Here is the honest number: a single, typical text question to ChatGPT uses roughly 0.34 watt-hours of electricity and around 0.32 milliliters of water, which is a fraction of a teaspoon. That is about what a modern LED bulb uses in a minute or two. So on a per-question basis, your individual AI use is tiny. The catch, and the reason this is a real debate, is scale: multiply a tiny number by a billion queries a day and you get data centers that strain local power grids and water supplies. Both things are true at once, and most headlines only tell you one of them.
The reason people see wildly different figures, one guide says a few drops of water and another says a whole bottle, is that they are measuring different things. This guide gives you the real ranges, explains why the numbers disagree, separates the part you control from the part you do not, and tells you the few habits that actually reduce your footprint.
Quick Answer: The Real Numbers Per Query
Rough per-query figures for 2026. They vary with model size, query length, and which data center answers you.
| What you do | Electricity | Rough comparison |
|---|---|---|
| Simple text question | ~0.3 Wh | About 10x a Google search; an LED bulb for a minute or two |
| Long or reasoning-heavy answer | 2 to 20+ Wh | Reasoning models think longer and cost much more |
| Generating one image | 2 to 5 Wh | Images cost far more than text |
| Generating video | 10 to 100x an image | By far the most intensive everyday task |
Why the Water Numbers Are So Confusing
Water is where the scary headlines live, and it is almost entirely a measurement problem. OpenAI's Sam Altman put an average query at about 0.32 milliliters of water, and Google reported a similar figure for a Gemini query, both counting only the water evaporated at the data center to cool the servers. A widely cited 2023 study from the University of California, Riverside, instead found that a 100-word GPT-4 response could use around 519 milliliters, roughly a full bottle. Both can be right, because they count different things. The small number is only the cooling water at the building. The big number also includes the water used at the power plants generating the electricity, and it reflected a worst-case, coal-heavy grid. The honest range for one query, counting everything, is somewhere between a fraction of a milliliter and a couple of tablespoons, depending on where the servers sit and how their power is made.
When you see any AI water or energy figure, ask one question: what does it include? A number under one milliliter per query is counting only on-site cooling. A number in the hundreds of milliliters is counting the entire electricity supply chain too. Neither is a lie; they are answering different questions.
The Part You Control vs the Part You Do Not
This is the point almost every viral post misses. Your individual choices barely move the total; the data center's design moves it enormously. The same query answered at a Finnish facility on renewable power with efficient cooling has a fraction of the footprint of one answered at an air-cooled site on a coal-heavy grid, and you have no say in which one handles your request. Analyses consistently find that infrastructure choices, the cooling technology and the electricity mix, matter something like ten to twenty-five times more than how often an individual uses AI. Cutting your own usage by 10 percent saves roughly 2 percent of your already-tiny footprint. That does not mean the total is nothing, it means the lever that matters is the industry building cleaner data centers, not you skipping a chatbot question.
Training vs Using: A Common Mix-Up
People often assume every question re-runs the enormous cost of building the model. It does not. Training a large model is a one-time expense, historically comparable to the annual electricity of some hundred-plus homes for a single model, but that cost is then spread across the billions of queries the model answers over its life. In fact, serving those queries, called inference, now overtakes training as the dominant energy cost within months of a model's release, simply because of volume. The takeaway: training sounds terrifying as a single number, but per use it is negligible; the real ongoing footprint is the sheer number of everyday queries, which is exactly why efficiency per query is improving so fast, with Google reporting a large efficiency gain in a single year.
How to Actually Shrink Your Footprint
If you want to use AI more responsibly without pretending your choices are decisive, a few habits genuinely help, and they mostly cost you nothing.
- Prefer text over images and video. A text answer can use tens or hundreds of times less energy than generating media.
- Do not reach for a heavy reasoning model for simple questions. Extended thinking multiplies the energy per answer.
- Batch and be specific. Getting a good answer in one clear prompt beats ten vague follow-ups that each cost energy.
- For private, low-footprint work, a small local model on your own computer uses essentially no water and modest power.
- Favor providers that run on renewable-heavy grids and publish their efficiency figures.
If you would rather not run a heavy model when a light one will do, a tool that lets you switch easily helps. A platform like LumiChats gives you many models in one place, so you can pick a smaller, faster model for simple questions and save the frontier models for the work that genuinely needs them, which is both cheaper and lighter on energy.
01How much energy does one ChatGPT question use?
A typical text question uses roughly 0.34 watt-hours, about ten times a Google search, or what an LED bulb draws in a minute or two. Long, reasoning-heavy answers can use several times more, and generating images or video uses far more still.
02How much water does AI use per query?
It depends on what you count. On-site cooling alone is about 0.32 milliliters per query, per OpenAI. Counting the water used to generate the electricity too, estimates range from under a milliliter to a couple of tablespoons, and one worst-case 2023 study put a 100-word answer near a full bottle.
03Is using AI bad for the environment?
Per use, the impact is small, far less than a short car trip or a single beef meal. At global scale it is significant, driving major new electricity and water demand. The biggest lever is how data centers are built and powered, which matters far more than any individual's usage.
04Does AI training use more energy than using it?
Training is a large one-time cost, but it is spread across billions of later queries. Serving those queries, called inference, actually overtakes training as the dominant energy cost within months of a model launching, because of sheer volume.
05What uses the most energy: text, images, or video?
Video generation is by far the most intensive, followed by image generation, with plain text the cheapest by a wide margin. Reasoning models that think for longer also use far more energy than standard text answers.
The bottom line: one AI question is genuinely trivial, roughly an LED bulb for a minute and a few drops of water, so you can stop feeling guilty about individual use. The real story is scale and infrastructure, and the honest response is to push the industry toward cleaner, more efficient data centers rather than to quietly count your own prompts. Read every dramatic figure by asking what it actually includes, and the panic usually dissolves into arithmetic.
