The stress tests would be performed starting at the safest level and gradually going up untill the device gets unstable, and then back down to the last detected stable level to confirm it is still stable, just like you would do if you were testing the device yourself.

I tried making a quick spreadsheet pulling the ratios of devices stable at a few different clocks out of my arse, with arbitrary prices proportional to the clock, and it says it's profitable, but that is pretty much meaningless without real data being used, and the calculations there assume there will be customers.