Procedural generation of terrain, to avoid storing huge meshes as suggested by
@Dictator.
No, the opposite would happen: By generating one large but detailed mesh of terrain, you need much more data then with kitbashing a terrain from a small number of instanced rocks (And overlapping models is not necessarily a performance problem, aside RT issues).
Thus Nanite is not well suited for procedural generation of large terrain, because such terrain would be unique everywhere, and can't be decomposed into a small set of building blocks.
This is quite a problem. As we see, all UE5 content shows pretty uniform terrain, meaning the base models have similar color and structure, so kitbashing can work. I perceive this as some comeback of UE3 days grey and brown.
I would experiment with modulating materials at runtime, to add some low frequency variance. Maybe driven from some painting or texture.
Regarding procedural generation, automatic placement of models (e.g. rocks over terrain) is where the most benefit seems to be.
For large open world games, i guess heightmaps remain pretty common, decorated with Nanite models.
Maybe they add an improved and detailed hightmap system, with support for dynamic deformation and procedural detail amplification at runtime. The game above suffers a bit from high detail rocks on low detail terrain.