As well as just using math, capacity tests are usually automated to verify the calculations. You use a daemon that properly simulates a user doing it and you push that to the breaking point. Of course, that means that a 100% valid QA environment is available, something that the zero-bug showed doesn't exist.