Energy
Wh or kWh as a range. Reported measurements take precedence, then token counts, finally a documented default assumption.
Environmental Ledger for AI events.
Azalee makes estimated energy, water and CO₂e values per logged AI event traceable. Not as an allegedly exact data-centre measurement, but with ranges, provenance and data quality.
The most important statement
Azalee estimates whenever no measured or operator-reported value exists. An environmental value is neither a certification nor proof of climate neutrality. The annual report documents methodology, limits and data coverage.
Water
ml or litres as a scenario for operational water demand. Without an operator value, no exact cooling technology is claimed.
CO₂e
g or kg CO₂e. A logged regional grid intensity is used; otherwise the range deliberately stays wide.
Data source priority
1. Measured
Energy, water or CO₂e value from a traceable measurement source.
2. Reported
Provider, infrastructure or customer value stored in the event with source and range.
3. Token-based
Estimate based on reported input and output tokens with versioned scenario factors.
4. Fallback
When no usage data exists, Azalee uses a clearly marked default assumption with low confidence.
Methodological system boundaries
| Area | Included by default? |
|---|---|
| Inference activity of the logged event | Yes, as a measurement or estimate |
| Model training | No |
| Network and end device | No, unless explicitly logged |
| Manufacturing / embodied emissions | No, unless explicitly logged |
| Offsets, avoided emissions, climate neutrality | No automatic statement |
Recommended event schema
{
"metadata": {
"environmental": {
"provider": "provider-name",
"model": "model-name",
"region": "eu",
"input_tokens": 420,
"output_tokens": 180,
"energy_wh": null,
"water_ml": null,
"carbon_intensity_g_per_kwh": null,
"source": "estimated"
}
}
}Raw prompts, model replies or personal content are not needed for the environmental calculation.
Annual Environmental Evidence Report
The export contains annual values, ranges, a breakdown per AI system, data coverage, methodology version and deliberate exclusions. HTML is printable and can be archived as PDF; JSON and CSV enable independent analysis.
Open to expert review: Azalee only names organisations, NGOs or standards as partners or reviewers once an actual collaboration has been agreed.
Send methodology feedback