I can't help but think that when numbers like this are tossed about they are tossed about just for marketing, and not because the system can actually effectively scale to those levels. Case in point, VMware claims they could of enabled 8-way SMP in ESX 3.5 (apparently you can now just with some undocumented feature), but the hypervisor wasn't efficient enough in handling 8-way VMs in 3.5 that they didn't want to turn it on, it wasn't worth the overhead, results wouldn't of been up to par.
You can see in this PDF vmware's own claims to scalability:
On page 8 they mention a 8-way VM can achieve 85% of native performance, I'd be willing to bet if they put out 16-way it'd probably drop to something like 60-70%. Page 55 shows SQL performance physical vs virtual on different # of vCPUs. You can see the gap increase significantly the more vCPUs are running.
Also find it difficult to believe that the linux cpu scheduler itself is efficient enough to handle so many cores. Traditionally it has not, sure you could run a lot but that doesn't equate to near linear performance improvement with more cores. I'm a long time linux fan but can acknowledge it's limitations. I haven't seen numbers posted recently perhaps it has improved exponentially in the past few years, if someone has some data I'd be interested in seeing it. Linux seems to have always been about horizontal scaling, rather than vertical(beyond say 8 CPUs).