Carbon-LLM/CSRD & Scope 3
Back

CSRD, ESRS E1, and Scope 3 for LLM usage

Purchased generative AI is a classic upstream indirect emission: activity data (tokens) × emission factor (gCO₂e per token), documented for auditors.

Where LLMs sit in the inventory
Most software companies buy inference from OpenAI, Anthropic, Google, Mistral, or Azure-hosted models. Those invoices are usually Scope 3 — value chain emissions you do not own directly.

Category 1 — Purchased goods and services is the typical anchor when you pay for API access to cloud LLMs: the service is a purchased input to your product or operations. carbon-llm produces month-level and tenant-level totals with a methodology section suitable for annexes and internal control testing — without claiming to replace your CSRD double materiality assessment.

ESRS E1 expects disclosure of gross Scope 3 where material; having consistent, API-derived numbers reduces reconciliation risk between engineering data and sustainability reporting.

What “good evidence” looks like
  • Token counts from provider APIs (not guessed from characters).
  • Model-level coefficients with named sources (LCA, benchmark paper, or disclosed methodology).
  • Versioned methodology text in downloadable reports.

Our methodology page lists references (Mistral LCA, Google Cloud inference blog, Ni et al. 2025, Frontiers 2025, etc.) and how confidence is labelled.