I just came across this article on bink.nu (brilliant site btw), here is a snippet:
Until a few months ago, the clearing and billing system for NYSE Group Inc.’s stock options exchange consisted of about 800 discrete Cobol programs running on an IBM mainframe. Today, the entire application set has migrated onto a pair of clustered, quadprocessor Windows servers. The recompiled programs remain in Cobol today, but they won't stay there for long.
If you dig into the links you get to some bold statements such as:
"the real-time clearing application runs three times faster than it did on the mainframe and the overnight batch processing of the reports and daily billing routines are 15 times faster. Plus, they can expand capacity by adding processors or servers at a modest incremental cost."
From a business perspective they were aware that maybe they'd lose a '9' or two compared to mainframe availability numbers but the cost-savings and performance improvements won over those objections.
The source of the bink.nu post is on blogs.msdn.com the specific blog was a new one to me, another one I'll have to monitor :-)