When I was working on Google Search what really astounded me is how we could leverage hundreds of machines in a single request and still have virtually no cost per search. The reason was that each search used a tiny amount of the total resources of those machines and for a very short time. A total search might have (made up numbers) one minute of computation time, but spread across 200 machines it only takes 300ms from start to finish.
That's the benefit the cloud will provide. You don't want to have a 1000-machine data center available at all times to store billions of possible documents and process your requests with low latency. If we went to a private-network model I fear that the turn-around time would be a lot closer to a human assistant. You'd ask it to do things and then it would get back to you sometime later (seconds? minutes? hours?) when it had finished it's research and come up with an answer.