Common questions
Good questions deserve careful answers.
AI uses water. The useful question is how much, where, and under which accounting boundary—not whether one dramatic headline wins.
The bottle claim
500 mL
A modeled amount for a group of responses—not a measured bottle poured out for every prompt.
Does one AI prompt use a bottle of water?
No. The 2023 study that popularized the image estimated that GPT-3 would consume about a 500 mL bottle across roughly 10–50 medium-length responses, depending on when and where the model ran.
It was a modeled operational estimate combining on-site cooling water with electricity-related water. It was not a direct measurement of one ChatGPT prompt, and it excluded the upstream water used to manufacture servers.
Newer company disclosures report much smaller per-query figures for newer systems. Google reports 0.26 mL for a median Gemini Apps text prompt; a 2025 public OpenAI estimate says about 0.000085 gallons, or 0.32 mL, for an average ChatGPT query. Those disclosures use different systems, periods, and boundaries, so they should not be treated as a like-for-like rebuttal to the older study.
The short version
Keep the boundary attached.
- 01
The familiar bottle headline compressed a 10–50 response range into “one prompt.”
- 02
Newer disclosures measure newer systems with different operational boundaries.
- 03
Small per-use estimates do not erase aggregate growth or local watershed pressure.
More questions
A calmer way to read the numbers
So does AI use water or not?
Yes. Data centers can consume water on-site for cooling, and the electricity that powers them can carry its own off-site water footprint. Chip and server manufacturing add an upstream footprint too. The amount changes with the model, hardware, cooling system, electric grid, weather, place, and time.
Why do per-prompt estimates disagree so much?
They often count different things. One estimate may model an older GPT-3 response and include electricity-related water. Another may report a median prompt on a newer production system and count operational data-center water. Different boundaries, models, locations, and years can produce very different numbers without making either number universal.
Can I compare a prompt with a shower, coffee, or a shirt?
Only with the scope clearly attached. A shower is mostly direct household water. A data-center figure may be on-site operational consumption. Coffee and cotton figures often describe embedded agricultural water across a supply chain. Those examples can teach scale, but they are not interchangeable measures of local impact.
Do small per-use numbers mean data centers do not matter?
No. A small per-use estimate can become large at high volume, and local impact depends on when and where water is used. Facility growth, drought, watershed stress, cooling choices, and the electric grid still matter. Per-use context and aggregate infrastructure impact answer different questions.
What should I check before sharing a water number?
Ask whether it is measured or modeled, withdrawal or consumption, on-site or electricity-related, per use or annual, and tied to which model, place, and year. If those details are missing, treat the number as a clue—not a conclusion.
Want to see how the boundaries fit together?
The method page explains the accounting terms with one visual flow.