I think you've pretty much got it, but what makes it work is that you're overprovisioned at each level. You don't have to set up an addressing scheme that's gua

I think you've pretty much got it, but what makes it work is that you're overprovisioned at each level. You don't have to set up an addressing scheme that's guaranteed to be able to satisfy every conceivable request as a contiguous range. The extra buffer means that if you run out at one level, you can get another allocation from the next level up without having to advertise it globally. Right now, if you need more IPv4 addresses, you have to buy a /22 or so from someone else. Because the address space is so hotly-contested, that block could come from almost anywhere, which means every core router on the internet needs to remember where it lives on the actual topology. With IPv6, maybe once in a while a customer comes along who needs more than a /56. No big deal; you just give them another (maybe non-contiguous, if you're unlucky) allocation from the ISP's block. The ISP needs to remember that both allocations belong to the same customer, and route packets accordingly. But everyone else can just route based on the ISP's all-encompassing prefix. That's what makes the system fundamentally more scalable.