Carbon-aware scheduling of Large Language Model training across multiple US power grid operators: a marginal emissions analysis

Abstract The rapid expansion of data center infrastructure, accelerated by the growing computational demands of Generative Artificial Intelligence (GenAI), has led to an unprecedented increase in electricity consumption. As AI developers race to deploy GenAI products, time-to-market pressure drives continuous model training that demands baseload power, often from high-CO₂ emissions sources like natural gas and coal instead of intermittent renewables like wind and solar. Carbon-aware strategies that shift model training in real time to periods or locations with cleaner power on the margin have emerged as a near-term decarbonization lever, but the relative effectiveness of temporal flexibility (pausing model training during high-carbon hours) versus spatial flexibility (migrating workloads to cleaner grid regions) remains poorly understood under real-world constraints. This study evaluates these strategies using 2025 marginal operating CO₂ emissions rate (MOER) data from WattTime across 43 subregions spanning four US grid operators: CAISO, ERCOT, MISO, and PJM. Using Meta's LLaMA 3.1 model family as representative workloads, we simulate four scenarios: business-as-usual, reflecting a carbon-informed siting choice with no real-time operational flexibility (S1, BAU); temporal flexibility with varying pause thresholds (S2); ideal spatial flexibility (S3); and spatial flexibility incorporating migration overhead (S4). We find that effective strategies are grid-dependent. In grids with high subregional MOER heterogeneity, spatial flexibility dominates: ERCOT achieves up to 36% emissions reduction via intraday alternation between wind-rich and solar-heavy zones, and the inter-ISO configuration achieves 67.3% reduction, with CAISO and ERCOT absorbing over 80% of training hours. In PJM, where subregional MOER profiles are largely homogeneous, spatial flexibility achieves only 0.3% emissions reduction while temporal flexibility reaches up to 17.2% lesser emissions than S1. The emissions–training time tradeoff under temporal flexibility is nonlinear: moderate reductions require marginal increases in training time, but near-zero emissions require nearly four times the baseline training duration. These findings provide grid-specific guidance for developers pursuing near-term decarbonization, and highlight opportunities for further improvement through empirical characterization of data migration overhead.