Personal research — AI & the environment
AI usage jumped fast enough that the infrastructure under it is now the story. What's actually true about the cost, what's an exaggerated meme, and what's actually being done about it.
grounded against the IEA's 2026 energy-and-AI analysis
Global data centre electricity consumption is on track to roughly double from 485 TWh in 2025 to about 945 TWh by 2030 — around 3% of all global electricity demand by then. That alone would be a big number. What makes it an AI story specifically: the servers doing AI work inside those data centres are growing far faster than the data centres themselves.
That gap between 15% and 30% is the actual headline: data centres in general are growing fast, but the AI-specific slice inside them is growing about twice as fast as that. It's not "computing got bigger" — it's "AI got bigger inside computing."
a 2026 snapshot, to make the scale concrete
In 2026 alone, the IEA projects roughly 260 TWh of data-centre electricity demand in the US and around 150 TWh in Europe. For a sense of scale: that US figure alone is in the same ballpark as the annual electricity consumption of a mid-sized country. The IEA's longer view has global data-centre demand reaching 1,200 TWh by 2035 — meaning the growth doesn't level off at 2030, it keeps compounding.
The honest uncertainty here: these are all forecasts built on today's adoption curve and today's hardware efficiency. They can come in lower if efficiency gains outpace demand, or higher if AI adoption keeps accelerating past what's already priced in. Nobody serious is claiming false precision on the exact number — the direction and rough magnitude are what's solid.
the claims that go viral, checked against the actual research behind them
| The claim | What's actually true |
|---|---|
| "One ChatGPT query wastes a whole bottle of water" | This is a misread of a real 2023 peer-reviewed estimate (Ren et al.), which said ~500mL per 10 to 50 medium-length responses — not per query. A single request is closer to single-digit millilitres, roughly 1% of a bottle. The researcher behind the original figure now puts a GPT-4-class prompt near 15mL full-lifecycle, about 5mL on-site. Your one prompt is a rounding error. The data centre serving millions of them is not. |
| "More efficient chips and models will fix this on their own" | Efficiency gains are real — better cooling design alone can cut cooling energy by up to 50%. But historically, efficiency gains in computing have been consistently outpaced by growth in how much of it people do (the same pattern economists call the Jevons paradox). Data centre electricity use is still climbing ~15% a year despite constant hardware efficiency improvements — efficiency has never been the whole answer by itself. |
| "Nobody's tracking or regulating any of this" | Partly true, changing fast. The EU AI Act now legally requires general-purpose AI providers to document known or estimated energy consumption in kWh, with fines up to €15 million or 3% of global turnover for non-compliance. It's a real, enforceable floor — but it has a real gap too: it doesn't yet account for grid carbon intensity, so a provider running on coal-heavy power and one running on hydro can report the same kWh figure and look identical on paper. |
None of these fully solve the growth problem on their own — see the efficiency-paradox point above — but each is a real, currently-deployed lever, not a hypothetical:
Hyperscalers are increasingly buying firm, carbon-free power directly. Adoption of nuclear as a data-centre power source has jumped from 11% of surveyed operators three years ago to 33% today.
Firm power, not intermittent — solves the "AI runs 24/7, solar doesn't" mismatchCooling is often as large a factor in a data centre's footprint as the compute itself. Newer direct-to-chip and immersion cooling designs are cutting that specific slice significantly.
Up to 50% reduction in cooling energy specificallyNot every task needs a frontier-scale model. Smaller, distilled, or task-specific models trade some raw capability for a real, direct cut in the energy cost of serving each request.
Cuts the "per query" number directly, not just the grid mix behind itThe EU AI Act's kWh-reporting requirement is the first real legal forcing-function for providers to know and publish their own number, rather than it being optional PR.
Turns "trust us" into an auditable, penalized-if-wrong figureOutside of climate circles, "responsible AI" usually gets used to mean bias, safety, or misuse. The environmental half of that word gets talked about far less — but it's now a real, binding compliance category, not just an ethics talking point. The EU AI Act's general-purpose-model providers now have a legal deadline (August 2027 for models that existed before August 2025) to document their energy consumption, with real financial penalties behind it.
The gap worth watching: disclosure isn't the same as accountability. A published kWh figure tells you the size of the bill, not whether it was paid with coal or hydro. The European Commission has already opened consultation on tightening this — expect the "known or estimated kWh" requirement to get more specific, not less, over the next few years.
For anyone actually building with AI day to day — which, working on this stack, is most of what I do — the practical version of "responsible" is less about individual guilt over asking a question, and more about the choices that actually move the number: picking a right-sized model instead of the biggest one by default, and caring which provider's disclosed number you're contributing to.
The individual query was never the problem. The infrastructure being built to serve billions of them, as fast as possible, is — and that's exactly the part that's measurable, regulatable, and still being decided right now.
The honest throughline — researched 2026-09-16, figures from the IEA and the EU AI Act, re-verify before citing further out