You mean the 'Netflix Model'? This is what they do all the time on AWS; when needed, spin up a bunch of VM's; test performance, drop the ones that perform worse and use the better ones.