That’s somewhat correct. But if your IB switch is full, you get two new IB switches and you swap them out 1 at a time while the cluster is up and running. No downtime is required. The same should apply to an Ethernet back end. Perhaps eventually they’ll add support for a leaf spine architecture or something like that, but to be honest when you can get 96 ports already in like a 2U switch it’s hard to imagine needing to go much denser. It would have to be a really big cluster to need that.