Next token is mostly bandwidth bound, prefill/ingest can process tokens in parallel and starts becoming more compute heavy. Next token(s) with speculative deco

Next token is mostly bandwidth bound, prefill/ingest can process tokens in parallel and starts becoming more compute heavy. Next token(s) with speculative decode/draft model also becomes compute heavy since it processes several in parallel and only rolls back on mispredict.