Don't underestimate the cost of just reading in all that text and parsing it. Modules allow you to cache some pre-parsed AST or similar internal representation of any exported template definitions which should greatly reduce the cost of instantiating them in the importing translation unit due to savings in IO and parsing time. Importantly, importing translation units never even have to look at definitions they don't use whereas currently they have to fully parse everything that gets dragged in by the headers they include even if they only use a tiny subset of the functionality.
A module enabled STL with well implemented compiler and linker support should greatly reduce build times for codebases that make heavy use of STL by simply replacing #include <header> with import std.header. The proposal eliminates all the unnecessary overhead of templates (re-parsing huge amounts of text) and leaves only the necessary / desirable overhead (instantiation from some internal representation with the specific types used). It's that final instantiation that gives templates their runtime performance advantages over other models (like C# generics) so you wouldn't want to lose that.
It's really an elegant solution considering the constraints they're working under.
[1] Preprocessor macros don't leak from includers to includees, so headers become independent of each other.
Borland came up with an approach that instantiated templates multiple times, and then weeded out duplicates. That was relatively easy to implement, so everyone else followed suit. Extern template makes it possible to make sure some instantiations are instantiated only once, so the compiler doesn't have to weed out extra instantiations (I wish I could use "instantiate" fewer times in that description, but I can't see how to pull that off).