The limits of Kubernetes for AI inference