WRT compiler overhead, this is an asymptotic improvement, not just a constant factor. If this can get my 30-second to 20-minute compiles (depending on what's changed since the last make) down to the compile-duration range of Go (which uses this kind of dependency management for compiler performance reasons) that completely changes the workflows that are possible.
But yeah - the standardisation issue is worrying, although at least it looks like they've thought about the issue with their transitional proposal.
Compile time is usually fine if you build your code so that dependencies are simple as intermediate .o files are only built in dependency chains. Link time is much higher on larger projects, which this still fails to solve.
Knowing your shit gets you further than language changes here.
For ref, I've been writing c since 1986 and I've seen proposals like this come and go lots and every one doesn't end up with an improvement.
I haven't used C++ recently, and never in truly large projects, but what I do remember is that that "build your code so that dependencies are simple" is not something you get for free, certainly not with C++.
I also think that Apple wants every compilation speed improvement it can find, so that it can improve syntax coloring and error highlighting in (almost) real time while you are editing.
Finally, I don't think this really is a proposal. Apple shows its intent earlier than they did with WebKit, but read that last slide: "clang implementation underway".
I expect they will listen to feedback and change this if people propose real improvements, but I think it is a given that this will be in clang soon, and also that it will be used in clang, LLVM, and WebKit. From there, chances are it will spread, either via tools that use WebKit or LLVM, or because of superior compilation speed.
How soon that 'soon' will be, I don't know. Implementation may not be as simple as it looks.
I've used Oberon, which has a module system quite similar to this, quite heavily, and the way to diminish link time is to not link, instead use dynamic loading: leaf modules are loadable and unloadable, nonleaf modules can be made leaf by unloading all their descendants. Gives you the possibility for immediacy/flow, that a good REPL does. For many cases, you get edit/compile/run cycles measured in (often single-digit) seconds.
But yeah - the standardisation issue is worrying, although at least it looks like they've thought about the issue with their transitional proposal.