On-premise will be a lot more difficult than just launching a few pods in Kubernetes. These AI tools (LLMs / vision models) will require some high powered gpus as well.