Methodology: score, electricity, water, CO₂

Overall score /100

Weighted mean of 1–5 sub-scores scaled to 100: Quality 35, Price 25, Security 20, Electricity 10, Water 10. Missing sub-score = excluded and weights renormalized ('partial data'). Below 35% weight available: no score.

The score never reads affiliation.

Electricity, water and CO₂ (estimates)

Never presented as service measurements. Each card shows a low/high range and an 'estimate' label.

Electricity: published baselines (e.g. Luccioni et al. 2023, arXiv 2311.16863: 1.38 Wh/image SD 1.5; 11.49 Wh/image SDXL on a lab GPU; Google 2025: 0.24 Wh median Gemini text prompt, arXiv 2508.15734). Scaled by modality in content/impact-estimates.json.

Water: mL ≈ Wh × WUE (L/kWh). Google Category 2 WUE = 1.15 L/kWh (same paper). Otherwise a documented industry-default average, explicitly labelled.

CO₂: gCO₂ ≈ (Wh / 1000) × grid intensity (gCO₂/kWh). Mix: Google market-based 94 g/kWh (2024, same paper); else Ember US/world averages when the region is undisclosed.

Hosting: provider and regions from official pages (e.g. OpenAI–Microsoft Azure partnership). If unknown: 'region not disclosed, default average used'.

Primary sources

Luccioni, Jernite, Strubell — Power Hungry Processing (arXiv 2311.16863).

Google — Measuring the environmental impact of delivering AI at Google Scale (arXiv 2508.15734).

OpenAI — Continuing Microsoft partnership (Azure hosting).

Ember Electricity Data Explorer — public grid intensities.