What's even the advantage of putting a lot of nodes in a single place? I bet if you add all of Amazon's computing capacity it's going to be much more than what a single localized supercomputer can do.

The short version is increasing by the performance of tightly coupled distributed software, often physics simulations. The more widely distributed the nodes, the slower the distributed algorithm will run, forcing smaller sizes to get results in a reasonable time frame. Denser machines allow for larger areas, higher resolutions or longer timescales.

You may run any simulation on it whose outcome will not contradict the parties plans or create disturbing rumors.

What good is the mightiest sword if you cripple the hand that wields it.

Are these programmed with Cray Chapel?

https://en.wikipedia.org/wiki/Chapel_(programming_language)

They are certainly pulling already in supercomputing drip. In the West, I think only MareNostrum looks that cool.