Transparency

How We Calculate AI Impact

AIrth estimates the energy, carbon, and water footprint of your AI usage using published research and industry data. This page explains our methodology in full.

Version 2025-10-v1 · Last updated October 2025

Core Formula

Every AI interaction produces three impact metrics:

Energy = Energy_raw × PUE

CO₂e (g) = Energy_kWh × Grid_Intensity (g/kWh)

Water (L) = Energy_kWh × WUE (L/kWh)

Where:

  • PUE (Power Usage Effectiveness) — ratio of total facility power to IT equipment power. Default: 1.56 (industry average).
  • Grid Intensity — carbon intensity of the electricity grid. Default: 124 gCO₂/kWh (UK average 2024).
  • WUE (Water Usage Effectiveness) — litres of water per kWh of IT power. Default: 1.9 L/kWh.

Energy Per Action

Text LLM (token-aware)

Energy_J = Overhead_J(model_class) + Tokens_total × J_per_token(model_class)
Model ClassJ/token (mid)Overhead (J)Range
Small (≤8B)1.0500.5–1.5
Medium (8–30B)2.21001.5–3.0
Large (30–70B)4.01503.0–6.0
Frontier (≥100B / MoE)9.02506.0–15.0

When token counts are unavailable, we use a conservative default of 0.30 Wh per text prompt (before PUE).

Image Generation

Standard image generation (512–768 px, ~30–50 steps): ~0.50 Wh/image before PUE. High-quality or large canvas generation can reach up to 3 Wh/image.

Concrete Examples

Using UK defaults: PUE = 1.56, Grid = 124 gCO₂/kWh, WUE = 1.9 L/kWh

Typical text prompt (no tokens)

Energy

0.000468 kWh

CO₂

0.058 g

Water

0.89 mL

Long reasoning (Large model, ~2,000 tokens)

Energy

0.00353 kWh

CO₂

0.44 g

Water

6.7 mL

1 image @ 768px, 40 steps

Energy

0.00078 kWh

CO₂

0.097 g

Water

1.5 mL

Platform Mapping

Each AI platform is mapped to a default model class based on their publicly known model sizes:

ChatGPTLarge
ClaudeMedium
GeminiLarge
PerplexityMedium
You.comMedium
Character.AIMedium
PoeMedium

User Overrides

For more accurate estimates, users can override the following defaults in the extension settings:

  • PUE — if your provider publishes region-specific PUE (e.g. 1.1–1.3 for hyperscale facilities)
  • Grid Intensity — varies significantly by country (e.g. France ~50 g/kWh vs Poland ~650 g/kWh)
  • WUE — some facilities achieve lower WUE depending on climate and cooling technology

Versioning

Every carbon log stores a factors_version identifier. When we update our coefficients, we bump the version (e.g. 2025-10-v12025-11-v1) to preserve reproducibility. Historical data always reflects the factors version used at the time of calculation.

Limitations

  • Estimates are approximations based on published averages, not direct measurements from AI providers
  • Token counts may be estimated when not directly available from the platform
  • Model class assignments are based on publicly known information and may not reflect real-time model routing
  • Water usage includes both direct cooling water and water embodied in electricity generation

We are committed to improving accuracy as more provider-specific data becomes available.

References

  • Uptime Institute, Global Data Center Survey (PUE trends)
  • Carbon Intensity (GB) API and UK electricity analyses (grid intensity)
  • Data center WUE guides and operator disclosures (typical WUE ranges)
  • Academic and industry measurement of LLM inference energy (token-wise energy, GPU power studies)

Glossary

PUE
Power Usage Effectiveness — ratio of total facility power to IT equipment power.
WUE
Water Usage Effectiveness — litres of water per kWh used by IT equipment.
Grid Intensity
Carbon intensity (gCO₂/kWh) of the electricity grid supplying compute.
kWh
Kilowatt-hour, a unit of energy.
MoE
Mixture of Experts — an architecture where only a subset of model parameters are active per query.