You run way better models, whatever is newest, for the five years it takes for the “spark cluster” to even reach cost parity with far worse performance.

Those things aren’t spectacular at inference… it’s not really why you drop that kind of money on them.