Study Maps Lower-Carbon LLM Training Across US Grids

Researchers Sudhir Ganesh Gopalakrishnan, Arvind P. Ravikumar and Sergio Castellanos evaluated how shifting large language model training across time or locations could reduce electricity-related CO₂ emissions. Their study, published in Environmental Research: Infrastructure and Sustainability in 2026, used 2025 marginal operating CO₂ emissions rate (MOER) data from WattTime across 43 subregions in four US grid operators: CAISO, ERCOT, MISO and PJM. Meta’s LLaMA 3.1 model family served as the representative workload.

The researchers simulated four scenarios: business as usual with carbon-informed siting but no real-time operational flexibility (S1); pausing training at varying emissions thresholds (S2); ideal workload migration between grid regions (S3); and migration that accounts for its overhead (S4). Results varied by grid. In ERCOT, shifting training between wind-rich and solar-heavy zones cut emissions by up to 36%; an inter-ISO setup cut them by 67.3%, with CAISO and ERCOT taking over 80% of training hours.

In PJM, where emissions profiles were similar across subregions, spatial shifts reduced emissions by only 0.3%, while pausing training cut emissions by up to 17.2% compared with S1. The emissions and training-time tradeoff for pausing was nonlinear: moderate emissions reductions required only marginally longer training, but approaching zero emissions took nearly four times the baseline training duration. The researchers said grid-specific strategies can guide near-term decarbonization and identified empirical measurement of data-migration overhead as an area for further improvement.

Researchers Sudhir Ganesh Gopalakrishnan, Arvind P. Ravikumar and Sergio Castellanos evaluated how shifting large language model training across time or locations could reduce electricity-related CO₂ emissions. Their study, published in Environmental Research: Infrastructure and Sustainability in 2026, used 2025 marginal operating CO₂ emissions rate (MOER) data from WattTime across 43 subregions in four US grid operators: CAISO, ERCOT, MISO and PJM. Meta’s LLaMA 3.1 model family served as the representative workload.

The researchers simulated four scenarios: business as usual with carbon-informed siting but no real-time operational flexibility (S1); pausing training at varying emissions thresholds (S2); ideal workload migration between grid regions (S3); and migration that accounts for its overhead (S4). Results varied by grid. In ERCOT, shifting training between wind-rich and solar-heavy zones cut emissions by up to 36%; an inter-ISO setup cut them by 67.3%, with CAISO and ERCOT taking over 80% of training hours.

In PJM, where emissions profiles were similar across subregions, spatial shifts reduced emissions by only 0.3%, while pausing training cut emissions by up to 17.2% compared with S1. The emissions and training-time tradeoff for pausing was nonlinear: moderate emissions reductions required only marginally longer training, but approaching zero emissions took nearly four times the baseline training duration. The researchers said grid-specific strategies can guide near-term decarbonization and identified empirical measurement of data-migration overhead as an area for further improvement.