AI Sustainability: SCI Report

This page estimates the Software Carbon Intensity (SCI) of AI calls made by ShoppingPartnerLab, using the Green Software Foundation method standardised as ISO/IEC 21031:2024: SCI = (E × I + M) per R. The figures are estimates and a self-assessment; they are not externally verified.

Functional unit

R = one AI API call. The score is shown as grams CO₂eq per call.

Coverage is incomplete

This is not a complete AI Factory footprint. The carbon estimate covers only calls written to the carbon-instrumented routing log. Julia answers and embeddings are not included. Direct Gateway calls and separately recorded agent runs, including work by Victor, Alexandra, content pipelines and voice shopping, can fall outside it. The sources overlap and have no shared request ID, so their totals must not be added together.

In the week 20–27 September 2026, the AI Gateway recorded 8,341 request attempts (8,310 successful, 30 client errors and 1 with another status), while the carbon log recorded 2 calls and 466 tokens. The routing-layer log's last call was 24 September.

AI Factory usage evidence

The carbon-instrumented routing log contains 11,417 calls and 9,178,511 tokens from 28 December 2025 through 24 September 2026. Under the current disclosed formula, that sample represents an estimated 1,611.26 g CO₂eq: 1,606.64 g from energy and 4.62 g from hardware.

Team meetings are included where they reached the carbon log: 50 calls between 28 December 2025 and 27 February 2026, with 948,378 tokens and an estimated 87.65 g CO₂eq under the current formula. The operational run log separately contains 92 completed team-meeting runs, but no tokens or model were stored for those rows, so no defensible emissions estimate can be calculated for all 92.

Across the AI Factory, the operational run log contains 16,872 runs and 2,311,993 recorded tokens; only 351 runs contain a non-zero token count. These are operational runs, not model requests, and are not added to the Gateway or carbon-log totals.

Method and assumptions

AI providers do not publish energy use per request. All values are estimates per model class, not measurements, and no uncertainty range is calculated yet.

  • Energy per token (E): 0.03 (lite), 0.10 (flash), 0.30 (standard), 0.80 (pro) and 2.0 (frontier) mWh per token, times a datacenter overhead (PUE) of 1.10. The code stores these as 0.000000030 to 0.000002000 kWh per token.
  • Grid intensity (I): 280 gCO₂eq/kWh (Netherlands/Ireland mix, Ember 2024) for every call. Providers may run in other regions.
  • Hardware share (M): 150,000 g CO₂eq divided by (4 × 365 × 24 × 3,600) = 0.001189 g per GPU-second (Gupta et al. 2022). We assume one GPU-equivalent and 0.05 (lite), 0.15 (flash), 0.40 (standard), 1.00 (pro) or 2.50 (frontier) GPU-seconds per non-cached call. A cache hit is assigned 0.000001 g.
  • No purchased carbon offsets are subtracted.

Estimated avoided emissions

The report also shows estimated avoided emissions from model choice and caching, compared with a hypothetical scenario in which every call runs on a large "pro"-class model. This is a comparison with an assumed scenario, not a measured reduction, and it is never subtracted from the SCI score.

Trend

Grams CO₂eq per call are shown per month from December 2025, split by the task types actually stored: chat_response, content_generation, fact_checking, strategic_analysis, seo, translation and other. These categories describe the measured sample, not individual AI Factory agents. Julia answers and embeddings are not included; direct Gateway requests can fall outside the log.

Figures are recalculated once per hour and load in the interactive version of this page.

See also: How we measure · AI transparency

ShoppingPartnerLab

ShoppingPartnerLab