I'm afraid without memory bandwidth they can't compete on inference either. Not on large scale. Which leaves local small models.