If you serve data from a DC in California to the west coast of Australia it will take 10 seconds of just network transfer time. Sometimes data has a high degree of cardinality and cannot be cached. Some companies cannot afford to, or for other reasons cannot, have regional instances of their compute surfaces. This is just one common scenario, there are many others.
> A middling back end system with median complexity is doing 10-100K requests per second per server and is near-perfectly horizontally scalable.
You know perfectly your SLAs without knowing anything about the business, data, or the user access pattern? You don't even need a budget? Impressive!