I've run raytracing clusters with 4000 cores. It basically scales linearly. So for this task, or any massively parallel compute-bound task, you pretty much can just add up the cores.

Also, I would rent them by the hour, because if a service has 500 machines available right now, I can get the job done in an hour, instead of 500 hours. Or 10 hours instead of 7 months.

Not something I can do with purchased hardware.