Local inference users are all about sampling, but users addicted to commercial inference services are wary of sampling, because they have to pay by the token.