Some comes from datasheets (HDD seek), some comes from pure mathematical calculations (network throughput), and some are kind of "today it might be true for a specific configuration, but in 20 minutes it will be just a random outdated number" (branch misprediction or fetching from memory - these are extrapolated from number of CPU cycles needed to execute them, which also come from datasheets). And of course there are some that can be easily measured with a stopwatch (or code, if you are that kind of person), like network latency.