The rapid expansion of high-performance computing (HPC), artificial intelligence (AI), and cloud-based services has significantly increased the energy demand of modern data centers. Among the key challenges in maintaining operational efficiency is managing excess heat from densely packed computing equipment. While liquid-cooling technologies offer superior thermal performance, they are often operated conservatively, leading to excessive energy use through overcooling. This study investigates overcooling in data centers using operational data from the Frontier supercomputer, currently the world’s fastest publicly available exascale system. A linear regression model was developed to predict baseline coolant return temperatures using compute power and waste heat as inputs, and its performance was validated using standard regression metrics (R² = 0.357, MAE = 2.76 °C). Overcooling was identified when actual return temperatures were at least 1.5 °C below the predicted baseline. The analysis revealed that approximately 6.9% of the cooling effort could be reduced without compromising thermal safety margins. The study also translates the energy implications of avoidable overcooling into public-scale usage equivalents, showing potential annual savings exceeding 2.5 million kilowatt-hours. These findings demonstrate the potential of AI-assisted thermal modeling as a lightweight and interpretable method to improve cooling efficiency, reduce operational costs, and support sustainable data center management.