Data Centre Water Usage 2026: What the Numbers Actually Mean

How much water do England's data centres use? The short answer

Before any number means anything, four different things get called 'water use' and they are not interchangeable. Withdrawal is water taken from a source — a river, aquifer or the mains. Consumption is the portion that doesn't return to that source, typically lost to evaporation. Direct use is water a facility pulls through its own cooling plant on site. Indirect use is the water consumed somewhere else, upstream, to generate the electricity the facility draws from the grid. A figure quoted without specifying which of these four it measures cannot be safely compared with another figure that measures something else.

With that distinction fixed, the answer for England is this: independent analysis — reported by the BBC — estimates data centres in England use approximately 1.88 billion litres (about 1.879 million cubic metres) of potable water annually, around 0.2% of England's non-household water usage. This is an estimate for England's non-household water market, not an audited UK-wide census, and it says nothing about how that demand is distributed across the country.

The concentration problem: a tiny national share, a large local issue

National averages flatter the picture. The British Resilience Index reports that the Environment Agency's 2025 framework indicates large data centres can consume millions of litres of water per day. That figure comes via BRI's account of the EA framework rather than from a document drawn directly from an EA publication, so treat the specific wording as secondary-sourced pending the primary document — but even allowing for that caveat, a single-site figure of this scale dwarfs what the 0.2% national share implies for any typical facility. A handful of high-demand sites can therefore account for a disproportionate share of the 1.88-billion-litre England total, while most facilities use comparatively little.

There is a second, more basic problem: the British Resilience Index also reports that no official national measurement of UK data-centre water use currently exists, and that the Environment Agency itself has acknowledged it lacks sufficient data to accurately model or predict future data-centre water needs. That gap applies across the UK, distinct from the England-specific 1.88-billion-litre estimate above, which covers England's non-household water market only. That combination — a plausible England estimate, acute site-level concentration, and no statutory UK measurement framework — is exactly why a national percentage cannot answer whether a specific proposed site poses a local risk.

Four metrics hiding behind 'water use'

Hyperscaler disclosures illustrate why the withdrawal/consumption/direct/indirect distinction matters in practice. Google reported 10.9 billion US gallons of total water consumption in 2025 and separately reported replenishing approximately 7.7 billion gallons through 165 projects across 97 watersheds, saying this covered about 78% of its freshwater consumption. Dividing 7.7bn by the 10.9bn headline total actually works out closer to 71%, which suggests Google's 78% figure is calculated against a narrower consumption baseline than the headline total quoted elsewhere in its disclosures. Readers should treat 71–78% as the defensible range for Google's replenishment coverage rather than quoting either number as precise, and in any case replenishment is not the same as avoided consumption: it restores water to a watershed elsewhere, it does not reduce the volume a facility actually used.

AWS disclosed approximately 2.5 billion gallons of data-centre water withdrawals and a 2025 WUE of about 0.12 L/kWh; confirm AWS's stated accounting boundary before comparing these figures with other operators' consumption or fleet-WUE metrics. Each of these numbers — Google's consumption, Google's replenishment, AWS's withdrawals — measures a different boundary, and none of them is interchangeable with a facility-level WUE figure.

Why published figures vary so dramatically

WUE is a widely used data-centre efficiency metric, normally expressed as litres of water per kilowatt-hour of IT energy; operators may calculate and disclose it using different boundaries, and Microsoft's own trajectory shows how much cooling-system design moves it. Microsoft reported an average fleet WUE of 0.27 L/kWh in 2025, down from 2.3 L/kWh in its first-generation data centres — a reduction the company describes as nearly 90%. That is an enormous range for the same company's own buildings, driven by differences in cooling architecture, climate and facility age rather than any single factor.

Even within a single year, regional engineering choices matter: Microsoft reported that its Phoenix data centres achieved a 23% year-over-year reduction in WUE in FY2025 through cooling improvements — a separate, more specific figure from the 0.27 L/kWh global fleet average, and the two should not be conflated. The underlying lesson for anyone comparing operator figures is that facility size, climate, cooling technology and whether the number covers a single site, a region or a global fleet will all move the answer independently of each other. This is the same engineering trade-off examined in our analysis of the liquid cooling threshold, where water and energy savings pull in different directions depending on rack density.

Worked example: turning WUE into daily and annual demand

To make WUE concrete, take an illustrative 10 MW IT load running continuously for a year — 8,760 hours — and label every assumption clearly, because this is a calculation, not a reported figure for any named facility. At 10,000 kW continuous draw, annual IT energy consumption is 87,600,000 kWh.

Apply Microsoft's reported 2025 fleet-average WUE of 0.27 L/kWh and that load implies roughly 23,652,000 litres — about 23,652 cubic metres, or 23.65 million litres — of water a year, averaging around 64,800 litres (64.8 m³) a day. Apply the same load to Microsoft's reported first-generation WUE of 2.3 L/kWh instead, and the annual figure rises to roughly 201,480,000 litres — illustrating, in a single comparison, why cooling-generation design rather than facility size alone can separate two sites by a factor of eight or more. Completing a genuine household-equivalent comparison requires local water-company consumption data that isn't part of this dataset; UK buyers can request that figure directly from their water company or model it using a server room cooling calculator alongside their own site's confirmed IT load.

Does AI use a fixed amount of water per prompt? Fact-checking the viral numbers

No. The widely shared 0.000085-US-gallon figure — about 0.32 mL per ChatGPT query, attributed to Sam Altman — was investigated by PolitiFact, which found Altman's post was the only source it could locate for the number and that it came with no independently verified methodology. That alone should disqualify it as a citable universal metric.

Independent estimates land in a very different range, and disagree with each other for traceable reasons. A 2026 estimate reported a broad range of roughly 0.6–17 mL per ChatGPT-4o prompt, depending on prompt length and whether direct and indirect water use are included; it is an estimate, not a universal measured rate. UC Riverside research, via Kymata Labs, estimated approximately 500 mL of water for every 20–50 queries once on-site cooling and indirect electricity-generation water were both included — a far broader boundary than a single on-site cooling figure. The same research estimated that training GPT-3 alone consumed approximately 700,000 litres of freshwater. None of these figures are comparable to each other because they measure different boundaries; none should be treated as a fixed, universal 'cost' of an AI query.

Cooling choices, local risk, and what to disclose before approving a UK site

Cooling-system choice is one of the biggest levers affecting potable-water demand, and the trade-offs are consistent across the industry even where the brief's data doesn't attach a specific litre figure to each one. Evaporative cooling rejects heat by evaporating water, so it uses more water but less mechanical energy. Air, closed-loop liquid and immersion systems can reduce or shift water demand relative to evaporative designs, but the water and energy outcomes depend on climate, heat rejection, rack density, equipment and the facility's complete cooling system — closed-loop liquid cooling recirculates a sealed fluid loop and immersion cooling submerges hardware directly in a dielectric fluid, and both can substantially reduce on-site evaporative water loss, although the complete facility may still use water for heat rejection, treatment or auxiliary systems; sourcing cooling water from reclaimed effluent or other non-potable supplies avoids drawing on drinking-water-grade mains regardless of which cooling method sits on top of it.

Judging whether a specific proposed UK development is a real local risk — rather than a rounding error against the 0.2% national share — means asking questions the national figure cannot answer: is the site in a water-stressed catchment or river basin already under abstraction pressure; has the local water company confirmed spare capacity in its area rather than assuming it; what happens to demand during drought restrictions; and, as noted above, is the facility one of the small number of high-consumption sites that can reportedly use millions of litres daily under the EA framework? Buyers can also ask whether the Environment Agency's 2025 framework classifies the site's catchment as water-stressed, and request direct confirmation from the local water company of spare capacity for that specific site — rather than relying on the national 0.2% figure. Given that, as noted above, the Environment Agency has acknowledged it lacks sufficient data to model future demand, and that no official UK measurement regime exists, the burden sits with the operator to disclose rather than with regulators to already know.

The following disclosure checklist gives UK buyers a practical template for requesting comparable figures from a proposed operator, because none of the national or hyperscaler figures above can be assumed to apply to a specific site:

• •WUE in L/kWh, with the reporting period and whether it is a single-site or fleet-wide average
• •Whether the figure covers potable withdrawal, consumption, or both, and whether indirect electricity-generation water is included
• •Annual water demand in cubic metres and litres, plus modelled peak-day demand for the specific site
• •Source water (mains, reclaimed, non-potable) and discharge arrangements
• •Cooling system type and backup/drought contingency arrangements

Methodology

This dataset compiles figures from hyperscaler sustainability disclosures (Microsoft, Google), secondary reporting on operator figures (AWS), UK media and sector analysis (BBC News, British Resilience Index), and an independent fact-check and academic estimate addressing AI-query water claims (PolitiFact, Kymata Labs / UC Riverside). Sources were published or updated between June and October 2026 and are cited individually in the text so each figure's scope, boundary and publication date can be checked against the original.

Every figure in this piece keeps the metric, geography, boundary and date stated by its original source; none were averaged, merged or extrapolated into a new number. The single illustrative worked example in this piece is explicitly labelled as a calculation applying a stated real-world WUE figure to a hypothetical facility size, not a measured result for any named site. Where the underlying dataset did not contain a verifiable figure — such as a universal cost-per-AI-query or a full UK household-equivalent water comparison — this piece says so explicitly rather than estimating one.

Sources

Every figure in this article traces to the sources below.

• •Microsoft — fleet WUE history and 2025 average, water-intensity goal progress
• •Microsoft — Phoenix FY2025 WUE year-over-year improvement
• •Google Sustainability — 2025 water consumption and replenishment data
• •PolitiFact — fact-check of the Sam Altman ChatGPT water-use claim
• •Kymata Labs — UC Riverside AI water consumption research summary
• •BBC News — England data-centre water use estimate and non-household share
• •British Resilience Index — UK data-centre water measurement gap and Environment Agency framework