> My guess is that they're routing a /64 to each physical machine, and then /116s to the virtual systems.
That doesn't really work, as the /116 is shared among all nodes in the same DC. It also doesn't explain how SLAAC works, as all nodes in the DC get an address from the same pool (2600:3c01::/64 for instance).
4096 addresses is much nicer than just one, so I'm thrilled to see this change. Switching over from a tunnel right now.