Unless things have changed dramatically*, there are rules that make it harder to use commercial cloud computing, as not all can guarantee that the services will only be hosted in the U.S.
Most agency cloud computing efforts are for internal number crunching (eg, scientific computing), not public facing websites. When they *have* gone and done it, they couldn't come up with a viable cost model for different groups to be willing to convert to the service. (Oh ... you can't tell me the price, because you need to break-even, and you don't know how many people will agree to use it? Okay, that's a decent price; it's not that much more than what we pay now ... oh wait, I have to pay for 3 VMs for prod / test / dev?)
The problem w/ building up a cluster to scale is that it means that you have inefficiencies of having idle machines; the way to get around this is to have lots of unrelated services running on the same system so that they shouldn't all need to max out at once.
In practice, it's often easier to switch to a 'low resource' version of the site when you start getting hit heavy -- drop all of the pretty images cluttering up pages, and just serve the basic content. Webserver tuning also helps dramatically ... as simple as splitting your static content off to a seperate server (so that you can repoint it at a CDN if necessary), while your local servers take the brunt of the dynamic requests. (and possibly make the site less 'interactive' in times of high load.)
* which wouldn't surprise me, as I work for a federal contractor and we seem to be the last ones to know about policy changes ... I once spent more than a year dealing with waiver paperwork only to find that by the time it had been granted that it had been allowed for 6+ months.