With the exception of pathological metaprogramming examples -- and even those have largely been fixed -- there's no way you could even measure this, let alone justify such a strong, broad opinion. You're using incomplete information to justify sloppy engineering and promoting it to others.
Templatizing/de-templatizing enough code to see a difference would be a significant effort on any non-trivial code base. But I'll spare you the trouble: instantiating a template is less work than parsing a duplicated file. Some of the early C++ compilers had problems but it hasn't been an issue in 20+ years. If you look at both the G++ and Clang test suites you'll see they verify performance, memory usage and correctness with complicated templates by doing basically this exercise for you.
