Most journals completely re-do the layout in some desktop publishing software and only copy over the text and figures. It is a fair bit of work to make everything consistent, for example formulas and references. I have only seen author-submitted PDFs in very low-cost publications, for example some conference proceedings.

One of the journals in my field is going through this at the moment. They were set up with an initial grant to cover editorial and typesetting, which is about to run out. Their two suggestions for carrying on are either to charge a couple of hundred dollars a paper for typesetting (which is apparently what it actually costs) or to require submissions in a standard LaTeX pipeline (which is not standard for my discipline but would enable it to be done automatically). I suppose a strict Word template would also suffice at the expense of annoying all the computational scholars (that's what another Diamond OA journal does in my field).

But somebody's doing the typesetting. Giving sufficiently technically competent authors that could be the authors, but otherwise some money has to be found from somewhere. I don't mind that coming from libraries or research grants if necessary: it's cheaper than Elsevier.

Competent infrastructure is very rarely that simple.

At least throw in some backups. And someone to manage them / know how they work. ;)

This does not match my experience in science publishing. Scientific journals usually have at least a few paid full-time and part-time editors. They also do their own copy editing and typesetting of the manuscripts. The differences between the arXiv versions and the published versions are usually quite noticeable.

"The APC to publish Gold Open Access in Nature is €9,750 / $11,690 / £8,490." https://www.nature.com/nature/for-authors/publishing-options Publishing closed access is still free.