I've tried to do lossy compression of epubs with some lines of bash scripts; i.e. removing the images and fonts that were not needed. Many epubs could be downsized to a third of their size, but then I found a book that needed the supplied fonts and gave up. When doing lossy compressions can not have those kind of bugs.

What I also found was that many of the images in the epubs were already unuseable and nothing like their counter parts in phsyical books.

Things can be rendered from compressed container files. For HTML with images, even slow-but-strong compression like LZMA is already fast enough to render pages as fast as you can click through them, even on fairly old hardware.

Kiwix .ZIM file format is a good example. The entire Gutenberg Library is a single ~65 Gb file, and you can read any book from it without unpacking anything.