AI Energy at Work
What each prompt actually costs, and the habits that shrink your organisation's footprint
The bottle of water hiding in every prompt
Before we say anything, let's watch the range live.
Pick one of the prompts below. Copy it. Paste it into whichever AI your team already uses (ChatGPT, Claude, Gemini, Copilot). Then run the next one. Then the next. Notice how long each response takes to come back, and how many words it produces.
Every one of those seconds is a data centre working. Every one of those words is a token, and every token has a small but real energy cost attached to it.
You just asked the AI to write a haiku, then a novel opening, then a spreadsheet analysis. The screen didn't hint at it, but each of those took a very different bite out of a data centre somewhere, and out of the water tank keeping that data centre cool.
For most of us the electricity bill and the water bill are somebody else's problem. They arrive quietly on the back of a monthly invoice, or they don't arrive at all because the AI vendor is absorbing them for now. That invisibility is the whole issue. It's very hard to be careful with a cost you never see.
Over the next hour we'll make the cost visible. We'll look at where the energy actually goes, how big one query really is, how it multiplies when a whole workforce reaches for AI on autopilot, and the small handful of habits that will meaningfully shrink your organisation's AI footprint without anyone having to give the tools up.
Write me a haiku about a Monday morning meeting.
Write the opening chapter of a novel set in a small English seaside town in November. Aim for about 800 words, with strong atmosphere and one main character introduced by the end.
I'm going to describe a fictional quarterly sales spreadsheet. Analyse the trends I mention, flag anything unusual, suggest three follow-up questions, and give me a 200-word summary suitable for a board pack. Quarter 1: Product A sold 12,000 units at an average price of £49, Product B sold 4,200 units at £128, Product C sold 900 units at £410. Quarter 2: A dropped to 9,500, B climbed to 5,800, C stayed flat. Quarter 3: A rebounded to 13,200, B fell to 3,100, C jumped to 1,400. Quarter 4: A ended at 14,000, B recovered to 5,600, C settled at 1,150.
Look at the three responses side by side and ask the room:
- Roughly how long did each response take to come back? Which was noticeably slower?
- Which one produced the most text? Roughly how many words?
- If your team ran the equivalent of Prompt C a thousand times a month, and Prompt A ten thousand times a month, which one is quietly dominating your organisation's AI footprint?
- Where does the extra work happen? On your laptop, or somewhere else?
Everyone in the room opens their AI of choice. On a signal, half the room runs a one-line prompt ("Give me a synonym for 'important'"), half runs a heavy one ("Analyse this 500-word memo and rewrite it in three different tones for three different audiences"). Compare how long each side waited. That gap is roughly proportional to the energy each request used. Now imagine those two habits scaled across 400 people, every working day for a year.
Where the energy actually goes
When the news talks about AI energy costs, two very different things usually get lumped together. It helps a lot to keep them separate.
The first cost is training. A large language model like the ones powering ChatGPT or Gemini learns by being run over vast amounts of text, tuning billions of internal parameters as it goes. Public estimates for training GPT-3 back in 2020 put the electricity bill at around 1,300 megawatt-hours, roughly what a small town uses in a month. Newer frontier models are bigger and hungrier. This is a one-off cost per model version, but a very large one.
The second cost is inference, which is the technical word for what happens every time someone types a prompt and presses enter. Compared to training, one query is tiny. But there are a lot of queries. Every request from every user, all day, every day. The daily inference total across a large AI provider now quietly rivals what a small country uses.
Both matter. But for you as a user or a team lead, it is inference you can actually influence. Training happens whether you sit at the keyboard or not. Inference happens because someone asked.
On a whiteboard, write training on the left and inference on the right. Ask the room where each of these belongs: the electricity bill for the data centre where the model was built; the cost of your daily prompts; the water used to cool the GPUs running the model right now; the carbon emissions reported once at model release; the vendor's monthly bill to you. The point is to notice that some costs happen once, some happen every single time, and only the second set is affected by how your people work.
How much is one query, really?
One query on its own is a hard thing to picture. It helps to compare it with things we have been doing for years without thinking twice.
The numbers in the cards below are ballpark estimates, not audited figures. They shift with the model you're using, the length of your prompt, and which data centre serves you. But they are accurate enough to build a working mental model on.
Before you tap the cards, have a guess. If a Google search were a candle, what would a ChatGPT query look like? A torch? A room light? A bonfire?
Tap each card to see the rough comparison.
A Google search
Roughly 0.3 watt-hours per search, based on Google's own published figures. That is about what it takes to power an LED bulb for a couple of minutes. This is our baseline.
A short ChatGPT text query
Independent estimates put a typical short exchange at roughly ten times a single Google search. Newer, better-tuned models will be cheaper than the earlier figures, but the ballpark still sits well above a search.
Generating an AI image
Peer-reviewed work has found one image generation can use roughly as much electricity as fully charging a smartphone. Video generation is dramatically higher again, several orders of magnitude above a text reply.
A long, context-heavy analysis
Pasting a whole document, PDF, or spreadsheet into a chat can multiply the cost by ten or more compared with a short prompt, because every token in the input has to be processed before the AI even starts writing.
None of these are a scandal on their own. It becomes a scandal only when we forget the scale, which is the next slide.
Ask each pair to guess, in advance, how many Google searches equal one ChatGPT query. Write the guesses on the board. Reveal the rough answer (roughly ten). Then ask a harder one: how many text queries equal one AI-generated image? Guesses will span three orders of magnitude. The reveal (often hundreds, sometimes thousands) tends to reset how people reach for the image tool casually for the rest of the session.
The water side
The electricity story gets most of the airtime. Water sits quietly next to it and matters just as much.
Modern data centres run hot. Racks of GPUs training and serving AI models generate enormous amounts of heat, and that heat has to go somewhere. In many facilities the answer is evaporative cooling: water is pumped through the building, absorbs heat, and evaporates into the air. Once evaporated, it is gone from the local supply. The data centre needs more.
Independent estimates suggest that a short conversation with a large AI model, perhaps twenty or thirty exchanges, is in the ballpark of half a litre of water once you count both the direct cooling and the water used upstream at the power station. Training the model that powers those conversations in the first place has been estimated at hundreds of thousands of litres. And in regions where water is already stressed, this stops being a metaphor and starts appearing on local council agendas.
Ask the room to name the AI tools their organisation uses. For each one, look up (or estimate) where the vendor's main data centres are physically located. Now cross-check that against a global water-stress map (WRI Aqueduct is a free public one). The result will surprise people. Cheap electricity and cheap land often mean data centres end up exactly where the local water table is already under pressure. That is a conversation worth taking to a procurement meeting.
The workforce multiplier
One query is easy to dismiss. It is roughly a light-bulb-minute of electricity and a mouthful of water. Nobody sensible loses sleep over that.
Now stretch it out. A workforce of 200 people, each asking the AI 30 questions across a working day, is 6,000 prompts a day. Add casual image generation and the occasional big-document analysis and the daily total balloons.
At the organisation level this shows up in three places. First on the vendor's invoice, because most AI providers now bill per token used. Second in the sustainability report that an increasing number of clients and regulators are asking for. Third in the wider infrastructure story, because every extra query is a small vote for building more data centres.
The point of this slide is not to make anyone feel guilty about typing into ChatGPT. It is to notice that a small individual choice, made 6,000 times a day across your organisation, is not really a small choice any more.
In small groups, take five real tasks from your team's actual week and place each one into one of three buckets: worth a full AI query, a smaller model would do, or do it without AI at all. Push each group to justify the two hardest calls to the room. The rule of thumb: use AI where it saves real thinking time, use a smaller model when the task is well-defined, and skip it entirely for anything a two-second manual action would fix.
Four habits that shrink your footprint
Personal habits matter here in a way they rarely do with other environmental costs, because your keyboard is where every AI query starts.
Before you tap the cards below, have a think. If you were coaching a junior colleague to use AI more responsibly, what would you tell them?
Tap each card to reveal.
Right-size the model
The biggest, slowest, most expensive model is rarely the right tool for everyday work. Most vendors now offer a smaller, faster, cheaper tier. For summarising a memo or rewriting a paragraph, that tier is plenty. Save the frontier model for the tasks that actually need it.
Write tighter prompts
Every token you send and every token the AI returns costs energy. A precise, well-structured prompt gets a better answer in fewer words. Cutting your average response from 400 tokens to 150 is a real saving when you do it thousands of times a month.
Batch and reuse
If you are going to ask the AI ten similar questions, ask them in one prompt with a numbered list. If a colleague already asked something you also need, ask them for the answer instead of running it again. Shared knowledge is free.
Notice what does not need AI
Not every task is an AI task. A calculator is thousands of times more efficient at maths. A spellchecker is thousands of times more efficient at fixing typos. Use the right tool for the job and save the AI for the jobs that genuinely need it.
None of these habits ask you to stop using AI. They just push you to use it deliberately, and at the right size.
Ask each person in the room to name one AI habit they've picked up in the last month that they would now like to change. Put the four habit cards on a wall. Ask them to walk over and stand next to the one that would fix theirs. It's a surprisingly good way to surface which habit the team is weakest on, which tells you where to focus any internal training you commission next.
Bad prompt, good prompt
Habits are a good start. Prompt hygiene is where they become concrete.
The two prompts below ask for the same thing: a short internal summary of a policy change aimed at line managers. Both work. One uses roughly four times the energy of the other, mostly because the padding forces the AI to produce a much longer, less useful response.
Hi there, hope you're doing well today. I was wondering if you might be so kind as to help me put together a really lovely, warm, comprehensive summary of the new remote-working policy for our line managers to read, ideally something that covers all the key points in a clear and friendly way and is easy for them to skim on a busy morning. Thanks so much in advance for your help with this, really appreciate it.
Around 90 tokens on the input, almost all of it padding. The AI mirrors the register back at length, so you tend to get a 500-token chatty response as well. Total: roughly 600 tokens per exchange. Run this 20 times a day across a team of 50 and the difference stops being trivial.
Summarise the new remote-working policy for line managers. 4 bullet points. Skimmable in 30 seconds. Include one line on the manager's own action.
Around 30 tokens on the input, and it produces roughly 120 tokens back because the constraints are tight. Same job done, roughly a quarter of the energy. Also easier for the reader, because the shape is fixed.
In pairs, take the following bloated prompt and cut it under 30 tokens without losing intent. Test the best rewrite live against your usual AI tool.
> Please could you kindly, when you have a moment, help me draft a comprehensive and thoughtful summary of the main risks in this quarter's financial report, ideally something that would be suitable for our senior leadership team to review at their next meeting?
Good rewrites will name the audience, the format, and the length. Bonus points for anything that also caps the length of the response.
So what would you do?
Time to put it into practice. You're a middle manager at a mid-sized professional services firm. The leadership team has just rolled out a paid ChatGPT licence for the whole staff of 400 people and encouraged everyone to use it liberally.
Two weeks in, someone in Finance flags that the AI vendor bill for the pilot fortnight was three times the forecast. Someone in Sustainability separately asks what this means for the firm's carbon reporting. Someone in HR worries that people will feel policed if you introduce limits.
You've been asked to write a one-page recommendation. You have the habits, the prompt trick, and a rough sense of the scale. You need a plan.
Read out the four options below. Ask everyone to vote by hand. Then pick three people to justify their choice: one for the strongest option, one for a tempting-but-weak option, one for whatever the room actually voted for. Push each person to name which habit or trick they are leaning on. Wrap by asking the room to vote again after hearing each other, not before.