How we calculate the score

Global constants

"Queries per employee per day" is an internal assumption pending public benchmarking — override-capable in a future advanced mode.

Score formula

avgWh   = Σ ( whPerQuery[m] × usageShare[m] )    // weighted by mix, not membership
avgEff  = Σ ( efficiencyScore[m] × usageShare[m] )

llmKwh        = employees × 40 × 230 × avgWh / 1000
serverKwh     = servers × 4500
deviceKwh     = employees × 750
totalKwh      = llmKwh + serverKwh + deviceKwh
tCO2e         = totalKwh × gridKgPerKwh / 1000          // gridKgPerKwh from region picker (UK default 0.207)

perEmpKwh        = totalKwh / employees
intensityScore   = clamp((3000 − perEmpKwh) / 2200, 0, 1)
gridDelta        = (UK_factor − gridKgPerKwh) / UK_factor
gridAdjust       = clamp(gridDelta × 12, −10, +10)       // ±10 pt cap, cleaner grids gain
score (0–100)    = round_to_5( (avgEff × 0.55 + intensityScore × 0.45) × 100 + gridAdjust )

grade  A ≥ 85 · B ≥ 70 · C ≥ 55 · D ≥ 40 · E otherwise

Efficiency weights: A = 1.00, B = 0.75, C = 0.45, D = 0.20. Weighting by usage share (not membership) means adding an efficient model to your selection only improves your score if you actually shift prompts to it.

Model catalogue

Per-query energy in watt-hours, assuming ~500-token responses. Confidence reflects how directly the figure traces to public data.

Grid emission factors by region

Location-based 2024 figures used by the region selector. Pick the closest match to where your compute runs — not where your office is. Cleaner-than-UK grids earn up to +10 score points; dirtier grids lose up to 10.

All figures are location-based (grid average). Market-based numbers using renewable energy certificates can be lower — disclose both if you publish externally. Where a provider's reported region figure differs from the underlying national grid, we use the provider figure (e.g. GCP europe-north1).

Sources

• epoch-2024Epoch AI — How much energy does ChatGPT use? (2024)(2024)
• hf-energy-2024Hugging Face AI Energy Score (2024)(2024)
• google-sust-2024Google 2024 Environmental Report(2024)
• iea-2024IEA — Electricity 2024, data-centre demand(2024)
• beis-2024UK DESNZ (BEIS) GHG conversion factors 2024(2024)
• uptime-pue-2024Uptime Institute Global Data Center Survey 2024 (PUE)(2024)
• woodland-2023UK Woodland Carbon Code — sequestration rates(2023)
• ember-2024Ember — Yearly Electricity Data 2024(2024)
• eia-2024US EIA — eGRID 2024 subregion emission rates(2024)
• aws-cust-carbonAWS Customer Carbon Footprint Tool methodology(2024)
• gcp-regionsGoogle Cloud — Carbon free energy by region(2024)
• azure-emissionsMicrosoft Azure — Emissions Impact Dashboard(2024)

Changelog

• v1.2 — 2026-06-06

  Added cloud-region grid selector with ~25 regions and a custom factor input. Grid factor is now an input to calculate(); UK remains the default. Score gains a capped ±10 point bonus/penalty so location materially affects the result without dominating model choice.

• v1.1 — 2026-06-06

  Switched LLM weighting from membership-based to usage-share-weighted. Added Wh ranges and confidence labels per model. Separated AI emissions from total digital emissions in the result card. Rounded score to nearest 5.

• v1.0 — 2025-09

  Initial public release with 11 models, UK grid factor, equal-weight LLM averaging.