Running LLM inference on Kubernetes usually means stitching together a serving engine, GPU scheduling, model storage, caching, and routing before you get a single token back.

For sovereign AI, another set of questions appears: Where do the model weights live? Where are inference requests processed? Can workloads move between cloud, neocloud, and on-premises infrastructure without rebuilding the entire platform around another provider’s API?

Modelplane is an open-source, Kubernetes-native control plane for AI inference. Installed in your own environment, it provides a declarative API for orchestrating models, serving engines, and the infrastructure underneath them.

In this hands-on walkthrough, we will explore the Modelplane API surface, see how its resources fit together, deploy a model end to end, and send a request through an OpenAI-compatible endpoint. Along the way, we will examine GPU scheduling, model caching, placement, and routing, including how infrastructure boundaries determine where models run and where requests are served.

We will also separate what a control plane can enforce from what remains the operator’s responsibility. Sovereignty is not a checkbox: it emerges from architecture, deployment choices, and operational control.

Attendees will leave with a practical understanding of the components behind self-hosted AI inference and how a Kubernetes-native control plane can make sovereign infrastructure usable without hiding it behind a proprietary managed service.