This concept was more for inference on-demand vs training a model. For example, if there was an API call that had a model call as part of it's workflow. Training even a small model still requires serious compute power.