ONLINEAGENT_OPS 2026.Q3 HOME ARTICLES CRAFT RECORD BLOG MAP HUBS FAQ SEARCH
HOMEBLOGEnergy And Water
BLOG · DATED PIECE

ENERGY AND WATER

Estimates of the water behind one AI query differ a thousandfold. Why they disagree, what operators disclose, and why where the water comes from matters most.

READ5 min
WORDS930
SECTIONS4
SOURCES7
TYPEDATED
CHECKED11 SEP 26
TL;DR — THE SHORT VERSION

Per-query AI water figures vary a thousandfold depending on method, so the firmer points are that use is rising, indirect water dominates, and location matters most.

  • Per-query figures disagree widely. Published estimates differ by about a thousandfold, because they count different things.
  • Indirect water is bigger. Electricity generation used more than ten times the water of direct cooling in US data centres.
  • Operators report rising use. Google's own reporting shows its water use still rising year on year.
  • Location matters more than totals. Evaporated water does not return to its source, so the same volume matters far more in a region already short of water.
  • Ask four questions first. Direct or total, consumed or withdrawn, which hardware and year, and where.
US DATA CENTRE WATER · 2023
Water used directly for cooling, against water used at power plants to make the electricity. Billions of litres.
Indirect, through electricity~800 bn L
Direct, for cooling66 bn L
Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report, December 2024, read at source 17 Sep 2026: “the 66-billion-liter total” and “nearly 800 billion liters” indirect.

The thousandfold spread

All of these are real, published, and about "an AI query". They disagree because they are answering different questions — the shape of problem auditing a vendor claim runs into constantly.

~0.3 mL
OPERATOR ESTIMATE
On-site cooling only, for a typical short text query, on current efficient hardware.Figure stated publicly by OpenAI’s chief executive, Sam Altman, The Gentle Singularity, June 2025, read at source 23 Sep 2026: an average query “also uses about 0.000085 gallons of water; roughly one fifteenth of a teaspoon.” The conversion to about 0.3 mL is this page’s arithmetic.
10–25 mL
FULL ACCOUNTING
One conversation, counting cooling and the water used to generate the electricity.Ren et al., University of California, Riverside, Making AI Less Thirsty, 2023, read at source 11 Sep 2026. Peer-reviewed in Communications of the ACM, July 2025.
~500 mL
WORST CASE, OFTEN QUOTED
A 100-word generated response, on older hardware, under specific conditions. This is the "bottle of water per email" figure in circulation.Same research group; the conditions are frequently dropped when the figure is repeated. The peer-reviewed version states the figure as “a 500ml bottle of water for roughly 10-50 medium-length responses” for GPT-3 — narrower than the “20 to 50 questions” that circulated in 2023 coverage, which is the drift this page is describing.

What changes between them: whether electricity generation is counted, how long the output is, which hardware generation, and which facility in which climate. None of those are details — each one moves the answer by an order of magnitude. A per-query water figure quoted without its method is close to meaningless, and that applies to the reassuring numbers as much as the alarming ones.

◈ THE FINDING THAT MATTERS MORE THAN ANY PER-QUERY NUMBER

A US national laboratory analysis found American data centres consumed roughly 66 billion litres directly for cooling in 2023 — while the electricity powering them accounted for approximately 800 billion litres indirectly. The indirect footprint is more than ten times the direct one. Which means a facility that switches to air cooling and announces near-zero water use has addressed the smaller share. Only changing the power source addresses both.Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report, December 2024, read at source 17 Sep 2026: direct use is “the 66-billion-liter total” for 2023, and “The total indirect water footprint of U.S. data centers is nearly 800 billion liters”.

TAKEAWAY

Before repeating a per-query water figure, find out what it counts. Without its method it can be made to look tiny or alarming.

What the operators themselves disclose

These are the firmest numbers on the page, because companies reporting their own rising consumption have no incentive to overstate it.

  • Google: 10.9 billion gallons in 2025, a 34% rise on 2024.Google, 2026 Environmental Report, page 35, read at source 16 Sep 2026: “In 2025, we consumed 10.9 billion gallons (41 billion liters or 41 million cubic meters) of water across our data centers and offices”, and “Our water consumption increased 34% from 2024 to 2025”. An earlier version of this line also called it more than double the 2021 figure and the steepest single-year increase in Google’s disclosed history; neither was checked, and both were removed.
  • Amazon: about 2.5 billion gallons in 2025, across its whole data-centre footprint.Amazon, Amazon’s data centers are 7x more water-efficient than the industry average, read at source 16 Sep 2026: “In 2025, we withdrew about 2.5 billion gallons across our entire global data center footprint during the whole year.” Amazon gives its efficiency as “0.12 liters of water per kilowatt-hour (L/kWh) in 2025”. An earlier version of this line called it the company’s first water disclosure; that was not checked, and was removed.
  • Microsoft: 0.27 litres per kilowatt-hour fleet-wide in FY25, down from 0.30 the year before, plus a cooling design that avoids evaporation.Microsoft, datacenter efficiency metrics (global WUE 0.30 for FY24, 0.27 for FY25) and Inside Microsoft’s two-decade push to cut water intensity, 24 Jun 2026: “Our average WUE has decreased from 2.3 L/kWh to 0.27 L/kWh in 2025”. On the design, Microsoft, 9 Dec 2024: “this design will avoid the need for more than 125 million liters of water per year per datacenter.” All read at source 16 Sep 2026. Corrected 16 Sep 2026 from 0.30, which is the FY24 figure.

Worth noting what these disclosures have in common: they are inconsistent in method. Google reports water consumed, Amazon water withdrawn, Microsoft litres per kilowatt-hour, so the three numbers above cannot be lined up against each other. Read each one in its own terms.

Where it is drawn matters more than how much

Water consumed by evaporation does not return to the source, so local scarcity — not global totals — determines the actual impact. The same volume means something very different in a wet region than in one rationing supply.

Two analyses are widely reported but were not read at source for this page, so treat their figures as reported, not verified: a Bloomberg News analysis said to find roughly two-thirds of data centres built since 2022 located in water-stressed regions, and an MSCI review of some 14,000 data centre assets said to conclude that about one in four may face increased water scarcity by 2050.Not read at source: Bloomberg News analysis of data centre siting since 2022, and MSCI review of ~14,000 data centre assets, 2025. No link to either was available on this page, and none is guessed. Marked 17 Sep 2026.

TAKEAWAY

Judge a data centre’s water use by the basin it draws from, not by its total alone.

The energy side, briefly

The energy figures are firmer than the water ones. The International Energy Agency projected global datacentre electricity demand roughly doubling by 2030 to around 945 terawatt-hours in its April 2025 Energy and AI report — just under 3% of world electricity demand, against about 415 TWh consumed in 2024 — and has since updated that central projection to about 950 TWh. The agency's own comparison: a large data centre can consume as much electricity as 100,000 households, and the largest under construction as much as 2 million.IEA, Energy and AI, April 2025, Base Case, read at source 23 Sep 2026: “projected to reach around 945 TWh by 2030 in the Base Case, representing just under 3% of total global electricity consumption in 2030. This is more than double the estimated approximately 415 TWh for 2024”; “A large data centre can consume as much electricity as 100 000 households. The largest currently under construction could consume as much as 2 million households.” Central projection subsequently updated. Full treatment in the Forge.

◈ HOW TO READ ANY FIGURE ON THIS TOPIC

Ask four questions before repeating a number. Direct or total? (the indirect share is ten times larger) · Consumption or withdrawal? (only consumed water is lost to the basin) · Which hardware, which year? (efficiency has moved fast) · Where? (the same litres are trivial in one basin and contested in another). A figure that cannot answer all four is not evidence, whichever direction it points.

◈ WHERE THIS SITE STANDS

Both the alarmed and the dismissive positions are being argued with unusable numbers. "Every email you send boils a bottle of water" drops the conditions that produced it; "a query uses a third of a millilitre" quietly omits the electricity that made it possible. The defensible position is narrower: consumption is rising fast and is now disclosed, the indirect footprint dominates the direct one, and siting decisions matter more than totals. That is less satisfying than a slogan, and it is what the evidence supports.

ABOUTMETHODVERIFYPRIVACYCONTACTINDEXAI PROMPT GENEER · EVERY ARTICLE CARRIES ITS OWN CHECKED DATE