Measuring the Carbon Intensity of AI in Cloud Instances

Data scientists have no easy or reliable access to the greenhouse-gas impact of their own AI workloads, which precludes developing actionable tactics to reduce it.

A framework for measuring SOFTWARE carbon intensity using LOCATION-BASED and TIME-SPECIFIC MARGINAL emissions data per unit of energy, rather than average grid intensity. Provides measured operational carbon intensity for a set of modern NLP and computer-vision models across model sizes, including pretraining of a 6.1-billion-parameter language model. Then evaluates three concrete reduction strategies on Microsoft Azure: choosing instances in different geographic regions, running at different TIMES OF DAY, and dynamically PAUSING instances when marginal carbon intensity exceeds a threshold.

Not a compute speedup. The measured findings: the GEOGRAPHIC REGION of the data centre plays a significant role in carbon intensity for a given cloud instance, and choosing an appropriate region can have the LARGEST operational-emissions reduction impact of the strategies evaluated; TIME OF DAY has notable impact as well; and dynamic pausing above a marginal-intensity threshold also helps. The headline is therefore that WHERE and WHEN you run is a bigger lever than what you run - a scheduling result, not a silicon result.

ADOPTION: in-production as methodology - marginal-emissions-based carbon accounting and region/time-aware scheduling are exactly what cloud sustainability tooling and carbon-aware schedulers implement, and the paper is the standard citation for the software-carbon-intensity measurement framing. VENUE: ACM FAccT 2022 per the arXiv comment field ('In ACM Conference on Fairness, Accountability, and Transparency (ACM FAccT) 2022'); the schema has no 'facct' enum value so venue is recorded as 'other' with the specific venue named in notes, which is the honest encoding rather than forcing a wrong enum. TIES DIRECTLY TO THE OTHER GREEN RECORD IN THIS SLICE: this paper's region-and-time result is the operating-point argument behind the carbon-aware routing record already in this repo (pwr2/xwind renewable-site inference router), and the two should be read together - this one establishes that marginal carbon intensity varies by region and hour, that one routes to renewable sites. Note the models measured are 2022-era; the per-model intensities are not transferable to current models, only the METHODOLOGY is. RESOLVED (2026-10-04 cross-reference repair): 'nvidia-v100-sxm2' -> nvidia-v100-sxm2-32gb, the canonical spelling of that slug and the only V100 record in the repo. Unambiguous: the old value was a prefix of exactly one id. AFFILIATIONS FILLED 2026-10-04 (paper-verification pass): from the Semantic Scholar graph API (paper/arXiv:2206.05229), which is the only machine-readable source that carries author affiliations — the arXiv Atom feed returns author names only, and no ORCID is exposed by either API. One known upstream typo corrected: "Tel Aviv Univesity" -> "Tel Aviv University". S2 affiliation coverage is partial (120 of 249 corpus records carry any), so these lists may be incomplete rather than exhaustive. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. Peer-reviewed venue confirmed as ACM FAccT 2022 (same). This is an additive track annotation; `venue` is unchanged at 'other' and no status change is implied by the track alone. VENUE RECLASSIFIED 2026-10-10 (crawler/w12-venue-rest): venue 'other' -> 'facct', venue_kind 'conference'. Confirmed by Crossref DOI 10.1145/3531146.3533234 ('Measuring the Carbon Intensity of AI in Cloud Instances', 2022 ACM Conference on Fairness, Accountability, and Transparency) and agreeing source record notes.