Explore how IBM Research built Vela, a cloud-native AI supercomputer, to train foundational models on Kubernetes at scale. Learn about the challenges faced in supporting multiple frameworks like PyTorch, Ray, and Spark for diverse research teams. Discover the role of Multi-Cluster App Dispatcher (MCAD) in queuing custom resources for large-scale AI training, and its interaction with the underlying Kubernetes scheduler. Gain insights into the implementation of gang priority, gang preemption, and fault tolerance for training processes that span hundreds of GPUs and run for extended periods. This conference talk provides valuable information on scaling AI workloads in a Kubernetes environment for researchers and developers working with foundation models.