Google publishes a WUE of 1.15 L/kWh, AWS of 0.12 L/kWh. That apparent factor of 10 is barely an efficiency gap at all: Google measures water consumed, evaporated for good, while AWS measures water withdrawn, part of which goes back to the environment.
PUE suffers from the same problem. Comparing a hyperscale fleet at 1.09 with a global survey at 1.54 mixes a method effect with a fleet effect. The correction is in the survey itself, which gives 1.44 for sites above 20 MW.
The first question to ask about a water figure
Before the year and before the boundary: withdrawal or consumption?
| Company | WUE | Basis | Year |
|---|---|---|---|
| 1.15 L/kWh | consumption, ISO category 2: water in minus water returned | 2023 and 2024, identical | |
| Meta | 0.19 L/kWh | withdrawal, "water withdrawal divided by IT electricity load" | 2024 |
| AWS | 0.12 L/kWh | withdrawal, "volume of water withdrawn per kWh of IT load" | 2025 |
| Microsoft | data absent | no WUE table in the 2026 fact sheet | n/a |
Litres of water per IT kilowatt-hour consumed. Two definitions circulate under the same acronym. Consumption WUE counts the water that never returns to the basin, essentially evaporation from the cooling towers. Withdrawal WUE counts everything that comes in, including what goes back out.
Without the site's discharge rate, the two cannot be converted into each other. Google states that it consumes 80% of the water it withdraws, which gives an order of magnitude but does not hold for another operator or another site.
The absolute volumes confirm it another way.
| Company | Withdrawal | Consumption |
|---|---|---|
| Google, 2025 | 14,689 million gallons | 10,869 million gallons |
| Microsoft, FY2025 | 13,266 ML | 8,170 ML, 48% of it in water-stressed areas |
| Meta, 2024 | 5,637 ML, 4,145 of it in data centres | 3,123 ML, 2,974 of it in data centres |
PUE also enters the formula for the emission factor per token.
A global volume hides the local impact
Structural limit of aggregated water figures, and Google itself gives you what you need to state it. Its freshwater withdrawals split into 72% in low-risk areas, 15% in medium risk, 13% in high risk.
A cubic metre withdrawn in Finland and a cubic metre withdrawn in Arizona are added into the same total and do not do the same damage. Microsoft publishes the same information in another form, with 48% of its consumption in water-stressed areas.
At sector level, the 2026 UNU-INWEH report puts 2025 at 189 million tonnes of CO₂e, 4,500 billion litres of water and 6,900 km² of land for the world's data centres. Its own contribution is methodological rather than numerical: no design wins on all three axes at once. Dry cooling saves water and doubles electricity consumption. A client moving a workload to a low-carbon region can worsen its water footprint without noticing.
Water per request, and what it leaves out
Google publishes 0.26 mL per median Gemini Apps text prompt in May 2025, on the full boundary, against 0.12 mL on the narrow one.
The formula is explicit: Water/prompt = (E_total − E_overhead) × WUE. Data centre overhead energy is removed before multiplying by WUE, to avoid counting cooling twice. The WUE used is the category 2 one, so on-site consumed water only.
What stays outside is decisive: the water evaporated upstream to generate the electricity is not counted. Most of the gap with other figures in the field comes from there. The EcoLogits methodology adds a PUE × off-site WUE term. Mistral reports 45 mL for a 400-token answer, on a life-cycle assessment boundary.
Google does position its 0.26 mL against "previous estimates of 45 to 50 mL". The two orders of magnitude do not measure the same thing, and putting them side by side misleads in both directions.
For litres per kWh by electricity generation technology, the reference is Macknick et al. for NREL, which separates withdrawal from consumption by technology and by cooling type. Those values were not verified for this course and are therefore not quoted here. The gap between withdrawal and consumption on once-through-cooled nuclear exceeds a factor of 50, and getting that point confused discredits the rest of the deliverable.
PUE, and why you do not compare it raw
| Value | Boundary | Source |
|---|---|---|
| 1.54 | self-declared, weighted average, 681 sites, all types | Uptime Institute, survey April-May 2025 |
| 1.48 | sites commissioned since roughly 2020 | same |
| 1.44 | data centres of 20 MW and above | same |
| 1.42 | average of around 160 French data centres, 23 operators | Arcep, 2024 survey |
| 1.36 | declared European weighted average, 829 sites | EED database, 2024 period |
| 1.09 | Google global fleet | 2026 environmental report, 2025 data |
| 1.14 | AWS fleet | AWS Cloud sustainability, 2025 |
| 1.08 | Meta data centres | Environmental Data Index 2025, 2024 data |
Total energy entering the data centre divided by the energy that reaches the servers. A PUE of 1.36 means 0.36 kWh of cooling, UPS and losses for every kWh of compute.
The ratio excludes two things the Uptime Institute states in black and white: water use, and IT efficiency itself. A PUE of 1.1 on servers running at 10% load is an excellent PUE and a bad data centre.
Comparing 1.09 with 1.54 mixes two effects. The Uptime survey is self-declared, one response per site, across a fleet where 27.5% of respondents work in facilities sixteen years old or more. Hyperscalers publish a fleet average weighted over their own sites, which are recent and large. The correction is in the survey itself: 1.44 for sites of 20 MW and above.
Internal proof that the two are not substitutable: for its category 8, the third-party data centres it does not control, Google writes that it uses "an average PUE based on the Uptime Institute Global Data Center Survey 2025 report". The hyperscaler uses the survey average as an approximation itself.
What EU reporting reveals
Under the Energy Efficiency Directive, European data centres of at least 500 kW report their indicators to a public Commission database. The aggregates for the 2024 period are accessible.
| Aggregate, 2024 period | Value |
|---|---|
| Reporting data centres | 829, across 22 countries |
| Energy | 16.4 TWh |
| IT power | 4.9 GW |
| Water | 5,768,310 m³ |
| Weighted average PUE | 1.36 |
| WUE | 0.45 |
| ERF, heat recovery | 0.0208 |
| REF, renewable share | 0.89 |
The figure that governs how you read all the others is in the technical report of July 2025: 770 reporting sites as of 20 June 2025, or 36% of the 2,161 data centres estimated in the Union. Six member states had supplied no data at all. Coverage ranges from 73% in Germany to 7% in Sweden, with 50% in France and 15% in Ireland.
So the published averages are averages of self-reporting sites, not fleet averages. Two further caveats come with that report: it states itself that it does not represent the Commission's official position, and it is not the report required by article 12(5) of the directive, which is still unpublished.
The same report confirms what the Uptime Institute says about the main driver of efficiency.
| Power band | PUE |
|---|---|
| 500 kW to 1 MW | 1.63 |
| 1 to 2 MW | 1.64 |
| 2 to 10 MW | 1.47 |
| Above 10 MW | 1.21 |
By type of operation: colocation 1.43, co-hosting 1.24, enterprise 1.31. Size is the first driver of efficiency, ahead of operator type.
Question
A client hosts in colocation, in an uninstrumented site whose operator publishes no PUE. Which value do you use by default?
Choisissez une réponse pour voir l'explication.
What to take away
On water, ask one question before all the others: withdrawal or consumption. The rest depends on location, since a cubic metre does not have the same effect in every basin.
On PUE, the default value depends on what you know about the site. A hyperscale fleet PUE is only usable with written confirmation from the provider. For uninstrumented colocation, take the survey average. And a good PUE says nothing at all about the efficiency of the servers it is cooling.
FAQ
Why do the WUE figures published by Google and AWS differ by a factor of 10?
Because they do not measure the same quantity. Google publishes 1.15 L/kWh of water consumed, meaning evaporated and permanently removed from the basin, following the ISO category 2 definition. AWS publishes 0.12 L/kWh of water withdrawn, part of which goes back to the environment. Without the discharge rate of the site concerned, the two cannot be converted into each other.
Which PUE value should you use when the provider does not publish one?
For uninstrumented colocation, 1.54 is the defensible default, the value from the Uptime Institute global survey of April-May 2025 across 681 sites of all types. Google itself applies that survey average to the third-party data centres it does not control. The hyperscale values of 1.08 to 1.14 are only usable if the workload actually runs at that operator and they confirm it.
How much water does an artificial intelligence request consume?
Google publishes 0.26 mL for a median Gemini Apps text prompt in May 2025, on the full boundary. That figure counts only water consumed on site for cooling and excludes the water evaporated upstream to generate the electricity. Mistral reports 45 mL for a 400-token answer on a wider life-cycle assessment boundary, which makes the two values non-comparable.
Does PUE measure the efficiency of a data centre?
Only partly. It divides total incoming energy by the energy reaching the servers, so it measures the overhead of cooling and electrical distribution. The Uptime Institute notes that it excludes water use and IT efficiency itself. A site reporting 1.1 with servers running at 10% load has an excellent PUE and very poor real efficiency.
Is EU data centre reporting reliable?
Its aggregates are public and usable, with one major caveat on coverage. The technical report of July 2025 counts 770 reporting sites as of 20 June 2025, or 36% of the 2,161 data centres estimated in the Union, with six member states supplying no data and coverage ranging from 73% in Germany to 7% in Sweden. The published averages are therefore averages of self-reporting sites.