As jtchang said, yes it comes down to planning. Ask the founders (or if you are one you should know this) what your daily active uniques are. Also ask what la

As jtchang said, yes it comes down to planning. Ask the founders (or if you are one you should know this) what your daily active uniques are. Also ask what language you are getting code in. Java needs more headroom than c++ for instance. Then plan +40% capacity of whatever they tell you. Always use an Aws autoscaling group to shave costs, but price as if you had always on on demand pricing. Don't forget to think about supporting infrastructure. Do you need memcached? If you do how much? Will you run your own or use elasticache? It all comes down to architecting your system on a whiteboard with needed resources (CPU, ram, etc) and then take that and apply it to the the instance type matrix and see where it fits My experience is based on three startups each with 2n or greater redundancy each handling at least 500 req/s into the frontend