We have very bursty traffic and run our baseload of ~ 30 severs (mostly m4.larges, with a few smaller utility servers as well as larger servers for back end services) most of the time, but when we need to scale, we ramp up to about 2 - 4 times that, but usually only for a few hours at a time. We also do some batch processing that takes a few high-memory servers for about a day each week.
We monitor costs pretty closely and our monthly AWS spend is still less than the coloc fees, and we are much more scalable now and have multi-datacenter (AZ) redundancy, and by the end of the year, we will have a cold spare in a second region (which will cost us very little since it will be mostly powered off)
We no longer need to vastly over-provision servers to handle our peak load and batch processing needs.
At some price point, if the prices are too high, it will be cheaper to hire extra people, and get a cage in a few datacenters and some physical hardware.