I'd bet the problem is maintenance scaling. TCO of IBM mainframe + support + staff vs lots and lots of servers + staff + hardware staff (+ probably porting some legacy mainframe code to something distributed).

If you're looking at a clean room build, it might not make sense often. But if you're starting with certain situations that accumulated over decades, maybe?

I know a lot of the people I work with are gradually porting stuff off mainframe onto more modern hardware stacks, but Rome wasn't built in a day, etc etc.