One ChatGPT text prompt uses only a little energy, about 0.34 watt-hours by OpenAI’s own estimate. That is roughly what a 10-watt LED bulb uses in two minutes. The bigger issue is scale. Data centers that run AI now draw a fast-growing share of the world’s electricity, and in some places they also use a lot of local water.
So the fair answer is this: per question, the footprint is small, but the total load is real and rising. Below are the best numbers available as of October 2026, and what they leave out.
How much energy one ChatGPT prompt uses
Still, OpenAI has published very little on this. The main figure comes from a June 2025 post by chief executive Sam Altman, The Gentle Singularity. He wrote that the average query uses about 0.34 watt-hours and about 0.000085 gallons of water. That is roughly 0.32 milliliters, or a few drops.
The figure has limits, though. OpenAI did not share its method or define an “average” query. It also did not say whether long reasoning tasks or model training were included.
Google has released a more detailed study. In an August 2025 paper, it measured its own systems serving the Gemini app in May 2025. The median text prompt used 0.24 watt-hours and 0.26 milliliters of water. It produced 0.03 grams of carbon dioxide equivalent, using market-based accounting.
| Source | Energy per text prompt | Water per prompt | What it leaves out |
|---|---|---|---|
| OpenAI (Altman, June 2025) | About 0.34 Wh | About 0.32 mL | Method not published |
| Google Gemini study (May 2025 data) | 0.24 Wh median | 0.26 mL median | Training, user devices, outside networks |
Simple text prompts are the cheapest kind of AI task. According to the International Energy Agency, reasoning models, video generation and AI agents can use hundreds to thousands of times more energy per request. So a quick question and a long research task are not the same thing.
Why scale matters more than a single prompt
A small number multiplied by a huge one is still big. ChatGPT alone has hundreds of millions of weekly users. As an illustration, a billion prompts at 0.34 Wh each would use about 340 megawatt-hours a day. Over a year, that comes to roughly 124 gigawatt-hours.
Meanwhile, the wider trend is clearer. In its 2026 report Key Questions on Energy and AI, the IEA estimates that data centers worldwide used 485 terawatt-hours in 2025. That was up 17% from 2024. AI-focused data centers grew even faster, by about 50%. The agency projects total data center use will roughly double to 950 terawatt-hours by 2030, about 3% of global electricity demand.
The United States also carries a large part of that load. A Department of Energy report prepared by Lawrence Berkeley National Laboratory found that data centers used 176 terawatt-hours in 2023. That was 4.4% of all US electricity. The report projects 6.7% to 12% by 2028, depending on how fast AI servers spread.
Efficiency is improving, though. The IEA says energy use per AI task has fallen by at least ten times a year in recent years. Google reported that its median prompt used 33 times less energy in May 2025 than a year earlier. Its carbon footprint per prompt fell 44-fold over the same period.
However, cheaper prompts tend to invite more use. So total demand can keep rising even as each request gets leaner. Economists call this a rebound effect, and it shows up in many technologies.
Training is the other piece. Building a large model takes a big one-time burst of computing. However, AI companies do not publish training energy for their newest models, so outside estimates remain rough.
Water: where the scary numbers come from
Data centers use water in two ways. First, many cool their servers with evaporative cooling towers on site. Second, power plants that supply their electricity also consume water, mostly for cooling.
The Berkeley Lab report put direct water use by US data centers at about 66 billion liters in 2023, or roughly 17.4 billion gallons. That was up from 21.2 billion liters in 2014. The indirect water tied to their electricity was far larger, at nearly 800 billion liters.
This is why per-prompt water estimates vary so widely. The OpenAI and Google figures count mainly on-site water. Studies that add power plant water arrive at much higher numbers. Neither approach is wrong; they simply measure different things.
Even so, location matters most. A data center in a dry region draws on scarce local supplies, while the same facility near abundant water has less impact. Some newer sites use closed-loop cooling, which recirculates water but tends to use more electricity. Better water tracking helps too, as an older archive story on finding water leaks with data showed.
Carbon depends on the power grid
Electricity is only as clean as its source. The Berkeley Lab report estimated that US data center power use in 2023 was linked to about 61 million metric tons of carbon dioxide equivalent. That assumed a national average of 0.34 kilograms per kilowatt-hour.
Company figures often look lower because of how they are counted. Google’s 0.03 grams per prompt uses market-based accounting, which credits clean power the company buys. Location-based numbers, which reflect the actual local grid, are usually higher.
The grid mix is shifting in both directions. The IEA expects renewables to supply a large share of new data center power. It also projects 15 to 27 gigawatts of on-site natural gas generation for data centers by 2030, mostly in the United States. For background on clean power sources, see our guide to what renewable energy is.
One prompt in everyday terms
Comparisons help, but only for simple text prompts. Using the US average grid factor above, a 0.34 Wh prompt works out to about 0.12 grams of carbon dioxide equivalent.
- A typical gasoline car emits about 400 grams of carbon dioxide per mile, according to the EPA. That equals roughly 3,400 simple prompts.
- A 1,000-watt microwave uses 0.34 Wh in about 1.2 seconds.
- A 10-watt LED bulb uses the same in about two minutes.
By these measures, an occasional chatbot question is a tiny part of a personal footprint. Heavy use of image, video or agent tools is a different story, since each task can cost many times more.
How to use AI with a smaller footprint
Individual choices matter less than how data centers are powered and cooled. Still, a few habits reduce waste:
- Write a clear prompt once instead of regenerating answers several times.
- Use a standard model for simple questions and save reasoning modes for hard problems.
- Skip AI image or video generation when a quick search would do.
- Turn off features that run AI in the background if you do not use them.
Energy is only one of the tradeoffs with these tools. Our overview of AI pros and cons covers the rest, and more explainers live in our science and technology guides.
Frequently asked questions
How much water does one ChatGPT query use?
OpenAI says about 0.000085 gallons, or roughly 0.32 milliliters, for an average query. That mainly covers data center cooling, not water used at power plants.
Is ChatGPT worse for the environment than a Google search?
A chatbot answer usually takes more computing than a classic search. Even so, the IEA estimates that if every web search were a simple AI text query, it would use under 4 terawatt-hours a year.
Is AI energy use still growing?
Yes. The IEA found that AI-focused data centers used about 50% more electricity in 2025 than in 2024.
Does training a model use more energy than using it?
Training is a large one-time cost, while use adds up every day. Because companies do not publish training figures for their newest models, the exact split is unknown.
Have a question about this topic? Ask in the comments below. Our editors read every comment and answer questions here, and we add useful answers to the guide.
Sources
- blog.samaltman.com/the-gentle-singularity
- arxiv.org/abs/2508.15734
- iea.org/reports/key-questions-on-energy-and-ai/executive-summary
- energy.gov/articles/doe-releases-new-report-evaluating-increase-ele…
- eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data…
- epa.gov/greenvehicles/greenhouse-gas-emissions-typical-passenger…