← Blog
blog

Datacentres, water, and the argument that doesn't hold any in the UK

Why the 'datacentres use huge amounts of water' argument doesn't really apply to the UK

/ 33 min read
#datacentres#sustainability#opinion#UK-tech

**The argument that “datacentres are using all of our water” circulates often in the UK press and on social media. If you live in places such as the US this may not be entirely untrue, but the UK is a different picture.

What is the argument?

You probably see it all the time. Open your social media of choice, or even a news site, and there is some article about datacentres (DCs) or AI, and the argument usually looks like this;

  • A new hyperscale (or even standard style) DC is announced in [insert some region].
  • Commentary/comments say something like datacentres “consume vast quantities of water,” “drain reservoirs,” “are worse than agriculture” for local water. etc.
  • Opposition groups cite global figures for water usage (often US data) as if they apply to every DC build.

The underlying anxiety is one I can understand. Water stress and consumption is a genuine concern in parts of England and parts of the South East for example, have not had a new major reservoir built in more than 30 years.

But what does the actual UK evidence have to say when it comes to DC water usage?

It says the argument, as applied to the UK, is largely inaccurate.

How a datacentre actually moves water

Before looking at UK numbers, it helps to really define what we mean by “datacentre water use” and how that might actually function like in reality. Datacentres don’t consume water the way a city or town does. They CAN (not always) “consume” it as part of heat rejection and removal (aka cooling), and the heat rejection technology chosen determines almost everything about how much is “used” and in what way this might be.

There are typically five styles of cooling solution in use today and only one of them is what the critics usually picture.

1. Open (evaporative) cooling towers This is the one that the critics like to draw on most often as an example of DC water usage. Water is sprayed through an outdoor tower and a portion evaporates and carries heat into the atmosphere. The remaining water picks up dissolved solids such as calcium (like rain water picking up dirt) as it cycles, so an amount is subsequently dumped to drain (blowdown) and replaced with fresh “make-up” water. In this case the total water used is the sum of evaporation + blowdown + drift (water typically suspended as particulate in the air). This is the system the “datacentres use lots of water” argument usually has in mind. To throw some statistics out there, this style of heat removal typically requires around 1.6-2.2 L/kWh, which can of course be higher in hot climates (The Green Grid / industry benchmark; Uptime Institute). For a 30 MW facility, an open tower system running at the 1.8 L/kWh industry average works out to roughly 1,294,611 litres of water a day. No matter which way you cut it, that’s a LOT.

To put 30MW into perspective, a typical non-GPU rack is likely to draw at maximum, 12.12kW (32A x2 if a Dual feed @230v) for around 40 1U servers and switching gear. This is 2500 RACKS of 40x 1U servers in a 32MW envelope. Further calculating out this works out to be approximately 100,000 regular servers including switching gear.

When we talk about GPU servers however, this might be something closer to 2000 servers with 16,000 H200 GPUs. At this scale things become a lot more power dense. The physical footprint is lower in terms of raw numbers of racks and servers.

2. Adiabatic systems Functionally a hybrid: an air-cooled heat exchanger runs dry most of the year, and a fine water mist sprays into the incoming air only during peak dry-bulb temperatures. Saturation efficiencies run 50–90%; the goal is to keep the heat-transfer coil dry to prevent scale. Compared to a full open tower, water reduction is up to 90% (DCD on adiabatic; SPX/CTI adiabatic white paper). WUE in cool climates sits around 0.2–0.9 L/kWh. The catch is that adiabatic cooling saves electricity by evaporating water on site. That’s a sensible trade in Arizona where you might want to use less electricity where water is being used to spin a turbine to generate power. In the UK, where generating that electricity barely uses water in the first place, you’re creating a local water demand to avoid one that hardly exists.

3. Closed loop, fill-and-retain liquid cooling Coolant (water/glycol mix, or a dielectric) is sealed inside a loop between the heat source and an external heat exchanger. Typically this is filled once, and recirculated almost indefinitely. The easiest to understand example of how this style of system works here, is a car radiator system, or central heating systems. Closed loop though is possibly the most efficient of the water options due to the constant cycling of the same fill in the system.

Closed loop cooling has a few sub-categories, and gets complicated quickly.

  • Direct-to-chip (DLC): cold plates or sealed cold blocks sit on the processors and a facility wide coolant loop carries heat to a heat exchanger and back. This is the new-build hyperscale design of choice. Pretty much like a typical PC watercooling setup you might find on an enthusiasts desk, on a much grander scale.

  • Rear-door heat exchangers (RDHx): A water-cooled radiator mounted on the back of each rack absorbs heat from exhaust air from the servers, is pumped and ultimately cooled via an external system and the facility water loop is closed. SemiAnalysis treats RDHx as conceptually “an in-rack CRAH” and notes that even with active fans it sits on a closed loop (SemiAnalysis, Datacenter Anatomy Part 2 - Cooling Systems). RDHx style deployments are accelerating in their adoption in Europe. Vertiv’s CoolLoop RDHx (March 2025) supports up to 80 kW/rack. nVent, Motivair, Schneider, Rittal are all expanding EU-portfolios (MarketsandMarkets - Data Center RDHx Market).

Microsoft’s August 2024 directive commits all new builds to zero-water-evaporation closed loop liquid cooling, which the company says avoids more than 125 million litres per facility per year. Oracle has described its 2026 AI datacentre builds (those in New Mexico, Michigan, Wisconsin, Texas) as “direct-to-chip, closed-loop, non-evaporative”. These systems are filled once via tanker, sealed, the cooling fluid remains in the system and isn’t “used up”. Even Nvidia’s Rubin generation is “100% liquid cooling - every chip, every networking component, cooled entirely by liquid in a closed loop with no fans anywhere in the system”, citing a design reduction from ~2.6 million gallons/MW/year to effectively zero on-site water use. Of course this could be simply marketing on their part, but I’d like to see it for myself.

Site water usage for closed-loop systems is around 0.0-0.5 L/kWh, but approaches zero usage when the external stage is air-cooled - which in the UK it usually is.

4. Traditional dry / free-air cooling.

  • CRAC / CRAH units (Computer Room Air Conditioner / Computer Room Air Handler) move heat from the data hall via a refrigerant cycle (CRAC) or via chilled water to a chiller (CRAH). This is pure mechanical cooling with no evaporation at the cooling stage (SemiAnalysis; Supermicro, RDHx glossary). Water is consumed only for humidification should it be in use. This is your traditional “big air con in a room cooling the servers” style setup.
  • Free air cooling draws outside air directly across the IT load when ambient conditions allow, with mechanical chillers as a fallback or additional support. This one is a great option in UK cooling architectures because quite simply, the climate permits it most of the year. Green Grid / LBNL air-side free cooling map shows that 99% of European locations can use free cooling all year round. For comparable central-European cities the TU Dresden / Rechenzentrum-Niedertemperatur study suggests ~6,400-6,650 free-cooling hours per year (out of 8,760) - That’s about 73-76% of the year running dry. UK sites are similarly capable of this percentile due to climate.
  • Hybrid free air cooling is a bit more nuanced but put simply is a mix of CRAH/CRAC, closed loop and free air cooling. This style of system often employs aircon style chillers and cycling heat via external radiators that are fan assisted and are closed loop systems - again like a car radiator. This doesn’t typically have any water usage other than the initial system fill and system maintenance cycles such as you would if you were to drain your central heating system.

Water usage on these systems is around 0.0-0.3 L/kWh. This is the dominant architecture when rack power usage is below ~20 kW/rack (and generally dominant in the UK), and the Uptime Institute Cooling Systems Survey 2025 finds that 75% of operator racks are still below 20 kW, putting them firmly in air-cooling territory. UK colocation operators like Datum in Manchester and Farnborough describe their design baseline as N+1 free-cooling chillers rated to 40°C ambient, with control systems explicitly maximising non-compressor hours.

5. Once-through cooling. Water is drawn from a river, lake or sea, run through the plant once for cooling, and discharged back at higher temperature. Very high withdrawal from water sources, but low consumption (the water is directly returned albeit warmer). Effectively this is never used in modern UK datacentres. It requires a large water body nearby to be effective.

The distinction the public conversation often misses is withdrawal vs consumption. Withdrawal is water removed from a source. Consumption is water lost to that source (mainly via evaporation).

Open towers and adiabatic systems both consume through evaporation and no direct return to source. Closed-loop and free-cooling systems withdraw almost nothing. (Clean Wisconsin - withdrawal vs consumption; ITIF, The Data Center Water Problem Is Soluble (Jul 2026)). Equinix’s own 2024 reporting illustrates the magnitude of the consumptive share for evaporative operators: 1.4 billion gallons withdrawn, 1.2 billion gallons consumed - 85% of withdrawn water is “consumed” (Equinix 2024 sustainability disclosure, via MOST Policy Initiative). If your system evaporates, almost all of the water you take is water you don’t return (though for simplicity sake, we wont get into the water cycle and how the evaporate is returned as rain).

What the UK data actually shows

The most useful data source on this is the techUK / Environment Agency survey of 73 commercial datacentres in England. This was published in August 2025 and has case studies from Ark Data Centres, Pure Data Centres Group, and NTT DATA.

The headline figures from this report are:

MetricFinding
Sites using waterless cooling51%
Sites using hybrid cooling (air + water-cooled chillers)44%
Sites using water-based cooling only5%
Sites using <10,000 m³ of water/year64% (less than a typical leisure centre)
Sites using >100,000 m³/yr4%
Sites that measure water use or use no water for cooling89%

(Source: The Ecologist, quoting the EA’s own dataset from the same survey.)

The 10,000 m³/year number is obviously a point of focus in the public conversation and is worth looking at a little deeper of course. 10,000 cubic metres is 10 million litres, its not a small number but is about what a UK leisure centre uses in a year. A 64% majority of surveyed datacentres sit below that threshold. 4% are heavy users.

Richard Thompson, the Environment Agency’s deputy director for water resources had this to day:

“I am encouraged by the work techUK have undertaken to better understand water usage. The findings suggest UK datacentres are utilising a range of cooling technologies and becoming more water conscious. It is vital the sector puts sustainability at its heart and minimises water use in line with evolving standards.”

That is not the voice of a regulator issuing a warning. It is the voice of a regulator who has looked at the data and concluded the baseline is already reasonable, with the caveat that growth must remain sustainable.

But the same dataset surfaces a counterpoint that techUK’s summary softens somewhat. The WRc analysis of the same 73-site dataset (extended to ~200 facilities and reported in July 2026) found that the water use is heavily concentrated. 6 consumers accounted for 65% of total water use, with the largest single site consuming more than a quarter of the 1,066 million litres used across the sample.

So in the UK, a small minority of sites drive the majority of consumption for DCs. A majority of sites use little or no water at all.

Why the UK is structurally different

The argument travels badly to the UK because the climate, the cooling technology mix, and the build profile are different from the American jurisdictions where the loudest warnings tend to originate.

1. Climate Most cited water-consumption figures come from the US West, where summer temperatures routinely exceed 25°C and adiabatic/evaporative cooling is the only cost-effective option for the climate and the scale. The UK’s climate is more temperate and summer temperatures across most of England sit in the 14–18°C range (though the recent heatwave would have you thinking otherwise). That makes free cooling and air cooling viable for most of the year as well as hybrid closed loop air cooling systems, which is why the techUK survey found 51% of the surveyed estates are on waterless systems and a further chunk of the 44% hybrid cohort uses water only on hot days.

2. The dominant UK architecture is closed-loop / air. As per the techUK survey, only 5% of sites use water-based cooling alone. 51% use no water. The other 44% are hybrid systems with air-cooling first and evaporative assistance in peak conditions. By contrast, the KETOS industry estimate for the global fleet is 75–90% water-based, dominated by mostly US samples.

The UK DC fleet uses a fundamentally different strategy.

3. The WUE numbers Water Usage Effectiveness (WUE) is litres of water per kWh of IT load. The industry target from the Climate Neutral Data Centre Pact is 0.4 L/kWh for cool climates.

Putting those numbers into perspective with some real world examples here (there is likely to be some bias however due to some of these being self reported data points):

The commonly-quoted “average of 1.8 L/kWh” comes from older US-dominated samples and intentionally includes high-WUE adiabatic sites. Applied to the UK market, it’s simply wrong by quite a margin due to the vastly different design choices employed here.

The 2024 US Data Center Energy Usage Report from Lawrence Berkeley National Lab (Shehabi et al., LBNL-2001637, Dec 2024) a US Department of Energy report, not a vendor disclosure, found that all US datacentres combined consumed 66 billion litres (17.4 billion gallons) of water in 2023, with 84% concentrated in hyperscale and colocation facilities. The US has the largest national DC fleet in the world. US hyperscale water demand is projected to reach 60–124 bn L/yr by 2028.

The UK DC market is roughly 1/30th the IT capacity of the US DCs. Even adjusting for cooling-mix differences, that sets the plausible upper boundary for UK annual datacentre water use to the low billions of litres. Thats quite small relative to 15.5 bn L/day entering the UK public network in 2024–25.

4. Scale context. The UK hosts around 348 live datacentre sites (CITP/Baxtel, end-2025) with ~2 GW of operational capacity. (20,000,00 W, approx. 123,762 racks or 4,950,480 servers using our earlier example) A 64% majority of those sites are using less than 10,000 m³ water/year, which is roughly 220 facilities consuming less water than a single leisure centre each, per year. Even the WRc’s full-sample figure of 1,066 million litres/year across ~200 facilities (5.3 million litres/day average) is small relative to total UK public water supply

Putting this into perspective datacentres in the UK are ~0.03% of total public water supply usage in the WRc-measured sample. Even tripling UK capacity by 2030 - which is the the government’s stated 20-fold sovereign-compute ambition, even if it lands as a 3x capacity increase at the operator level, would still leave datacentres as just a tiny fraction of total English demand.

Per day the worst water usage offender is loss through leaky pipes in the networks. England and Wales alone leak 2.9 billion litres a day. That’s a fifth of everything they pump. A 30 MW datacentre on evaporative cooling might consume 1-2 million litres a day at the worst. You could run a thousand such DCs on the water that is leaked through the water network and never reaches a drinking water tap!

Where the concern is reasonable

Its not ALL noise and conflated numbers of course. There are some instances in which there is reasonable concern of course.

1. Concentration risk in water-stressed catchments. The WRc finding that 6 facilities = 65% of UK datacentre water use matters less as a national statistic than as a local one. If those six sites are spread across six different catchments, each one might be manageable. BUT If they’re clustered, and the techUK report’s own language is that operators “choose cooling technologies according to local conditions”, then a hot summer in the South East could put one or two sites into direct competition with households and agriculture in already water stressed areas. The Environment Agency’s Catchment Abstraction Management Strategy (CAMS) process exists for exactly this reason. The new Equinix Slough abstraction licence at 367,500 m³/year shows it being applied to a real hyperscale bid, though it was withdrawn (Withdrawn Equinix UK Limited application, GOV.UK).

2. Hyperscale AI sites are a real engineering conundrum The CITP pipeline shows roughly 10.88 GW in announced or permitted capacity which is a 5.8x expansion if fully built. The big AI sites are liquid-cooled, but they are also larger than the entire existing UK estate and are concentrated geographically. Just look at Crawley, South Mimms, Blyth, the M25 corridor and Great Torrington. The latest hyperscale designs such as Microsoft’s zero-water liquid design, AWS’s closed-loop adiabatic and Nvidia’s GB200 liquid-cooled NVL72 design architecture all claim WUE near zero, but whether those claims hold up at 100MW+ density is the real question, and one the techUK report explicitly flags.

3. The 3-million-litre single-site problem. In July 2026, Water UK told MPs that individual proposed datacentres are requesting up to 3 million litres of water a day for cooling. This is roughly what 7,500 households require per day based on UK average water usage (Water UK in the linked article says this is 3,500 households but that’s a gross over-estimate of how much an average household uses and would imply that the average is 857L per day, but in reality it would be closer to 450L per day for a 4 person household. The 857L per day figure might be possible in peak hot weather however for a household).

The Environment Agency’s National Framework for Water Resources 2025 forecasts a public water supply deficit of up to 5 billion litres/day by 2055 and is explicit that its forecasts exclude emerging demands like datacentres from the non-public-supply category. This is the sort of planning gap that lets a single hyperscale water request cause some real issues. The framework’s own pressures chapter lists cooling datacentres alongside a UK population above 71 million by 2055 but hasn’t put a number on it yet.

4. The water vs energy trade-off. Going dry costs electricity. The Mokkapati & Das siting study found adiabatic saves 30–41% cooling electricity at the cost of 30–210 L of on-site water per kWh saved which is an order of magnitude worse than the upstream water it displaces. And as The Ecologist noted quoting the WRc research, “waterless” datacentres carry a hidden water cost at the power-generation source. So “no water at the site” doesn’t mean “no water footprint” it means the water requirement is moved somewhere else on the ledger, assuming that the electricity production requires water at some point in the generation process. Renewables such as solar and wind of course help significantly here as the reliance on creation of steam go generate power by burning fuels is negated somewhat.

What the UK in-build and in-development pipeline actually looks like

If the public conversation was right that hyperscale builds are designed around water-hungry evaporative cooling, then you’d expect the largest UK sites currently in planning or construction to specify open cooling towers. They don’t. Spot-checking the major named schemes:

The list of designs for new DCs continues onwards, all with the same indicated patterns, including those from Microsoft and Equinix.

The pattern is quite clear. Opt for closed-loop primary, dry cooler or air-side economiser as the cooling technology, evaporative assist is reserved for the small fraction of hours when ambient conditions demand it.

In 2026 specifically, what’s actually been filed, approved, and started this year?

For readers wanting to check the argument against the freshest activity, here is the 2026 specific picture drawn from planning portals, court records, and trade press through September 2026:

Approved in 2026:

Submitted in 2026 (decisions pending):

Construction scheduled to start in 2026:

Setbacks and rejections in 2026:

  • Greystoke / Altrad Iver (Buckinghamshire, 72,000 sqm, ~£1 bn) - government conceded a legal challenge in January 2026, accepting that its planning permission (originally overturned by then–Housing Secretary Angela Rayner in 2025) contained a “serious logical error” because mitigation measures cited in granting approval were never legally secured. Permission expected to be quashed by consent; the developer Greystoke refused to sign a consent order, so the case continues to a substantive hearing. The government’s planning inspectorate had concluded an Environmental Impact Assessment wasn’t needed because the site would use air-cooled servers with minimal water - the same claim that holds up in the technical reality of every other UK hyperscale scheme on this list, but here failed the legal test because no binding mitigation existed. (Computer Weekly, Jan 2026; Mondaq legal analysis; The Guardian, Jan 2026; DCD; Foxglove press release).
  • Carbon3.ai Amlwch Port (Anglesey, former Octel chemical works) This was rejected by Anglesey County Council April 2026 (Baxtel, “Data Center Plans in Wales Turned Down by Council”). Anglesey is one of the five designated AI Growth Zones, so this is a case where the zone designation has not yet translated into approvals.
  • AI Growth Zones “yet to accelerate planning activity” - Data Centre Review (Aug 2026), citing FOI data: since the first zone (Culham, Jan 2025) was announced, only one project in any AI Growth Zone has actually been approved (Data Centre Review, Aug 2026; Barbour ABI analysis). The pipeline is real, but most AI Growth Zone projects are still in pre-application consultation, not construction.

The 2026 TL;DR Every 2026 planning decision and submission that actually names a cooling technology points in only one direction: Closed-loop / air-cooled / heat-reuse designs with water-use targets at or near 0. Where projects were rejected or quashed in 2026, the reason was legal or EIA process and not because anyone built or designed an evaporative site the regulator couldn’t handle.

Closing thoughts

Global figures should not be imported wholesale. A US West datacentre quoted as “the average” does not describe a Yorkshire data hall. The “global average WUE 1.8 L/kWh” claim is the perfect example. It may well be true as a global weighted mean, but false as a UK descriptive statistic. Anyone making the case in the UK press should be able to point to a UK-specific figure and not the global average. The techUK/EA dataset exists for exactly this purpose.

The current UK datacentre estate uses substantially less water than the “they’re draining our reservoirs” doom and gloom comments section implies, and the bulk of the estate is on a trajectory toward zero water consumption regardless of regulation. The heavy users exist, are identifiable to the regulator, and should be. They are being regulated at the site level through the existing abstraction licensing regime, But the argument that is commonly cited, that UK datacentres are a major water problem is not supported by the regulator’s own survey of the industry, and is inconsistent with the technology mix actually deployed here.

That isn’t a reason to stop scrutinising new builds. It is a reason to scrutinise them with UK data, on UK geography, against UK water-stress classifications, with proper accounting for concentration risk, rather than with imported headlines that conflate an Oregon cooling tower with a Manchester air-cooled hall.

Sources:

Peer-reviewed and academic

Government, regulator, and intergovernmental

UK site data (planning, permits, cooling specifications etc)

Independent think tanks and academic institutes

Independent trade press and data

NGO / advocacy

Industry surveys, trade bodies, analyst firms (may have possible bias)

Vendor / commercial sources (self-reported, no independent audit)

| *Sources: 50 sources - 7 peer-reviewed, 8 government/regulator/IEA, 23 UK site planning documents and trade press (in-build and 2026-specific pipeline), 4 independent think tank, 1 NGO counterpoint, 4 industry surveys, 9 vendor/commercial.

Last updated: September 2026