This feels like the 2007 iPhone playbook applied to AI infrastructure. "Devs" build against MLX, optimize for Apple Silicon, and suddenly their models run best on Apple's stack.