My first thought when reading about this is I know NVIDIA makes rack mounted devices with their "Tesla" GPU for HPC applications like this. I would feel more confident with hardware designed for what the Airforce is trying to do and it seams like the Tesla servers would have a lot less wasted hardware and may be more efficient in a server environment.
So I googled it and found out IBM already moved from Cell to Tesla last May.
http://arstechnica.com/business/news/2010/05/ibm-cheats-on-c...
I'm also curious how AMD's new APU will affect this market.
http://news.ycombinator.com/item?id=2067057