How Much Water Does ChatGPT Really Use? A Developer Investigates the 0.32 mL Claim

Two things matter here. The accelerator is the dominant consumer but it is not the only one, and the amount of work it does is a function of tokens, not requests. A request is a billing unit. A token is a physics unit.

GPUs do not drink water

This is the part that gets lost in headlines about GPU water consumption. A GPU consumes electricity and emits heat. That is the entire interaction. Water shows up in three places around it, and only one of them is inside the building.

Direct versus indirect water consumption

This distinction decides almost every disagreement about AI water consumption, so it is worth being precise.

Direct, on site, scope 1. Water evaporated in the facility's cooling towers to reject heat into the atmosphere. This is measured with water usage effectiveness, WUE, in litres per kilowatt-hour of IT energy. Google reported a fleet-wide WUE of 1.15 L/kWh for 2023 and 2024 ^[4].

Indirect, off site, scope 2. Water consumed at the power plants generating the electricity the facility draws. Thermoelectric plants evaporate water for cooling, and hydroelectric reservoirs lose water to evaporation. The researchers behind Making AI Less Thirsty reported that training GPT-3 consumed about 0.7 million litres on site, and about 5.4 million litres once off-site electricity water was counted ^[5].

That ratio is the thing to hold on to. In that study, indirect water was several times larger than direct water. So a per-query figure that counts only cooling is measuring the smaller half of the problem, and a figure that counts both is measuring a different quantity entirely. Neither is wrong. They are answers to different questions, and putting them in the same sentence is where the public conversation breaks.

Why there cannot be one universal number for AI water consumption

Water per prompt is not a constant, it is the output of a chain of multiplications where every term is a variable.

Some concrete evidence for how wide that variation gets. Epoch AI estimated about 0.3 Wh for a typical GPT-4o query assuming 500 output tokens on H100 hardware, and in the same analysis put a 10,000 token input at roughly 2.5 Wh and a 100,000 token input at roughly 40 Wh ^[6]. That is a factor of more than 100 within one model. Benchmarking across 30 models, Jegham and colleagues found the heaviest models exceeded 29 Wh on a long prompt, over 65 times the lightest ^[7]. Reasoning models generate several times more tokens than a direct answer, and you pay for every one.

Location adds another dimension. In the Making AI Less Thirsty tables, the same workload came out at roughly 7.6 mL per request in Texas and roughly 47.6 mL in Washington, driven by differences in cooling conditions and the water intensity of the local grid ^[5]. Same model, same prompt, six times the water.

Comparing the published estimates side by side

Here is every credible published figure I could find for inference, with what each one actually covers.

Read the right-hand columns rather than the numbers. 0.26 mL and 17 mL are not in conflict. One is a median on-site cooling figure from the operator's own telemetry, the other is a modelled upper end that includes more of the system. The spread across that table is not scientific disagreement, it is boundary disagreement.

Reverse engineering the 0.32 mL figure

Since both of Altman's numbers were given together, they constrain each other. Google's published water formula is worth borrowing because it is the only one we can see:

water per prompt = (total energy - facility overhead energy) x WUE

Checking it against Google's own figures: (0.24 Wh - 0.02 Wh) x 1.15 L/kWh = 0.253 mL, which rounds to the 0.26 mL they published ^[4]. The formula reproduces their result, so it is a fair template.

Apply it to OpenAI's pair. If 0.34 Wh is the full facility figure and overhead is a similar 8%, IT energy is about 0.313 Wh:

0.32176 mL / 0.313 Wh  =  1.03 litres per kWh implied WUE

An implied WUE near 1.0 L/kWh is entirely believable for a well-run modern facility, sitting just under Google's reported 1.15. So the figure may be plausible under certain assumptions, specifically the assumption that it counts on-site cooling only and excludes the water consumed generating those 0.34 watt-hours.

What happens if you include that second half? Using the direct-to-total ratio from the GPT-3 training figures in Making AI Less Thirsty, 5.4 divided by 0.7, as a rough scaling factor:

0.32 mL x 7.7  =  about 2.5 mL per query

I want to be careful here. That is an illustration of sensitivity, not a measurement. The ratio comes from a different model on different hardware in different locations. The point is only that the choice of boundary moves the answer by a multiple, and 0.32 mL is the number you get from the narrower of the two choices.

Training and inference are different budgets

Every serious paper separates these, and almost every viral post merges them.

Training is a one-time capital cost. It is enormous and it is finite. The GPT-3 figures above, 0.7 million litres on site, are training numbers ^[5]. As a developer the closest analogue is a CI pipeline that builds a container image: expensive once, then amortised across every request the image serves.

Inference is the marginal cost per request, and it is what the 0.32 mL claim is about. At OpenAI's scale, inference long ago overtook training in total consumption simply through volume. A rough sense of scale, using the company's own figure at a widely cited 700 million queries per day:

700,000,000 x 0.32 mL = 225,000 litres per day
                       = about 82 million litres per year
                       = roughly 33 Olympic swimming pools

Notice how differently the same number reads at two scales. Per query it is a rounding error. Annually it is a municipal water account, and that account is concentrated in a handful of locations rather than spread across the world. Both readings come from the same 0.32 mL.

The almond comparison, checked separately

The 38,000 queries per almond claim is the clearest example of a boundary mismatch I have seen, and it is worth walking through because the error is not in the multiplication.

PolitiFact traced the almond side to a 2019 study finding about 3.2 gallons, roughly 12 litres, of water per California almond. Of that total, only about 6 litres is direct fresh irrigation water. The rest is rainfall that would have fallen anyway and water accounted for dilution of pollutants ^[3].

Reproducing Altman's arithmetic:

12.1 litres / 0.32176 mL  =  about 37,600 queries  (rounded to 38,000)

The multiplication is correct. The comparison is not, because the two sides use different definitions of the word "water". The almond side uses a full water footprint including rainfall and dilution allowances. The ChatGPT side, by the reverse engineering above, appears to count only water evaporated in cooling. Putting the direct irrigation figure against a broader per-prompt estimate gives a very different answer:

6,000 mL / 0.6 mL   = 10,000 prompts
6,000 mL / 17 mL    = about 350 prompts

PolitiFact landed on roughly 1,000 to 10,000 prompts per almond, which is between one quarter and one fortieth of the quoted figure ^[3]. Shaolei Ren of UC Riverside, one of the authors of the water footprint research, put the underlying problem plainly: "There is very little public information about how today's ChatGPT systems operate and what resources they use." OpenAI did not respond to PolitiFact's requests.

System boundaries, explained as an API latency problem

If you have ever argued about p99 latency in a sprint review, you already understand this failure mode.

When someone tells you an endpoint responds in 12 ms, the first question is where the timer started. If one team instruments the database call and another instruments the full round trip including TLS handshake, authentication, database time and JSON serialisation, both can be honest and the numbers can differ by an order of magnitude. Nobody accuses anyone of lying. Everyone asks for the trace.

AI water figures need the same discipline and do not yet have it. Latency at least has conventions like p50 and p99 that force you to say which statistic you mean. For water per prompt there is no agreed boundary, no agreed statistic, and in OpenAI's case no published trace.

Why the older bottle of water claim was also oversimplified

Before 0.32 mL went around, the viral figure went the other way: one conversation with ChatGPT drinks a 500 mL bottle of water.

That also traces to real research. Making AI Less Thirsty stated that GPT-3 "needs to drink a 500 ml bottle of water for roughly 10 to 50 medium-length responses", depending on where and when it runs ^[5]. The paper is careful. The headline was not, because the conditions dropped away in transmission:

• The model is GPT-3, not a current production model on current hardware.
• A "response" assumes roughly 800 words of input and 150 to 300 words of output, far longer than a typical chat turn.
• The figures are modelled from public disclosures, not measured in the serving fleet.
• The 10 to 50 range spans locations, and the paper reports it as a range for exactly that reason.

So both viral numbers come from the same pattern. Take a figure computed under stated conditions, drop the conditions, and present the remainder as a universal constant. 0.32 mL and 500 mL are the same mistake pointing in opposite directions.

What developers should ask before trusting an AI environmental metric

This is the checklist I now run before repeating any AI water or energy figure. It works the same way as reviewing a benchmark someone posted in a pull request.

1. What model was tested? Model family, size, whether it is a reasoning model, and whether it is the one actually serving production traffic.
2. How many input and output tokens? Without token counts the figure has no unit of work behind it. A 50 token prompt and a 50,000 token prompt are different jobs.
3. What hardware? H100, TPU generation, something else. Efficiency per token changes by generation and the numbers age fast.
4. Which data center? Climate, season and local water stress change the answer by multiples, not percentages.
5. What cooling system? Evaporative cooling towers, closed-loop direct-to-chip, air cooled, or a hybrid that switches by outdoor temperature.
6. Was indirect electricity water included? The single most important question. Scope 1 only, or scope 1 plus scope 2. If the source does not say, treat the number as on-site only.
7. What measurement period? A week in January and a week in July at the same facility give different WUE.
8. Average, median or modelled? A mean is pulled upward by long prompts, a median hides them, and a model is only as good as its assumptions. Google says median. Altman says average and defines neither the average nor the population.
9. Can it be reproduced independently? Is there a formula, an input dataset, a stated WUE, anything a third party could recompute? Google published a formula I could check in one line. OpenAI published a sentence.

If a source cannot answer at least six of these, the figure is a talking point rather than a measurement, whichever direction it points.

FAQ

Conclusion

After going through the sources, my position on ChatGPT water usage is unsatisfying in a way I have made peace with. The 0.32 mL figure is not fabricated. The unit conversion is correct, the teaspoon cross-check agrees, and the implied water usage effectiveness sits in a believable range for a modern data center. It may be plausible under certain assumptions.

It is also not verifiable, and it is not a constant. The public methodology is limited to two sentences on a personal blog. The figure appears to cover on-site cooling only, which is the smaller half of the water story. It varies with tokens, hardware, location, season and load in ways that published research puts at more than an order of magnitude.

The argument worth having is not whether a prompt costs 0.32 mL, five drops, or a bottle. It is whether the methodology, the infrastructure assumptions and the accounting boundaries are transparent enough for anyone to interpret the number at all. Google published a formula I could verify against their own result in a single line of arithmetic. That is the standard. Until more of the industry meets it, treat every per-prompt water figure the way you would treat a benchmark with no test harness attached: interesting, possibly honest, and not yet evidence.

Sources and references

1. Sam Altman, The Gentle Singularity, blog.samaltman.com, 10 June 2025. Primary source for 0.34 Wh and 0.000085 gallons per query.
2. Data Center Dynamics, Sam Altman: ChatGPT queries consume 0.34 watt-hours of electricity and 0.000085 gallons of water, June 2025. Contemporary reporting on the absence of peer review and methodology.
3. PolitiFact, Mostly False: Sam Altman says growing a single almond equals 38,000 ChatGPT queries, 4 September 2026. Almond water figures, the 0.6 to 17 mL range, and the Shaolei Ren quotation.
4. Elsworth, Huang, Patterson, Schneider, Sedivy, Goodman, Townsend, Ranganathan, Dean, Vahdat, Gomes and Manyika (Google), Measuring the environmental impact of delivering AI at Google Scale, arXiv:2508.15734, 21 August 2025. Median 0.24 Wh, 0.26 mL, 0.03 gCO2e per Gemini Apps text prompt, energy component split, WUE 1.15 L/kWh, published water formula.
5. Pengfei Li, Jianyi Yang, Mohammad A. Islam and Shaolei Ren, Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models, arXiv:2304.03271, April 2023, revised March 2025. Scope 1 and scope 2 definitions, GPT-3 training at 0.7 million litres on site and 5.4 million litres total, and the 10 to 50 responses per 500 mL range.
6. Epoch AI, How much energy does ChatGPT use?, 7 February 2025. 0.3 Wh per typical GPT-4o query, and the scaling to 2.5 Wh and 40 Wh for long inputs.
7. Nidhal Jegham, Marwen Abdelatti, Lassad Elmoubarki and Abdeltawab Hendawi, How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference, arXiv:2505.09598, May 2025. Cross-model benchmark showing a 65x spread on long prompts.
8. The Green Grid, White Paper #35: Water Usage Effectiveness. The standard definition of WUE in litres per kilowatt-hour.

Zubair Hussain is a full stack developer working in Next.js, React and Node.js. If you spot an error in the arithmetic above, email thezubairh@gmail.com and I will correct it and say so.