These are thoughtful comments, here's a couple more
> his basically limits this "supercomputer" to embarrassingly parallel problems that have very small specifications
note support for map-reduce so it goes a step beyond "embarrassing" (map).
> It's hard to imagine a workload in which this would be the truly best solution.
Biology offers a few examples - typically (alignment-free) sequence analysis is limited by memory, not by communication, so being able to reach a large number of volunteers may be more significant that the time it takes for each of them to communicate their piece of the problem. Here's an example: http://www.almob.org/content/7/1/12