I thought about this too, but the throughput over a network is incredibly slow. It’s not usable for interactive use.

That isn’t true. llama RPC is incredibly slow but staged splits in skippy are orders of magnitude faster.