I created https://github.com/bep/create-a-million-pages some weeks ago with this in mind.

Currently, if you want to build a Hugo site with "millions of pages" you will need some serious hardware, and even then you will most hit the memory ceiling.

My current working theory is that, if we kept the metadata in memory, but streamed the page content (via a LRU cache of some sort), the above should be doable.