This isn't our experience when focusing the application logic towards the client-side whenever possible and I don't agree GAE is only for prototyping/small apps. We started on EC2, two instances over a load balancer. We had to go for 2 medium instances as smalls couldn't scale fast enough when the traffic spiked.

We had to create the logic to deploy the application simultaneously to all running instances and writing monitoring code to ensure everything was scaling properly (the load balancer triggers just weren't up to it). Daily traffic 2k users, cost 200USD a month. About 4 non-trivial outages in the 5 months we were on AWS.

With GAE we push to git, deploy done. No setting up deployment, no complex monitoring for scaling. Daily traffic 15k users, costs 30USD a month. 1 non-trivial outage in around 20 months now, with 100% uptime during 2013.

We have run a reasonable scale, reasonable traffic production grade application for at least 5 months on both, long enough to make an extensive comparison for our application. GAE isn't perfect, but the cost savings in terms of our developer time make the charge-for cost savings (200 vs 30) look like very small fry.