That claim is a bit surprising. It would be interesting to have that confirmed/denied by some independent testers (meaning someone not getting paid by Microsoft).
I would assume that W2K virtual machines still perform better but then I'm not familiar with any possible changes in Linux kernel thread model that might dip the balance. I found it interesting though that JBoss was the only J2EE app server that had the best performance under Linux. All others got their best marks on W2K. It almost seems as if someone wanted to have a predefined order in the report.
But it would be nice if someone could confirm or deny any performance changes on Linux JVMs.
Also it seems JBoss scales a bit better than the others when more CPU's are thrown in. Which is also nice :-)