How Much Can Intelligent Control Decarbonize Liquid-Cooled AI Data Centers?

AI workloads are increasing data-center electricity demand, but the climate benefit of coordinated operational control remains poorly quantified. We evaluate carbon- and water-aware control for liquid-cooled AI data centers using a digital twin that couples thermo-fluid dynamics, workload heat generation, weather, and grid signals. Constrained multi-objective Soft Actor-Critic (CMO-SAC) jointly controls cooling and, in joint environments, workload dispatch under thermal and service constraints. Across 500 evaluation episodes, CMO-SAC lowers DataCenter PUE from 1.18 for tuned PID to 1.04 while reducing cooling power from 9.8 to 4.8kW. At equal IT work, this PUE change reduces facility electricity by 11.9\%. Under explicit assumptions for a representative 10-MW IT load, the efficiency result corresponds to 12.3~GWh, 4.9~ktCO$_2$e, and 18.4 million liters avoided annually, or 0.49~ktCO$_2$e per MW-year. Separately, a grid-aware policy reduces on-peak load by approximately 40\%; an illustrative marginal-carbon scenario values this flexibility at an additional 1.49ktCO$_2$e per year. The experiments distinguish measured outcomes from scenario-derived impacts and show that PUE-only optimization violates thermal and service constraints.