Engineering is always about making the right compromises. Why does it have to be 100% dedicated or 100% cloud? What most people don't realize is that it really is a continuum.

For example, why not run the disk performance sensitive DB server on a dedicated machine, while fronting the whole arrangement with proxies and app-servers hosted in the cloud? Ok, so there are latency considerations to be made, but you can see that mixed architectures can make sense.

I think what's stopping people from considering this is that there haven't been good cross-provider network virtualization solutions available. But if you could create your own network topology and your own layer 2 broadcast domains, no matter where your machines are located, things are starting to look up.

There are a number of network virtualization providers out there now, which you might want to look at to see what's possible. Disclaimer: I work for vCider ( http://vcider.com ), which provides solutions for on-demand virtualized networks, which can span providers and data centers.