> So we can get rid of STUN/ICE/etc and make peer-to-peer easier. All you're left with is hole punching through your SPI firewall (no worse than today).

There’s (IMO) a much better reason: so your endpoints can have globally unique addresses. If I have a service in Google Cloud and a service in AWS and an on-prem service and a customer service and ten different upstream services from ten different hosted providers, I want these services to have different addresses, TYVM. And I want to configure my networking to them without needing to worry about whether all parties actually agree on the addressing scheme.

This is far more important than actually making anything globally routable. If my AWS thingy is 192.168.3.7, I don’t need whatever random third party service I’m using to be able to route to 192.168.3.7 (and I’d probably prefer that it not be able to route there). But I don’t want to worry about the service provider using 192.168.3.7 for another purpose.

I’d also like traceroute to work without having a bunch of intermediate hops where the addresses simply cannot be mapped to their owners without extra side knowledge of the network topology.