Guys,
Isn't there a slight possibility that he (or better his company) got hold of some experimental new Intels with the 64bit busses (Intel is developping them as we write and is also know of supplying beta-versions of upcoming procs to some of their customers)?. That would mean there's twice the throughput, and twice the number of blocks crunched, right? A server with 8 procs of that kind might be able to get such a keyrate, if optimization is 100%.
I can't believe he got hold of them, but still! It wouldn't take him much to optimize the linux kernel to get such a keyrate.
Also I can't believe he rewrote the Dnet-client's code, since the staff of Dnet would jump all over him, if they found out someone send in blocks with broken integrity.