In any highly skilled profession, 95% of the time someone with very little of your skill could do your job. The valuable part is that you can also be counted on for that 5% as well.

Most of my time at Google hasn't involved any substantial theoretical work. One time though, I did have to come up with a algorithm for computing connected components that could run in reasonable parallel-time in mapreduce. The resulting algorithm ran in log n parallel time passing n log n messages over its lifetime. (This probably is in the literature somewhere considering that a coworker and I had to solve it for unrelated reasons and independently came up with the same thing.)

I'd like to expect my coworkers to be able to do that sort of work in a domain that they're familiar with, because it takes certain problems and moves them across the boundary from impossible to possible. Working at Google scale means that you'll inevitably have to solve some problems that have never been solved before, at least until a lot of universities find it worthwhile to build a thousand node cluster for research purposes.