This is only really true for compute instances that don't have big caches or long warmup times.
Even then, I think you'll find you want the bare metal the instances to run on to have high uptimes (on the order of years, not minutes), since the hardware with optimal $/perf can fit more and more workloads per machine (I think this is all Moore's law is doing to help compute these days). That means you need a decreasing number of physical machines to hold your workload. At some point your "cloud" has 10 nodes instead of 1000.
Fun exercise: "cloud scale" code is typically 5-100 slower per node than single machine scale up code.
How much money would you save by consolidating smaller workloads to big machines? More importantly, how much developer productivity would you gain by eliminating network latency / marshalling for internal requests?
I think you'll see an increase of developers "coding around" devops over the next few years. I could be wrong, of course.
