The short answer
AI uses real electricity and real water, and demand is growing quickly. But most of the per-question figures circulating are unreliable, including the flattering ones and the alarming ones. The honest position is that the direction is clear, the individual numbers mostly are not, and your personal chatbot use is a very small part of it.
Two different costs, often muddled together
Training a model is a large, one-off expenditure of energy. It happens once, takes months, and is the figure usually quoted in headlines.
Answering your question is a tiny cost by comparison — but it happens billions of times a day. Over a model's life, serving it generally uses far more energy in total than training it did.
Water enters because data centres are cooled, and some cooling evaporates fresh water. How much depends enormously on the site, the local climate and the cooling design, which is exactly why single global figures are so unsatisfying.
BEEP / THE MYTH-BUSTER
I have seen my own energy use described as trivial and as catastrophic in the same week, both with a confident number attached. Somebody is wrong, and it might be everybody.
Why the numbers you have seen are shaky
Providers publish very little. Most public per-query figures are estimates built on assumptions about hardware, model size and utilisation that the estimator could not verify.
The variables are enormous. A short answer from a small model and a long answer from a large reasoning model can differ by orders of magnitude, so any single average hides more than it reveals.
Location matters as much as technology. The same query on a grid running on wind has a fraction of the carbon of one on a grid burning coal. A global average silently blends the two.
Comparisons to everyday things — a kettle, a mile driven, a web search — are almost always doing this blending, and usually in whichever direction supports the article.
What is actually known
Data centre electricity demand is rising, and AI is a significant part of that rise. That much is well established by energy agencies, and it is the part worth taking seriously.
It is also true that the growth is concentrated in a small number of very large operators, several of which buy substantial renewable capacity. Both of those facts are true at once, which is why the argument never resolves.
What does not hold up is the idea that an individual asking a chatbot a few questions a day is making a meaningful difference either way.
Where the impact actually sits
- Large
- Training frontier models; running them at global scale; building the data centres.
- Moderate
- Heavy automated use — agents, bulk processing, anything looping unattended.
- Very small
- One person asking questions, drafting and summarising during a normal day.
The shape of where the energy goes, not a measurement. Credible per-query figures are not publicly available.
What to actually do
- Use a smaller model when a smaller model will do. This is the one lever with a real effect.
- Do not run reasoning models or agents on jobs that do not need them.
- Treat any confident per-question figure, in either direction, as an estimate with wide error bars.
- Ask providers for disclosure. Individual restraint matters less than published numbers.
The honest summary
The direction of travel is real and worth watching. The individual guilt is largely misplaced, and the precise numbers are not currently knowable from outside the companies.
Anyone giving you a confident figure for the footprint of a single question is telling you more about their argument than about the world.
ORBIT / THE NAVIGATOR
This is the section where I would like to give you a tidy number and cannot. Nobody outside those companies can. That is the finding, not a failure to research it.
See it for yourself
Reading about it only gets you so far. This one is written to show you the thing described above, rather than describe it again.
What actually uses the energy when I send you a message? Then tell me which parts of that answer are measured and which are estimates.
Then paste it into whichever assistant you use — ChatGPT, Claude, Gemini or Copilot. Not sure which? Meet them all.
If you remember nothing else
- Training is a big one-off cost; serving answers is small but happens constantly.
- Public per-query figures are estimates with very wide error bars.
- Where the data centre is matters as much as what the model is.
- Data centre demand is genuinely rising, and AI is a real part of that.
- Choosing a smaller model is the one individual choice that measurably helps.
Where this came from
- Nature, Light bulbs have energy ratings — so why can’t AI chatbots?checked 26 September 2026
- International Energy Agency, Energy and AIchecked 26 September 2026
Anything here that turns out to be wrong is worth telling us about. Figures that move — prices, sizes, energy use — are kept in the registry with their dates, rather than written into the prose where they would quietly go stale.
Questions about “What AI costs the planet”Asked by readers, answered by us. Open to read them.
Open to load them.
Ask about this page and we will answer it here, where the next person can read it too.
Nobody else can answer you, and you cannot answer anybody else. This is not a forum — we would rather say that than run a room we cannot watch at three in the morning.
Your question goes up with the answer and without your name. Nothing identifies you, ever.
If we cannot answer it we will tell you why, and that one stays between us.