Cooking on Gas: The Hidden Cost of Your AI Habit

A sprinkle of AI here, a splash of AI there, and we're cooking on gas. It's brilliant. I lean on it daily and I'm not about to stop.

But cooking on any fuel comes at a cost — and AI quietly bills you twice. Once on the invoice, and once on the environment. Most of us only ever see the first one. Plenty of businesses don't even see that.

I've written plenty about the financial side. This time I want to look at the other meter running in the background — the energy and water behind every prompt — because it turns out the two are the same problem wearing different clothes.

Every query draws real power

Each time you fire off a prompt, a data centre somewhere spins up serious hardware to answer it. A single ChatGPT-style query uses several times the electricity of a plain web search — estimates put it at roughly three to ten times, depending on the model and the setup.

On its own, that's nothing. One query is a rounding error. But multiply it across a business running AI through a dozen tools all day, every day, and the rounding errors become a line item — for the grid as much as for your card statement.

Then there's water. Data centres run hot, and many are cooled with water. A long AI session quietly accounts for a surprising amount of it — one estimate puts a hundred-query session at around half a litre in the thirstiest regions. Microsoft alone reported millions of cubic metres of water use in a single year, and the figures have only climbed since.

The numbers are big, and getting bigger

Globally, data centres burned somewhere around 460 to 490 terawatt-hours of electricity in 2025 — roughly the annual consumption of a mid-sized country. AI-focused facilities drove a 50% jump in a single year.

Where they cluster, it's stark. Data centres already account for around a fifth of Ireland's entire electricity use, with forecasts pushing towards a third. In parts of the US, the cost of feeding them is landing directly on ordinary households' power bills. By 2030 the projections roughly double — and AI is the single biggest driver.

Here's the twist: per query, it's getting more efficient

This is the part that surprised me. The energy needed for a single AI task is falling fast — by something like an order of magnitude a year, as chips and models improve. On a per-task basis, AI is getting dramatically leaner.

So why is the total still climbing? Because we're adopting it far faster than it's getting efficient. Every efficiency gain gets eaten by a new, hungrier use. The growth now isn't simple chatbot queries — it's reasoning models that think before they answer, agents that fire dozens of calls per task, and video generation. The heavy, expensive stuff.

Sound familiar? It's the exact shape of the cost problem. Cheaper per call, far more calls, bigger bill.

Which means the fix is the same for both

And this is the genuinely good news for anyone running a business on AI: the waste that inflates your invoice is the same waste burning needless power and water.

Re-sending the entire conversation history on every call. Running a frontier model on work a small one would handle. Letting responses ramble with no cap. Agentic loops firing ten calls where two would do. Every one of those costs you money and burns energy you never needed to spend.

Tighten how you use AI and you win twice — a smaller bill and a smaller footprint, from a single set of changes. I wrote up the practical levers recently in Stop Burning Tokens: cap your output, trim the context, cache what repeats, right-size the model. Every one of them is an environmental win as much as a financial one.

You can't cut what you can't see

The catch is the same too. You can't reduce a number you're not measuring — and almost nobody is measuring either one. Most businesses have no real idea what their AI costs them, let alone what it's burning.

Visibility is the starting point for both. Once you can see what each model, each tool, and each workflow is actually consuming, you can do something about it. That's what SpendLil is for — it sits in front of your AI and tracks every call, so the waste has nowhere to hide. Optimise with the meter running and the savings show up in black and white.

Because here's the thing I keep coming back to: spend visibility and footprint visibility are the same dashboard. Manage one and you've largely managed the other.

The bottom line

Cooking on gas is fine. I'm not suggesting anyone goes back to doing it all by hand.

But every hob you leave running when you've only got one pan on costs you twice — on the bill and on the planet. The fix was never to use less AI. It's to stop wasting the AI you use.

Same output. Smaller bill. Lighter footprint. There's no version of that worth ignoring.