What happens when your high-priority RL training job and your mission-critical LLM inference spike hit the same cluster at the same time? Building a multi-tenant AI platform is a game of constant compromise. In this session, we’ll go beyond the "perfect" reference architecture to discuss the messy reality of multi-cluster orchestration. We'll share peer-to-peer insights on setting preemption policies that don’t alienate researchers, managing "goodput" during traffic surges, and the architectural shifts required to turn a collection of Kubernetes clusters into a unified, efficient AI factory.