1000 C++ servers vs. 100,000 Python servers is a big reason to avoid Python.
To keep iPhone apps responsive, they saw problems with apps using GC.
They removed tracing GC from Objective-C, because it was a failure when dealing with C semantics.
Automating retain/release calls was a much easier and safer approach, hence plan B.
In any case, reference counting is a GC algorithm.
Another may be that it simplifies the runtime for system code such as kernels and device drivers.
But people do use GC on their phones, insofar as they like to use web browsers.
.NET interop with COM is a good example of how mixing a tracing GC with a reference counted runtime requires an higher engineering budget that Swift team most likely wasn't willing to spend resources on.
It has been proven multiple times that reference counting is slower and has performance issues when dealing with large graphs, and also stop-the-world issues when cascade deletions occur.
If only pro-RC folks would bother to read CS papers and benchmarks.