The exponential growth of digital services and artificial intelligence (AI) workloads has positioned data centers as critical infrastructure while simultaneously rendering them significant contributors to global energy consumption. Traditional monitoring and control paradigms, reliant on static thresholds and reactive rule-based systems, are increasingly inadequate for managing the dynamic complexity of modern facilities. This article presents a comprehensive examination of AI-driven methodologies for data center power monitoring and optimization. We synthesize contemporary research across multiple domains: machine learning-based power consumption modeling, non-intrusive workload disaggregation, predictive cooling control, and grid-interactive flexible operations. Drawing upon recent advances in deep learning, anomaly detection, and reinforcement learning, we analyze how AI transforms raw telemetry into actionable intelligence. The article evaluates architectural frameworks, algorithmic approaches including LSTM networks, Transformer models, and Isolation Forests, and presents validation results from production deployments. We demonstrate that integrated AI monitoring systems can reduce cooling energy by 15 40%, improve Power Usage Effectiveness (PUE) by up to 15%, and enable dynamic power shedding of 25 40% during grid stress events while preserving critical workload performance. Furthermore, we examine emerging metrics such as Power Compute Effectiveness (PCE) and the role of AI-driven digital twins in optimizing behind-the-meter renewable energy integration. Finally, we discuss implementation considerations, explainability requirements, and future research directions toward autonomous, self-optimizing data center infrastructures.