Where I currently work (a stealth mode startup), we currently have a data aggregation engine running on between 20 and 120 instances. We take advantage of the

Where I currently work (a stealth mode startup), we currently have a data aggregation engine running on between 20 and 120 instances. We take advantage of the ability to spin up instances quickly, not in response to a spike in traffic (since we don't have any yet), but in response to the size of the queue of jobs (which varies depending on what we've collected). We only need 120 running for an hour or so, then we terminate them. The cost savings we get is from not having to run 120 physical boxes all the time, and there is no downside if the queue grows a little while new instances spin up.