Besides fill factor being hard to match, there is also scaling - you can't scale local inference 10x for a spike, but you can with cloud inference.