Public cloud supercomputing
The thing that many people forget is that supercomputing is as much about the interconnect speed and latency as it is about the basic CPU grunt.
Public cloud can stand up vast amounts of CPU resource, but if you can't shunt the data between the nodes fast enough, significant computing problems that rely on stepped iteration between many parallel compute engines will stall every communication phase of each iteration.
This is the reason why Cray have the Aries Interconnect, and IBM developed their HFI, each running elegant and complicated node-to-node topologies like Dragonfly.
Many people think that fat tree Infiniband is the bees knees when it comes to interconnects, but these people just don't understand the total combined bandwidth of these more esoteric interconnects.