AI and the environment, with the figures
AI uses energy, and its growth weighs on power grids. But a request weighs far less than people think, and where it runs makes a big difference. Here is what the studies say, and what we do.
What a request weighs
In August 2025, Google published the full measurement of a median text prompt to Gemini: chips, servers, idle machines and cooling included. The result: 0.24 Wh, less than nine seconds of television. In one year, the energy per prompt was divided by 33.
Mistral AI ran the first full life-cycle analysis of a language model with ADEME, the French ecological transition agency, and Carbone 4, reviewed by two independent firms. A 400-token answer from its assistant emits 1.14 g of CO₂e.
These values vary with the model, the length of the answer and the electricity behind it: a long answer or a reasoning model uses more. The order of magnitude stays the same: a few tenths of a watt-hour, a few grams of CO₂e at most.
- per median Gemini text prompt, everything included
- 0.24 Wh
- Google, August 2025
- per 400-token answer from Le Chat
- 1.14 g CO₂e
- Mistral AI, ADEME and Carbone 4, July 2025
- the energy of a Gemini prompt, from May 2024 to May 2025
- ÷ 33
- Google, August 2025
What the whole weighs
Worldwide, data centres used 415 TWh in 2024, 1.5% of the planet’s electricity. The International Energy Agency expects more than twice as much by 2030, driven first by AI.
In France, digital technology accounts for 4.4% of the national carbon footprint, and the share of data centres within it is growing fast: 46% in 2022, according to ADEME and Arcep, the French telecoms regulator.
These figures are real. One request is frugal, billions of requests are not, and cheap AI encourages using more of it. So the right question isn’t only what a request weighs, but what it is for.
- used by the world’s data centres in 2024, 1.5% of global electricity
- 415 TWh
- International Energy Agency, April 2025
- of France’s carbon footprint comes from digital technology (2022)
- 4.4%
- ADEME and Arcep, January 2025
A machine next to human work
A study published in 2024 in Scientific Reports compares the emissions of an AI and of a person doing the same task. To write a page, the AI emits 130 to 1,500 times less CO₂e.
The authors say it themselves: an AI doesn’t replace a person, and time saved isn’t carbon avoided if it goes into producing more. What can be said is that a repetitive task handed to an agent costs the planet very little, compared with the hours it used to take.
- less CO₂e for a page written by an AI than by a person
- 130 to 1,500 ×
- Tomlinson et al., Scientific Reports, 2024
Where Glove runs
The Glove application is hosted on Google Cloud in Belgium (the europe-west1 region). In 2025, 79% of the electricity used hour by hour in this region was carbon-free, according to Google. Its AI models run in European Union data centres through Vertex AI, in the regions where Google spreads the load.
For scale: the Belgian grid emits 126 g of CO₂e per kWh, against 343 g in Northern Virginia, the world’s largest data centre hub. These figures apply to the region that hosts the application, not to each request to the models, whose footprint we don’t publish yet. And if you prefer, Glove installs on your own servers.
- carbon-free electricity, hour by hour, in the region that hosts the application
- 79%
- Google Cloud, 2025 data
- for the Belgian grid, against 343 g in Northern Virginia
- 126 g CO₂e/kWh
- Google Cloud, 2025 data
What we do
We haven’t measured Glove’s own footprint yet: we will, and we will publish the result here. In the meantime, here is what already limits it.
A frugal model by default
Glove uses a fast model from the Gemini Flash family by default. A more powerful model is chosen when the task warrants it, not for every question.
Agents that don’t run idle
An agent starts on a request, a schedule or a specific event, does its job, then stops: no loop running for nothing.
No training on your data
Glove uses existing models. Your documents are never used to train new ones, neither for us nor for a provider.
Hosted in Europe
The application runs in one of Google Cloud’s lowest-carbon regions in Europe, and the “Data in Europe only” setting blocks any provider outside Europe.
On your premises, if you wish
Installed on your servers, Glove runs on the French grid’s electricity, one of the lowest-carbon in Europe.
A site without trackers
This site is made of static pages, with no cookie or third-party script: nothing loads that you didn’t ask for.
Sources
Every figure on this page comes from one of these publications.
- Measuring the environmental impact of delivering AI at Google Scale
Google, August 2025
- Our contribution to a global environmental standard for AI
Mistral AI, with ADEME and Carbone 4, July 2025
- Energy and AI
International Energy Agency, April 2025
- Observatoire des impacts environnementaux du numérique (in French)
ADEME and Arcep, January 2025
- The carbon emissions of writing and illustrating are lower for AI than for humans
Tomlinson, Black, Patterson and Torrance, Scientific Reports, February 2024
- Carbon free energy for Google Cloud regions
Google Cloud, 2025 data
Useful AI, where it counts
Let’s look together at the tasks where Glove would save you the most time, and only those.