Amor
Preparing your next read…
Loading stories.
Justin · Amor · Visual essay
Is using AI really a waste of water?
A reply appears on your screen. Somewhere, a building did the work. Follow the water behind it—and find out whether the concern matches the cost.
↓ Scroll to follow one answerOur reading of the evidence
The evidence gives ordinary text users less reason for guilt—and communities good reason to scrutinize expansion.
01 / Repair the unit
The research behind the familiar comparison estimated a 500 mL bottle across 20–50 questions and answers.
That was a modeled conversation using assumptions about GPT-3—not a meter reading attached to every response.
The unit matters. A conversation became a slogan about a prompt.
The estimate cannot tell you the water cost of everything someone does with AI today. A short text answer, an image and an extended automated workflow are different tasks.
That correction changes the accusation. Now let’s look at the infrastructure.
02 / Open the cloud
U.S. data centers consumed an estimated 66 billion liters directly in 2023.
Consumption means water not returned for immediate reuse, including evaporation.
Counting the water associated with their electricity adds nearly 800 billion liters in the same analysis.
The power connection serves the computing as well as the cooling. It brings water used elsewhere into the accounting.
Together: approximately 866 billion liters. This includes all data-center workloads, not just AI.
866 billion liters sounds enormous. Compared with what?
Different years; all data-center workloads. These are comparison volumes, not shares of one water supply.
03 / Find the scale
Crop irrigation in the lower 48 states consumed approximately 108.6 trillion liters in water year 2020.
Both measures count consumption. Neither is simply the amount that passed through a pipe.
Even including electricity-related water, the 2023 U.S. data-center estimate is about one one-hundred-and-twenty-fifth of that volume.
About 0.8% of the irrigation volume in this comparison.
This is perspective, not a water budget shared by two industries. The figures cover different years, and the data-center total includes much more than AI.
But the comparison changes what the enormous number means. Size alone does not establish that an ordinary use is extravagant.
Our reading of the evidence
The bottle shorthand overstates what the original research said about one response. The national comparison puts the industry’s demand in perspective.
That does not make every task worthwhile. It does give us reason to reject automatic guilt over ordinary text use.
Pumped and delivered water is a different measure from consumption.
05 / West Des Moines, Iowa
The national comparison tells us about scale. A utility has to deliver water to the customers connected to its system.
In West Des Moines, Microsoft received 2.4% of the water pumped through the local utility in 2025. The utility reports monthly shares of 2–7%.
It also says the 2025 watering ban and May 2026 alert came from river nitrates—not data-center demand.
A restriction and a nearby data center can make a convincing story. Here, that was not the cause the utility identified.
Clearing up one accusation does not settle the next city’s plans.
Consumption is part of withdrawal.
Additional supply needed under these assumptions. The plan’s 2024 baseline shows a surplus.
06 / The Dalles, Oregon
At The Dalles, Google reports consuming 469 million gallons in 2025, from 654.3 million gallons withdrawn.
The city’s 2024 water plan projects an annual supply deficit of about 90 million gallons by 2034, or 140 million under its climate assumption. Its 2024 baseline shows a surplus.
Those are different records: reported company use and a citywide planning scenario. The plan includes other growth, and it does not establish that Google has caused household shortages.
The useful question is whether future demand has a credible supply plan.
And the demand those plans must serve is growing.
07 / Growth changes the planning question
Google reports global operational water consumption of 8.135 billion gallons in 2024 and 10.869 billion in 2025—an increase of approximately 34%.
That includes data centers and offices, across workloads. It is not an AI-only total.
Better efficiency can reduce the cost of each task while expanding activity increases the total. Growth gives us a reason to plan now; it does not change what the original bottle claim meant.
The next facilities have not all been built. Their demands are decisions we can still examine.
08 / Before construction
Decision 01
What supply will serve it, and what else depends on that supply? A proposed connection needs more than a line on a map: it needs an explanation of the resource it will draw on.
Decision 02
An annual total cannot describe every requirement of a hot day. The plan should account for the moments when the facility and the surrounding community both need reliable service.
Decision 03
Water source, cooling and electricity choices should be considered together. Ask what each option changes across the whole system, rather than treating one favorable metric as a verdict on the design.
Decision 04
The plan should explain which upgrades are needed and how the project’s share will be funded. Required capacity, project contribution and public commitments belong in the discussion before construction.
09 / That is where the water argument should lead
The evidence does not justify treating ordinary text use as an extraordinary act of water waste.
It does justify asking hard questions about expanding infrastructure: where the demand lands, whether the supply can meet it, and who pays for the capacity.
A prompt is a poor unit of blame. A new facility is a consequential decision.
The cloud has a hometown.
It should be a good neighbor.
The open notebook
Sources, definitions, and the boundaries of the comparison.
Note 01
The April 2023 version of Making AI Less Thirsty estimated 500 mL for 20–50 questions and answers. Its March 2025 revision describes roughly 10–50 medium responses. Both concern modeled GPT-3 operation, including onsite and electricity-related water, rather than direct measurements of today’s ChatGPT. The original version explains the comparison’s origin; the revision does not make it a fixed rate for every prompt.
Note 02
| Disclosure | Reported result | Boundary |
|---|---|---|
| Google, Gemini Apps text May 2025 data | 0.26 mL for its reported median prompt | Onsite freshwater cooling consumption derived from serving energy and fleet water intensity; excludes electricity-generation and manufacturing water. |
| Mistral, Le Chat July 2025 disclosure | 45 mL for a 400-token response | A provider-commissioned life-cycle assessment with upstream impacts; the public disclosure does not give a detailed water-stage breakdown. |
These results are not a clean efficiency ranking or the endpoints of a universal range. Google’s median is constructed from models’ average serving energy and prompt distribution; multiplying it by all prompts would not recover total consumption. Mistral describes external participation and review, but its public detail is limited. Text, image, video and long automated workflows also represent different tasks.
Note 03
The LBNL estimate combines 66 billion liters of direct U.S. data-center consumption with nearly 800 billion attributed to electricity in 2023. Electricity attribution includes reservoir evaporation and uses regional grid mixes. These are all-data-center figures, not AI-only figures or the marginal water cost of another prompt.
USGS estimates lower-48 crop-irrigation consumption at 78,400 million gallons per day in water year 2020. Multiplying by 3.785411784 liters per U.S. gallon and 366 days gives approximately 108.6 trillion liters. Dividing 866 billion by that value gives about 0.8%.
The comparison uses different years and slightly different geographic coverage. It compares consumption volumes, not shares of one national pool. Irrigation here does not include all rainfall supporting crops.
Note 04
West Des Moines Water Works reports Microsoft’s deliveries as 2.4% of local water pumped in 2025. Its explanation attributes the cited watering restrictions to nitrates, not data-center demand. Pumped and delivered water are distinct from net consumptive losses.
For The Dalles, Google’s 2025 consumption record and the city’s projected 2034 deficit are separate measures. The citywide scenario includes demand beyond Google, and the city’s FAQ says its two largest replacement projects would be needed without Google. We do not treat those records as proof of company-caused household shortages.
Note 05
Google’s displayed time series uses the internally consistent 2024 and 2025 global operational consumption figures in its 2026 report. They cover data centers and other facilities, not AI alone.
LBNL’s June 2026 update estimates 192 TWh of U.S. data-center electricity in 2024 and a reference case of 649 TWh in 2030, with scenario bounds of 521–843 TWh. It does not replace the national water estimate used here. These electricity scenarios are not measured water forecasts.
Other global estimates count differently. IEA reports about 560 billion liters annually for data centers and 1.2 trillion in its 2030 base case. UNU-INWEH’s June 2026 release reports 9.3 trillion liters associated with data-center electricity in 2030. A 2026 Patterns paper estimates 312.5–764.6 billion liters for AI globally in 2025. These cannot be combined into one trend or treated as a statistical confidence interval.
The full UNU PDF and Patterns paper were inaccessible in this research; their publisher/author summaries were available. Their detailed assumptions have not been reconciled for this piece.
Note 06
Our conclusion is an editorial judgment about proportionality. A water footprint cannot establish the value of every task, and a national comparison cannot decide whether a local project is suitable. The evidence supports rejecting automatic water-waste stigma for ordinary text use while scrutinizing the demands of expansion. We are not claiming a definitive present-day AI-only water total or judging AI’s full social and environmental value.