Alessandro Stefouli-Vozza

The next bottleneck in your data center isn't compute. It's the power feed. As GPU clusters multiply, energy is becoming the scarcest resource in the AI-era data center — and the one Kubernetes knows least about. The scheduler that packs pods by CPU and memory is blind to watts, heat, and carbon. This talk makes energy a first-class scheduling signal. Using open-source tooling, we'll measure per-workload power draw with Kepler, shift batch and training jobs to cheaper, cleaner hours with Kueue and KEDA, and pack GPUs for maximum utilization-per-watt. You'll learn how to measure what every pod actually costs in watts, schedule around power caps, prices, and carbon intensity, raise GPU utilization instead of buying more GPUs, and avoid the failure modes: thermal throttling and starving batch queues. Same clusters, same workloads — just running where and when the watts are right.