But at odd sizes, density becomes harder to achieve. Local disk I/O tends to be harder to deal with with small machines.
It's a lot easier to satisfy 4 8-VCPU guests than 32 1-VCPU guests if you've only got a handful of spinning disks...
Network bandwidth is also in this category. If they are trying to make a fixed amount of bandwidth available to each virtual instance, there is a maximum number of machines that can run behind a physical network adapter.
I suspect cheapskates will take however little you give them in terms of I/O. I'm opposed to local disk, but that's a different discussion...
I'm curious what you would replace local disk with. EBS? Now all the root devices in your system are dependent on a functioning network, greatly magnifying the impact of network failures. I am all for I/O optimized instances with relatively little CPU or RAM. I like having my disks local to my instances though.