Is AI bad for the environment? The numbers, both ways

If you search this question you will find two confident answers. One camp says a single query is a rounding error and the worry is overblown, the other says AI is draining rivers and keeping gas plants open, and both quote real numbers. They are not actually disagreeing, because one is talking about a single use and the other about the total, and you need both to answer the question properly.

What "bad for the environment" has to mean

AI affects the environment in four ways. The models run on electricity, which carries the carbon of whatever generated it. The data centres are cooled with water, and more water is consumed at the power plants behind that electricity. The hardware carried embodied carbon before it served a single token, and so did the training run that produced the model. And there is a credit side: what AI might save elsewhere, in grids, buildings and industry, which the IEA takes seriously enough to put a number on.

The first three can be measured per token. The fourth can only be projected for now, so this post sticks to what has been measured and flags the places where a projection is being quoted.

The numbers per use

Per query these are small amounts. The electricity is roughly what an oven draws in a second, and the water is about a fifteenth of a teaspoon. Under CrbonFree’s per-token factors a short chat on a medium tier model lands at 0.32 millilitres of water, the same figure OpenAI published, from an independent set of factors. Longer exchanges and reasoning models run several times higher, because output tokens cost more than input tokens and reasoning models produce far more of them.

The numbers in total

In total the amounts are large. Data centres as a category still sit below 1.5% of energy-sector emissions, which sounds modest until you read the IEA’s next sentence, that they are among the fastest growing sources of emissions, and Berkeley Lab attributes the doubling of US data centre demand between 2017 and 2023 largely to AI servers. The totals include everything data centres do, but AI is what is moving them.

Why both are true

Energy per query is falling fast: Google reports a 33 times reduction for a median Gemini prompt in one year, and the older 3 watt-hour estimate for ChatGPT is now about ten times too high. Usage is rising faster. Our own case study metered a year of traffic through the OpenRouter gateway and found it growing many times over, and every one of those tokens is a fraction of a watt-hour that nobody was counting. Both trends are real, and for now the growth in the number of answers is outrunning the fall in energy per answer.

Where it is worse, and where it is better

The same query is a different number depending on where it runs. On Google’s fleet the grid intensity is 0.345 kilograms of CO₂ per kilowatt-hour and the power usage effectiveness 1.09; the default factors in CrbonFree’s methodology assume 0.42 and 1.2. That is a third more carbon per token for the same work, before anything about the model changes. Water swings further: evaporative cooling in a hot, dry region can consume many times the water of a closed-loop site somewhere cool, which is the point the UC Riverside team made in 2023 and the reason their per-response range spans a factor of five.

The model matters as much as the location. Under our factors a small tier does a typical task at a third of the energy of a medium tier, and a reasoning model at more than twice it. Put the three together and the per-token answer to this question depends on which model you called, on which grid, in which building.

What AI might give back

The IEA estimates that broad application of existing AI-led solutions could cut emissions equivalent to around 5% of energy-related emissions in 2035, more than data centres are projected to emit. That figure is a projection of what could be deployed rather than a measurement of what has been, and it does not cancel anyone’s own footprint. It does mean that a fair account of AI and the environment has entries on both sides.

Is there an eco-friendly way to use AI?

For a person, the levers are the ones that cut tokens: shorter prompts, shorter answers, a small model for routine work, reasoning modes only when the task needs reasoning. For a company, the first lever is measurement. Usage per model is the number every other decision depends on, and a spend-based guess cannot tell an efficient model from a wasteful one. With the number in hand, caching the fixed part of a prompt cuts its input energy by ten times, routing routine tasks to a small tier cuts a typical task from 0.59 to 0.21 watt-hours, and retiring verified credits against the measured footprint is a separate line, reported alongside it rather than subtracted from it.

How to know for your own usage

The three posts this one summarises each have a calculator that runs the methodology for any exchange size, model tier and number of queries a day:

• How much water does ChatGPT use?, with the water calculator.
• How much energy does AI use?, with the energy calculator.
• AI carbon footprint: what it is and how to measure it, with the three-layer calculator.

CrbonFree meters the tokens themselves, per model and per provider, from a provider key, the CLI or the browser extension, and the sample account shows a year of real traffic metered that way with no login required.