Depends on what you scaling. Actually there is no borders in such approach. If you app overloaded with template parsing you could move it to another machine, if problem in "database" thing same solutions came as for NoSQL. There is no limit what to scale and how. You could have everything in one process exept something that need scaling and this would be more efficient than dividing things on databases, cgi and balancers.
What makes you think that IPC is the dominating cost?? It's usually network or disk bandwidth, not IPC. And once you scale past one machine, you can't avoid network chatter by definition.
