SOMA AI's four-tier method slashes AI inference emissions overestimates by 10–40x

A new preprint from SOMA AI researcher Guillermo Llopis tackles a gap in corporate GHG accounting: AI inference services (API subscriptions, enterprise chat tools, SaaS with embedded AI) fall under Scope 3 Category 1 per CSRD, but no standardized methodology exists. Current practice either omits the category or uses generic ICT sector EEIO factors, which overestimate AI inference emissions by 10–40x relative to physically derived alternatives. Llopis proposes a four-tier framework matching estimation precision to available data, from direct token-based physical estimation (using GPU energy benchmarks from the MLCommons Leaderboard v3 and regional grid carbon intensities from EPA eGRID 2023 and Ember 2023) down to a spend-based EEIO fallback for services without usage data. Applied to a 200-person European firm, the framework yields total emissions below 1 tCO2e, demonstrating that the compliance challenge is methodological rather than magnitude-driven.

The paper also documents a water-carbon trade-off absent from current ESG tools: Sweden's hydro-dominated grid delivers the lowest carbon intensity in the dataset but the highest water footprint, directly impacting data center location strategy. The methodology uses peer-reviewed GPU benchmarks, confirmed grid carbon intensities, and published water use effectiveness data (Li et al., 2025). For companies forced to report AI inference emissions under CSRD (fiscal years starting Jan 2024), this framework provides a practical, accurate path to compliance while surfacing hidden sustainability trade-offs. The preprint is available on arXiv (2606.10660) with a data repository.