But llms are relatively rarely used, and on the other hand, perf/latency is important to ux, and perf is variable(simple question, complex question, visual work).

Those demand are better fullfiled at the cloud.