The short version: training a model has a huge environmental cost compared with a single prompt. But inference, actually using the models, adds up at today’s scale, and over a model’s lifetime it can outweigh the training.
I first wrote this up for work, to get my thoughts straight on AI, energy and the environment. This version adds a calculator so you can try out your own usage.
Two easy-to-read pieces from 2025 shaped most of my thinking: MIT’s 
Explained: Generative AI’s environmental impact and Epoch AI’s 
How much energy does ChatGPT use?.
Training uses a huge amount of energy
MIT cites an estimate of 1,287 MWh to train GPT-3, producing around 552 tonnes of CO₂. That’s the yearly electricity of about 120 US homes, or nearly 500 UK ones.
Newer models are far bigger. Epoch’s data suggests the largest training runs now draw over 100 MW. A run that size lasting a few months would use something like 200 GWh, enough for around 75,000 UK homes for a year.
How much energy one prompt uses
Epoch estimates a typical GPT-4o query uses around 0.3 Wh, about a 10W LED bulb running for two minutes.
Google has since 
measured its own: about 0.24 Wh for a typical Gemini text prompt, counting idle machines and cooling as well as the chips. It also says that figure fell 33 times over the year to May 2025. Nobody outside Google has checked the numbers, but they land close to Epoch’s.
Energy per query climbs quickly with longer input, because the model has to work through everything you give it. Input is measured in tokens, chunks of text that average out at roughly three quarters of a word each:
| Query | Energy | Same as a 10W LED for |
|---|---|---|
| Short chat | 0.3 Wh | 2 minutes |
| ~10k tokens in, a short paper | 2.5 Wh | 15 minutes |
| ~100k tokens in, about 200 pages | 40 Wh | 4 hours |
Reasoning models and agents push it further. In Epoch’s testing, o1 produced around 2.5 times as many tokens as GPT-4o for the same question, and an agent can make dozens of calls for one request.
Inference never stops
Training is a one-off. Inference runs continuously across millions of users for as long as the model is in service, and MIT notes it’s a growing share of AI’s energy use.
In July 2025, OpenAI told Axios that ChatGPT was handling about 2.5 billion prompts a day (
via TechCrunch). Even if every one were a short chat, that’s as much electricity as training GPT-3 roughly every two days.
Try it yourself
Set how many prompts a day, how heavy they are and whose grid they run on. The presets go from light use up to the whole of ChatGPT.
- a 10W bulb left on all year: 87.6 kWh a year
- a typical UK home for a year: 2.7 MWh a year
- training GPT-3: 1.29 GWh a year
- a 100 MW training run lasting a few months: 200 GWh a year
- every data centre in the world, 2022: 460 TWh a year
100 short chat prompts a day on the US grid: 10.9 kWh and 4.38 kg of CO₂ a year.
How it's worked out
- Energy per prompt from Epoch AI. Short chats only; reasoning and agents use more.
- CO₂ per kWh from Ember's 2025 data, with the US's 384 g rounded to 400.
- A kettle boil is 0.15 kWh, a UK home 2,700 kWh a year, a petrol car 170 g of CO₂ per km.
- Cooling water is MIT's estimate of about 2 litres per kWh.
- None of it includes building the hardware or the data centres.
Electricity isn’t the whole cost
MIT also covers what a per-prompt figure misses. Cooling uses roughly two litres of water for every kWh. Making GPUs and building data centres has its own footprint too, and there’s more hardware every year: GPU shipments to data centres rose from 2.67 million in 2022 to 3.85 million in 2023.
Where the power comes from matters as well. Going by 
Ember’s grid data, the same kWh on France’s mostly nuclear grid produces about a tenth of the CO₂ it would in the US.
MIT’s early-2025 article also cites a projection that data centres could use around 1,050 TWh in 2026, which is this year. That now looks far too high. The IEA’s 
2026 report puts 2025 at about 485 TWh, up 17% on the year before, and doesn’t expect data centres to reach 950 TWh until 2030. The AI-focused part is growing much faster, though: up 50% in 2025.
My takeaway
Total use is still climbing, but I only see the cost per task going down: chips, models and data centres keep getting more efficient, and cleaner grids make each kWh cost less CO₂.
So it isn’t something I go out of my way to cut back on. AI is a big part of my day-to-day work now, and I think it’s worth the energy it uses. Most people don’t weigh up the carbon cost of jumping on a flight for a holiday, and even if you do, you’re nowhere near the worst offender (looking at you, Taylor Swift ✈️).
It does have an impact, and it’s worth making good choices for the environment overall. But I think the “AI is bad for the environment” story is overblown. A year of 100 short prompts a day comes out at about the same CO₂ as driving 26 km. Oil spills, single-use plastics and companies like Shell do far more damage.
A lot of this gets framed as personal responsibility, but what one person does with their prompts barely moves the needle. The big companies and governments are the ones who can actually fix it.