for local LLM, obvious choice:

idle = doesn't have to be H200 speeds

uses macs = local LLM can use all the ram, be smarter

as for hosted models, potato is tasty