For highly parallelizable and distributable applications (anything where mapreduce makes sense), commodity clouds are awesome. Just like relying on the subway in a city. You'll have a bit of a hassle when you need to go shopping, but your normal use case is such that it is worth the hassle.
Now let's say you have a code where your tasks require much tighter coupling due to communication (basically "Science codes"). In these cases, what you want is the interconnect. You want to be able to get a physical block of nodes in a very short amount of time for a somewhat long amount of time. In terms of cars, imagine that you live outside of a city or tend to do a lot of shopping or long distance driving.
And that is basically the logic here. It sounds like the Japanese rig is intended to be similar to the US's ORNL Titan in that it is primarily a capacity machine for science in general. So people with those kinds of jobs can buy time and run their large simulations with ease while still relying on stuff like AWS for postprocessing and the like. In that case, even though this is a 130 petaflop machine, it will likely only ever run one or two jobs at that scale and will instead operate as a "cloud" with a focus on large contiguous allocations.
Whereas the other form of SuperComputer are the ones meant for large jobs for a small subset of users. Those are the ones that tend to be owned directly by governments and involve a fair amount of effort to get access to. Those run a mixture of the above kinds of jobs (for the target community) as well as actual large scale simulations. As for what they are simulating: Think about what would need really detailed simulations using codes and data that probably should never be near Amazon's servers.