How Much Energy and Water Does One AI Prompt Use?

You may have read a headline saying a chatbot "drinks a bottle of water" for every few questions. Then you may have read another saying one prompt uses only "five drops". Both come from real studies. So which one is right?

This lesson gives you the best current numbers for a single AI prompt, explains why the estimates disagree so much, and compares them with things you already know, like a light bulb or a kettle.

What You'll Learn

• The best published figures for the energy and water used by one text prompt
• Why different studies give such different answers
• How a prompt compares with everyday things like a light bulb, a phone charge and a kettle
• Why images, video and long "thinking" answers cost much more than a short text reply

The two units you need

Two simple units are enough for this whole course.

• Watt-hours (Wh) measure energy. A 10-watt LED bulb left on for one hour uses 10 Wh.
• Milliliters (mL) measure water. A teaspoon holds about 5 mL. A small water bottle holds 500 mL.

The best current figures for one text prompt

Only a few AI companies have published per-prompt numbers. Here are the main ones.

The energy numbers agree quite well: roughly 0.2 to 0.4 Wh for a typical text prompt to a large chatbot. The water numbers do not agree at all. They range from a fraction of a milliliter to tens of milliliters.

Why the water numbers disagree

The gap is mostly about what gets counted, not about who is lying.

1. Water used on site. Many data centers cool their chips by letting water evaporate. Google's figure of about 0.26 mL counts this water.
2. Water used to make the electricity. Many power plants also use water for cooling. If you add that, the number grows.
3. Water used to make the hardware. Making chips and servers uses water too. Mistral's life cycle study counts this, which is one reason its figure is much higher.

An earlier academic study, "Making AI Less Thirsty", estimated that an older model used about a 500 mL bottle of water for every 10 to 50 medium-length answers. That works out to roughly 10 to 50 mL per answer, and it also counts power-plant water. The authors said the real number depends heavily on where and when the data center runs.

So when you see a water figure, ask one question first: does it count only cooling water, or also the water behind the electricity and the hardware?

Comparing a prompt with everyday things

Numbers like 0.3 Wh are hard to picture. Here is the same energy next to things you know. These are simple calculations using about 0.3 Wh per prompt.

• A 10-watt LED bulb uses 0.3 Wh in about two minutes.
• Charging a phone fully takes very roughly 15 to 20 Wh for a typical battery. That is about 50 to 70 text prompts.
• Boiling one liter of water in a kettle needs about 100 Wh. That is roughly 300 text prompts.
• A web search. In 2009, Google said one search used about 0.3 Wh. That old figure is often quoted as "the same as a chatbot prompt", but search has changed a lot since then, so treat the comparison as a rough guide only.

For one person sending a few short text prompts a day, the energy is small. It is less than leaving a light on for a few extra minutes. The bigger picture changes when you multiply by billions of prompts a day, and that is what the rest of this course is about.

Not all prompts are equal

The figures above are for a short text reply. Other tasks can use far more.

• Long inputs. Pasting a whole report or a long chat history means the model has to read much more text first. Epoch AI notes that very long inputs can raise the energy cost many times over.
• Reasoning ("thinking") models. These write many hidden steps before they answer, so they can use much more energy than a quick reply.
• Images. A 2023 study by researchers at Hugging Face and Carnegie Mellon found image generation used far more energy than text tasks. In their tests, one image from a large model came close to the energy of a full phone charge. Newer tests with smaller image sizes found much lower numbers, so estimates vary.
• Video. A 2025 MIT Technology Review analysis measured one open video model and found a five-second clip used about as much energy as running a microwave for an hour. Other analysts think that figure is high. Either way, video is in a different league from text.

Why these numbers keep changing

Per-prompt figures are snapshots. Google reported that the energy used by its median text prompt fell sharply over one year, thanks to better software and hardware. New models, longer answers and more image and video use can push the numbers the other way. When you read any figure, check the date, the model and what it counts.

Key Takeaways

• A typical text prompt to a large chatbot uses roughly 0.2 to 0.4 Wh of energy, based on the few published figures.
• Water estimates range from under 1 mL to tens of mL per answer, mostly because studies count different things.
• Always ask whether a water figure counts only cooling water, or also electricity and hardware.
• For one person, a few text prompts use about as much energy as a light bulb left on for a few minutes.
• Long inputs, reasoning, images and especially video use much more than a short text reply.

Sources

• Google, Measuring the environmental impact of delivering AI at Google scale (2025)
• Sam Altman, The Gentle Singularity (2025)
• Epoch AI, How much energy does ChatGPT use? (2025)
• Mistral AI, Our contribution to a global environmental standard for AI (2025)
• Li, Yang, Islam and Ren, Making AI Less "Thirsty"
• Google, Powering a Google search (2009)
• Luccioni, Jernite and Strubell, Power Hungry Processing (2023)
• MIT Technology Review, We did the math on AI's energy footprint (2025)