Running AI at scale has an invisible footprint: both in cost and carbon. This session quantifies them—and shows how to reduce both without compromising performance. We’ll explore GPU scheduling efficiency, workload bin-packing, autoscaling strategies, and real-time CO₂e estimation using OpenTelemetry + FinOps data. You’ll see how a few configuration tweaks in Kubernetes (topology-aware scheduling, node affinity, mixed GPU pools) can yield double-digit savings in power and euros. We’ll finish with an open-source toolkit to track emissions per request and integrate sustainability directly into your CI/CD pipelines.