I've had a fair bit of experience with large book archives (university libraries), and have a pretty good sense of what "a million books" looks like (a large on-campus multi-storey building might house 1--5 million books). There's a physical progression, from the words on a page (~250 typewritten, about 500 typeset), to pages in a chapter, chapters in a book, books on a shelf, bookcase, aisle, library floor, building, campus, etc.
A well-organised library gives structure to navigating that space, logically, metaphorically, and literally. And there's a lot of hard work that goes into that organisation.
Simply tossing a 3-D overlay onto a file manager ... isn't that. And most of the projects I've encountered either don't realise that, or don't give that fact its due.
For that matter, the concepts of search and locality (as in, works near each other) have entirely different manifestations online and in a physical archive. Shelf-reading is still one of my favourite pastimes.
I think part of what made 3d and other very visual file managers so bad (or at best, very slow like Treemappers) is the effort to render everything accurately at once, which is bound to be slow across terabyte scales.
Maybe we would be better off exploring Voronoi/blob/fisheye network diagrams for representing context, with the size indicating the number of files nested within. The Carrot2 search engine implements a variation of this idea which is sometimes very useful for topic exploration.