For what it's worth, we're regularly working with 100K-1MM element graphs over at github.com/graphistry/pygraphistry , e.g., for analyzing IT operations & security events. Most tools force analysts to start from ~scratch and don't scale, but that's solvable :)
I'm the cofounder of a systems and network mapping service (kumu.io) and I can't agree more. Although our primary focus is on systems mapping we also work with a lot of networks. We're constantly reminding our users that the network itself is not the deliverable. Large networks must be reduced to communities / metrics or unfolded piece-by-piece if you want people to actually gain something from them. We have a built in presentation builder to help with that process but a lot of people are just too lazy. I cringe when customers propose 3D as the solution instead.