Amazon's scale encounters limits that most companies don't ever have to consider. There's a parallel from 20ish years ago: Amazon eschewed software performance engineering with the mantra "if it can be solved with a credit card [i.e. you can buy more hardware] don't worry about it now". That worked up to the day they called the company that made their database server and asked to upgrade and were told "you already ha…
In single master architecture this is inevitable. At some point you wont be able to throw more hardware. Thats why multi-master architecture was created in the first place.
Otoh, maximum database server today is a ton bigger. Looking at just Supermicro, you can get 8 TB ram in a dual Epyc server, and 24 TB in exotic, but still off the shelf systems. You can get a lot of qps with 24 TB of ram and a whole bunch of cores.
If you go with Oracle, they'll sell you a Sparc server with 48 TB of ram, and 384 CPU cores.