ARC doesn't work just as well. At least theoretically, ARC retains a lot more garbage because the compiler has to be much more conservative about reachability than the garbage collector. ARC has theoretical failure modes that can cause huge amounts of retained garbage, and that's less well-understood than the failure modes of GC. Doing reference count updates is also generally slower in terms of CPU cycles than doing garbage collections. Oh, and ARC doesn't handle cycles.

That said, I think reference counting warrants more attention than it has gotten in memory management research. If you're willing to punt on the leaking cycles issue, you can get a lot of mileage out of using a reference counting collector as a tenured generation with a copying nursery.