High Performance Computing at Amazon: A Cost Study

I've done similar cost calculations at various times. Assuming your application(s) can run on cost effective commodity hardware, co-location becomes cheaper once you can occupy roughly half a rack. This assumes a middle of the market data center, 1 full time sysadmin at 100k or equivalent smart hands hours and a 2:1 peak to average traffic level and dynamically scaling your ec2 instances accordingly.

If your traffic is peaky, but those peaks can be anticipated with at least a fraction of an hour's time, then ec2 can be a win at a larger host count.

And of course, it almost aways makes since for batch jobs like scraping.

Another interesting way to use ec2 is as a disaster recovery plan: keep your application configuration current in ami's and keep a db slave going there. If your primary co-location fails catastrophically you can be up at scale on ec2 as fast as your DNS propagates.

Overall hybrid models seem to be the most effective way to use ec2.