ChatGPT Water Usage: I Tried to Reproduce the 0.32 mL Claim

FACT CHECK · AI INFRASTRUCTURE · 18 SEPTEMBER 2026

I tried to reproduce the 0.32 mL ChatGPT water number. Here is what I found

Q

How much water does one ChatGPT query actually use?

ZH

The arithmetic checks out. The public methodology does not exist. The figure is consistent with counting only the water evaporated in data center cooling, which is the smaller half of the water story.

Plausible under stated assumptions. Not verifiable. Not a constant.

0.32 mL Sam Altman. Boundary not stated.

0.26 mL Google median. Cooling only, formula published.

0.6 to 17 mL Independent range. Wider boundary.

A figure has been circulating for over a year: one ChatGPT query uses about 0.32 millilitres of water, roughly one fifteenth of a teaspoon. It resurfaced in September 2026 when Sam Altman used it to claim that 38,000 ChatGPT queries use as much water as growing a single almond.

I build web applications for a living, so my instinct with any performance number is the same. What was measured, where did the measurement start and stop, and can I reproduce it? I spent a few evenings on this one.

The arithmetic is fine

Altman's original blog post said 0.000085 US gallons. One US gallon is 3,785.41 mL, and his teaspoon comparison agrees with the conversion.

unit conversion

0.000085 gal x 3785.411784 mL/gal = 0.32176 mL
1 US teaspoon = 4.92892 mL, and 4.92892 / 15 = 0.3286 mL

Two statements in the same sentence that agree with each other. The arithmetic was never the weak point.

What a prompt actually touches

To reason about where water could possibly enter, you need the request path. Here is the version a backend developer would recognise, stripped to the parts that draw power.

Figure 1. The accelerator dominates, but it is not the only consumer. Work scales with tokens, not requests.

GPUs do not drink water

This is the part headlines skip. A GPU consumes electricity and emits heat. That is the whole interaction. Water is consumed around it, in three places, and only one of them is inside the building.

Figure 2. Direct water evaporates on site. Indirect water is consumed at the power station. The silicon consumes neither.

A closed-loop direct-to-chip cooling circuit is not consumption either, because the same coolant returns. Consumption happens where water evaporates or is discharged unrecovered.

Where the number probably comes from

OpenAI has not published what the 0.32 mL covers. Google has, so I borrowed their formula. Water equals total energy minus facility overhead, multiplied by water usage effectiveness. Checking it against Google's own published numbers reproduces their result in one line, so it is a fair template.

reverse engineering

Google check:      (0.24 Wh - 0.02 Wh) x 1.15 L/kWh  =  0.253 mL  ->  published 0.26 mL
Applied to OpenAI: 0.32176 mL / 0.313 Wh             =  1.03 L/kWh implied WUE

An implied water usage effectiveness near 1.0 litres per kWh is realistic for a modern facility, sitting just under Google's reported 1.15. That only works if the number counts on-site cooling and nothing else.

In the peer-reviewed research, training GPT-3 consumed about 0.7 million litres on site and about 5.4 million litres once the water used to generate its electricity was counted. A per-query figure that leaves out the second number is measuring the smaller half.

Why there cannot be one universal number

Water per prompt is not a constant. It is the output of a chain of multiplications where every term is a variable, and the last term is optional.

Figure 3. Change any single term and the answer moves. Change the boundary and it moves by a multiple.

Epoch AI estimated about 0.3 Wh for a typical GPT-4o query, 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. That is a factor of more than 100 within one model. Location matters just as much: in the peer-reviewed tables the same workload came out at roughly 7.6 mL per request in Texas and roughly 47.6 mL in Washington.

The estimates only look contradictory

Figure 5. Log scale. 0.26 mL and 47.6 mL are not in conflict. They answer different questions.

Read the labels rather than the bars. The spread across published research is mostly disagreement about the accounting boundary, not about physics. Estimates that count only on-site cooling cluster below 1 mL. Estimates that add electricity generation water sit between 7 and 48 mL.

The boundary problem, as a latency bug

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

Figure 4. Both teams measured correctly. The numbers differ by 15x because the brackets are in different places.

Nobody accuses anyone of lying about latency. Everyone asks for the trace. AI water figures need the same discipline and do not yet have it, and unlike latency there is no shared convention about where the brackets go.

And the almond

PolitiFact rated the 38,000 queries per almond comparison Mostly False in September 2026. The multiplication is correct, and I reproduced it. The comparison is not, because the almond side uses a full water footprint of about 12 litres including rainfall and dilution water, while the ChatGPT side appears to count cooling only.

the almond, both ways

Altman's version:  12.1 L / 0.32176 mL  =  about 37,600 queries  (rounded to 38,000)
Direct water only: 6,000 mL / 0.6 mL    =  10,000 prompts
                   6,000 mL / 17 mL     =  about 350 prompts

Compare direct water with direct water and you land somewhere between 1,000 and 10,000 prompts per almond, which is what PolitiFact concluded.

THE FULL TECHNICAL VERSION

The long write-up lives on my portfolio, with everything this post left out.

It covers the full request path end to end, direct versus indirect water with the scope 1 and scope 2 definitions, every variable that moves the number, a side-by-side comparison table of all six published estimates with their boundaries and statistics, the training versus inference split, the system boundary problem explained as an API latency bug, why the older "one prompt equals a bottle of water" claim was also oversimplified, a nine-point checklist called What Developers Should Ask Before Trusting an AI Environmental Metric, an FAQ, and every source with dates and links. It is also published in Urdu, German, Spanish, French, Portuguese, Japanese and Chinese.

Read the full investigation →

ZH

Zubair Hussain is a full stack developer working in Next.js, React and Node.js. He writes about AI infrastructure, web performance, and the numbers people quote about both.

Spot an error in the arithmetic? Email thezubairh@gmail.com and I will correct it and say so.