I’ll give it a whirl.
Just re-installed my work rig, so it’ll take a bit.
Maybe it’s because of how they are spawned under the hood that you get better performance?
Instead of instance copy/paste?
And yea, I red that as baked light map rather than “shadow maps”.
Doesn’t really change too much in the sense that the items being set to movable prevents the light from baking anyway so I guess it’s a moot point.
There’s a starting cost for the initial render, and an additional to every tris. It’s not a 1:1 ratio (though according to epic docs it seemed like a direct ratio.
I read this just last month and I can’t find the in question via Google, so It was likely edited? O_o)
Also performance wise Directional Lights and PointLight have some wierd differences.
If you were to make your pointlight large enough to shadow all items the same, you’d expect same coats. but it’s not so… because you know, unreal consistency? 
You also aren’t really looking at all the spheres. There’s maybe what 100 on screen and most of the rest is culled out and/or has no need to render a shadow?
Freeze rendering and see what was actually being computed.
Perhaps the difference was that, the way I test it all the cubs are on screen at the same time, there’s no culling involved at all.
Either way I’ll give a look at your shared thing.
I’ll give that a whirl too.
There wasn’t one in .25
Basically setting the CSM to something stupid small (1.5m? I think?) Was the only way to get around 60fps steady on a heavy scene.
But the distance range was definitely affecting performance, so if now it’s not that’s a regression.