Second its often (not always) a lot more work to test scalability with benchmarks than it is to theoretically analyze.
Especially because people are often worried about both average case and worst case. The input that takes down your site is not the one you expect, and simulating realistic load is difficult.
It can also be pretty hard to compare benchmarks. You probably dont want to test every algorithm known to man yourself, but two separate benchmarks on diff hardware and diff conditions cannot be reasonably compared. More abstract analysis can (to an extent anyways)
to a screen that
What was that game that took forever to start up because the code for parsing the config files was accidentally quadratic or exponential?
Edit: also, the standard library of any programming language tends to have quite a bit of theoretical computer science inside.
Sampling theorem is also case in point. Also compressed sensing. Information theory will tell you there is no point trying any harder at compression when you reached epsilon closer to the theoretical limit.
Anything requiring extremely optimized algorithms will generally start in academia and work its way outward as the techniques are adopted in commercial software. Sorting and compression algorithms are two examples that come to mind. Even games development which has a history of non-academic roots now borrows heavily from academic research to refine things like culling algorithms or optimized computational methods for specialized physics calculations.