Sorry, I haven’t notice that I was answering in the meetings archive, here goes my comment about it so we can have a talk in a more proper place.
Regarding this… could be interesting to have this inside Blender, to avoid Blender cluttering the system ram with a big scene.
Right now we have different duplicates of the scene in memory when a render is done AFAIK, and that’s why the ram usage of Blender sometimes is a bit in the clouds:
1.- The blender scene itself, several Gb
2.- the experted scene to Cycles, more or less the same amount, may be a bit less because some optimizations
3.- the scene in GPU memory, unavoidable obviously and nothing wrong with that.
But the point 1 and 2 elevates the RAM requirements to be able to render some scenes, because a big scene that requires around 20Gb of RAM to render it actually needs around 40Gb of RAM because Blender keeps eating narly the same RAM when Cycles renders it.
Maybe using that Hydra delegate (if I understood it correctly) we can render USD’s ready to render withoutthe need of Blender loading in the scene memory that information, freeing a lot of RAM for Cycles.