Asynchronous design just removes some of the memory pressure and a chunk of CPU loss from context switching. If your thread stacks are short and you have plentiful memory, you're still CPU-bound.

But yes, if you have to hit a serialized disk somewhere for each and every request, then you really are screwed. But I don't think that's really often the case, is it? If it were, it would be a miracle that most web pages ever come up in under a second.