Reduced ad absurdum I thinks that is pretty much how it works. There is an acceleration structure that is traversed for each ray, to reduce the search space for ray-geometry intersections (probably a bounding volume hierarchy of some sorts), and using instancing only a very small subset of the total of all grains of sand actually need to be stored and processed. I imagine the sand is modeled using some particle system where each particle is actually a small scene of sand grain models in itself, with its own BVH etc, and the raytracer somehow reuses whatever happens inside them for every other particle that has the same lighting conditions.

It’s probably way way more complicated than that though. Extremely interesting stuff.