How Ecoia Measures the Footprint of Every AI Request
This is the complete, sourced methodology behind every carbon, water and energy figure Ecoia reports. It is written to be audited. Every factor below is published with its source and version, the formulas are shown in full, and what we do not yet count is stated plainly. These are modelled estimates from published factors, with a confidence band of about ±40%, not metered readings.
Tokens → energy → carbon & water
Every request follows the same short chain. Only the first step depends on the model.
energy (Wh) = tokens × energy-per-token (Wh/token) × PUE
CO₂ (g) = energy (kWh) × grid intensity (0.395 kg CO₂/kWh) × 1000
water (L) = energy (kWh) × water intensity (3.4 L/kWh)
Image models are billed per image, not per token, so the first line becomes energy (Wh) = images × Wh-per-image × PUE. Everything downstream is identical.
Energy per token, by model class
Models are mapped to a size class; each class has a median energy-per-token drawn from published inference benchmarks. Larger and reasoning models compute more per token.
| Model class | Example models | Wh / token | × vs mid-size |
|---|---|---|---|
| Small / fast | Haiku, GPT mini, Flash, Nano | 0.0008 | 0.53× |
| Mid-size | Sonnet, GPT-4/5 | 0.0015 | 1.00× |
| Frontier / large | Opus, GPT-5.5 | 0.0038 | 2.53× |
| Reasoning-heavy | o3, R1, reasoning models | 0.0040 | 2.67× |
| Standard image | Image generation, 1024×1024 | 2.00 Wh / image | — |
| Higher-quality image | Image generation, 1024×1024 | 3.50 Wh / image | — |
Source: aggregated public inference-energy measurements, normalised per token at the median. Values are deliberately conservative (mid-to-high) so estimates are unlikely to understate impact.
Infrastructure constants
These do not vary by model. Each is a published figure with a named source and version.
PUE 1.56
Power Usage Effectiveness — the data-center overhead multiplier for cooling and power distribution on top of chip energy.
Source: Uptime Institute Global Data Center Survey, 2024
0.395 kg CO₂ / kWh
Grid carbon intensity applied to electricity used. A US-average factor; regional grids vary widely.
Source: US EPA eGRID 2022 (US average)
0.3 L / kWh on-site
Water evaporated by data-center cooling per kWh of IT load.
Source: Microsoft environmental disclosures, FY24
3.1 L / kWh off-site
Water consumed generating the electricity itself (thermoelectric + hydro).
Source: Ren et al., 2023
Total water intensity = 0.3 + 3.1 = 3.4 L/kWh.
A 700-token answer on a mid-size model
Take a typical chat exchange of about 700 tokens (prompt read + answer written) on a mid-size model at 0.0015 Wh/token:
- Energy: 700 × 0.0015 × 1.56 = 1.64 Wh
- CO₂: 0.00164 kWh × 0.395 × 1000 = 0.65 g CO₂
- Water: 0.00164 kWh × 3.4 = 0.006 L
The dashboard runs exactly this arithmetic on the real token count of every request, using the energy factor for the specific model you chose.
What is and isn’t included
Included
- Inference (serving) energy per request
- Data-center overhead via PUE
- Grid carbon of that electricity
- On-site cooling water + off-site generation water
Not yet included
- Model training (amortised across requests)
- Hardware manufacturing / embodied carbon
- Networking to the end user’s device
- The user’s own device energy
We disclose exclusions rather than bury them. Because Ecoia offsets over 200% of the includedfootprint, the margin above 100% is designed partly to cover these known gaps and the ±40% confidence band.
Why “more than 200%”
Neutral means retiring offsets equal to 100% of estimated impact. Ecoia retires verified offsets and water restoration for more than 200% — over two units retired for every unit estimated.
Two units, not one, because: (1) it makes usage genuinely net-negative rather than break-even, and (2) the extra headroom absorbs the ±40% uncertainty in the estimate and the excluded scopes above. Offsets and restoration are retired against measured usage, not purchased once a year against a guess.
Where offsets go and how retirement is documented
Offsets are only credible if the retirement is traceable to a registry. This section lists the projects and the documentation trail.
Offset & restoration partners: verified carbon and clean-water partners (e.g. charity: water for water restoration). Each retirement is recorded against measured usage.
Registry & retirement records: [to be published — registry name, project IDs, serial numbers and retirement dates]. Publishing serials here lets anyone verify the credits were retired and not resold.
Additionally, 10% of all revenue funds conservation, on top of per-request offsetting.
Methodology FAQ
Are these measured or estimated figures?
They are modelled estimates built from published, peer-reviewed and government factors, applied to the real tokens and images of each request. Ecoia meters usage (tokens and images) precisely; it does not have a power meter on the specific GPU serving a request, so the energy, carbon and water figures are estimates with a stated confidence band of about ±40%.
Why does the energy per request change by model?
Only the energy factor varies by model. Larger and reasoning-heavy models do more computation per token, so they use more energy per token. Overhead (PUE), grid carbon intensity and water intensity are infrastructure constants applied on top of that energy.
What does "more than 200% offset" actually mean?
For every unit of carbon and water Ecoia estimates for your usage, it retires verified offsets and water restoration for more than two units. Retiring more than 100% is what makes usage net-negative rather than merely neutral; Ecoia targets over 200% to leave a margin above the estimate’s uncertainty.
What is not included in the estimate?
The current model covers inference energy plus data-center overhead, grid carbon and cooling/generation water. It does not yet allocate model training, hardware manufacturing (embodied carbon), networking to the end user, or the user’s own device. These are disclosed as exclusions rather than hidden.
Can I audit the numbers myself?
Yes. Every factor on this page is published with its source and version date, and the worked examples show the full arithmetic. The per-request figures in your dashboard use exactly these factors.
Keep reading
Use AI that gives back more than it takes
Ecoia.ai runs Claude, GPT & Gemini for chat, images and an API, and offsets over 200% of the water usage and carbon emissions your AI creates.