Each extra "nine" of uptime — 99% → 99.9% → 99.99% — costs roughly an order of magnitude more effort.
99% allows 3.7 days of downtime a year; 99.99% allows 52 minutes. Chasing nines past what users notice is expensive vanity.
Not below the level your dependencies can deliver. Four nines of application uptime on a three-nines cloud region is unreachable at any price, so the cost curve does not rise — it goes vertical. And the first nines are often nearly free, so the multiplier only describes the middle.
Each extra "nine" of uptime — 99% → 99.9% → 99.99% — costs roughly an order of magnitude more effort. 99% allows 3.7 days of downtime a year; 99.99% allows 52 minutes.