A number of people I know are uneasy about using AI. Not about what it writes. About what it costs to run.

They have made environmental commitments they take seriously. Net zero targets, B Corp status, a sustainability page on the website that means something. And they have read the headlines about data centres, water use and power demand. So every time they type a prompt, a small question sits in the back of their mind: am I quietly undoing the work?

I have been digging into the numbers. The honest answer is more interesting than the headlines, and it splits into two halves that point in opposite directions.

What your own use actually costs

Start with the part you control: your prompts.

For years the figures on this were guesswork, and most of the viral ones were wrong. The widely repeated claim that a short conversation with ChatGPT uses a bottle of water came from a 2023 academic estimate based on older systems and a broader definition of water use. It estimated roughly one 500ml bottle for between 10 and 50 responses, not for every question. Then in August 2025 Google published one of the most detailed provider disclosures yet: measured figures for its own service, covering electricity, water and carbon.

A typical text prompt to Gemini uses about 0.24 watt hours of electricity and about 0.26 millilitres of water. That is nine seconds of watching television, and five drops of water. Google also put the carbon at 0.03 grams of CO₂e per prompt, though that figure leans heavily on how the company buys and accounts for its energy. Sam Altman has said the average ChatGPT query uses around 0.34 watt hours, although OpenAI has not published the same level of supporting methodology.

Put that into business terms. Say someone on your team uses AI heavily: 50 ordinary text prompts on each of 250 working days. That is roughly 3 kilowatt hours of electricity a year. Boiling a kettle uses about a tenth of a kilowatt hour, so a year of heavy AI use costs about the same electricity as thirty kettle boils. The office kettle gets through that in a week. That is a useful benchmark for ordinary text use, not a universal figure: long reasoning tasks, image generation and agentic work can require considerably more computing power.

One more thing worth knowing: Google also reported that the energy used by its median Gemini prompt had fallen 33-fold in just twelve months. That is how quickly these systems are changing, and why any per-prompt statistic needs a model, a date and a methodology attached to it.

One prompt is small. Scale is a different beast

So why the alarming headlines? Because a single prompt is not the same thing as an industry serving billions of them.

Training frontier models requires enormous, concentrated bursts of computing power: the biggest training runs now draw over 100 megawatts of power for months at a time. Running those models for hundreds of millions of people creates a different but continuous demand. Add the two together, along with cooling, networking and the manufacture of specialist chips, and the industry-level picture looks very different from the footprint of one person typing into a chatbot.

And the industry is scaling fast. Data centres of all kinds, not just AI, used about 1.5% of the world's electricity in 2024. The International Energy Agency projects that to reach just under 3% by 2030, roughly what the whole of Japan uses today, with AI a major driver of the growth. Researchers at Epoch AI project that the largest individual training runs could need between 4 and 16 gigawatts by 2030. That is a projection, not a certainty, but it is power station territory, for one model.

Both halves are true at once. Your use is a rounding error. The industry's use is not.

The argument on our doorstep

This is not an abstract global story. It is now a live Scottish one.

Scotland generates renewable electricity equivalent to more than its entire electricity consumption, and that is exactly why data centre operators want to build here. Proposals include a 300 megawatt facility near Falkirk, a 600 megawatt site in Fife, and a potential 500 megawatt campus in North Lanarkshire's AI Growth Zone. The SNP motion cited 24 proposed hyperscale projects in total and argued that, built to their stated capacities, they could require more than one and a half times Scotland's peak electricity demand.

That has forced a real political fight. In July the SNP's National Council passed a motion calling for a temporary moratorium on new data centres, citing pressure on the grid, water and climate targets. The Scottish Government said no to a formal moratorium but promised national planning guidance by Christmas, and in September MSPs, its own among them, backed a motion saying no planning decisions should be made on hyperscale data centres until that guidance is published. Since 17 September every new application for a data centre over 50 megawatts has also needed a full environmental impact assessment. Westminster, meanwhile, is actively courting data centre investment as economic growth. Holyrood and Westminster are pulling in opposite directions, and the argument will run for years.

I don't think there is an obviously right answer here. Where clean generation and grid connections are constrained, a large new load competes with the capacity needed to decarbonise transport, heating and industry. The counterargument is that data centre investment can fund new generation and grid upgrades. The North Lanarkshire proposal, for instance, is pitched alongside over a gigawatt of privately wired renewables. Whether that generation is genuinely additional is exactly the question worth asking of each project. And the jobs, investment and digital capacity are real too. What I do think is that this decision deserves more attention than it is getting, because it will shape where Scotland's clean power goes for a generation.

What to say when someone asks

If a client, a board member or a B Corp assessor asks about your firm's AI footprint, you can now answer honestly and specifically. Three statements, all defensible:

For ordinary text use, our direct footprint is measured in kettle boils per year, not tonnes of carbon: around 3 kilowatt hours for a heavy user, on the best published figures. We don't treat that as covering every model or every task.

The real environmental question sits at industry level, in training, data centre build-out and serving AI at global scale, and we watch it. It is legitimate to say the aggregate picture concerns you while your own use does not.

We prefer providers that publish measured figures. Google has. Others are starting to. Transparency is one of the things worth rewarding with your business, alongside energy mix, water practices and where the power actually comes from, because it is the only way the numbers stay honest.

That answer is better than guilt and better than a shrug. It shows you have actually looked.

Worth using and worth understanding

I think these tools are worth using. The gains for a small business are real. And for a small business, giving up ordinary text use would make no discernible difference to the problem. Choosing providers carefully and using the heavier tools deliberately would.

The costs are real, concentrated, and mostly invisible to the person typing. They sit in training runs drawing the power of a small power station and in planning applications half an hour up the road.

Worth using and worth understanding are not the same thing. A business that takes its environmental commitments seriously owes itself both.

Figures correct as of August 2026; the per-prompt figures and the Scottish planning position were re-checked in October 2026. Per-prompt figures are Google's published measurements for a median Gemini text prompt; other providers' figures may differ, and all of them are moving quickly.

Sources

Written with Claude, then steered, cut and checked by me. I answer for it.

Finlay Carmichael
Finlay Carmichael
Finlay Carmichael is the founder of The AI Lab, helping teams put AI to practical use through hands-on training and clear advice. He brings nearly 30 years' experience building and leading a Microsoft Business Solutions partner, working with hundreds of clients across the UK. Request a call

Reflect on how you use AI at work with the free Personal AI Index. It is in Research beta and open to everyone, with no training booking required.

Take the Personal AI Index More lab notes