How Much CO2 and Water Does One AI Token Really Use?

When you use an AI model through an API, you are not billed by the question or by the minute. You are billed by the token. A token is a small chunk of text, roughly three quarters of a word, and every model prices its work per million tokens. That makes the token the natural unit for measuring environmental impact too. If you know what one token costs in energy, you can scale it up to a prompt, a product feature, or a company's entire annual usage. This piece works through the numbers: how much electricity a token draws, how that turns into carbon, and how it turns into water.

Why the token is the right unit

Most headlines measure AI in queries: "a ChatGPT question uses ten times more energy than a Google search." That is a useful shock statistic, but it hides a lot. A query can be three words or three thousand. A model can reply in a sentence or in five pages. The token strips that ambiguity out. It counts the actual work done, both the tokens you send in and the tokens the model generates back.

It is also the number you already have. Every provider dashboard, from OpenAI to Anthropic to AWS Bedrock, reports token counts on your bill. So unlike vague per-query estimates, a per-token figure lets you calculate impact from data you can actually export.

Energy per token

Published research from the IEA and Epoch AI, combined with GPU inference cost analysis, puts a large frontier model somewhere around 3 to 4 watt-hours per thousand tokens. Divide that down and a single token lands in the range of a few thousandths of a watt-hour. Tiny on its own, which is exactly why it adds up quietly across millions of calls.

Treat these as informed estimates, not stopwatch readings. Providers rarely publish exact per-token power draw, so every figure here is reconstructed from hardware specs and independent research. The gap between model tiers, though, is real and large. Choosing a light model where it will do is the single biggest lever you have.

From energy to carbon

Energy alone tells you nothing about carbon until you know where the electricity came from. The same token processed in two different data centers can differ by a factor of twenty in emissions, because the grid behind each one is different. Carbon intensity is measured in grams of CO2 per kilowatt-hour, and it swings enormously by region.

So a single frontier-model token produces somewhere between roughly 0.0000001 and 0.0000026 kg of CO2 depending on the grid. The point is not the decimal places. It is that region choice moves your footprint more than almost any other single decision, and it is often a setting you can change in your provider console.

From energy to water

Data centers drink water in two ways. On site, evaporative cooling turns water to vapour to shed heat from the servers. Off site, the power stations feeding the grid use water for their own cooling. Together these are captured by a figure called water usage effectiveness, and a common rule of thumb lands around 1.8 to 2 litres of water per kilowatt-hour of IT energy once both are counted.

Apply that to the token. A million frontier tokens draw roughly 3.7 kilowatt-hours, which pulls somewhere near 7 litres of water through the cooling chain. That is why the familiar line, "a long ChatGPT conversation drinks about a 500ml bottle of water," holds up. A chatty session can run to tens of thousands of tokens, and the millilitres accumulate fast.

A worked example

Say a small product feature runs 15 million tokens a month through a frontier model on a US grid. Put the numbers together:

• Energy: about 55 kWh a month, roughly what a fridge uses.
• Carbon: about 20 kg of CO2 a month, near enough to driving 80 km in a petrol car.
• Water: about 100 litres a month through the cooling chain.

Move that same workload to a Nordic region and the carbon drops by more than ninety percent, while the energy and water barely change. Switch the light-touch parts to a smaller model and the whole bill falls again. None of this requires using AI less. It requires using it deliberately.

Work out your own number

The figures above are averages. Your real footprint depends on your exact model mix, your regions, and your token volume. Our free AI Impact Calculator does the arithmetic for you: pick a model, set your tokens, choose a region, and it returns carbon, energy, and water with the equivalents spelled out.

If you are doing this for a company rather than out of curiosity, and you need a figure you can put in front of an auditor, the calculator alone will not close it. That is what the business tool is for. Drop in your provider billing CSV and it maps every token to its grid region on the GHG Protocol basis. When you need it formalised, we prepare an audit-ready ESG report from your figures, cited and ready for SEC or CSRD workpapers, emailed within one business day.

Common questions