These are completely disconnected from commodity hardware. They are descendents of System/360, with a completely custom architecture and "mainframe processors" which only run specific types of code (Java, XML, DB2).
Or that's what I gather from wikipedia anyway.
In the old days you had various modules you connected with dedicated CPUs for certain tasks, you'd pay more for them. These days they have a bunch of CPUs and the amount of money you pay IBM determines what those CPUs can do.
There's no physical difference between the Linux CPU, the DB2 CPU and the Java CPU, they're all the same as the main CPUs. IBM just charge you more for general purpose CPUs.
Last time I did it, Dell took several weeks to deliver the new machine.
And it's not that IBM deliver an entire server like in your case. For most practical applications, it's merely the case of adding CPUs to the pool of CPUs available to the PowerVM hypervisor. If you need to provision additional memory or IO, that needs to be accounted in as well.
That hypervisor is weird on it's own. It's part firmware part software.
If you have big box serving your users who solely rely on CPU cycles then yes, it's a sweet deal. And consider yourself very lucky if you have users who will immediately see better performance if you throw more CPU cycles at them.
That was in the early 90's and a lot could have changed since then.
http://www.winestockwebdesign.com/Essays/Eternal_Mainframe.h...
Plug: I wrote the above.