#llm-carbon-estimation

Consumer-side tools must assume model architecture and a worldwide-average grid because providers publish almost nothing about model size, hardware, or data-centre efficiency. Every downstream estimate therefore carries wide, irreducible error bars.

Benchmark models on the same hardware for the same tasks, then publish results as star-band ratings on a leaderboard with a shareable label. An ENERGY STAR-style signal makes efficiency legible to non-experts and usable in procurement, pressuring providers to disclose.

Every tool in this space estimates differently, so numbers are not comparable across teams or vendors. A shared specification expresses software carbon as a rate — emissions per functional unit such as a request, user, or training run — so results can be compared and tracked over

When a generative model is called through a provider's API, hardware cannot be metered. Estimate per-request energy and CO2 from model size, token counts, and data-centre assumptions using a life-cycle method, giving API consumers an impact figure they otherwise cannot obtain.

AI Energy Score, launched in February 2025 at the Paris AI Action Summit by Hugging Face, Salesforce, Cohere and Carnegie Mellon University, benchmarks models on standardised NVIDIA H100 hardware across 10 tasks and ass…

The Software Carbon Intensity (SCI) specification reached ISO accreditation as ISO/IEC 21031:2024 in Q1 2024. It defines software carbon as a rate: SCI = (E × I) + M per functional unit R, using a rate rather than a tot…

EcoLogits is an open-source Python library that estimates energy and multi-criteria environmental impacts of generative AI API calls. It patches provider Python client libraries (OpenAI, Anthropic, Mistral AI, Cohere, G…