AI Energy and Environment: Where the Electricity Goes

Every AI question costs real electricity and, indirectly, real water

Filter by key stage or audienceClick one or more chips to filter the sections below.

FAQ for teachers

Common questions teachers ask when running this lesson.

Why should a school lesson care about AI energy at all?

Because your pupils are already heavy AI users, the energy is invisible to them, and the infrastructure decisions made this decade will define the next thirty years of UK grid and water use. Schools that treat AI as free electricity produce citizens who cannot argue back when a hyperscaler proposes a data centre near their town. Pupils leaving class knowing that a prompt costs kilowatt-hours can hold that judgement later.

How much energy does a single ChatGPT query actually use?

The honest answer is that nobody outside the frontier labs knows precisely. Published estimates for a single text query cluster around 0.3 to 3 watt-hours, roughly ten times a Google search. Image and video generation sit an order of magnitude higher. Teach the range rather than a single figure; the pedagogical point is order of magnitude, not the second decimal. Cite the IEA's 2024 electricity outlook if asked.

How much water does training a large AI model actually consume?

Estimates vary widely, but a leaked internal figure for training GPT-4-class models sits around 700,000 litres of clean fresh water, mostly for cooling servers during the multi-week training run. Every user query then adds a small ongoing draw, roughly a shot glass of water per few dozen prompts. Water use is site-specific: hotter climates cool with more water, cooler climates cool with more air.

How do I explain data centres to KS1 without scaring them?

Call it 'the room where the questions live'. Show a photograph of a data-centre hall; there are plenty online, no logos needed. Tell the class that when they ask a talking computer a question, the answer is thought about in a room full of humming boxes, and those boxes get hot and need cooling like a poorly friend. Avoid climate framing at KS1; focus on wonder and scale, not blame.

Aren't the frontier labs already switching to renewable power?

They are buying renewable-energy certificates at scale, which is not the same as running on renewables. The certificates offset annual totals; the servers themselves draw from whichever grid they sit on, moment by moment. When Slough draws 500 MW at 6pm on a still winter evening, that power is gas. Google, Microsoft and Meta have all conceded their grid-carbon numbers have risen since 2023 as AI load outpaces clean supply.

Is it worse to use AI or to use Google search?

Per query, AI is worse: published estimates put a ChatGPT-style prompt at roughly ten times a Google search. But per task, the answer depends. If the AI reply saves you three follow-up searches, you may come out even. The honest teaching move is to reject the false binary. Show pupils that the real choice is not 'AI versus Google' but 'AI versus the cheapest tool that actually answers the question'.

What should a workforce learner in Care, Construction or Manufacturing take away?

Three things. First, AI in these sectors is being pitched as free efficiency; the electricity is real and someone pays for it, usually the employer via a subscription and the grid via load. Second, sector-specific AI has a bigger footprint than a chatbot because it runs constantly. Third, the frontline worker is often best placed to spot wasteful AI use, such as queries that run overnight or dashboards nobody reads.

Does training a model or running a model use more energy overall?

Running wins, and it is not close. Training a frontier model burns roughly the annual electricity of a small town, once. Inference, meaning every user query answered by that trained model, adds a fraction of that per query but happens billions of times per day, for years. Industry estimates now put inference at over 80 percent of an AI system's lifetime energy footprint, and rising.

What is the Jevons-paradox concern and is it real?

Jevons observed that as coal engines became more efficient, Britain used more coal, not less, because efficiency made the fuel cheaper to use everywhere. Applied to AI: every efficiency gain in model inference has so far been swallowed by rising query volume. Is it real for AI? The 2024 to 2026 numbers say yes: total AI energy use has grown roughly 40% year on year despite per-query improvements.

Is on-device AI genuinely more efficient than cloud AI?

For the query itself, usually yes: your phone runs a small model on a low-power chip and skips the round trip to Slough. For the fleet, it depends. Thirty million phones running local inference dwarf a single data-centre GPU but come from batteries charged from the same grid. The manufacturing footprint of an AI-capable phone is roughly 70kg CO2 per handset. On-device is a tool, not an absolution.

Common misconceptions

What pupils tend to think, and what to say back.

Pupils often say

AI is virtual, so it doesn't really use any energy.

It's actually

Every AI answer runs on a physical chip in a physical building, drawing electricity from the grid and cooling water from local supply. The 'cloud' is a shed in Berkshire, not a metaphor. If it truly used no energy, nobody would need to build new power stations to keep up with it.

Try asking

If it used no energy, why do the companies running these models sign contracts for entire power stations?

Pupils often say

One query is nothing, so it doesn't matter.

It's actually

One query is small; a billion queries a day is a small country's electricity demand. Individual AI use scales like a tap left running, unnoticed until you multiply it out. The point of the lesson is not the single prompt, it is the sum of everyone's single prompts.

Try asking

If a tap dripped once a second, how much water would you lose in a year?

Pupils often say

Renewable-powered data centres are carbon-free.

It's actually

Very few data centres are 100 percent renewable in real time. Most match their annual use with renewable certificates bought elsewhere, while still pulling coal or gas from the local grid when the wind drops and the sun sets. The certificate is an accounting move, not a physical fact.

Try asking

What is the difference between 'powered by renewables' and 'matched annually with renewable certificates'?

Pupils often say

The water at a data centre is only for cooling, so it goes back.

It's actually

Most cooling systems evaporate the water into the air. It leaves the local supply as vapour and does not return to the same reservoir. In a dry UK summer, that pipe of drinking-quality water is in direct competition with the same one filling household kettles.

Try asking

Where does the steam from your kettle end up, and how easy would it be to get it back into the tap?

Pupils often say

Training is the only expensive part; using it is basically free.

It's actually

Training is a one-off spike. Inference, meaning every reply the model gives for the rest of its life, is a permanent tap that has now overtaken training in total energy at company scale. Once a model is live, most of its lifetime footprint comes from answering, not from being built.

Try asking

Which uses more petrol over its lifetime, building the car or driving it?

Pupils often say

Small local models are always greener.

It's actually

Running a model on your own device usually saves the data-centre trip, which helps. But if pupils use it ten times more often because it feels free, or if they still fall back to the big cloud model whenever a task is hard, the total footprint can go up, not down.

Try asking

When does making something cheaper accidentally make people use much more of it?

Pupils often say

AI companies would tell us if it was bad for the planet.

It's actually

Reporting standards for AI energy and water use are voluntary, patchy and often years behind. Several of the biggest firms have quietly raised their annual emissions figures since launching mainstream chatbots. What we know about the real cost, we know because researchers and journalists dug for it.

Try asking

Who has an incentive to publish a number that looks bad, and who does not?

Pupils often say

Streaming a video is way worse than asking an AI, so I might as well ask.

It's actually

A minute of high-definition video is roughly the same electricity as a long chatbot answer, and much less than one AI-generated image. Video is not the villain many people assume, and AI is not the bargain many people assume. They sit in the same order of magnitude, so the choice between them is a real one.

Try asking

What would you actually need to measure to compare them fairly?

Pupils often say

If I close the chatbot tab, that saves the model energy.

It's actually

The heavy work happens in the second the model generates the answer, not while the tab sits open afterwards. Closing the tab saves your own device a little screen and processor power; it does not switch anything off in the data centre. The saving is on your side, not theirs.

Try asking

If you unplug your headphones, does the song stop playing at the radio station?

Pupils often say

The environmental cost of AI is exaggerated by activists.

It's actually

The figures used in this lesson come mostly from the AI companies' own sustainability reports, the International Energy Agency, and peer-reviewed academic papers. The direction of every recent revision has been the same: higher than last year's estimate, not lower.

Try asking

Whose interests are served by the published number being too low?

5-minute prep

Five ready-to-run ways to open this lesson. Pick one, copy the prompt, paste it into ChatGPT or Copilot.

Starter5-8 min

Ask the AI how much it just cost

The fastest way to make AI energy real. Ask the model to price its own reply and translate it into a kettle you can picture.

I'd like to show a class what one of your answers actually costs in electricity. Please estimate the energy used to generate a single, average-length reply from you, given as a single number in watt-hours. Then convert that number into two everyday equivalents a UK classroom will feel: how many seconds a standard 3 kilowatt kettle would boil for on the same electricity, and how many LED bedroom lamps you could run for one full minute. Present the answer as three short lines the teacher can read straight off the screen, and finish with one sentence naming the single biggest source of uncertainty in your estimate.

Compare8-10 min

Cloud brain versus pocket brain

Same request, two very different places to run it. Line up a big cloud model against a small on-device one and watch the energy gap open.

Please compare, on the same simple request, the energy cost of two answers: one from a large cloud-based chatbot running on GPUs in a data centre in Slough, and one from a small on-device model running on a modern smartphone chip. Assume the request is 'write a two-sentence definition of photosynthesis'. Give a single watt-hour figure for each, express the ratio between them in plain English, and finish with one line about which kinds of task should always be sent to the cloud and which should stay on the device. Please do not hedge; commit to a best-estimate number.

Explain10-12 min

Why is asking cheaper than teaching?

Get the AI to explain the training-versus-inference gap in plain classroom English, and to say honestly where the cooling water actually ends up.

Imagine you are talking to a curious class of pupils who have just found out that AI uses electricity and water. Please explain, in three short paragraphs of plain classroom English, why the one-off electricity cost of TRAINING a large model is enormous, while the cost of a single reply (INFERENCE) is small in comparison but now dwarfs training in total. Cover: (1) what training actually does with all that electricity, (2) why one reply is cheap but the total is not, (3) where the cooling water at a UK data centre physically goes after use. Do not apologise or hedge, and give one concrete UK figure the class can hold on to.

Check12-15 min

A week of prompts, in kettles

Turn a real class's honest AI use into a weekly energy budget on the board. Numbers pupils recognise from their own hands.

I'm running a classroom exercise turning our own AI use into a weekly energy budget. Please build a small table for a class of 30 pupils each sending 5 chatbot prompts per school day at roughly 3 watt-hours per prompt, over one school week. Give the daily total in watt-hours, the weekly total, the equivalent number of full UK kettle boils (assume 100 watt-hours per boil), and the equivalent litres of data-centre cooling water at 0.5 litres per prompt. Finish with one line naming the single biggest lever the class could pull to cut the weekly figure in half. Use British sources and figures where possible.

Repair12-18 min

Write our energy-lean prompting guide

The advanced move. Ask the AI to co-author the rules the class will live by for a term, then judge which rules will actually change habits.

Please write a one-page 'energy-lean prompting' guide a class could adopt for the rest of the term. Structure it as seven numbered rules, each one sentence long, that measurably reduce the electricity and water cost of the class's AI use without losing the learning benefit. Cover: prompt length, requested answer length, one clear ask versus long back-and-forth, choosing between a search, a dictionary, a calculator and a chatbot, when a small on-device model is enough, when to batch questions, and when not to ask at all. Finish by naming the two rules you predict pupils will find hardest to follow, and why.