As far as I know, a TCP/IP stack can only maintain an absolute maximum of 65535 concurrent connections; while reading the white-paper for some F5 BigIP load balancers, I see that these can hold open millions of concurrent connections;
I am specifically looking at the lowest end model, the 2000s, which states that it can maintain 5 million concurrent connections.
At an OS level, how is this managed? Can there only be 65535 per copper / fiber link? Would this mean it would require 75+ uplink connections in order to even hit this limit?
Nope, you really don't need to have 75+ uplinks. One connection is identified by 4 parameters:
There is limit 65535 concurent connection, but not for all those parameters combined but for each of them. So you can have 65535 concurrent connections per source IP address, but you can have milions of connection distributed between thousands of clients (source IPs).