Modern mainframes are just cloud compute. These are services and microservices written in proprietary programming languages, using possibly-proprietary CPU and RAM. They're equivalent to a few racks of Windows or Linux servers behind load balancers. Or a small fortune in AWS services. Cray, IBM, et al sell computing nostalgia and laziness. Their biggest product is the need to not rewrite code in an open system.
Say you have a 30-year-old codebase. Odds are there's not completely accurate documentation for everything. Even if you have an accurate spec on paper, it may not encode every gimmick and undefined behavior that external consumers depend on. And if you bollix the switchover up, you're burning tens of thousands of dollars per minute of unavailability.
In that situation, saying "pay $5M for a new mainframe, which buys you another ten years of predictability, 20% more performance, and 20% less power consumption" is a solid alternative to "spend five years and several million dollars in developer effort trying to build something modern that's a drop-in replacement and then praying it doesn't catch fire or have hideous real-world performance when you go live"