don't know about training, but small-scale on-premise inference setups are certainly easier to build and operate than to stock gpus with all the cooling etc :)