Pre-clustering is the optimization. Shoot a ray, hit a cluster, draw only the clusters facing the screen, and only the ones that are enough to fill the screen 100%, close to no overdraw. Done.
There is where ‘just chuck a bunch of polys in that model’ comes into play since you are only going to draw just-about enough to see what you can see, and nothing else.
It’s not necessarily optimized for performance such as it is ‘endless’ detail. Sure, it’s a bit of a cost-bump to get up/into it, but once you are in that compute-space, the upper-limit isn’t so bounded. And at some point, enough is enough if the detail is smaller than a pixel.