OPEN LINK ↗
// 1h agoOPENSOURCE RELEASE
Berget AI open-sources per-request CO₂ method
Berget AI has released an MIT-licensed method and calculator for estimating the carbon footprint of individual AI requests. Its Kimi K3 example estimates 0.38 grams of CO₂e in Sweden versus 3.2 grams in Texas on identical hardware, driven largely by regional grid intensity.
// ANALYSIS
Per-request emissions reporting could make AI infrastructure location a procurement metric instead of a marketing claim, though the figures remain methodology-dependent estimates.
- –Calculates emissions using model workload, hardware, cooling, grid intensity, and embodied hardware impact
- –Shows open-weight model users can reduce emissions by choosing cleaner inference regions
- –Follows the Green Software Foundation’s SCI for AI specification and publishes its assumptions openly
- –Excludes training emissions and currently relies on average rather than marginal grid intensity
- –Gives developers a practical way to compare sustainability alongside latency, cost, privacy, and model quality
// TAGS
berget-aiinferenceobservabilityopen-sourceresearchinfrastructure
DISCOVERED
1h ago
2026-10-23
PUBLISHED
3h ago
2026-10-23
RELEVANCE
8/ 10
AUTHOR
jlundberg