{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"the-electricity-and-the-water","title":"The electricity and the water","url":"https://addaly.com/learn/ai-actually-explained/the-electricity-and-the-water","position":87,"of":92,"minutes":10,"answer":"Data centres drew about one and a half per cent of the world's electricity in 2024 and are expected to double by 2030 with AI the fastest-growing part.","keyIdea":"Data centres drew about one and a half per cent of the world's electricity in 2024 and are expected to double by 2030 with AI the fastest-growing part, and the per-query figures are small because they leave out the grid, the water behind the power and the concentration in a few places."},"body":{"format":"markdown","text":"## The number to anchor on\n\nThe agency that tracks the world's energy estimated that data centres consumed around 415 terawatt-hours of electricity in 2024, about one and a half per cent of all electricity generated on earth, and projected that this would more than double, to around 945 terawatt-hours, by 2030. The fastest-growing share is the servers built for AI. For scale, the 2030 figure is somewhat more than the whole of Japan's electricity consumption today.\n\nThat is the honest headline number, and every other figure in this lesson is a way of understanding what it is made of and what it hides.\n\n> Figure: Data-centre electricity, against a country. Bars (terawatt-hours a year): All data centres, 2024 415; All of Japan, today 900; All data centres, 2030 projection 945. About one and a half per cent of all electricity generated on earth in 2024, projected to more than double by 2030 with AI-built servers the fastest-growing part. The strain is not on the world's supply but on particular grids: data centres took around a fifth of Ireland's electricity and a quarter of Virginia's, and both have restricted new connections.\n\n## Why the per-query number is small and the total is not\n\nThe module on putting it to work gave you the per-query figure from the labs' own 2025 measurements: about a quarter to a third of a watt-hour for a median text prompt. Multiply by the billions of prompts a day and add images, video, reasoning and training, and you have a large part of the growth above. Both numbers are true. A person who quotes only the first is telling you the system is cheap; a person who quotes only the second is telling you it is ruinous; neither has told you the mechanism, which is that a small thing done billions of times is a large thing.\n\n## What the per-query figure leaves out\n\n**The accounting boundary.** The labs' figures measure the electricity at the chip, the server around it, and the building's cooling and overhead. They do not count the water used to generate that electricity at the power station, which in many regions is larger than the water used on site. A 2023 university estimate that included the water behind the power put a conversation of a few dozen prompts at around half a litre; the lab's own 2025 figure, counting only on-site cooling, was a fraction of a millilitre per prompt. The two are not in contradiction. They count different things, and neither says which.\n\n**Training.** A frontier training run draws tens of gigawatt-hours over months. Meta reported the training of its largest 2024 model at around eleven thousand tonnes of carbon dioxide on the basis of where the electricity was generated, and zero on the basis of the renewable-energy certificates it had bought. Both numbers appear in the same report. The first is what the grid emitted; the second is an accounting convention. When a lab says its model was trained on clean energy, ask which of the two it means.\n\n**Where it happens.** Electricity is not global; it is local. A data centre draws from one grid and one aquifer. Data centres used around a fifth of Ireland's electricity in 2023 and around a quarter of the American state of Virginia's; both have since restricted new connections. The strain is not on the world's supply. It is on particular places that did not plan for it.\n\n## Water\n\nData centres use water in two ways: directly, to cool the chips, and indirectly, through the water used to generate their electricity. One large cloud company's direct water use rose by about a third in a single year, 2022, as its AI building began; another reported consuming over twenty billion litres in 2023, most of it for cooling. Cooling can be done with less water and more electricity, or more water and less electricity, and the choice is made per site. In a water-stressed city the trade is not a technical one.\n\n## India\n\nIndia's data-centre capacity was around a gigawatt in 2024 and several times that was under construction, concentrated around Mumbai, Chennai, Hyderabad and the capital region. Two facts make the Indian case its own. The grid is around seventy per cent coal, so a query answered in India carries several times the carbon of one answered in a hydro-rich region: the same electrons, a different source. And several of the cities being built in are water-stressed, so the cooling choice above is a choice about municipal supply. The national compute programme from the first lesson of this module will draw from the same grids.\n\n## What the industry is doing\n\nBuying power. In 2024 one cloud company contracted to restart a closed nuclear reactor for its exclusive use; others signed for small reactors that do not yet exist, and for gas turbines that do. Several labs have said that electricity, not chips, is now the constraint on the next generation of models. The grid connections being sought in some regions exceed the region's current capacity to build.\n\n## How to read an energy claim\n\nPer query, per day, or per year? Chip only, or building and grid? On-site water, or the water behind the power? Location-based carbon or market-based? And where: because the same query costs a different amount of coal in Mumbai and in Quebec. A claim that answers all five is rare and worth trusting. A claim that answers one is true about one thing."},"check":{"question":"A lab reports that a query uses a fraction of a millilitre of water, and a university study reports that a short conversation uses about half a litre. Why do the figures differ so much?","options":[{"text":"The university study was done on an older, less efficient model that needed far more cooling for the same answer.","correct":false,"tempting":"Reaches for model efficiency; the gap is far larger than any efficiency change and comes from what was counted."},{"text":"The lab counts only on-site cooling, while the study also counts the water behind the electricity.","correct":true},{"text":"The lab's figure is per token and the study's is per conversation.","correct":false,"tempting":"Misplaces the unit; even scaled to the same conversation the on-site figure stays far below the study's."},{"text":"One of the two must be wrong, since water use is a physical fact.","correct":false,"tempting":"Assumes a single boundary; both are physically correct measurements of different boundaries."}]},"prev":{"slug":"who-owns-what-it-makes","title":"Who owns what it makes","url":"https://addaly.com/learn/ai-actually-explained/who-owns-what-it-makes"},"next":{"slug":"the-safety-argument","title":"The safety argument, both sides fairly","url":"https://addaly.com/learn/ai-actually-explained/the-safety-argument"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'The electricity and the water', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/the-electricity-and-the-water (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"The electricity and the water","courseTitle":"AI, Actually Explained","module":{"index":9,"slug":"who-decides","title":"Who controls it, and what is unsettled"},"answer":null,"keyIdea":"Data centres drew about one and a half per cent of the world’s electricity in 2024 and are expected to double by 2030 with AI the fastest-growing part, and the per-query figures are small because they leave out the grid, the water behind the power and the concentration in a few places.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"bars","unit":"terawatt-hours a year","title":"Data-centre electricity, against a country","series":[{"label":"All data centres, 2024","value":415},{"label":"All of Japan, today","value":900},{"label":"All data centres, 2030 projection","value":945,"highlight":true}],"caption":"About one and a half per cent of all electricity generated on earth in 2024, projected to more than double by 2030 with AI-built servers the fastest-growing part. The strain is not on the world’s supply but on particular grids: data centres took around a fifth of Ireland’s electricity and a quarter of Virginia’s, and both have restricted new connections."},"alt":"Figure: Data-centre electricity, against a country. Bars (terawatt-hours a year): All data centres, 2024 415; All of Japan, today 900; All data centres, 2030 projection 945. About one and a half per cent of all electricity generated on earth in 2024, projected to more than double by 2030 with AI-built servers the fastest-growing part. The strain is not on the world’s supply but on particular grids: data centres took around a fifth of Ireland’s electricity and a quarter of Virginia’s, and both have restricted new connections."}},"blocks":[{"t":"h2","id":"the-number-to-anchor-on","runs":[{"text":"The number to anchor on"}]},{"t":"p","runs":[{"text":"The agency that tracks the world’s energy estimated that data centres consumed around 415 terawatt-hours of electricity in 2024, about one and a half per cent of all electricity generated on earth, and projected that this would more than double, to around 945 terawatt-hours, by 2030. The fastest-growing share is the servers built for AI. For scale, the 2030 figure is somewhat more than the whole of Japan’s electricity consumption today."}]},{"t":"p","runs":[{"text":"That is the honest headline number, and every other figure in this lesson is a way of understanding what it is made of and what it hides."}]},{"t":"h2","id":"why-the-per-query-number-is-small-and-the-total-is-not","runs":[{"text":"Why the per-query number is small and the total is not"}]},{"t":"p","runs":[{"text":"The module on putting it to work gave you the per-query figure from the labs’ own 2025 measurements: about a quarter to a third of a watt-hour for a median text prompt. Multiply by the billions of prompts a day and add images, video, reasoning and training, and you have a large part of the growth above. Both numbers are true. A person who quotes only the first is telling you the system is cheap; a person who quotes only the second is telling you it is ruinous; neither has told you the mechanism, which is that a small thing done billions of times is a large thing."}]},{"t":"h2","id":"what-the-per-query-figure-leaves-out","runs":[{"text":"What the per-query figure leaves out"}]},{"t":"p","runs":[{"text":"The accounting boundary.","b":true},{"text":" The labs’ figures measure the electricity at the chip, the server around it, and the building’s cooling and overhead. They do not count the water used to generate that electricity at the power station, which in many regions is larger than the water used on site. A 2023 university estimate that included the water behind the power put a conversation of a few dozen prompts at around half a litre; the lab’s own 2025 figure, counting only on-site cooling, was a fraction of a millilitre per prompt. The two are not in contradiction. They count different things, and neither says which."}]},{"t":"p","runs":[{"text":"Training.","b":true},{"text":" A frontier training run draws tens of gigawatt-hours over months. Meta reported the training of its largest 2024 model at around eleven thousand tonnes of carbon dioxide on the basis of where the electricity was generated, and zero on the basis of the renewable-energy certificates it had bought. Both numbers appear in the same report. The first is what the grid emitted; the second is an accounting convention. When a lab says its model was trained on clean energy, ask which of the two it means."}]},{"t":"p","runs":[{"text":"Where it happens.","b":true},{"text":" Electricity is not global; it is local. A data centre draws from one grid and one aquifer. Data centres used around a fifth of Ireland’s electricity in 2023 and around a quarter of the American state of Virginia’s; both have since restricted new connections. The strain is not on the world’s supply. It is on particular places that did not plan for it."}]},{"t":"h2","id":"water","runs":[{"text":"Water"}]},{"t":"p","runs":[{"text":"Data centres use water in two ways: directly, to cool the chips, and indirectly, through the water used to generate their electricity. One large cloud company’s direct water use rose by about a third in a single year, 2022, as its AI building began; another reported consuming over twenty billion litres in 2023, most of it for cooling. Cooling can be done with less water and more electricity, or more water and less electricity, and the choice is made per site. In a water-stressed city the trade is not a technical one."}]},{"t":"h2","id":"india","runs":[{"text":"India"}]},{"t":"p","runs":[{"text":"India’s data-centre capacity was around a gigawatt in 2024 and several times that was under construction, concentrated around Mumbai, Chennai, Hyderabad and the capital region. Two facts make the Indian case its own. The grid is around seventy per cent coal, so a query answered in India carries several times the carbon of one answered in a hydro-rich region: the same electrons, a different source. And several of the cities being built in are water-stressed, so the cooling choice above is a choice about municipal supply. The national compute programme from the first lesson of this module will draw from the same grids."}]},{"t":"h2","id":"what-the-industry-is-doing","runs":[{"text":"What the industry is doing"}]},{"t":"p","runs":[{"text":"Buying power. In 2024 one cloud company contracted to restart a closed nuclear reactor for its exclusive use; others signed for small reactors that do not yet exist, and for gas turbines that do. Several labs have said that electricity, not chips, is now the constraint on the next generation of models. The grid connections being sought in some regions exceed the region’s current capacity to build."}]},{"t":"h2","id":"how-to-read-an-energy-claim","runs":[{"text":"How to read an energy claim"}]},{"t":"p","runs":[{"text":"Per query, per day, or per year? Chip only, or building and grid? On-site water, or the water behind the power? Location-based carbon or market-based? And where: because the same query costs a different amount of coal in Mumbai and in Quebec. A claim that answers all five is rare and worth trusting. A claim that answers one is true about one thing."}]}],"check":{"question":[{"text":"A lab reports that a query uses a fraction of a millilitre of water, and a university study reports that a short conversation uses about half a litre. Why do the figures differ so much?"}],"options":[{"text":[{"text":"The university study was done on an older, less efficient model that needed far more cooling for the same answer."}],"correct":false,"tempting":[{"text":"Reaches for model efficiency; the gap is far larger than any efficiency change and comes from what was counted."}]},{"text":[{"text":"The lab counts only on-site cooling, while the study also counts the water behind the electricity."}],"correct":true,"tempting":null},{"text":[{"text":"The lab’s figure is per token and the study’s is per conversation."}],"correct":false,"tempting":[{"text":"Misplaces the unit; even scaled to the same conversation the on-site figure stays far below the study’s."}]},{"text":[{"text":"One of the two must be wrong, since water use is a physical fact."}],"correct":false,"tempting":[{"text":"Assumes a single boundary; both are physically correct measurements of different boundaries."}]}],"order":[3,0,2,1],"why":null},"prev":{"slug":"who-owns-what-it-makes","title":"Who owns what it makes","minutes":9},"next":{"slug":"the-safety-argument","title":"The safety argument, both sides fairly","minutes":10}}}