The supercomputer nodes cost far more, per node, than AWS machines- 15% or more of the budget was spent on interconnect. SUpercomputers don't partition CPUs like that, because interference causes small performance degradation. Unfortunately, CPU is not perfectly compressible- in particular, the cache is shared by all processes that run on the CPU, so if you run another job on the same node, you will see slower performance due to higher cache replacement (this is measured using Cycles Per Instruction).