You can scale PG horizontally with read replicas. Many frameworks support this (rails, for example: https://dev.to/schwad/how-to-use-horizontal-sharding-in-rail... )

I don't know your business but scaling to 1TB of data (not just files, but actual rows in a database) seems weird. Feel free to prove me wrong (I'm just some person on the internet, after all), but I've seen many startups fool around with over architecting instead of building features users want.

If, on the other hand, you're positive you're going to need to handle that kind of data from the start, I'd read up on real life architectures that are similar to yours: http://highscalability.squarespace.com/blog/category/example