That's not to say it destroys freedom, but its not ideal for a certain subset of software. The GPL is also not magical in this regard, since you have projects where things don't get up streamed.
Cost. If your advantage was perpetual and didn't increase your future support costs, then sure, you wouldn't contribute it back (ignoring belief / culture).
It isn't though. Someone will do something similar and even if they don't, each new version requires you to not upgrade or spend the time integrating your wonder code into the mainline again.
Long term cost of staying current is going to be an issue. Each change to your fork makes bringing in new features from the original harder.
Plus, in my mind, it is a nice self selected pre-screening filter for companies / individuals that would be good to work with.
There have been issues with GCC for over a decade that go unfixed and unchanged because it doesn't make the most business sense to contribute to it. In the category of interoperation in particular, the GCC project proved to be openly hostile to the notion; they left a wide gap in the market for Clang/LLVM to fill.
Does it matter if my competitor can take advantage of my bug fix in a compiler? I don't think copy left licenses are a good fit for every project, but there is certain infrastructure in the world where ideally everyone can benefit.
Which they achieved at the cost of correctness: https://code.google.com/p/android/issues/detail?id=58726#c10
And vendors do contribute performance patches back to AOSP. ARM's Linaro group has updated & NEON-enabled some libraries, Intel contributed SSE optimizations and x86 support in general, Nvidia contributed NEON optimizations, Imgtec for MIPS support, etc...
There's a cost to maintaining a downstream fork, and a slight competitive edge tends to not actually be significant enough to justify the engineering cost.