Every YC batch should have at least 1-2 big wins. They're betting that it's more important not to miss out on those than to save a little money trying to pick the winners. A great strategy if you can afford it.
Can you quantify a big win?
So either the homeruns have to be further out of the park than Heroku or (as is likely) the median return needs to keep things in balance a bit more.
I don't believe that's been borne out yet.