1. Our approach
At Appliance Labs, our job is to design and deploy private, local artificial intelligence solutions. We are not climate or life-cycle assessment specialists.
But we think it matters to try to understand the environmental impact of the technologies we deploy. So we started digging into the subject.
We studied the work published by Artificial Analysis, Hugging Face, ML.ENERGY, ADEME and other recognized organizations. We tried to relate the capabilities of AI models to their energy use, where measurements exist.
From that data, we built our own estimation method, with its assumptions, its extrapolations and, inevitably, its limits. We don’t claim to have all the answers.
The figures shown are neither a certified life-cycle assessment nor a real measurement of the cloud services you use. They let you explore different scenarios and better understand what drives the footprint of AI.
2. What we compare
Two ways of giving a team of up to 50 people the same amount of AI work for three years: AI services in the cloud, or a single Appliance installed on your premises.
A request is one complete piece of work handed to the AI. It may contain several model calls, reasoning and retries: all of it is already counted in its size.
| Use | Default share | Size of a request | Open reference model | Benchmark conditions |
|---|---|---|---|---|
| AI assistantA question, a summary, a text to write or translate. | 70% | 3,000 tokens read, 1,000 tokens generated | Qwen3-32B ↗ | lm-arena-chat 1 × NVIDIA B200, BF16 |
| AI at workA long document to analyze, a synthesis or a complete deliverable. | 20% | 15,000 tokens read, 5,000 tokens generated | Qwen3-235B-A22B-Instruct-2507 ↗ | lm-arena-chat 4 × NVIDIA B200, BF16 |
| AI agentsA long multi-step task, with search and tools. | 10% | 150,000 tokens read, 30,000 tokens generated | DeepSeek-R1-0528 ↗ | GPQA 8 × NVIDIA B200, FP8 |
The split is a split of requests, not of people. Pairing a kind of use with a model is illustrative: GPQA is not a real agent run, and lm-arena-chat is not your daily work. No equivalence with a proprietary model is validated. Request sizes are assumptions, not observed usage.
What is counted, and what is not
Cloud AI
Counted
- Electricity of the servers and of the data center
Unknown, absent from the totals
- Manufacturing of the servers
- Tools called by the AI
- Training of the models
One Appliance
Counted
- An initial footprint: manufacturing, delivery, installation
- The electricity used every year
Unknown, absent from the totals
- Accessories
- End of life of the device
An unknown item is not zero. The result therefore compares subtotals: it is not a certified emissions reduction.
3. How we calculate
1. The volume of requests
Number of people × requests per day × 220 working days a year. Yearly growth applies from the second year. Requests are then split across the three kinds of use.
2. Energy on the cloud side
We multiply the generated text by the energy ML.ENERGY measured for the open reference model. That energy already covers every GPU and request batching: we do not multiply it a second time. We add 15% for the other server components, then a factor of 1.20 for the data center. Applying a benchmark to our kinds of use is an extrapolation, not a measurement.
3. Emissions on the cloud side
The resulting electricity is multiplied by the carbon intensity chosen for the cloud’s power grid. It is an assumption: we do not know your provider’s.
4. The Appliance
255 kg CO₂e counted from the start, not spread over three years. Then electricity: 38 W idle and 140 W for 8 hours a day, or 630.72 kWh a year, multiplied by 0.029 kg CO₂e/kWh. This budget is fixed: it does not depend on the number of requests.
5. The comparison
We add up both sides year after year. The crossover point is when the cloud total reaches the Appliance’s. It may not exist: with low-carbon cloud electricity, the cloud stays below.
The sources used in the calculation
| Source | What it provides | What it is not |
|---|---|---|
| ML.ENERGY Leaderboard ↗ | The only source of energy figures: measurements published on three open models. | It is neither a measurement of a proprietary model nor a replication by us. |
| RTE, Bilan électrique 2025 ↗ | Carbon intensity of the Appliance’s electricity: French 2025 production, life cycle. | It is not the exact factor of delivered electricity, imports and losses included. |
| NVIDIA DGX Spark ↗ | Declared idle power of the DGX Spark: 38 W. The declared maximum is 233.2 W. | No manufacturing footprint. The 140 W active power is our assumption, not an NVIDIA figure. |
Sources consulted on October 11, 2026. ML.ENERGY’s measurements are cited and linked to their source. We do not republish their data tables.
The reference scenario
50 people, 20 requests per day, a 70 / 20 / 10 split, flat usage, one Appliance, over three years. The calculator opens on carbon-intensive electricity; all three assumptions are below.
| Cloud electricity | Accounted cloud | Accounted Appliance | Difference (cloud − Appliance) | Crossover point |
|---|---|---|---|---|
| Low-carbon (0.029 kg CO₂e/kWh) | 142 kg CO₂e | 310 kg CO₂e | -168 kg CO₂e | None |
| Intermediate (0.200 kg CO₂e/kWh) | 980 kg CO₂e | 310 kg CO₂e | 670 kg CO₂e | about 10 months |
| Carbon-intensive (0.400 kg CO₂e/kWh) | 1,959 kg CO₂e | 310 kg CO₂e | 1,649 kg CO₂e | about 5 months |
These figures do not demonstrate an emissions reduction and describe no commercial service. The variants do not rank providers, and their range is not a confidence interval.
4. Our assumptions and limits
Every value carries its status. Two of them, cloud electricity and the manufacturing footprint, can be changed in the calculator’s advanced settings.
- AssumptionAppliance initial footprint: 255 kg CO₂e
Manufacturing, delivery and installation. Unverified provision: it is not an NVIDIA figure. Editable in the advanced settings.
- AssumptionActive power: 140 W
8 hours a day, 365 days a year. Assumption, not a measurement under this load.
- Declared by NVIDIAIdle power: 38 W
Value declared by NVIDIA for the DGX Spark.
- 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.
What the method cannot do
- The Appliance’s capacity is not verified. The reference scenario asks it to generate about 163 tokens per second on average over eight working hours, with no margin for peaks. Several devices might be needed, which would increase the local footprint accordingly.
- Pairing our three kinds of use with the three reference models is illustrative. The effect of long inputs is not modeled.
- The 140 W active power and the 255 kg CO₂e initial footprint are not verified. The Appliance’s fixed electricity budget is not a load law.
- The emission factors of the cloud and of the Appliance do not yet rest on harmonized boundaries.
- The variants are scenarios, not probabilities. Some put the cloud below the Appliance, others above.
- A renewable electricity contract does not automatically bring physical consumption down to zero emissions.
- The method has not yet been through an independent review.
5. A method that keeps improving
As we deploy our appliances, we will try to check our estimates against field data, notably with CodeCarbon. We are also studying EcoLogits to enrich our method.
To date, these tools have produced no field measurement: no figure on this page comes from a measurement made by Appliance Labs.
We will share our results as they come, including when they are not favorable to Appliance Labs.
And if you work on these topics, we would be glad to talk. A better source, an error you spot or constructive criticism will help us improve this approach.
The work we are studying for the next versions
| Work | What we look for in it | Where we stand |
|---|---|---|
| Artificial Analysis ↗ | Compare the capabilities of models, to relate comparable services. | Research in progress. No score enters the calculation, no equivalence is validated. |
| Hugging Face AI Energy Score ↗ | Study the energy use of open models, under known measurement conditions. | Research in progress. No coefficient used; stars are not watt-hours. |
| CodeCarbon ↗ | Check our estimates against field data, on deployed Appliances. | Not deployed yet. No field measurement to date. |
| EcoLogits ↗ | Enrich the method with another estimate of cloud services. | Under study. Tool not run; its results are not added to ours. |
| ADEME ↗ | Better estimate manufacturing, from the life-cycle assessment of AI GPUs. | Under study. Its value is not transferred to the Appliance in the current calculation. |
Estimation method v0.1.0 · sources consulted on October 11, 2026