I think that you are missing the value of the opposite proposition - that you are able to turn it off and stop paying. This is an obvious value to someone who needs to be able to scale sideways, but having essentially no cost for provisioning and de-provisioning allows you to do things for development, testing and operations you cannot do without great expense outside of it. I worked at a number of banks 10 years ago and we can get the same functionality now for a few bucks that we did then by buying far too many spare E3500s.
We have a decent sized EC2 architecture right now and EBS snapshotting + provisioning on demand means that we can get all of the benefits of multiply redundant hardware at a fraction of the cost.
- We never need to upgrade software in place. Software gets freshly installed on a new cluster that duplicates our existing cluster and gets a increasing portion of traffic. Failover is as simple as a proxy cutover. The freedom to do config changes or 3rd party software upgrades and be able to just swap instances in and out enables us to experiment often and safely with production changes.
- Slave gets corrupted or falls out of sync with the master? A machine suffers a degraded disk? Bad net interface? Don't fix, replace.
- We can spin up an entire test stack that precisely mirrors our prod stack to stress test with significant cash outlay. Since this cost is low, we can stress test often and test the effect of s/w and config changes on performance easily. Stress tests the H/W outlays associated with them are traditionally hugely expensive if you want to test exactly like you run.
- Significant database migrations can take place on a slave that is easily promoted, snapshotted and cloned without having to deal with any long term ping ponging between instances.
- Copies of databases can be instantly provisioned and destroyed for data analysis needs. If analysis needs peak at 5x in a week, no one has to wait. Run 5 copies and shut them all off when you're done.
These are just a few quick things that we do with regularity. And I'm sure there are lots more opportunities to exploit these capabilities than we are thinking of.
Also - in terms of cost - don't forget to compare against reserved instance pricing.