On a hot day they had an air condition problem in the data center and had to turn off a lot of machines. They tried to turn off the mainframe using the big button, but it just kept running. Then they decided to pull the power plug. System still kept running. It turned out it had an built-in UPS. After some more attempts they finally managed to turn it off.
IIRC, that's still true. I think IBM mainframes are typically leased, and the customer is billed on usage.
There's some complexity to this, where I think the metering only applies to programs in the "traditional" mainframe environment, while things running on the mainframe Linux environment are exempt from metered billing.
One of the interesting features I heard when they launched z15 was that, if one needed to restart a partition, it could tap unlicensed CPUs to start faster and then go back to its licensed capacity.
But at what cost? Instead of getting suppliers to out-bid one another you are now stuck in a single supplier situation. [0] Building the redundancy in software allows you to run on commodity hardware that can be purchased from anyone.
One of my older greybeard coworkers tells stories of how sometimes an IBM employee would show up at 6 AM to replace a water pump (or some other component) that had phoned home the night before to report that it was beginning to fail. The hardware is super reliable and resilient but also heavily instrumented to detect & resolve failures before they caused problems. And they would send their techs out to pre-empt those problems before the customer ever noticed anything.
Not that it would have been too much a problem if it had failed - because there was so much hardware redundancy built into those machines.
One of his other stories, is a tech demo that IBM did for his company, where they put a mainframe in the back of a trailer and would drive it around to customer sites. The demo consisted of starting a huge computation on the machine, and pulling up a screen that showed the available hardware resources, then a tech would reach into the machine and start pulling out CPUs and RAM, and you'd watch live as it detected that the hardware had disappeared and shuffled the jobs around to other hardware resources without crashing. And then the tech would start swapping hardware back in, and it would add the resources back into the pool and start doing work with them once again.
All handled by the hardware and the OS.