How much water and energy does your AI use?

Q: Does one ChatGPT question really use a bottle of water?

An older widely shared figure claimed roughly 500 ml for a short conversation with an early model. Newer estimates for efficient models are far lower, often a few milliliters or less per question. It varies by model and data center, so a range is more honest than a single number.

Q: Is AI worse for the environment than everyday things I already do?

Per query, usually not. A single AI question is often comparable to a web search or a few seconds of video streaming. The concern is the speed at which total AI use is growing, not any one query.

Every question you ask a chatbot runs on a data center that draws electricity and burns through water to stay cool. Individually it is tiny. Over a year, it adds up. Here is a rough estimate for your own use.

By Rahul Nayak · Last reviewed September 2026

What kind of AI use is it?

Everyday use: normal back-and-forth with a standard assistant.

How many questions a day?

8 a day is the average for a regular ChatGPT user.

Where are the data centers?

The local power mix changes how much carbon the same electricity creates.

For how many people?

Water, per year —

Electricity, per year —

Carbon, per year —

These are estimates, not exact readings. Published figures for a single AI query range from about 0.3 to 50 milliliters of water depending on the model, the data center, and how it is cooled. This tool uses mid-range public figures and shows a low-to-high band so you can see the uncertainty rather than a false exact number.

How this is calculated (and sources)

The math is deliberately simple and transparent:

Energy per question depends on the type of use: roughly 0.3 Wh for a quick question on an efficient model, ~1.75 Wh for everyday use, and up to ~19 Wh for long, complex "reasoning" answers.
Water is estimated at about 1.8 milliliters per watt-hour, covering both the water used to cool servers and the water used at power plants to make the electricity.
Carbon uses the local grid: about 0.71 kg CO₂ per kWh in India, 0.37 in the US, and 0.48 as a global average.
Daily figures are multiplied by 365, and by the number of people, to get the yearly total.

Figures drawn from public disclosures and research including Google's 2025 AI inference report, OpenAI's per-query figures, Li et al. "Making AI Less Thirsty," and MIT Technology Review's analysis. Efficiency is improving fast, so treat these as a snapshot, not a permanent number.

Why AI uses water at all

It surprises most people. The link is heat. Answering your question means a rack of powerful computer chips running hard for a moment, and those chips get hot. Data centers pump water through cooling systems to carry that heat away, and more water is used back at the power station generating the electricity. So "water use" is really the shadow of energy use.

The reason it matters is scale. One question is nothing. But data centers already consumed roughly 4.4% of all US electricity in 2023, and that share is projected to reach 6.7% to 12% by 2028 as AI use grows. Multiply a small number by billions of queries a day and it becomes a real strain on power grids and local water supplies, especially in places that are already short of both.

How to shrink your AI footprint

You do not have to stop using AI. A few habits cut most of the waste:

Match the model to the job. Using a smaller, faster model for simple tasks instead of a heavy "reasoning" model is the single biggest lever, often 10 to 50 times less energy.
Be direct. One clear question uses less than five vague ones that each need a follow-up.
Do not regenerate needlessly. Each "try again" is a fresh query with a fresh cost.
Batch related questions into one conversation rather than reopening the tool all day.

Common questions

Does one ChatGPT question really use a bottle of water?

An older, widely shared figure claimed roughly 500 ml for a short conversation with an early model. Newer estimates for efficient models are far lower, often a few milliliters or less per question. The honest answer is that it varies enormously by model and data center, which is why this tool shows a range.

Is AI worse for the environment than everyday things I already do?

Per query, usually no. A single AI question is often comparable to a web search or a few seconds of video streaming. The concern is not one query, it is the speed at which total AI use is growing.

Why does location change the carbon number?

The same electricity creates different amounts of carbon depending on how it was generated. A grid leaning on coal produces far more CO₂ per unit than one running on hydro, nuclear, or solar, which is why India's figure is higher than the US average.