We shouldn’t underestimate some companies’ desire to keep running the same software. However, as for stuff they legitimately do "better", they’re mind boggling

We shouldn’t underestimate some companies’ desire to keep running the same software. However, as for stuff they legitimately do "better", they’re mind bogglingly reliable. We’re talking two or three seconds of downtime per year. They’re also accurate - every instruction they perform is checked for correctness and if something goes wrong, that task can be migrated to another CPU core and resumed transparently. Cores can fail, socket can fail, etc. without the system going down. Hardware can be replaced without downtime. Processes and data are just migrated elsewhere. They recover from hardware faults and resume at the same instruction within milliseconds with no data loss (cache contents can be mirrored to other machines in the data center, etc.) The hyperscaler model would be to eventually recognize tasks sent to some machine aren’t returning and reschedule them from the beginning. VMs would resume from the last snapshot, etc. It takes a few seconds at best. For 99.9% of the world, that’s perfectly reasonable and acceptable. We work off of eventually consistent models. But there still remains a small audience for which that model isn’t acceptable, and they’ll pay through the nose because there’s not many companies making machines that way anymore.