Benchmarks are nice but what's the latency at scale? That's what actually matters for production.