PCIe 4.0 is in the order of Gigabytes per second. Hundreds of billions of triangles would demand hundreds of Gigabytes. There is no way they can calculate this stuff in real time (and tbf, they never claimed that they did).

So presumably, the engine now does that kind of stuff at load-time. The bottleneck is then the place on the SSD and the scene change pace. This fits quite nicely with the repetitive statures, but it leaves a question mark for that final flight.