I'm saying that a rack full of servers is all hooked up to a single controller. You then have multiple, interconnected controllers, each one connecting a rack of servers. This is meant to give you HA and resiliency.
However, that is only true if the switch control plane is separate from the x86 hardware running in the controller (because, here we're getting HA and redundancy against the risk of controller hardware failure). I'd hope or expect that to be the case, because Nebula has smart people, but I don't see any material to that effect, yet.
(Typically a DC server farm will have two switches per rack of servers, with a dual-homed server having a connection into each. Therefore if one switch dies, you still have a redundant data path (though, only half the bandwidth)).