IPv6 has so many addresses that the default is for /64 subnets to be the minimum network size, allowing devices to choose their addresses randomly within that range. If they do that, 2^64 is big enough that it's not practical to enumerate all devices on a given network. Attempting to do so would be a bandwidth DoS attack: attempting to ping 2^64 devices requires 2^64 * 64 bytes = 1024EiB, roughly equal to total global internet traffic per year.

Now, sometimes NAT is needed anyway: QubesOS supports IPv6 using NAT because it splits things up into many different VMs on one computer. But that's a pretty rare case.

To be fair that info is already available w/ port sequence analysis, the same machine will pick from the same set of ephemeral ports for outgoing connections. I'd say the time for vanity privacy arguments is over, I'm all-in for team globally routable addresses w/ firewalls just saying "yes" or "no" to connections instead of routers guessing at port mappings.

That's what the privacy extensions are for. You can have a regularly changing, randomly generated IP address on each machine.

My network would look like it has dozens of machines, since everything is periodically rotating its IP randomly around the subnet, and often have multiple IP addresses at once with privacy extensions…

Your machines could change their IPs for every second of every day and you wouldn't come close to exhausting your /64 subnet for the next billion years.

Yeah, I had an idea to have an IP for each container (which means almost every "containerized" application soon, possibly even every browser window).