[eqvinox]

FYI: After skimming the article, I'm downvoting because you conflated what the article calls "This 96-bit NAT address space is a highly theoretic ceiling, […]" became a plain unqualified "96 bits of address space" in your comment. Which you're then comparing with the "56 bits of usable address space" in IPv6.

Also, that sentence continues in the article with: "[…] but the pragmatic question is how much of this space can be exploited in a cost-effective manner such that the marginal cost of exploitation is lower than the cost of an IPv6 deployment."

[anecdotal1]

Pretty much anything is lower than the cost of IPv6 deployment if you already have a large IPv4 deployment.

[lmm]

CGNAT routers aren't cheap. At some point you're spending more keeping your IPv4 running than it would take to replace it.

[j16sdiz]

Except you can't really replace it. You need to keep BOTH for the good old internet to work.

[lmm]

You probably need some amount of CGNAT or equivalent (e.g. DS-Lite). But if you offload the vast majority of your traffic onto v6 (which is easier than it sounds, because most traffic is to a few big sites and most of them are v6-enabled) then you can reduce the number of connections you have to track by orders of magnitude, which saves you real money.

[ThatPlayer]

Yeah I remember seeing that in a talk about an ISP adding IPv6 support: https://youtu.be/75h4gm7t1oI . They say that 30% of traffic is IPv6, which means 30% less CGNAT hardware needed.