I do have one argument FOR storing mesh data on disk. I assume your mesh data is designed to be compact (perhaps using single bytes for coords, expanded in shaders, and so on).
Levels of detail. If you were to save meshes to disk, including levels of detail, you could quickly load the low LoDs for distant chunks, without having to load the expensive chunk data!
With the normal approach, you have two options for LoDs:
A) Load the full chunk, generate low detail mesh (theres TONS of distant chunks and chunks are not cheap to load and process so this is really slow)
B) Load low LoD of the actual chunk data and use that to generate the low detail mesh. The problem with this is that the mesh will probably not be as pretty as it could.
If you actually save the mesh and the LoDs to disk, you can generate the low detail meshes from the full chunk data to get the best quality, and then later you only need to load the low detail mesh to quickly render distant chunks.
It might make sense to store the low detail meshes separately from their chunks somehow, so you dont end up loading data you dont need if its all in a big blob (dont know how this stuff works)
SSDs getting more popular increase the advantages of this approach.
The high detail meshes should still probably be generated on the fly because you already have the chunk data loaded with high probability. But again, if you dont need to load the chunk data itself, loading just the mesh could be better (since the mesh only contains the visible surface, it could be more compact. Not necessarily though.)