Whats the problem? The LLM runs on a big fat server, the devices only have to call APIs, no?