Of course Cray cannot live up to their promises, but this is familiar ground for them and I am sure they can reach an agreement with the gov when the time comes. When this much money is at stake, failure looks bad. And darpa does not want to look bad.
The issue for Cray is whether they can turn this "opportunity" into a real product, or if it will just be something for their engineers to uselessly beaver at for a few years. So far it has been the latter.
All of Cray's products since 2000, except the xt3, have more or less been disasters. Cray management is probably looking for a way to give its customer the parallelism it wants without taking on more engineering risk, which is what is killing them. Intel's larabee, or whatever its called, is attractive.