Why does AI use so much energy and water?

Last updated: July 2026

Typing a question into a chatbot feels weightless. But somewhere, a building full of computers is drawing power and evaporating water to answer it. Individually the cost is tiny. Multiplied by billions of queries a day, it adds up to something the size of a small country's demand.

When you ask an AI a question, the request travels to a data centre: a warehouse packed with specialised computer chips. Those chips do a huge number of calculations to produce your answer, and doing calculations uses electricity. The chips also get hot, and keeping them cool often uses water. None of this is visible to you, which is exactly why the scale of it surprises people when they see the numbers.

Let's build the picture from the ground up, starting with a single query and zooming out to the whole industry.

The cost of a single query

How much does one AI question actually cost in energy? The honest answer is that estimates vary, partly because the companies haven't always been transparent about it.¹ But the figures have started to converge.

Google has published a figure of around 0.24 watt-hours for a median text query on its Gemini model, and OpenAI's Sam Altman has mentioned roughly 0.34 watt-hours for an average ChatGPT query.¹ Independent estimates put the range wider, from about 0.3 up to 3 watt-hours for larger queries, which is very roughly a few times the energy of a traditional web search.²

On its own, that's genuinely small. A single watt-hour wouldn't boil a spoonful of water. The catch is what happens when you multiply it.

Where the water comes in

The water side surprises people more than the electricity, because the connection isn't obvious. Data centres run hot, and many are cooled by evaporating water in cooling towers, the same basic principle as sweating. On top of that, the power plants supplying the electricity consume their own water. So every query has a water cost in two places: cooling the chips, and generating the power.³

A widely cited 2023 study from the University of California, Riverside worked out the numbers. By its estimate, a 100-word response from ChatGPT can use around 500 millilitres of water, a small bottle's worth, once both cooling and electricity generation are counted.³ Other estimates land lower, at roughly 10 to 25 millilitres per prompt, depending on assumptions and the efficiency of the specific data centre.⁴ Either way, the point holds: an invisible action has a physical, measurable water cost.

Why the total is so large

Here's where the small numbers turn into big ones. ChatGPT alone is estimated to handle around 2.5 billion prompts every single day.⁵ Multiply even a tiny per-query cost by that, every day, across many AI products, and you arrive at industrial scale.

The International Energy Agency estimated that global data centres used around 415 to 460 terawatt-hours of electricity in 2024 to 2025, and projected that this could roughly double to about 945 terawatt-hours by 2030, with AI driving most of the growth.² To put 945 terawatt-hours in perspective, that's more electricity than the entire country of Japan uses in a year. A United Nations University report projected an associated annual water footprint on the order of 9.3 trillion litres and warned that data centres are becoming country-scale consumers of electricity, water, and land.⁵

The surprising part: using AI costs more than building it

You might assume the huge energy cost is in training these models, the massive one-off process of building them. Training is indeed enormous: one estimate put the training of GPT-4 at somewhere between 50 and 70 gigawatt-hours of electricity.⁵

But it turns out that's not where most of the energy goes. The UN University report found that inference, the everyday running of a finished model to answer user prompts, accounts for an estimated 80 to 90 percent of an AI system's total energy use.⁵ Training happens once. Answering questions happens billions of times a day, forever. The daily grind of use, not the dramatic one-off of training, is the real story.

This also means the cost scales directly with how much we all use AI. Every additional query, image, and video adds to the total. And the heavier tasks cost far more: the same UN report estimated that generating an AI image can use around 1,450 times the energy of a basic text task.⁵ If you've ever wondered why an AI video feels more "expensive" than a chatbot reply, that's why. For more on what those image and video systems are doing under the hood, see our explainer on how AI generates video.

Is anything being done about it?

Yes, though it's a moving target. The big technology companies have made efficiency and sustainability commitments. Google, for instance, reported replenishing 64% of its freshwater consumption in 2024, up sharply from 18% the year before, as part of a goal to be "water positive" by 2030.⁴ Chips are getting more efficient, and cooling methods are improving.

But efficiency gains are running up against sheer growth in demand. Even as each query gets cheaper, the number of queries is exploding, and heavier uses like video generation are becoming common. Whether efficiency can outpace demand is one of the genuinely open questions about AI's environmental future.

What to take away

The resource cost of any single thing you ask an AI is small, and there's no need to feel guilty about a chatbot question the way you might about a long flight. But the aggregate is real and large, driven by billions of daily uses rather than by any one of them, and dominated by everyday use rather than by model training.

The useful frame isn't "AI is an environmental disaster" or "AI's footprint is trivial". It's more precise than either: a tiny individual cost, a vast collective one, and a technology whose appetite for power and water is growing fast enough that it's now shaping decisions about electricity grids and water supplies in the places these data centres are built.

Sources

1. Hannah Ritchie, "How much electricity does AI consume?" (2026), citing Google and OpenAI figures — hannahritchie.substack.com
2. Presenc AI, "AI Data Center Energy Consumption Statistics 2026," citing the IEA Energy and AI report — presenc.ai
3. Li et al., "Making AI Less 'Thirsty'," University of California, Riverside, arXiv:2304.03271 (2023), as reported by The Washington Post — businessenergyuk.com
4. Axis Intelligence, "AI Data Center Water Usage Statistics 2026," citing UC Riverside and Google's 2025 Environmental Report — axis-intelligence.com
5. United Nations University, "AI Is Threatening Natural Resources for Billions" (2026) — unu.edu