The PALS system optimizes power usage in large language model inference by treating GPU power caps as adjustable parameters, enhancing energy efficiency by up to 26.3% without requiring model retraining. This innovation could lead to more sustainable AI operations in data centers, addressing both energy consumption and performance quality of service.