A little skimming, and I think this article is a good answer too…
I think the tl;dr is: Clipping happens in the gear when the “top” of the how-load-is-it for a sample is reached. So 32-bit lifts that “top” insanely high… instead of clipping, you get the quality of an over-driven mic. It’s still “too hot” but at least isn’t not completely missing.