AI assistant
Questions, summaries, writing, translation.
Popular models for this use
- GPT-6 Luna
- Claude Haiku 5.5
- Gemini 3.8 Flash
- Mistral Small 4
CARBON CALCULATOR / TO UNDERSTAND, NOT TO CERTIFY
How much does your AI weigh?
Run the numbers.
Estimate the carbon footprint of three years of AI for your team: in the cloud, or on an Appliance on your premises.
You describe your usage. We show every assumption.
Your team
Describe how your team uses AI.
Advanced settings
Two assumptions weigh heavily on the result. You can change them.
YOUR SCENARIO / 3 YEARS / ONE APPLIANCE
Carbon crossover point
after about 6 months.
With these assumptions, the accounted cloud total overtakes the Appliance’s at that point. This is a conditional result, not a verified reduction.
Cloud electricity assumption: Carbon-intensive. Editable in the advanced settings.
Cloud AI, 3-year total1,959 kg CO₂eOne Appliance, 3-year total359 kg CO₂e
Who emits what, year after year?
Cumulative emissions of the accounted items, in kg CO₂e
Cloud AIOne Appliance
See the values year by year
| Cumulative, in kg CO₂e | Start | Year 1 | Year 2 | Year 3 |
|---|---|---|---|---|
| Cloud AI | 0 | 653 | 1,306 | 1,959 |
| One Appliance | 310 | 326 | 343 | 359 |
| Difference (cloud − Appliance) | -310 | 327 | 963 | 1,600 |
The difference is a difference between accounted items. It is not “avoided emissions.” A negative difference means the cloud is lower.
What this calculation does not know
These items are unknown. They are not zero, and they are in none of the totals above.
The ability of a single Appliance to carry this load is not verified either. If several devices were needed, the local footprint would grow accordingly.
- Cloud sideManufacturing of the servers
- Unknown
- Cloud sideTools called by the AI: search, storage, network
- Unknown
- Cloud sideTraining of the models
- Unknown
- Appliance sideAccessories (base, cables) and installation
- Unknown
This calculator is an invitation to understand and question the footprint of AI, not an environmental certification.
Conditional estimate, estimation method v0.2.0, updated on October 13, 2026. It is not a life-cycle assessment, not a verified emissions reduction, and not the footprint of any cloud provider. The calculation relies on the measurements ML.ENERGY published on three open models and on official data from the Appliance’s manufacturer; Appliance Labs’ field measurements will come with the first deployments. Read the methodology
THREE KINDS OF USE
Each kind of use has its models.
And an energy landmark.
Popular models illustrate each kind of use. The calculation itself uses the energy of an open model measured by ML.ENERGY.
AI workflow
Long documents, analyses, complete deliverables.
Popular models for this use
- GPT-6 Terra
- Claude Sonnet 5.5
- Gemini 3.1 Pro
- Mistral Large 4
- Grok 4.7
AI agents
Long multi-step tasks, with search and tools.
Popular models for this use
- GPT-6 Sol
- GPT-6 Astra
- Claude Opus 5.5
- Claude Fable 5.1
- Gemini 4 Argon
- Grok 4.7
Model names checked on October 13, 2026. The energy use of these popular models is not published: we do not measure it and we attribute no figure to them. The energy landmark of each kind of use is an open model, with no validated quality equivalence. This calculation ranks no brand.
OUR APPROACH
How did we build
this estimate?
Three simple steps, with public sources. We show our work, limits included.
- 1
Compare capabilities
We compare the capabilities of models with Artificial Analysis, to place those that deliver a comparable service.
Artificial Analysis - 2
Look for measured energy
We look for published energy data, notably from Hugging Face and ML.ENERGY, on open models. For the Appliance, we rely on official data from its manufacturer.
Hugging Face AI Energy Score, ML.ENERGY, official manufacturer data - 3
Build scenarios
We build scenarios from these references and three usage profiles, spelling out the extrapolations they require.
Appliance Labs estimation method
Where we stand: the current calculator uses the measurements published by ML.ENERGY on three open models. For the Appliance, we obtained official manufacturer data: its carbon footprint over the whole life cycle and its power draw measured in the laboratory. Cross-referencing Artificial Analysis and AI Energy Score is still research in progress: it does not enter the calculations, and no equivalence is validated.
Read the methodologyASSUMPTIONS
What is assumed.
And what is declared.
Every default carries its status. Changing an assumption can reverse the result.
- Manufacturer dataAppliance initial footprint: 310 kg CO₂e
Manufacturing, transportation and end of life. Value set by Appliance Labs from the official carbon footprint published by the manufacturer, which covers the device’s whole life cycle. Editable in the advanced settings.
- AssumptionActive power: 140 W
8 hours a day, 365 days a year. Our working assumption, to be confirmed by measurements on deployed Appliances.
- Manufacturer dataIdle power: 27 W
Measured in the laboratory by the manufacturer, under the European ecodesign regulation. Rounded value.
- Published by RTEAppliance electricity: 0.029 kg CO₂e/kWh
French 2025 production, life cycle, published by RTE. A proxy, not the exact factor of the electricity delivered.
- AssumptionCloud electricity: 0.029, 0.200 or 0.400 kg CO₂e/kWh
Three assumptions to choose from in the advanced settings; the calculator opens on the most carbon-intensive one. None is the value of a provider.
- AssumptionData center overhead (PUE): 1.20
Calculation assumption, not the value of a provider.
- AssumptionOther server components: 15% of the GPU energy
Calculation assumption.
- AssumptionVolumes: 220 days
Working days per year, over 3 years. The size of a request in each kind of use is an assumption, not observed usage.