I worked at IBM for about 4 years in & after college system testing Linux and middleware on the mainframe. When I first started, I wasn't particularly impressed with them compared to commodity x86 servers but over time I started to recognize some of the power and positive aspects that they provided over racks full of commodity servers.

First off, they are I/O monsters. They were and are still widely used for transaction processing. It helps that IBM develops either specialized hardware instructions or separate hardware components to accelerate or offload workloads. Virtualization support has been baked into the hardware for decades, including partitioning resources at the hardware level. Crypto, networking, and other I/O can be offloaded to "cheaper" processors or SoCs, leaving the million dollar CPs (central processors) for your application. Over the past 15 years since I moved on they've added additional specialized hardware and capabilities to stay abreast of current computing needs/trends.

Folks seem to not want to think about this in the age of microservices, but there's a significant cost to scaling out in terms of complexity, performance, and HW/SW maintenance. It can be more convenient, simpler, and possibly less expensive to scale up, particularly when IBM lets you connect them together in a parallel sysplex configuration for even more redundancy. It definitely takes big piles money to do so, but it could provide a decent ROI if you don't need clusters or data centers worth of gear and as many people to keep it going – think "data center in a box" or "cloud in a box" (albeit expensive).

(I left IBM to go work on x86 virtualization and cloud computing, so I've also got plenty of experience at the opposite end of the spectrum too.)

IBM is more than willing to pay it. When I worked there the company was fond of pointing out that they make around a billion dollars a year just from their patent portfolio.