To be more precise, the lion's share of memory goes into harfbuzz structures, which is a text shaping library that Godot uses. Plus the icu4c lib and Godot's own structures contribute too.

I intuit that the number of strings is more of a problem than the amount of text. Perhaps converging these thousands of strings into fewer may help reduce the footprint somehow? Some sly way to shape them together before splitting during render.