EC2 shines for very small, very large or very dynamic deployments (jitter of at least +/- 20 servers).
Most startups outgrow "very small" (2-4 Servers) quickly and then stick around in the mid-range (5-30 servers) for an eternity.
During this mid-range phase EC2 is usually 2x-3x more expensive than rented dedicated servers.
2.) For small systems, don't forget hybrids! EC2 is a _steal_ as a disaster contingency option. Run your own stuff. Replicate state to EC2. Keep your configuration managed and up to date both places. Build your routing horizon (dns most likely, perhaps anycast) to be able to shift traffic between the two fluidly and quickly.
Moron at your data center kills the power? No big deal, spin up instances and keep users happy. Got a big but occasional data processing job? No problem, and less ways the job can affect the production system.
The open source tooling for doing this kinda sucks but it's getting better fast. It's a huge opportunity for smaller businesses to level advantage vs large organizations with multiple in house data centers.
That expense might be worth it, though, if your startup goes unexpectedly viral and the site goes down for hours or days while you wait for your host to provision new dedicated servers for you.