dedicated hardware. common. that might have been an option in the 90s. i want to be able to use as many cpus as i want when i need them. not have a bunch of machines rotting away idle.

I would actually recommend a hybrid approach. If you don't need too much memory there is a cpu heavy Amazon instance that you should scale up and down and have a VPN between a dedicated co-lo machine that handles the base load. If you need memory and cpu, its going to be costly to use EC2 unless you pay upfront for dedicated instances. In this case though the economics are probably not as favorable as getting some co-lo servers and letting them rot away. In short you will need to do a cost benefit analysis to see what makes sense.

For our service we are going to use a co-lo server for our processing with an Elastic Beanstalk frontend and use a VPC + OpenVPN setup to bridge the two. We will incur some bandwidth charges because of this, but the cross talk between the boxes is actually minimal, since the client will post directly to the co-lo box(es) when it needs to upload, etc.

I worked on the LHC's CMS data team about 2-3 years ago. We had thousands of machines crunching data. Comparing the cost to Amazon, we laughed at the deal we were getting with disposable hardware (compared to EC2 pricing). Even buying in bulk from Amazon, the dedicated hardware was cheaper.

Amazon is a huge premium over physical hardware. You use them when a) you want to scale immediately, b) you have financial reasons for not buying your own gear for long-term use (1-3 years), and c) you don't mind paying the premium.

The bigger instances (like high CPU) tend to be virtualized less aggressively, so they have good performance.

It's really cheap to try and see if it's good enough for you.

so you ask for a question saying you do "a lot of number crunching", then complain at the fact that dedicated hardware will do what you need if you actually process a lot of data?

ass.