Some folks look at mainframes as though they were dinosaurs. They may favour the newer open-source and distributed computing techniques and may opt for server virtualization to lower costs of doing IT business. But these folks overlook the fact that, open-source technologies are relatively new and are under development, when compared to solid mainframe technologies that have stayed developed for decades (and still being actively enhanced and supported by IBM).
In distributed computing, you still rely on hardware parts from unstable sources. Lets say your PCI bus firmware vendor has decided to save a few dollars by opting for cheap equipment made in Taiwan or China. These are small problems with the potential to erupt big when your lifeline of business depends on IT to get things done.
With open-source and distributed computing, you quickly grow your server farm to hundreds of servers that support hundreds of failover strategies and you have to invest in the support mechanisms for all these. When you use z/OS mainframe you only need solid reliable doable support for a few servers. You can use LPAR (logical partitioning) in mainframes to create hosted OS environments that are insulated from one another.
From a performance stand-point, mainframes handle increasing workloads very well. Even better than the distributed computing crowd. Mainframe performance is measured in MIPS (million instructions per second). MIPS around 12 are common with mainframes. That is far better than the open-source hosted crowd that uses server virtualization with VMWare.
In short, you cannot go wrong with your MIPS when you put money on big iron.