Batching cost with SRP batcher

I'm trying to determine how significant batches are in HDRP when using SRP batcher. For the use case where it can't actually batch multiple draw calls. Unique meshes.

The data persists on the gpu regardless, and I'm not seeing a significant hit from lots of batches. Within reason a few thousand. It seems like it can go either way depending on context. Like if I'm combining meshes where I end up with a few thousand, vs a dozen instanced tens of thousands of times. It seems like other parts of the pipeline have a per render cost that scales in favor of the higher batch counts in this case. Shadows seemed to be one thing that works like that, although I haven't done enough detailed debugging to know for sure.

Any general advice on what or what not to do here, especially with rendering at large scale, would be helpful. A bit concerned about what I might not know yet and end up shooting myself in the foot. There are a few significant areas of engineering that hinge on the right answer here.