At a recent public talk, the first answer from the room to “what is AI?” wasn’t about chatbots or robots. It was “it uses a lot of water.”
That’s a fair place to start, but it bundles two different things together. Energy and water are connected (data centres use water partly to shed the heat their electricity creates), yet they raise different questions and have different answers. So this page takes them one at a time, then gets to the question that matters more than either number: is the work worth the cost?
Energy: how much electricity does one AI request use?
A short text answer uses a fraction of a watt-hour.
- Google measured the median Gemini text prompt at 0.24 watt-hours in May 2025. That includes idle machines kept ready for demand and the data centre’s cooling and power overhead, not just the chips doing the work.
- Independent estimates land in the same range. Epoch AI put a typical GPT-4o response at about 0.3 watt-hours, and a 2026 study in Joule models short requests at a median of 0.31 watt-hours.
The number moves a lot with the kind of work you ask for. Measured benchmarks on open models in the ML.ENERGY leaderboard show:
- Images: roughly 0.1 to 2 watt-hours per image, depending on the model.
- Video: roughly 7 to over 300 watt-hours for a single short clip.
- Long reasoning: a request where the model “thinks” through thousands of words can use many times what a quick answer does. The same 2026 study models long requests at a median of 3.9 watt-hours.
No company publishes an audited, current energy figure for its latest models, so treat all of these as good estimates rather than meter readings. The pattern is reliable even when the exact figure isn’t: text is cheap, images cost more, video and long reasoning cost the most.
How does that compare with things we already do?
This is where most people’s sense of scale gets corrected. Using about 0.3 watt-hours for a short text answer:
- One minute of a 1,000-watt microwave uses about 16.7 watt-hours, or roughly 55 short AI answers.
- One hour of streaming video uses about 77 watt-hours by the IEA’s estimate, counting your device and the network. That’s around 250 short answers.
- Fully charging a smartphone takes about 22 watt-hours, or roughly 75 short answers.
- Running a 3-kilowatt central air conditioner for one hour uses 3,000 watt-hours, the same as about 10,000 short answers. Run it all day through a heat wave and the gap gets very large.
None of this means AI use is free. It means AI belongs in the same conversation as everything else we plug in, most of which we never think twice about.
Want to run your own numbers? The AI Energy Explorer lets you compare AI tasks like for like against streaming, microwaves, air conditioning, lighting and a working day at your desk, with every figure linked to its source. You can get the link sent to you below.
Is AI driving up electricity demand?
Yes, and this is the honest other half. Small per request, multiplied by billions of requests a day, is a large number.
The International Energy Agency estimates data centres used about 485 terawatt-hours of electricity in 2025, and forecasts roughly 950 by 2030, around 3% of the world’s electricity. AI is the main reason for the growth, though not all data-centre electricity is AI: the same buildings run streaming, storage, email and ordinary business software.
It also helps to see AI in the wider picture. The IEA’s 2025 analysis found that data centres account for around one-tenth of global electricity demand growth to 2030, less than industrial motors, air conditioning in homes and offices, or electric vehicles. AI has put a bright spotlight on a grid that was already under pressure. It didn’t create the problem on its own, and it isn’t the only thing making it more pressing.
Water: where does the water actually go?
Water isn’t destroyed when a data centre uses it. That’s worth saying clearly, because “AI drinks water” makes it sound like the water is gone for good.
Hydrologists separate two things:
- Withdrawal is water taken from a river, lake, aquifer or the municipal supply.
- Consumption is the part of that withdrawal that isn’t returned to the local source. The US Geological Survey defines consumptive use as water that is evaporated, transpired, built into products or otherwise removed from the immediate water environment.
Most cooling towers work by evaporation. That water rises as vapour, joins the natural water cycle (hydrologists call it the hydrologic cycle) and eventually falls again as rain or snow. Some cooling water isn’t evaporated at all: it’s discharged back to treatment plants, or kept in closed-loop systems that recirculate the same water over and over.
So the water isn’t lost to the planet. The catch is that it may come down hundreds of kilometres away, in a different watershed, weeks later. For the community the data centre draws from, evaporated water is gone from their supply. That’s why the useful water question is less “how much” and more “from where, and when.”
How much water does one AI request use?
Far less than the viral figure suggests, for a short text answer on today’s systems.
The “bottle of water per email” figure comes from research at UC Riverside, first released in 2023, which estimated about half a litre for every 10 to 50 responses from GPT-3, an older model on older hardware. It also counted the water used by power plants to generate the electricity, not just cooling at the data centre.
Newer published figures are much smaller, and they measure different things:
- Google: 0.26 millilitres of on-site cooling water for its median Gemini text prompt, about five drops.
- Mistral: 45 millilitres for a typical 400-token response, from a lifecycle study that also spreads the water cost of training the model and making the servers across every answer.
So depending on what you count, one text answer sits somewhere between a few drops and a few tablespoons. Not a bottle.
One detail most coverage misses: the bigger share of data-centre water usually sits at the power plant, not the data centre. The US Department of Energy’s lab at Berkeley estimates US data centres used about 66 billion litres of water directly in 2023, and roughly 800 billion litres indirectly through the electricity they bought. That figure covers all data centres, not AI alone. It means the cleanest way to cut AI’s water footprint is often the same as cutting its carbon footprint: cleaner electricity.
Why does location matter so much for water?
Because a litre in a desert town during a drought is not the same as a litre drawn from the Great Lakes in spring.
The same workload can have almost no local water impact in one place and a real one in another, depending on:
- The cooling design. Closed-loop and dry cooling can sharply reduce freshwater use, usually at the cost of using more electricity.
- The water source. Some sites use recycled or non-drinking water instead of the municipal supply.
- The local climate and supply. Evaporative cooling in a water-stressed region competes directly with homes and farms.
- The power grid. Plants that burn fuel or use nuclear power need water for cooling too. Wind and solar need very little.
Large providers now publish water replenishment goals. Google reports replenishing 78% of the freshwater it consumed in 2025. That’s worth something, but replenishing water in one basin doesn’t fix stress in another. Where data centres get built is a community decision as much as a technical one.
Can AI help solve the problems it adds to?
It can, and this is the part of the story that rarely makes the headlines.
AI is one of the most powerful tools we have for designing more efficient systems, modelling the climate, and finding and testing new solutions faster than people could on their own. The IEA estimates that widespread use of AI applications that already exist, in areas like industry, buildings and transport, could cut about 1,400 million tonnes of CO2 emissions in 2035. That’s around three times the total emissions it projects for data centres even in its fast-growth scenario, and it doesn’t include breakthroughs AI might help produce along the way.
The IEA is also clear that none of this happens automatically. Without deliberate adoption, it says the benefit could be marginal, and some gains could be cancelled out if efficiency simply leads to more use. The tool that adds to demand is also one of the best tools for reducing it, but only if people put it to that work.
Is the work worth the cost?
When this came up at the talk, the answer Teryl gave was four questions. They work for AI exactly the way they work for anything else that plugs in:
- How much? Text is cheap. Images, video and long reasoning cost more. Know which one you’re doing.
- From what source? The same task has a very different footprint on a clean grid than on a coal-heavy one, and a very different water impact in a wet region than a dry one.
- For what purpose? Is this solving a real problem, or is it novelty?
- Who benefits, and who pays?
Then the question underneath all four: is the work worth the cost?
That applies to the AI request, and it applies just as much to the air conditioner running in an empty house, the show playing in the background while nobody watches, and the second fridge in the garage. For a business, being intentional with AI usually looks like this:
- Describe the problem well the first time. A lot of wasted AI use is regeneration: asking five times because the first request was vague. Spelling out what goes in, what the rules are and what the finished output should look like gets you there in one or two attempts instead of ten.
- Match the tool to the job. Don’t generate an image when a sentence would do, and don’t send a quick lookup to the heaviest reasoning model you have.
- Don’t put AI in the product if it doesn’t need to be there. AI can be the tool you build with without being the engine that runs every day. A workflow that follows your rules as ordinary software, with no model in the loop, uses almost nothing once it’s built.
- Weigh it against what it replaces. A job that took a team weeks of screen time, travel and meetings and now takes an afternoon isn’t automatically a net win or a net loss. That’s the comparison to make, honestly, case by case.
Some uses are worth it and some aren’t. The point is to decide on purpose, for AI and for everything else.
If you want help working out where AI genuinely earns its place in your business, and where it doesn’t, book a free readiness call. Twenty minutes, no pitch.