How much water does artificial intelligence actually use?

1Summary for decision-makers

Where the water of an AI query goes

GPT-3 estimate (Li et al.): the bulk is not the cooling of the data centre, but the water consumed to produce its electricity.

Off-site — electricity generation On-site — cooling

See the data

Item	mL / query	Share
Off-site (scope 2)	14.704	87%
On-site (scope 1)	2.200	13%

Li, Yang, Islam, Ren — Making AI Less Thirsty, US average. Confidence: high on the dominance of the indirect share.

2Essential concepts

Without these four distinctions, no figure in this report can be interpreted.

2.1Withdrawal (prélèvement) vs consumption (consommation)

• Withdrawal: the volume taken from a river, an aquifer or a mains network. Most of it can be returned to the environment.
• Consumption: the fraction that does not return to the basin — evaporated, incorporated into the product, or discharged elsewhere.

In a data centre cooling tower supplied with good-quality water, about 80% of the withdrawal is evaporated, and therefore consumed (Li et al., section 2.2; ratio independently confirmed by Google). For air cooling with evaporative assistance, Meta reports ~70%. Unlike an open-loop power plant or domestic use, a data centre is therefore a facility where withdrawal and consumption are of the same order of magnitude. HIGH

2.2On-site water (scope 1) vs off-site water (scope 2)

• On-site: the cooling of the building.
• Off-site: the water consumed to produce the electricity the data centre draws (power plant cooling, evaporation from hydroelectric reservoirs).

Scope 3 — water embodied in the manufacture of chips and servers — is excluded from almost all public figures, for lack of data. Li et al. exclude it explicitly.

Direct consequence: a low WUE (on-site efficiency) does not mean a low water impact. Meta posts 0.19 L/kWh thanks to air cooling, a technique that consumes more electricity — and therefore more off-site water. The impact is shifted, not removed. HIGH

2.3Green, blue and grey water (Water Footprint Assessment Manual, Hoekstra et al., 2011)

Data centre water is almost exclusively evaporated blue water. Any comparison with a green+blue+grey total is a category error. HIGH

2.4WUE (Water Usage Effectiveness)

Litres of water per kWh of IT load. Major pitfall: there are two non-comparable variants.

Comparing AWS's 0.12 with Google's 1.15 is a category error. Moreover, the origin of the “industry average of 0.84 L/kWh” put forward by Amazon is not sourced, and the reference year of its figure varies from one Amazon page to another. Confidence on the incomparability: HIGH.

A notable fact: the 2026 Google environmental report no longer publishes a fleet-average WUE (only a PUE of 1.09). The 1.15 L/kWh value exists only in the August 2025 technical paper.

Withdrawing is not consuming

On one and the same object, the gap between the two notions reaches a factor of 291. It is the leading source of error in the public debate — and the scale below is logarithmic, without which the consumption bars would be invisible.

Withdrawn — mostly returned to the environment Consumed — does not return

See the data

Object	Withdrawn	Consumed	Factor
Nuclear reactor, open loop (1 kWh)	168 L	1.02 L	165×
Fairphone 5 smartphone (life cycle)	92,900 L	319 L	291×

Macknick et al. (2012); Fraunhofer IZM for Fairphone (2024). Logarithmic scale — the exact values are in the table.

3What AI consumes

3.1Training

GPT-3 method, coefficient by coefficient — each applies to a different energy figure, and that is what explains the result:

Decisive limitations. The actual location of GPT-3's training is not public: the authors sweep across Microsoft sites. The 700,000 L figure is the most cited in the world and the most systematically taken out of context — it is scope 1 alone, i.e. 13% of the estimated total. Water embodied in the chips (scope 3) is absent. Finally, Mistral's 281,000 m³ is not comparable with GPT-3's 5,439 m³: LCA scope against operational scope, and 18 months of cumulative use against training alone.

What can be retained: HIGH — training a large model represents a water volume on the order of a few thousand to a few hundred thousand cubic metres depending on the scope. LOW — any precise value.

What “the water of a query” means, depending on who is measuring

Three published figures, a factor of 173 between them — and none measures the same thing. Scope weighs more than technology.

See the data

Source	mL / query	Scope
Google (August 2025)	0.26	Scope 1 — on-site cooling
Li et al. (GPT-3)	16.904	Scope 1 + 2
Mistral (July 2025)	45	Full life-cycle assessment

Company reports and Li et al. These three values are not comparable with one another — which is precisely what the chart shows.

3.2Inference

This is where most of the public debate plays out, and this is where the scopes diverge most.

How to read this gap from 0.26 mL to 45 mL (factor 173).

This is not a disagreement about physics but about scope:

• Google measures only the cooling of the building, and excludes the water used for electricity generation.
• Li et al. measure cooling + electricity generation, the latter accounting for ~87%.
• Mistral measures a life cycle, including upstream stages and infrastructure. The page consulted does not explicitly settle whether the 45 mL includes electrical scope 2.

Two further reservations, symmetrical.

• Li et al. themselves describe their estimate as “conservative”: it rests on 0.004 kWh per query, whereas a Llama-3-70B already consumes ~0.010 kWh and a Falcon-180B ~0.016 kWh. GPT-3 is a 2020 model, not representative of current models.
• Google publishes a median, not a mean, and states that the distribution is strongly skewed. The figure covers text-only prompts (no image, no video) and a single point in time.

3.3Company-level volumes (all activities, AI not isolated)

Three comparability warnings — Confidence: HIGH. 1. Microsoft retroactively restated all its historical series in 2026 after having “identified previously unreported water volumes”. Earlier publications (including the 29 billion L / 23 billion L of 2023) are not comparable as they stand with the recent figures. 2. Amazon does not publish a consumed volume, only a withdrawal: no comparison with Google or Microsoft is possible. 3. The 2025 Google metrics are subject to limited assurance by a third party — a low level of assurance, not an audit.

On growth: HIGH. Google +34% in consumption in one year, Microsoft +22%. On its attribution to AI: undetermined. Google attributes it to “growing demand for digital services” without quantifying the AI share.

3.4Global aggregates

On Li et al.'s 2027 projection: it is a withdrawal, not a consumption; 97–98% comes from scope 2, so the value depends almost entirely on an average US electricity factor (43.83 L/kWh) applied to the entire world; and the energy demand projection (93.5–147.4 TWh) comes from a third-party source. LOW

An inconsistency worth flagging. The ~560 billion L/year attributed to the IEA for “today” sits poorly with the 513 billion L (140 + 373) attributed to the IEA for 2023 and relayed by de Vries-Gao. The scopes probably differ, but this could not be confirmed, as the original report is inaccessible.

4Putting the figures in perspective

4.1How to read the tables that follow

Three rules, all of HIGH confidence: 1. Only compare columns of the same nature: blue water consumed with blue water consumed. 2. The domestic uses listed are withdrawals, not consumption: most of it goes down the drain, is treated and returned. 3. The reference food footprints cover 1996–2005 and have not been recalculated globally since. As yields have improved, the per-kg values are probably 10 to 25% lower today for cereals — but no published reassessment allows this to be quantified. Confidence on the order of magnitude: HIGH; on the 2026 value: MEDIUM.

4.2Food — the headline total vs the genuinely comparable component

Three takeaways — Confidence: HIGH.

• Chocolate, the champion of shock figures, has a smaller blue footprint than a kilo of wheat (198 L against 342 L). The total misleads as to the real pressure on the resource.
• Extensive grazing has the largest total footprint for beef (21,829 L) and the smallest blue footprint. Counter-intuitive, and documented.
• For soda, 99.7 to 99.8% of the footprint is upstream: the plant's operational footprint is only 0.5 L — exactly the water that becomes an ingredient. The same upstream/on-site gap applies to AI, in the opposite direction.

The figure that circulates, and the one that counts

Spectacular food footprints are mostly rainwater. The component comparable to a data centre's water is blue water — withdrawn from rivers and aquifers.

Total footprint (green + blue + grey) Blue water only — the comparable component

See the data

Product (1 kg)	Total (L)	Blue water (L)	Green share
Chocolate	17,196	198	98%
Beef	15,415	550	93%
Wheat	1,827	342	70%
Rice	1,673	341	68%

Mekonnen & Hoekstra (2011, 2012), 1996–2005 data. Chocolate has a smaller blue footprint than a kilo of wheat.

4.3Everyday life (withdrawals unless stated)

4.4Textiles and objects

4.5Electricity — the key to AI's scope 2

Operational values only: excluding the upstream fuel cycle and equipment manufacturing.

Four readings — Confidence: HIGH unless stated.

• Once-through nuclear is the textbook case of the confusion: a 165× gap on the same plant. Thermal discharge remains a real ecological impact, but it is not water consumption.
• Wet-cooled concentrated solar consumes more water per kWh than nuclear or coal, and it is generally sited in arid areas.
• The zeros for PV and wind are operational zeros, not life-cycle zeros: silicon purification, steel and concrete are out of scope. Saying “solar consumes no water” on that basis is unwarranted. Confidence on the full life cycle: UNKNOWN, from this source.
• Hydropower is by far the largest consumer per kWh, with a variability of a factor of 2,800 — but allocating reservoir evaporation to electricity is the subject of an open methodological controversy that the sources consulted for this work do not document (a dam also serves irrigation, flood control and recreation: charging all of the evaporation to it is a choice, not a given). Confidence in the central value of 245 L/kWh: LOW.

The water consumed to produce a kilowatt-hour

Since most of AI's water is its power plant's water, the electricity mix of the country where the data centre runs decides almost everything.

See the data

Source	Consumed (L/kWh)	Range
Hydropower	245	1.1 – 3,046 depending on the dam
Concentrated solar (wet-cooled)	3.43	2.74 – 4.20
Coal, cooling tower	2.60	1.82 – 4.16
Nuclear, cooling tower	2.54	2.20 – 3.20
Combined-cycle gas, tower	0.78	0.49 – 1.14
Photovoltaic	0.004	0 – 0.019
Wind	0	0 – 0

Macknick et al. (2012); Mekonnen & Hoekstra (2012) for hydropower. Operational values: excluding equipment manufacturing. The zeros for solar and wind are operational zeros, not life-cycle zeros.

4.6Orders of magnitude — honest comparisons

The conversions below are derived calculations by the author; they are illustrative only. They use only references of blue water consumed, the only quantities comparable to a data centre's water.

What this table shows — Confidence: HIGH. The gap between the three columns (a factor of 173) is larger than the gap between the rows. In other words: the choice of scope weighs more on the result than the use being compared. That is why these equivalences, whichever camp brandishes them, settle nothing.

What this table does not show. A shower is an individual act; AI is an infrastructure. Comparing a query to a shower answers the question “is my personal use heavy?” — not the question “is the concentration of data centres in this watershed sustainable?”. These are two different questions, and the second is the one facing communities.

5What actually determines the impact

The global volume is the wrong indicator. Water is not a fungible resource: a cubic metre evaporated in Ireland and a cubic metre evaporated in Arizona do not have the same effect.

5.1 Location, first. At Microsoft, two sites illustrate the range: Singapore, 423 ML withdrawn for 338,845 MWh; Dublin, 18 ML for 1,308,581 MWh — an intensity gap of more than 90× between two sites of the same company, in the same year. (Confidence: HIGH; these are withdrawals, consumption per site is not published.)

5.2 The local electricity mix. The water intensity of US grids ranges from 0.68 to 11.98 L/kWh (a factor of 17.6). Washington State's EWIF is 9.501 L/kWh, that of Texas 1.287. Since companies do not disclose where their AI workloads run, this margin of error cannot be reduced. (Confidence: HIGH.)

5.3 Cooling technology, and its communicating-vessels effect. The on-site WUE of Microsoft sites ranges from 0.000 L/kWh in India (dry cooling) to 1.900 L/kWh in Indonesia. But India, at 0.000 on site, shows 6.340 M L in total for training GPT-3, because of scope 2. Dry cooling does not remove the water: it moves it to the power plant.

5.4 Season and weather. On-site evaporation ranges from 1 L/kWh (Google's global annualised average) to 9 L/kWh (large commercial data centre in Arizona, in summer): a factor of 9. The upper bound is a local seasonal peak, not an average.

5.5 Local water stress. This is the metric that counts, and the environmental reports published in 2025 and 2026 document it: depending on the fleet, the share of water withdrawn in water-stressed areas ranges from 13% to 50%, and one operator reports 1,704 ML drawn from areas of high or extreme stress, up 25% in a year. These percentages do not report the same volumes, nor against the same bases, nor under the same stress thresholds — no ranking follows from them. What they establish in common: a significant fraction of data centre water is withdrawn where it is scarce.

5.6 The case of “replenishment”. Google reports having replenished 7,717 Mgal in 2025, i.e. 78% of its freshwater consumption (2022→2025 series: 6% / 18% / 63% / 78%, target 120% by 2030). Methodological point, confidence HIGH: replenishment is not water returned to the watershed used. It is a volumetric credit obtained elsewhere, often in another watershed, and no external standard governs this calculation. It therefore does not offset a local impact.

Section conclusion. “How many litres per query?” is a badly posed question, not because the answer would be awkward, but because the determining variable is not the query: it is the watershed, the month of the year and the electricity mix of the grid powering the machine.

6Anatomy of the viral 500 mL figure

What the source says. The abstract of Li, Yang, Islam and Ren (Making AI Less Thirsty, arXiv:2304.03271) states: a 500 mL bottle for roughly 10 to 50 responses of average length, depending on where and when the model is deployed.

What the wording becomes as it circulates. The sentence is commonly turned into “one ChatGPT query = one bottle of water” or “500 mL per question” — a formulation taken up in the general press, on social networks and in corporate presentations since 2023. We do not cite a specific occurrence here: no dated reuse was collected and verified in the course of this work. Confidence on the existence of the distortion: HIGH (it can be observed by any search on the phrase); confidence on its scale or on the precise vectors: UNKNOWN.

A nuance within the source itself. The abstract announces “10 to 50”; Table 1 in fact gives 10.5 (Washington) to 70.4 (Ireland) queries per 500 mL. The authors therefore round their upper bound downwards — the actual published range is wider than the one put forward.

Where the range comes from. From Table 1: from 10.5 queries per 500 mL (Washington State, 47.506 mL/query) to 70.4 queries per 500 mL (Ireland, 7.107 mL/query). The central “US average” value is 29.6 queries per 500 mL, i.e. 16.904 mL per query.

What the figure includes. Scope 1 (2.200 mL) and scope 2 (14.704 mL): 87% of the value is the water used to produce the electricity, not the data centre.

The three most frequent distortions.

What to take from it. The figure is neither a lie nor a measurement: it is a modelled estimate, with a broad scope, on an old model, with a geographic spread of a factor of 6.7 explicitly published by its authors. Its main defect in the public debate is not being wrong — it is being quoted without its scope and without its range.

9Recent models: what can be calculated, and what cannot

The question that naturally follows everything above: what about today's models? This report's reference figure concerns GPT-3, a 2020 model. Since then, generation after generation has arrived — GPT-4 and 5, Gemini, Llama, Claude, DeepSeek, Mistral. How much water do they consume ?

The honest answer fits in one sentence: for most of them, nobody can calculate it, and not for want of method — for want of input data.

9.1What stopped being published

GPT-3 still serves as the world's reference point for a precise reason: it is the last large model whose training energy was published — 1,287 MWh, by a team from Google and Berkeley. After it, the door closed.

GPT-4's parameter count has never been confirmed by its publisher: the figure of 1,760 billion that circulates everywhere goes back to an unsourced verbal statement, subsequently passed along from one account to the next. For the Claude models — including the 2026 generations — a check of the official technical documents finds no parameter count, no compute budget, no energy consumption, no water volume. The only mention of water ever published is qualitative: "water-efficient cooling", with no metric.

One notable exception: an operator published a production measurement in 2025 for its consumer assistant — 0.24 Wh and 0.26 mL of water per median text query, with its methodology. It is to date the only official datum of this kind on a closed model, and it covers on-site cooling only.

9.2What can be measured: open models

Where a model's weights are public, any laboratory can run it on instrumented hardware and measure the energy with a wattmeter instead of estimating it. Those values are solid — and they give the order of magnitude of a generation.

Two lessons. First, the gap between the measured open models and the 2023 assumption is a factor of 20 to 150: efficiency per query has improved fast, and the historical assumption of 4 Wh is probably well above today's reality. Second, model size does not decide everything: a 20-billion-parameter model can consume more than a 405-billion one, depending on quantisation, hardware and response length. — MEDIUM

9.3Calculate it yourself

Converting energy into water, for its part, is no secret. If you have an estimate of energy per query that you trust, the tool below applies the reference formula, factor by factor, and shows you each step of the calculation.

9.4What this implies

The calculator above makes visible what the whole report demonstrates: changing only the location and the cooling, at identical energy, makes the result vary by a factor greater than 10. The model's technology is not the dominant variable — the infrastructure and its geography are.

And that is precisely why no "water per model" table appears in this report. Publishing one would require knowing, for each model, the actual energy per query, the electricity mix of every data centre serving it, and their cooling technique. None of these three data points is public for a closed model. Such a table would not be a measurement: it would be a series of assumptions presented as a result — exactly what this report sets out not to do. — HIGH

7What we do not know

This section is as important as the preceding ones.

1. No AI / non-AI breakdown exists. No operator publishes a water metric by workload. Every "AI" estimate is derived from assumptions about installed capacity. (Confidence: HIGH on the observation.) 2. No company publishes its indirect (scope 2) water consumption across all of its operations — that is, precisely the item that dominates. 3. The water intensity of electricity is the dominant uncertainty: 1.04 L/kWh (IEA assumption) versus 3.40–5.3 L/kWh (reported data, Siddik et al.), a factor of 3.8 to 5. Li et al. further note that an LBNL study gives 4.35 L/kWh against the 3.14 they adopt (+38 %). 4. The location of AI workloads is not disclosed. That is what makes the margin of error irreducible. 5. Scope 3 (chip and server manufacturing) is absent everywhere, for lack of public data. Mistral explicitly flags the absence of public impact factors for GPUs and describes its own work as a "first approximation". 6. Two central sources could not be opened: the IEA report Energy and AI (HTTP 403) and Sam Altman's post. The values that depend on them (~560 billion L/year, ~1,200 billion L in 2030, 2 M L/day for a 100 MW DC, 0.3 mL/prompt) are to be treated as unverified. 7. Historical series are not stable. Microsoft recalculated all its previous years in 2026. Multi-year comparisons published before that date are void. 8. Meta excludes 1,019 ML of withdrawal linked to the construction of its data centres (+18 % on its total) — a real item, left out of the count. 9. Google no longer publishes a fleet WUE in its 2026 report. Visibility on that metric has decreased, not increased. 10. The rebound effect is addressed by no source. Google documents a 33-fold reduction in energy per prompt while recording +34 % in absolute water consumption: the link between unit efficiency and total volume is analysed nowhere in the sources consulted. 11. On the comparative side: food footprints have not been recalculated since 1996–2005; there is, to the author's knowledge, no peer-reviewed and publicly accessible water LCA of bottled water; the viral figures of "12,000 L per smartphone" and "400,000 L per car" have no openable primary source and are set aside in this report.

8Sources

Peer-reviewed or preprint academic literature

Corporate publications (self-reported)

Public agencies and organisations

◆ Sources cited but NOT opened — not to be used without verification

Final note to the reader

This report concludes neither that AI's water consumption is negligible nor that it is alarming. It establishes three verifiable things:

• The figures circulating on both sides rest on incompatible scopes, and the gap between scopes (a factor of 173 on one and the same query) exceeds the gap between the uses being compared.
• The datum that would settle the matter — water consumption attributable to AI, by site and by watershed — is published by no one.
• The relevant question is not "how many litres per query" but "what pressure on which watershed, in which season, with which electricity mix" — a question that current publications answer only very partially, and only where companies publish site by site.

Readers who wish to form their own view should, faced with any figure on water and AI, systematically ask four questions: Consumption or withdrawal? On-site or total? Which location? Which date, and for which model? Without those four answers, the figure is not interpretable.