Before spending some time in a wireless sensor net research lab as an undergrad, I had a difficult time imagining the scale of an application that could justify such a massive network address space...probably because I was too focused on trying to visualize what it might look like without considering the possibility that it would be invisible to the human eye.
[1] http://www.eecs.berkeley.edu/~pister/SmartDust/SmartDustBAA9...
there are 3.4 * 10^38 addresses in the IPv6 space
we could individual address every water molecule on earth and still have an enormous amount of space left over
so even if DARPA is planning on replacing the oceans with dust (and somehow managing to not exhaust a number of critical resources in this drive to make a truly staggering volume of dust), it's still not an issue
I was thinking more along the lines of an architecture adopted today which survives many millennia into the future that accounts for exponential human population growth on a galactic scale, isn't limited by the natural resources of this planet, where increment() will be the only pragmatic address assignment operation, meaningful archaeologic insight might be attained by discovering a microscopic 128-bit invariant, and so on.
path MTU discovery algorithms aren't going to be efficient when packet transit times are measured in decades, for one thing