Your AI Prompt Uses About Five Drops of Water. That Is Not the Number You Should Be Worried About.

For two years, the honest answer to “how much energy does an AI chatbot use?” was that nobody outside the companies knew, and the companies were not saying.

Then in August 2025 Google published a measurement. A median text prompt to its Gemini app, it said, uses 0.24 watt-hours of electricity, emits about 0.03 grams of carbon dioxide equivalent, and consumes 0.26 millilitres of water. Roughly five drops, and about as much power as running a microwave for one second.

If you were expecting a scandal, that is not one. Which is exactly why it is the wrong number to fixate on.

Why the per-prompt number feels reassuring

Small numbers land differently from big ones. Five drops of water sounds like a rounding error next to a shower, and it is.

The trouble is that the interesting quantity is not the cost of one prompt. It is the cost of all of them, plus the machines waiting to serve the next one, plus the training runs that produced the model in the first place. And a per-prompt median is a poor tool for reaching that total, for a few reasons that Google itself has been fairly open about.

A median hides the tail. Asking a model for the capital of France and asking it to read forty documents and write a synthesis are both “a prompt”, and they are not remotely the same amount of computing. Image and video generation, which are far heavier, are outside the figure entirely.

The carbon figure is also market-based, meaning it credits Google for the clean electricity it buys rather than measuring the grid the machines are physically plugged into. Reporting against grid averages instead would raise it by roughly a third.

And the one thing you need to turn a per-prompt number into a national number is missing. Researchers pointed this out immediately: Google did not say how many prompts Gemini serves. Without a denominator, the arithmetic cannot be done from outside.

The number that actually matters

So use the meter that measures the whole building instead.

The International Energy Agency, in its Energy and AI report, puts global data centre electricity consumption at about 415 terawatt-hours in 2024, roughly 1.5 per cent of the world’s electricity. That total has been growing about 12 per cent a year since 2017, more than four times the growth rate of electricity demand overall.

By 2030 the IEA’s base case has it more than doubling, to around 945 TWh. To give that a shape: it is a little more than the entire electricity consumption of Japan today.

AI is the main driver. Demand from AI-optimised data centres is projected to more than quadruple over the same stretch. AI has accounted for something like 5 to 15 per cent of data centre power in recent years, and the IEA expects that to reach 35 to 50 per cent by 2030.

The load is also concentrated. In 2024 the United States accounted for about 45 per cent of global data centre electricity use, China about 25 per cent, and Europe about 15 per cent.

The American grid is where this gets awkward

In the United States the picture is sharp enough that the Department of Energy commissioned a study of it.

Lawrence Berkeley National Laboratory’s 2024 report found that US data centres used about 4.4 per cent of the country’s electricity in 2023, up from 58 terawatt-hours in 2014 to 176 TWh in 2023. Its projection for 2028 is a range rather than a point, between 325 and 580 TWh, or somewhere between 6.7 and 12 per cent of American electricity.

That range is the most honest thing in the report. A spread that wide, five years out, means the modellers do not know, and are declining to pretend otherwise.

The IEA is similarly frank about the long view: its scenarios for 2035 run from about 700 TWh to about 1,700 TWh. Anyone quoting a single confident figure for 2035 is quoting a scenario and dropping the label.

Where the electricity comes from

A doubling of demand is only a climate problem if it is met with fossil fuels, and the projected mix is more mixed than either camp usually admits.

The IEA expects renewables to supply over 450 TWh of the growth to 2035, about half of it, with natural gas adding around 175 TWh and nuclear roughly the same. Small modular reactors are pencilled in from about 2030, which is a technology forecast rather than an observation.

Emissions from data centres sit at about 180 million tonnes of CO2 today, and the IEA’s base case has them reaching 300 million tonnes by 2035. That is a real increase, and in the context of global energy emissions it stays below 1.5 per cent of the total.

The other side of the ledger

There is a genuine counterweight, and it deserves to be stated at its real size rather than its marketing size.

The IEA estimates that AI applications that already exist, used across industry, transport, buildings and the power system itself, could cut emissions by an amount equivalent to around 5 per cent of energy-related emissions in 2035. That is considerably more than data centres are projected to emit.

But the report attaches an important caution, and it is worth repeating rather than burying. That saving is far smaller than what tackling climate change actually requires, and it will not happen automatically. It describes a technical possibility, not a forecast. Nothing about training a model causes a steel plant to become more efficient.

What to take away

Your individual prompt costs almost nothing, and the guilt some people attach to it is misplaced. Skipping a chatbot query to save the planet is like skipping a single sheet of paper to save a forest.

The industry as a whole, meanwhile, is adding something like a Japan’s worth of electricity demand to the world’s grids inside a decade, at a speed that grids are historically bad at matching. How much of it gets met with fossil fuels will be settled by decisions about transmission lines, generation and where data centres are allowed to plug in.