What are the environmental costs of AI?

Every AI answer runs on computers in a data center, a building full of servers that draw electricity and need cooling. The environmental cost comes in two phases. Training, when a model learns from huge amounts of text over weeks or months, is a large one-time cost. Inference, each time the finished model answers someone, is small per request but repeats billions of times. On top of electricity there is water, often used to cool the machines, and the emissions and minerals involved in manufacturing chips and servers.

Measuring all this is hard, and only a few companies publish detailed numbers. Here is what the main disclosures say, and what each one leaves out.

What has been measured

Google, per prompt. In August 2025, Google researchers published a study of serving the Gemini assistant in production. They counted the AI chips, the host computers around them, idle machines kept ready, and the building's overhead such as cooling. Their result: the median Gemini Apps text prompt used 0.24 watt-hours of electricity and 0.26 milliliters of water. They put it in everyday terms as less energy than watching television for nine seconds, and about five drops of water. They also report that efficiency work and clean-energy purchases cut the energy per median prompt 33-fold, and its carbon footprint 44-fold, over a year.

Mistral, over a model's life. In July 2025, the French company Mistral AI published a lifecycle analysis, a method that adds up impacts from manufacturing through use, done with the consultancy Carbone 4 and the French environment agency ADEME. Training Mistral Large 2 plus its first 18 months of use came to 20.4 kilotonnes of CO₂ equivalent and 281,000 cubic meters of water. For one 400-token answer from its Le Chat assistant (a token is a word or a piece of a word), it reports 1.14 grams of CO₂ equivalent and 45 milliliters of water, about three tablespoons. Those figures include upstream impacts, such as building the servers.

Data centers as a whole. A report from Lawrence Berkeley National Laboratory, released by the U.S. Department of Energy in December 2024, found that data centers used about 4.4% of U.S. electricity in 2023, up from 58 to 176 terawatt-hours since 2014. It projects 6.7% to 12% by 2028. Not all of that is AI, but the department names AI as one of the drivers.

Why the numbers differ, and their limits

The Google and Mistral figures cannot be compared directly. Google measured the energy and water involved in serving a prompt; Mistral counted a wider lifecycle that includes manufacturing, and measured a different model and a different answer length. Neither number covers what your own phone or laptop uses.

Each has limits. Google's figure is a median, the middle value, for text prompts in one app; long reasoning answers, images and video are likely to cost more, and the study was done by the company itself. Mistral calls its own study a first approximation, because no industry standard exists and the impact of making GPUs, the chips that run AI, had to be estimated. And per-prompt numbers say nothing about training, or about how quickly total demand is growing.

One pattern does hold across disclosures: size matters. Mistral reports that impacts rise roughly in proportion to model size, so a model ten times larger has about ten times the footprint for the same output.

A worked example: your own use

Using Google's median as a rough guide, here is an illustrative estimate. It applies only to short text prompts on a similar system.

Illustrative arithmetic: prompts × 0.24 Wh × 365 days. For one person doing text chat, the footprint is modest. The larger impacts come from scale, from heavier tasks, and from the companies' choices about models, hardware and power sources.

How can you use AI with a lighter footprint?

Start with the biggest lever: model size. Many services offer a small, fast model next to a large one. Use the small one for routine work such as rewording, short summaries and simple questions, and save the largest or "thinking" modes for problems that need them. Mistral's report recommends exactly this, matching the model to the job.

Second, ask fewer, better questions. A clear prompt with the context included often gets a usable answer in one try instead of five. Mistral also suggests grouping requests: one message asking for three things uses less computing than three separate conversations.

Third, be deliberate with images and video. Generate a few drafts rather than dozens of variations, and reuse what works.

Finally, use your voice as a customer. Prefer providers that publish measured figures and their methods, and ask the ones that don't. Comparable disclosures are what will let everyone, from individual users to public buyers, choose the less wasteful option.