01 / The Scale
One prompt is tiny. Billions are not.
Start with the number that surprises most people: a typical AI text prompt uses roughly a quarter of a watt-hour of electricity — about the same as a web search, and far less than early estimates claimed. If you've felt guilty about asking a chatbot a question, the honest answer is that the individual act is small.
The problem is aggregation. Data centres consumed about 1.5% of the world's electricity in 2024, and the International Energy Agency projects that figure roughly doubling by 2030 — to more electricity than all of Japan uses today — with AI the main driver. The electricity used by AI-focused data centres grew about 50% in 2025 alone. No single prompt matters much; the sum of all of them is reshaping power grids on three continents.
02 / Carbon
The tech giants' own numbers tell the story.
You don't need critics to make the carbon case — the companies' own environmental reports do it. Google has disclosed that its emissions rose 48% between 2019 and 2023, and they have kept rising since; Microsoft's grew roughly a quarter since 2020. Both point to the same cause: building and powering AI infrastructure, and manufacturing the hardware inside it.
The build-out is also changing what powers the grid. In parts of the United States, gas plants are staying online and coal retirements have been delayed to feed data-centre demand. The flip side: AI has triggered the biggest wave of nuclear investment in decades, including the restart of the Three Mile Island reactor under a twenty-year deal with Microsoft. Whether the race to power AI speeds up clean energy or entrenches fossil fuels is still being decided.
03 / Water
The hidden cost is wet.
Data centres drink — directly, to cool the servers, and indirectly, through the power plants that feed them. You may have heard that one AI conversation "costs a bottle of water." That early research estimate has since been revised sharply downward for modern systems — current figures range from a few drops to a few millilitres per prompt, depending on what's counted. Once again, the individual act is small.
But once again, the aggregate isn't. Google's data centres consumed around eight billion gallons in 2024. A single large facility can use millions of gallons a day — the draw of a small city — and roughly one in five American data centres sits in a water-stressed region. That's why the sharpest conflicts over AI have been local water fights: a town in Oregon where Google's use grew to roughly a third of the city's supply, and protests in Uruguay, during the country's worst drought in decades, under the slogan "It's not drought, it's pillage."
04 / Where It Concentrates
The damage has addresses.
AI's environmental cost isn't spread evenly across the world — it piles up in particular places, at particular stages of the supply chain. These are the ones we kept returning to:
Canada is joining this map quickly: about 90% of planned Canadian data-centre capacity targets Alberta — much of it in high-water-stress areas — while Quebec is rationing its hydro and roughly doubling data-centre power rates, and Ontario passed legislation to screen which projects get grid connections.
05 / Music
What about making a song?
Generating a track with an AI music tool is a bounded cost — seconds to minutes of computing — comparable in spirit to the prompt numbers above; though the environmental cost may be small per individual act, when you consider the number of people on each platform doing thousands, even millions of prompts a day, it quickly adds up. And, unlike other AI platforms, nobody outside the companies knows the real figures. Suno and Udio, used by millions, publish no environmental data at all, and no independent study has measured a per-track footprint. That silence is worth noticing.
06 / The Other Side
Hope, caution, and paradox — with a grain of salt.
The energy per AI prompt has fallen roughly tenfold since 2023; Google reports its per-prompt energy dropped 33-fold in a single recent year; and machine learning cut the cooling energy of Google's own data centres by up to 40%. AI can also be useful for the environment — forecasting renewables, optimizing grids, modelling climate, discovering better battery materials.
But two cautions keep the ledger honest. First, an old economic pattern called the Jevons paradox: when a technology gets more efficient, we don't use less of it — we use more, and total consumption rises anyway. This can be illustrated perfectly in LED lights — when they were first invented, they promised a world with inexpensive, high-output, low-consumption lights that last thousands of times longer than conventional lightbulbs. However, what actually happened is that we started using LEDs frivolously, and any gains that could have been made were lost because we started using LED lights everywhere — yes, in households and offices, but also in toys, shoes, and countless throwaway objects. Every efficiency gain so far has been outrun by growth. As well, the industry's green claims deserve scrutiny: a 2026 independent review of 154 such claims found that three-quarters were unproven, and could not identify a single verified case of consumer AI tools producing substantial emissions cuts. The potential is real; the proof, so far, is thin.
07 / What Could Fix It
The solutions exist. The pressure doesn't — yet.
None of this is unsolvable, and the fixes are known. Data centres can be required to run on genuinely clean power around the clock, rather than buying paper credits — Ireland now demands new facilities bring their own. Cooling can go liquid and closed-loop instead of evaporating a town's water. Models can be smaller: routing everyday questions to lightweight systems instead of frontier giants is the single easiest saving available, and it's already happening. Siting rules can keep facilities out of water-stressed regions, and honest pricing can make companies pay what power and water actually cost.
The one fix underneath all the others is transparency. Nearly everything on this page had to be prised out of companies by researchers, journalists, and lawsuits. Standardized, mandatory disclosure — how much energy, how much water, from which grid, per facility and per prompt — would let the public see the real costs and reward the builders who shrink them. It's hard to fix what no one is allowed to measure.
The individual prompt is nearly free. The system behind it is not. Knowing the difference — and asking who pays, and where — is what it means to use these tools with your eyes open.