Ask for short
Say how long you want the answer; “in five lines” is often enough.
The low-carbon transition of AI
The carbon footprint of your AI, finally visible.
A one-page answer: about 1% of a kettle. A 10-second video: 1 to 2 loads of laundry.
CarboneWatch estimates the electricity, carbon and water of your exchanges with AI services, across your whole computer, without ever reading what you write. Sourced orders of magnitude, not a precision no tool has. Free, no account.
Download See the calculation on an answer See the shared counter
Try it live
Write whatever you like, or ask for an image or a video. The counter follows the answer as it arrives.
Your text stays in your browser and does not change the calculation: it is the length of the answer that weighs.
The answer will appear here. Nothing you type leaves the page.
This answer, so far
0 g
of CO₂e · 0 Wh of electricity
0 mL of water consumed, estimated from fleet averages: the actual data centre location is unknown
For an online service, it is the provider's electricity grid that counts, not yours: most large services compute in the United States.
Simulated answers, images and videos: nothing is sent anywhere. Calculated with CarboneWatch's public rates, data-centre overhead included; an order of magnitude, not a measurement. The method
Three habits to try
Start a fresh conversation
With every message, the model rereads the whole conversation. Changing the subject? Open a new one.
Get it right the first time
Giving the context in the first question avoids answers that miss the point. Every regenerated answer counts again.
Choosing your tools
Beyond how you ask, the service and the model matter too. For each one, what CarboneWatch can see of it.
Save images and video for when they matter
A 10-second video uses as much as 600 to 1,300 one-page answers.
In CarboneWatch: counted for the image and video services it measures; the others are flagged “seen, not measured”.
Pick a lighter model when it is enough
Reasoning modes and the largest models use more for every answer, according to a PEReN study for Arcep (May 2026). For a simple question, the fast model is often enough.
In CarboneWatch: invisible. It does not know which model answers, nor whether the answer reasoned.
Look at European services
Vibe (formerly Le Chat), by the French company Mistral, or DeepL, a German company, for translation. Both compute in Europe, where electricity is on average less carbon-intensive than in the United States. But “European” does not mean “in France”: Mistral spreads its computing across several EU countries, and the French grid, one of the least carbon-intensive in Europe, is only part of it. That place is read in the provider's documents, not inferred from its nationality.
In CarboneWatch: partly. The carbon follows the region its registry assigns to each service. For Mistral, the EU average applies from 0.2.23 on; 0.2.22 still counts it at the French mix.
Try a model on your own computer
With Ollama or LM Studio, the computing happens on your computer, with your electricity. It is not necessarily less, nor necessarily more: it depends on the machine and the model.
In CarboneWatch: counted, from your computer's computing time and the electricity zone set in its settings.
How it works
- You install the application.
- It counts the volume of your exchanges with the AI services it knows. Never the content.
- It converts that volume into electricity, carbon and water, as an order of magnitude.
Estimated from the installations that share
The shared counter
CO₂e of online services
—
Machines that measure
—
Median per machine per day
—
CO₂e per day
By kind of use
What that amounts to
Sharing is off by default. Once on, it is pseudonymised. A service is only named once 5 installations use it.
Three questions of trust
Does CarboneWatch read what I write?
No: it counts volumes, not words.
Do I need an account?
No.
Do my figures go anywhere?
Not by default. If you turn sharing on: per hour and per service, pseudonymised, with no IP address and no machine name.
A complete solution for the low-carbon transition of AI
Measure to understand, understand to act, act to reduce. For your fleet, per department, never per person.