There's many apps that need say, a 100GB database in 2022. Those apps also needed 100GB database in 2007, when horizontal scaling was the hottest thing around.
Nowadays however, Moore's law has steadily overtaken an order of magnitude of use cases from 2007.
Maybe 1/10th of the 2007 use cases still need that kind of enormous, big tech scaling.
TIL there are machines out there that have 64TB of RAM!
8 socket NUMA is a bit of a silly one, but it has been around in enterprise tech for a LONG time.
Scale horizontally and you're often constraining and or complicating your design/architecture, but you can handle outages far better, if not seamlessly. It's much harder to get right, so it's usually best for more senior teams that have specific need for it.
Limited horizontal replication is not the same as PAXOS.
How fatal[1]? Not all outages are fatal. Scale vertically and you'll have some downtime if any node between (and in) your user and your system goes down. For 999 out of a thousand businesses, it won't be fatal.
If your business is so fragile that the first 24 hour outage kills it, then it would have been killed anyway in the near future.
[1] I know, this is liking asking "how pregnant?"...
When folks talk about scaling up for stateful resources (database), this typically includes having stand-by replicas to which one can fail over within seconds.
Scaling up the DB and scaling out stateless compute nodes is more than enough for many, many use cases. Having a DB machine with one 1TB RAM isn't anything special these days.
Setting up something like a read replica on RDS (or frankly even a DB you manage yourself is pretty trivial, especially if you're using it solely for availability and not actually reading from it.