ARC fits better with the C world of quickly tuning bottlenecks to register speed. Have you ever debugged C code with custom object lifetimes under a garbage collector? It's a nightmare.
But that's, statistically, almost no people out of 'people who use GC'. There aren't that many allocators in common use that even have much in the way of tunable parameters (I can only think of the JVM ones, offhand), the people who spend weeks tuning them have very particular needs - high throughput, low latency, high-enough complexity to warrant a high-level language. These are the sort of people who also do things like start at a PHP page and end up chasing performance improvements in the bowels of a network kernel driver. It's done, but it's unrepresentative.
ARC fits better with the C world
It seems like a much more central reason ARC is more suitable than a GC for Objective C than 'GC tuning is occasionally a dark art'.
I'm only reacting to the idea that (pp) "because it's 2011, nobody should be using ARC". Well, a statistically tiny number of apps may require the bare-metal performance that hand allocation provides, but they're also disproportionately important apps.
And now MS shows WinRT that is not .Net. Hmmm... :)
(Neither, though, does C++, and smart pointers don't help there either. ;) )
The fact that Obj-C takes C and adds objects on top is its biggest flaw, just like with C++. Worse than C++ though, Obj-C doesn't even attempt to fix at least some of C's more glaring flaws... you get the whole language, quirks and all.