How is NAT Port Exhaustion handled?

"Port address translation (PAT) resolves conflicts that arise when multiple hosts happen to use the same source port number to establish different external connections at the same time. PAT may then assign the connection a port number from a pool of available ports, inserting this port number in the source port field."

https://en.wikipedia.org/wiki/Network_address_translation

What if we have 10s of highly scalable servers which accepts thousands of requests ? Port exhaustion seems a very basic limit(64k) which we will easily reach.

How does industry handle such scenarios?

At the end of the day NAT really just boils down to a lookup table on the NAT device containing the mappings for what connections go where. These table entries however are created on a per-session basis.

Ex. If I have a host at 10.10.10.10, wanting to communicate with a server on the internet at 8.8.8.8, through my internet gateway whose public IP is 1.1.1.1, when that communication happens my internet gateway might create a pair of entries looking something like this:

While it may appear that 1.1.1.1:40000 is now occupied for this session, now leaving ~63999 ports, this is not exactly true. Since these entries are on a per-session basis, it would be perfectly reasonable for subsequent communications to occur and build out the table like so:

So on the public interface (which is typically where you would be concerned about NAT exhaustion), we can see that there are effectively 4 sessions:

1.1.1.1:40000 --> 8.8.8.8:80
1.1.1.1:40000 --> 8.8.8.8:81
1.1.1.1:40000 --> 8.8.4.4:80
1.1.1.1:40000 --> 8.8.4.4:81

but these 4 sessions are not totally consuming 4 ports, and the NAT device knows exactly where traffic belonging to any of these 4 sessions should be routed.

So while it may seem like there is a 64k port limit, it's really more like 64k per destination socket (with some caveats). Not inexhaustible, but certainly more scalable than a flat 64k.