through which UK water companies project their ability to meet customer demand. It provides water supply-demand balance for at least the next 25 years by water resource zone (the local network), water company and geographical region, for England only.
Data is split between “baseline” – what would be the case were no improvements made to infrastructure – and a “final planning” dataset that reflects water company forecasts of supply and demand after they have implemented infrastructure improvement during the planning period.
The GDSA appears to get its “five billion litres a day” figure by adding up the WRMP baseline datasets supply-demand balance for 2049-2050. It equals a deficit of 4,947.7 million litres per day in the dataset updated on 7 April 2026.
What the GDSA report does not cite, however, is the final planning dataset. In that, the April figures show a positive supply-demand balance for 2049-2050 of 895 million litres per day.
So, it appears the report cherry-picked figures to show a grimmer picture than if it had also shown the WRMP final planning data that predicts a water surplus.
Having said that, to account for future datacentre demand over that period and its effect on water supplies is no easy matter.
Many datacentres are in water-stressed areas
What we can see from the WRMP data in the above map (left) is that of the 101 total water resource zones, 76% of them are projected to face some level of water supply deficit by 2045-2046 if no infrastructure improvements take place. We can categorise 67 of these as being areas of minor deficit (between zero and 100 million litres per day), eight as areas of moderate deficit (-300 to -100 Ml/d) and two as severe (< -300 Ml/d).
Then, if we plot the 119 planned datacentre projects to these areas of water stress, 59% of them are within zones currently projected to face a water deficit. These map to:
- Minor deficit: 29 datacentres.
- Moderate deficit: 17 datacentres.
- Severe deficit: 24 datacentres.
Provisos here are that this is the “do nothing” scenario, and that datacentres mapped to areas of water stress are only the 119 currently in planning stages, with the furthest window into the future reaching only to 2037 (the 1GW Elsham Tech Park in North Lincolnshire), while WRMP projections go to 2050 and beyond.
If it seems counter-intuitive that high-rainfall regions like Yorkshire, Lancashire and the Lake District should show as water-stressed, that’s because water supply infrastructure is impacted by environmental protection. Water companies need legal permission to take water from rivers and aquifers to maintain healthy river flows and protect ecosystems. There are also climate change impacts, where water companies now plan for one-in-500-year drought events, and population growth.
While that was the do nothing scenario, the second version of the same map (right) – with water companies’ final planning for infrastructure upgrades included – shows no water region in deficit and so should be able to cope with datacentres located there.
But that doesn’t account for future datacentre development.
Based on the growth trend from existing data (1.6GW in 2024 to 8GW in 2037), datacentre capacity in the UK is expanding at an annual growth rate of just under 14%. If we project those figures, we get to about 7GW by 2036 (4.5 times what it is now) and 24GW by 2046 (15 times the current capacity).
The question is, does the WRMP data take into account this volume of growth?
The other question is, how much does that matter, given what we know about datacentre water usage? It’s an area where it’s important to get terminology right and a field of rapid change right now, so it’s worth looking at.
Cooling vs heat rejection
According to the GDSA report, “evaporative cooling” is a “common method” in datacentres, and requires “continuous replenishment with fresh water”, “utilises potable (drinking quality) water” and can contribute to “significantly lower water levels and impact fish and other aquatic ecosystems in the surrounding areas”.
All those things – bar one – are undoubtedly true.
Before we look at that – and for the purposes of clarity – let’s distinguish between cooling and heat rejection in datacentre design.
In datacentre design, “cooling” refers to the process of transferring heat away from IT equipment to a secondary medium (usually air or liquid), while “heat rejection” is the final step of transferring that heat out of the facility and into the external environment.
While cooling often uses closed-loop systems to circulate water or refrigerant internally, the heat rejection phase is the point where facilities decide between “dry” methods, such as mechanical refrigeration and heat exchangers, or water-consumptive methods, such as evaporative or adiabatic cooling.
To return to the GDSA report’s points, evaporative cooling – or more strictly heat rejection – is incredibly wasteful of water. Its modus operandi is via water piped to an evaporative cooling tower and to a heat exchanger, where a constant supply of water evaporates and carries away heat.
It has been a popular method of heat rejection, notably in the hotter regions of the US. Evaporative heat rejection places less demand on electricity and is well-suited to dry regions where the air can absorb a lot of moisture.
The question is, however, to what extent evaporative cooling is used in the UK.
There is no dataset to refer to here.
Is evaporative cooling common?
What we can say is that many older and smaller datacentres – which make up a large proportion of the UK’s installed base – are far more likely to rely on air-conditioning/refrigerant methods of heat rejection. Of 190 existing datacentres in the UK, 100 are of less than 5MW capacity and 40 more are smaller than 10MW.
We also know that the current stage of evolution in chip design for AI graphics processing units (GPUs) means that direct-to-chip cooling is now mandatory. That’s because the power draw of the new generations of GPU processing is such that it operates at unprecedented wattages and levels of heat generation and at higher temperatures at the chip than were hitherto normal.
In direct liquid cooling (DLC), water is held in two closed loops similar to a car radiator: one for cooling, one for heat rejection. They are filled once. That might mean a few hundred thousand gallons – at a fill volume of roughly 10,000 to 20,000 litres (10 to 20 cubic metres) per 1MW of compute capacity – but it’s not ongoing consumption. Once it is in the system, that’s it.
In the latest datacentre designs, the heat rejection loop goes out to dry coolers that blow ambient temperature air across heat exchangers.
That also works because the latest power-hungry chips can run at much higher temperatures, and so the heat delta between cooling water and the outside air doesn’t need to be as steep to bring the temperature down sufficiently.
That said, systems built with dry coolers sometimes use so-called adiabatic cooling for heat rejection – where water is used to boost the heat exchange process – during particularly warm weather to deal with temperature spikes.
As for purely evaporative cooling, it’s not known how prevalent this is in the UK. Most legacy datacentres rely on liquid cooling with air/refrigerant heat rejection, while newer and larger AI-capable datacentres will likely rely on DLC and dry heat rejection, with some adiabatic during peak temperature periods.
The prevalence of “evaporative cooling” in the UK is therefore speculative due to insufficient reporting data. Accordingly, comparing the projected water footprint of new UK facilities to the consumption profiles of desert-based US datacentres is tenuous.
Opaque planning applications
It’s actually really difficult to determine what methods of cooling will be used in datacentres that are currently planned or under construction in the UK because the planning process doesn’t mandate it, or at least such details are not obligatory until a project is well past initial approvals.
We looked at the top 10 largest datacentre projects in data supplied by construction planning tracking specialist Barbour ABI, and then looked at documents submitted with planning applications.
Of these, only the largest – the aforementioned 1GW Elsham Tech Park – has any kind of information about water use attached to its planning application.
Most of the rest lacked much in the way of documents submitted at all, not least when it came to water use. That’s probably largely because it’s not a requirement to set out anything about water use at proposed datacentres until the detailed stages of local authority approval.
That’s unlike energy or traffic impact assessments, which have clear triggers. So, information on water is sometimes provided during the detailed design phase or in private discussions with utility providers rather than being a transparent, publicly available metric during initial public consultation and planning phases.
To find out more, we made contact with those behind two datacentres on the list – QTS’s 500MW site at Cambois and Amazon’s 185MW (estimated) works at the former Didcot power station.
Elsham, Cambois and Didcot
Elsham Tech Park provided what looked like a comprehensive Water requirements report, in which it told us it will use “‘dry’ systems, with water only being required for the initial filling”.
Inside the datacentre, it is proposed that a chilled water system will keep temperatures down, with refrigerants taking over should temperatures rise beyond a certain level. Meanwhile, GPUs – this is an AI datacentre in an AI growth zone, remember – will be cooled via DLC with the secondary (heat rejection) loop going to dry coolers outside the building.
It said that in its “proposed design … it is likely that for 99% of the year, using dry coolers alone will be sufficient for all the heat rejection from the direct to chip cooling system… For the remaining 1% of the year, chillers will be used to provide any additional cooling that is required. There is no additional water required for cooling, other than the initial filling of the closed system.”
It said annual water use would be 14,416m3, which includes offices for up to 900 people, and is the equivalent of about 96 households’ yearly usage in the region. It contrasts that to what it said would be water use in an adiabatic system per annum of 465,416m3 and in evaporative cooling of 2.017 million m3 yearly.
Regarding QTS’s Cambois datacentre project, the developer said: “In 2025 alone, the QTS closed-loop cooling system saved an estimated 28 billion litres of water worldwide when compared to evaporative cooling.”
Meanwhile, projected water use for the “initial phase” – do they mean the fill? It’s not clear – at Cambois is approximately 2,268m3 and will use air-cooled chillers for heat rejection without adiabatic or evaporative cooling.
Amazon Data Service’s Didcot planning application supporting documents stated: “No water is used to cool the servers for 97% of the year. It is calculated that just 96 hours of evaporative cooling are required annually. Furthermore, no potable water supply is used for industrial cooling purposes, with all industrial water for cooling purposes supplied from an external water treatment facility on the wider Didcot power station site.”
The latter appears to have been necessary because Thames Water “identified an inability of the existing water network to accommodate the needs of this development” in November 2025. In April 2026, it was announced that water treatment firm Gradiant had been awarded a contract to design and deliver a water treatment facility for the datacentre site.
Arup assessment: Trending away from evaporative and adiabatic
In the absence of reliable information about the cooling and heat rejection systems in place and likely to be deployed in planned UK datacentres, we spoke to some industry experts.
Gareth Williams, datacentre business lead at design and architecture firm Arup, said most UK datacentres already use closed-loop systems and reject heat via refrigeration rather than evaporation, that the trend is away from adiabatic and evaporative cooling, and that the UK’s climate means water-assisted methods are only likely to consume water for a small fraction of the year.
“The majority of UK datacentres already use closed-loop water systems. These facilities use water internally but do not consume it through evaporation. In that sense, ‘dry cooling’, where there is no net water consumption for heat rejection, is already well established across the UK market,” said Williams.
“The heat from these closed-loop systems is typically solved through refrigeration rather than through water consumption. We would expect new UK facilities to continue favouring these approaches. Broadly, there has been a move away from adiabatic systems in new UK datacentre designs, though they have not disappeared entirely, and legacy systems remain in operation,” he added.
“Many of the very high consumption figures are based on older legacy systems in the US, where the climate is significantly different. Well-designed facilities operating in a UK climate have materially different performance.”
Patchy and fragmented planning information
The three datacentre projects we reviewed illustrate a broader problem. Data regarding water usage – and indeed, most details necessary for large-scale infrastructure planning – is patchy and fragmented. Among the 10 largest datacentres currently in the planning or construction pipeline, substantial documentation was only available for Elsham, Cambois and Didcot in the planning portals of the relevant local authorities.
Even within this small sample, transparency is inconsistent. Only Elsham disclosed its anticipated water usage upfront. For the Cambois site, we had to contact the developers directly, while the Didcot planning documents provided usage figures only in percentages, rather than absolute volumes. Furthermore, developers do not universally disclose project capacity in megawatts – while Elsham provided this figure, we had to infer it for the other two sites based on the proposed floor area.
This opacity is compounded by the administrative landscape. Planning applications reside on separate, disparate council portals, with supporting information buried in unconnected PDF files – although that’s set to change. Because the minimum information requirements for proposed datacentres are sparse and largely unstructured, collating this data for sector-wide reporting or national infrastructure planning remains an arduous, if not impossible, task.
Lack of transparency needs addressing
The debate over datacentre water consumption is hampered by a significant data gap. While the GDSA report correctly identifies the need for greater industry accountability, its reliance on a “do nothing” baseline and its application of high-consumption US cooling models to the UK market risks skewing the picture.
But the industry’s lack of transparency makes such debate inevitable. Because water usage remains largely invisible until the final stages of the planning process, if at all, there is no reliable way to verify the actual cooling strategies of the UK’s planned infrastructure pipeline.
Meanwhile, there is a temporal mismatch. While water companies attempt to forecast demand decades into the future, the datacentre industry evolves much more rapidly.
The goal should be to ensure water infrastructure planning can keep pace with digital growth. As the report recommends, this requires a mandatory, standardised framework to report operational water usage at the point of application. Without that, the true impact of the UK’s datacentre expansion will remain a matter of speculation rather than rigorous planning.
