And not even a major cost. 50bn sounds like a lot to a layperson (and makes for great clickbait). We spend 700bn on military every year.
It's more jobs, especially different jobs. It's fewer imports (possibly not more exports with this isolationist trend we're seeing globally). There's more trust in the supply chain.
Ths article/propaganda is very short-sighted, as per usual.
It's jobs that Americans can't do (or at least not in sufficient volume). These are high-skill jobs; you can't just grab some uneducated person and expect them to do the work necessary to keep a place like TSMC running.
That's why these chips are made in Taiwan, and not in places like Cambodia, El Salvador, or Rwanda. America also doesn't have the skill necessary to replace TSMC. What chipmaking capacity it has is already tied up with Intel (which doesn't make chips for most things outside of PCs and servers). America doesn't have an excess of engineers that could be trained for this work; they're already busy doing other stuff (usually software), and there's already a shortage of them anyway.
This isn't a problem you can just throw money at to solve. There simply aren't enough skilled workers available to do what the leaders want here.
Plus, money DOES allow you to hire outside consultants and teachers.
Decentralising the chip foundries is a necessary step, arguably one that should've been made long ago.
Hiring consultants and teachers (from where?) isn't going to help with the complete lack of engineers. It takes many, many years to build a workforce of engineers needed for what's being discussed here, and even then it's not guaranteed because it relies on getting young people to go to engineering school and get degrees in relevant specialties, and then work in industry for a while to gain relevant experience. This isn't something you can just throw some money at to acquire or develop.
Path dependence is a thing. It might be perfectly economical once you get there, it was just simpler to establish in Taiwan for whatever historical reasons.
Given how long it takes foundries to get up and running, this seems like something you can solve in parallel.