Middle-ware still sucks, but is catching up. Fewer and fewer need one big expensive thing to stay up 24x7; more and more can tolerate many cheap white-boxen being down at a time without (apparent) loss of service. Google, Facebook, Amazon, Rackspace [admittedly, references needed], etc., spec out stateless commodity servers that are hot-(un)plugged when the racks fall below a percentage of up/online machines. They don't even bother opening the dead boxes to find out what went wrong anymore -- that takes skills (and usually a screwdriver).
Midrange servers and Middleware are two totally different things. Middleware is software designed to broker connections between system, like JBoss and Websphere. Midrange servers are commodity servers, usually x86.
I've seen services built like that. 400 boxes, and they just run them until they die. However, none of those are ultra critical, life or death type stuff. No one is going to die because Facebook glitched, and they can't upload a cat pic. Those services can have a little downtime. People will be inconvenienced, but they'll get over it.
Mainframes are designed to never go down and to be extremely paranoid about errors. They are so paranoid that they will run the same computation on two different CPUs then compare the two to make sure there are absolutely no errors.
Midrange server setups already look like mainframes setups. There are compute nodes that are nothing but NICs, RAM, CPU, and a just enough storage to boot an OS or hypervisor, and the bulk of the storage will be on a SAN connected via an HBA. This is exactly how a mainframe is setup, with different terminology. The difference is in the reliability of the software and hardware. Midrange stuff is just not built to the same standards as mainframe stuff.
Midrange stuff can start co-opting mainframe features, but then it would be mainframe stuff.