I liked the tenant-rent analogy xlarge:large:small::house:room:couch.
tl;dr version: The author is insinuating that Reddit doesn't have enough money to rent a house, so its renting a couch, and so it can't complain if the other occupants have an all-night rave party.
Long, but well reasoned article on EBS. Basically comes down to-know your load, and be darned sure you have adequate metrics in place for tracking latency.
My take-away message from this: if you want predictable predictable performance using AWS you need to use largest instances and EBS block sizes. This suggests Amazon scales up but not down. As in, if you're using the smaller instances (because you're a startup, say, and don't have a great deal of traffic) you're hosed. On the other hand, a lot of the calculations I've seen suggest maintaining your own server infrastructure is cheaper than AWS and you get more predictable performance, so I wonder why anyone who has a lot of traffic would use AWS (except for handing traffic spikes or batch jobs). So to me it seems that AWS doesn't have a great use case. I'm interested in other viewpoints.
It depends on the application(s) you wish to run on AWS. If your EC2 instance runs an HTTP server that is backed by an external database like RDS, your I/O bottlenecks is not the EC2 instance's disk.
The article points out rather successfully that once you know at least something about the internals of each AWS service, you can fit them into your particular needs.
tl;dr version: The author is insinuating that Reddit doesn't have enough money to rent a house, so its renting a couch, and so it can't complain if the other occupants have an all-night rave party.