I think some people only see one side of the trade-offs here.
A counter example would be a start-up I worked with that started out on a smaller provider. When we launched a beta of the application, we unexpectedly got invitations for national TV coverage with less than 24 hours notice. We asked the provider for everything they could give us, but they only had a limited amount of unused capacity themselves so there was little they could do. Of course we couldn't come near to soaking up the traffic spike, and lost a huge potential number of positive user experiences.
That sort of thing doesn't happen to you on Amazon or the other largest on demand providers.
That's why you start with Amazon. And then once you reach scale you slowly start building your own private cloud. Its much more efficient and costs a lot less. We have been doing the same in our startup.
Something else I think people fail to consider is what I would call a hybrid approach.
For example, even if you mostly operate on your own infrastructure, keep a replica of your data tier going on EC2, along with freshly configured images for the rest of your moving pieces. This way you have an instant pop up B site for disaster scenarios.
There are a lot of other possible concepts. Keeping a dataset at amazon to use for map reduce jobs is convenient. Or even if you want to keep your infrastructure virtual, being able to price arbitrage across a couple providers could be a powerful cost reduction. The price of entry for all this is devops automation, but I think people are now mostly convinced that's worth it.
I think it's the best way to go, but also for a resilience approach, I want to be able to ditch a datacenter in 15min. If you can ditch the datacenter, why not ditch the whole provider?
So we tried to keep the deployment scripts ready for a competitor at all times, having backups we could restore there etc. And if you grow big enough, this becomes a bargaining chip with your provider too: I can ditch you in 15 min.
A counter example would be a start-up I worked with that started out on a smaller provider. When we launched a beta of the application, we unexpectedly got invitations for national TV coverage with less than 24 hours notice. We asked the provider for everything they could give us, but they only had a limited amount of unused capacity themselves so there was little they could do. Of course we couldn't come near to soaking up the traffic spike, and lost a huge potential number of positive user experiences.
That sort of thing doesn't happen to you on Amazon or the other largest on demand providers.