GPU Core Utilization is the most critical metric for identifying underutilized GPUs in an AI data center. This metric, accessible via NVIDIA's nvidia-smi or DCGM, measures the percentage of time GPU cores are actively processing tasks, directly indicating whether GPUs are underperforming due to idle time or poor workload distribution. Low core utilization suggests inefficient task scheduling or bottlenecks elsewhere (e.g., CPU, I/O). Option B (memory usage) is important but secondary, as high memory use doesn't guarantee core activity. Option C (network bandwidth) affects distributed workloads, not local GPU use. Option D (uptime) ensures availability, not utilization. NVIDIA's monitoring guidelines prioritize core utilization for performance diagnostics.