The Berkeley RAD Group paper 'Above the Clouds: A Berkeley View of Cloud Computing has a nice overview of the directions they think cloud computing will follow. It is interesting to compare it to the Xenoservers project (Jan 2003) at Cambridge, which basically kicked off the whole server virtualisation market with Xen, the VM hypervisor.
Six years has given us some more data points, and it seems reasonable that owning server farms and data centres will turn into a real estate play (high capital costs, low margins) rather than a technology-driven, high margin industry. We might expect that the folks who are good at the real estate business will be successful in building cloud data centres.
The Berkeley paper suggests that there are certain applications where the cloud model doesn't work, citing real-time stock trading as an example where WAN latency is too high, or scientific computing as an example where it might be necessary to FedEx physical media to the Cloud. I think this misses a potential direction.
It is true that a bank would not be able to port their (low latency) stock trading systems to Amazon Elastic cloud, but there is no reason that an entrepreneur would not start offering cloud computing in an environment with a low latency connection to major stock exchanges such as in Canary Wharf, London. It would surely cost a lot more than CPU time in a consumer-grade data centre, and there would have to be higher levels of security in place, but none of that seems to fundamentally change the market for utility computing.
The other example that was cited is when there is a large amount of data being created. For the scientific community it might be hard to move the experiment to the data centre. In these cases we might expect to find that smaller data centres are built close to the thing generating the data. For a university this might mean that a new physics lab would be designed with space for a few containers of servers. A science park that wanted to attract Biotech start-ups would want to offer a local cloud with very high bandwidth connections to the offices. A Hollywood film might rent a container load of servers and bring it to the filming location to handle the data from digital film cameras.
The utility model for computation works in all these cases, not just the 'cheapest page hit' that has characterised the current offerings.