Trivial reference-counting can lead to memory leaks when you have an object graph that’s disjoint from an RC root object.

Non-trivial reference counting starts to look like mark-and-sweep GC.

And even then, when a ref count drops to zero, the runtime cost of destruction and deallocation of particular objects can be expensive. Stop-the-world GC is bad, but “make a blocking call because you can’t use async APIs in destructors” is worse - especially when the destructor runs in a program thread instead of a GC thread.

I feel that languages which provide support for defined object ownership and lifetime (Rust, any others?) will become the future because in many scenarios object-lifetime can be nailed-down by the compiler and thus eliminate the need for GC or ARC.

What makes you think reference counting leads to shorter pauses than tracing? While it is absolutely true that sophisticated implementations of garbage collection via ref counting can exhibit similar performance to sophisticated tracing algorithms[1], most implementations of ref counting aren't nearly as sophisticated as modern implementations of tracing, and tend to be worse.

[1]: https://www.cs.virginia.edu/~cs415/reading/bacon-garbage.pdf