There is severe hindsight bias in play here. How many tables weren't problematic despite all those joins and just chugged along scaling fine? How many startups were bogged down by unnecessary complexity or trying to keep a semblance of consistency instead of shipping features and just didn't reach the scale when performance becomes a problem?

Joins are good, actually, and normalisation should be the default. It's much easier to keep things consistent by default rather than trying to make them somewhat consistent afterwards. When something becomes a problem there is a bunch of tools (from denormalisation or materialised views to specialised datastores) to deal with that, but before that it's just premature optimisation.