I used Instruments' System Trace — watching ANE driver activity (H11ANEIn) while firing 1, 2, 3, 4 concurrent requests. With 1-2, ANE dispatch returned immediately. At 3+, the third blocked indefinitely waiting for a slot. The thread timeline made the 2-slot ceiling very clear. Re serial queue: agreed it's simpler. The actor was for my case where I needed 2 slots saturated for max throughput while queuing the rest.