The water bottle was per 10 to 50 responses

TL;DR. The claim that an AI response consumes a 500 ml bottle of water comes from a 2023 UC Riverside study. The study says a bottle per 10 to 50 responses, which is 10 to 50 ml each, and includes water consumed generating the electricity as well as on-site cooling. Google's August 2025 measurement, from May 2025 production data and including cooling, idle capacity, CPU, RAM and overhead, puts a median Gemini text prompt at 0.26 ml of direct water. So the circulated figure is off by 10 to 50 times from its own source, and by roughly a thousand times from the on-site number, because a qualifier was dropped and two different scopes were merged. And the underlying concern survives all of it. Google used 10.9 billion gallons in 2025, up 34% year on year, and unlike carbon, water is local: one facility in Memphis draws around a million gallons a day.

---

Status: established. Primary sources: Li et al., Making AI Less Thirsty, arXiv:2304.03271, 2023, expanded in Communications of the ACM 2025; Google's environmental reporting including its August 2025 measurement; and the EU Energy Efficiency Directive with its delegated regulation. This article corrects a number and does not dispute that data centre water use is a serious issue.

---

What the study said

*Li and colleagues at UC Riverside and UT Arlington published Making AI Less Thirsty in 2023. It found that GPT-3 inference consumed roughly a 500 ml bottle of freshwater for every 10 to 50 responses*.

That is 10 to 50 ml per response, and it counts two things: scope 1, the water evaporated in on-site cooling, and scope 2, the water consumed generating the electricity the data centre uses.

Both are real water. Thermoelectric generation consumes water at scale, and excluding it understates a facility's footprint on the wider system.

What travelled

A bottle of water per email.

The qualifier, per 10 to 50 responses, did not survive the journey. A widely seen newspaper graphic made the offsite component clear in its text, and many people who saw only the graphic understood the whole bottle to be used inside the data centre.

Two errors compounded. The per-response divisor vanished, multiplying the figure by 10 to 50 times. And scope 2 was read as scope 1, which moves it by roughly another thousand.

One analysis tracing the chain calls it the most consequential mistake in the history of writing on AI and the environment, and notes that correcting the errors brings the on-site figure close to Google's published number.

This is citation decay in its cleanest documented form: the number survived, the divisor and the scope did not, and the result is stated as fact by people who have never seen the study and would not recognise it if they did.

What is measured

Google published a technical measurement in August 2025 using production data from May 2025. A median Gemini text prompt: 0.24 Wh of energy, 0.03 g CO2e, and 0.26 ml of water, described as about five drops.

The method is not a best case. It includes cooling, idle reserve capacity, CPU and RAM, and data-centre overhead. It is scope 1, direct water only, and Google says so.

OpenAI has published no comparable breakdown. Its chief executive stated approximately 0.3 to 0.32 ml per query in June 2025, also direct cooling only.

Independent lifecycle work lands higher. Mistral published an analysis in July 2025, audited with Carbone 4 and France's ADEME, putting a roughly 400-token reply at about 45 ml of water and 1.14 g CO2e on a full lifecycle basis.

Which is the whole point. 0.26 ml, 0.32 ml, 45 ml and 10 to 50 ml are not contradictory. They measure direct cooling, direct cooling, full lifecycle, and cooling plus generation respectively. A number without its scope is not a measurement of anything, and every one of these travels without it.

And the concern is justified anyway

This is where a debunk usually stops, and stopping here would be wrong.

Google consumed 10.9 billion gallons of water in 2025, up 34% year on year and more than double its 2021 level. Amazon disclosed 2.5 billion gallons at a water usage effectiveness of 0.12 L/kWh; Microsoft reports 0.30 L/kWh fleet-wide. Global AI data centre direct water consumption reached roughly 560 billion litres in 2025.

US facilities used about 17.4 billion gallons directly, with an estimated 211 billion gallons indirectly through electricity generation. The indirect figure is more than twelve times the direct one, which is exactly why scope matters and exactly why the scope-2 number is not a rhetorical trick.

And water is local in a way carbon is not. A tonne of CO2 has the same effect wherever it is emitted. A million gallons a day matters entirely differently in Iowa than in Arizona.

One Google facility in Council Bluffs, Iowa consumed around 4,900 megalitres of potable water in 2024. A supercomputer facility in Memphis draws around a million gallons a day with projected demand near five million, comparable to a town of up to fifty thousand people. MSCI analysis of 14,000 data centre assets found one in four may face increased water scarcity by 2050.

The aggregate is manageable. The basin is where it binds, which is the same shape as the electricity finding and for the same reason.

Which is why the correction matters

A wrong number produces the wrong policy.

If the problem is per-query consumption by individuals, the response is to use AI less. If the problem is facility siting in water-stressed basins, the response is siting rules, cooling design mandates and disclosure, and individual restraint does almost nothing.

Regulators are acting on the second reading. The EU Energy Efficiency Directive with its 2024 delegated regulation requires data centres above 500 kW to report annually on 24 sustainability indicators including total water input and water usage effectiveness, with first reports due September 2024. The Netherlands banned hyperscale facilities above 70 MW from January 2024. Singapore set a target water usage effectiveness of 2.0 m³/MWh. More than 190 data centre bills were introduced across US state legislatures in 2025, and Arizona municipalities have imposed caps that pushed developers toward zero-water cooling designs.

None of those measures would follow from a bottle-per-email framing, and all of them follow from a basin-level one.

Three things this establishes

A number can be wrong by a factor of a thousand and still point at something real. The bottle figure was wrong and data centre water use is a genuine and growing constraint. Correcting the first does not settle the second, and treating a debunk as a resolution is its own error.

Scope is the load-bearing choice. 0.26 ml and 10 to 50 ml differ almost entirely by what is counted, not by disagreement about the world. Any water or carbon figure without a stated scope is uninterpretable, and almost all of them are quoted without one.

And locality changes what the aggregate means. Global water consumption is a number; basin-level withdrawal is a constraint. A framing that reports only the first will systematically miss where the problem actually is.

What it does not establish

That the UC Riverside study was wrong. It was careful, stated its scope, and its figure is defensible for what it measured. The distortion happened downstream of it.

That Google's number is complete. It is scope 1 by design and Google says so. Researchers note such figures exclude water in electricity generation and in semiconductor manufacturing, both of which are real.

That per-query figures are useless. They are the right measure for the marginal question and the wrong one for the total, which is the same distinction as with electricity.

And nothing about whether any particular facility should be built. That depends on basin conditions this article has not examined.

What is unresolved

Whether disclosure becomes standard. The EU regime is the first mandatory one at scale, and its data has not yet produced minimum performance standards.

What the full lifecycle number actually is. Mistral's audited 45 ml for a 400-token reply is the most complete published figure and covers one model on one infrastructure. Nobody has done it across providers on a comparable basis.

How much zero-water cooling costs. It trades water for electricity, and where that trade is favourable depends on the local grid and the local basin, which is a per-site calculation nobody has published in aggregate.

And what the indirect figure really is. The 211 billion gallon US estimate for water in electricity generation rests on grid-average intensities that vary enormously by region and by hour.

The counter-argument

Correcting the figure serves the industry's interest and should be read with that in mind. The most precise number available comes from a company with an interest in it being small, and this article gives it prominence while dismissing a figure produced by academics with no such interest. That is a real asymmetry, even if the arithmetic error in the circulated claim is genuine and demonstrable.

Scope 2 is arguably the honest default. Water consumed generating electricity is consumed because the data centre demanded the electricity. Calling only on-site cooling "the" water figure lets a facility export its footprint to a power station and report a smaller number.

The bottle framing worked. It is wrong by a factor of tens and it produced public attention, regulatory interest and disclosure requirements that a technically precise 0.26 ml would not have. Whether a useful falsehood is preferable to an ignored truth is a real question, and this site's answer is no, but the position deserves stating rather than assuming.

And the individual-versus-siting framing may be a false choice. Aggregate demand is the sum of individual queries, and telling people their usage is irrelevant is itself a claim with policy consequences.

The short version

The 500 ml bottle comes from a 2023 UC Riverside study that says a bottle per 10 to 50 responses, which is 10 to 50 ml each, counting both on-site cooling and the water used to generate the electricity.

Google's August 2025 measurement, on May 2025 production data and including cooling, idle capacity, CPU, RAM and overhead, puts a median Gemini text prompt at 0.26 ml of direct water. The circulated claim is off by 10 to 50 times from its own source and by around a thousand from the on-site figure, because a divisor was dropped and two scopes were merged.

0.26 ml, 0.32 ml, 45 ml and 10 to 50 ml are all defensible and all measure different things. A figure without its scope is not a measurement.

And the concern survives the correction entirely. Google used 10.9 billion gallons in 2025, up 34% in a year. US data centres used about 17.4 billion gallons directly and an estimated 211 billion indirectly, which is more than twelve times as much and is exactly why scope matters. Water is local: one Memphis facility draws around a million gallons a day against projected demand near five million, and one in four of 14,000 data centre assets may face increased water scarcity by 2050.

Which is why the correction matters rather than being pedantry. A bottle-per-email framing points at individual restraint. A basin-level framing points at siting rules, cooling mandates and disclosure, and that is what regulators are actually doing.

Common questions

Does an AI response really use a bottle of water? No. The 500 ml figure comes from a 2023 UC Riverside study which found roughly a 500 ml bottle for every 10 to 50 responses, which is 10 to 50 ml each, and which counts both on-site cooling and the water consumed generating the electricity. The circulated version dropped the divisor and was widely read as on-site consumption, which compounds two errors: a factor of 10 to 50 from the missing divisor, and roughly a thousand from reading a combined figure as a cooling-only one.

What is the measured figure? Google published a technical measurement in August 2025 based on May 2025 production data: a median Gemini text prompt uses 0.24 Wh of energy, emits 0.03 g CO2e and consumes 0.26 ml of water, about five drops. The method includes cooling, idle reserve capacity, CPU and RAM, and data-centre overhead, so it is not a best case, and it is direct water only, which Google states.

Why do the different figures vary so much? Because they measure different things, not because anyone disagrees about the world. 0.26 ml and about 0.32 ml are direct on-site cooling. Mistral's audited lifecycle analysis puts a roughly 400-token reply at about 45 ml. The UC Riverside range of 10 to 50 ml covers cooling plus water consumed generating electricity. Each is defensible for its scope, and a figure quoted without its scope is uninterpretable.

So is AI water use not a problem? It is a problem, and the correction does not touch that. Google consumed 10.9 billion gallons in 2025, up 34% year on year and more than double its 2021 level. Global AI data centre direct water consumption reached roughly 560 billion litres in 2025. US facilities used about 17.4 billion gallons directly plus an estimated 211 billion gallons indirectly through electricity generation.

Why does it matter where a data centre is? Because water is local in a way carbon is not. A tonne of CO2 has the same effect wherever it is emitted; a million gallons a day means something entirely different in Iowa than in Arizona. One Google facility in Council Bluffs consumed around 4,900 megalitres of potable water in 2024, and a Memphis supercomputer facility draws around a million gallons a day with projected demand near five million, comparable to a town of up to fifty thousand people. MSCI found one in four of 14,000 data centre assets may face increased water scarcity by 2050.

What are regulators doing? Acting on the basin-level framing rather than the per-query one. The EU Energy Efficiency Directive with its 2024 delegated regulation requires data centres above 500 kW to report annually on 24 indicators including total water input and water usage effectiveness. The Netherlands banned hyperscale facilities above 70 MW from January 2024. Singapore set a target water usage effectiveness of 2.0 m³/MWh. More than 190 data centre bills were introduced across US state legislatures in 2025, and Arizona municipalities have imposed caps that pushed developers toward zero-water cooling.

Should I use AI less to save water? That is the response the wrong framing implies, and it is close to irrelevant at 0.26 ml per prompt. The response the evidence supports is siting rules, cooling design requirements and mandatory disclosure, because the constraint is where facilities draw from rather than how many queries individuals send. The counter-position, that aggregate demand is the sum of individual queries, is worth stating and does not change which lever is effective.

Is it suspicious that the smallest number comes from a company? It is worth weighting, and this article says so. The most precise figure available is published by a party with an interest in it being small, while the larger figure came from academics with no such interest. That asymmetry is real. What makes the correction stand regardless is that the arithmetic error in the circulated claim is demonstrable from the original study's own wording, independently of who published anything afterwards.