No, not really, current MoE limit the computation, not memory requirements. Router experts are not "sticky" enough to achieve what robrenaud describes - they'd have to be chosen per prompt, or at least per chunk, not per token.