We need to keep the size of the global routing table in check, which means "wasting" a good deal of address space.
That's not untrue. However, as the size of the global routing table is limited, we don't need so much number of bits for routing.
61 bits, allowing 4 layers of routing each with 32K entries, is a lot more than enough.
Even in IPv4, where addresses are considered at least somewhat scarce, a significant part of all possible addresses is lost because of this.
Only 20 bits or so for routing is, certainly, no good.
If we want to keep stateless autoconfig and be modestly future-proof we need at least a little over 80 bits. 96 would have been a good number, but I have no idea what the tradeoffs are in using a broken power of two. If we assume at least 96 bits are necessary, IPv6 only wastes 2 x 32 bits = 8 bytes per packet, or about 0,5% of a maximum size packet. Not a huge deal. And there's always header compression.
Stateless autoconfig is mostly useless feature applicable only to hosts within a private IP network that 64 bits could have worked.
128 bit is here to enable separation of 64 bit structured ID and 64 bit locator.
Masataka Ohta
