How do you control LLM inference costs when usage scales 10x?

We launched an AI feature that's getting heavy adoption. Inference costs have gone from predictable to alarming. We've looked at caching, smaller models for classification steps, and batching — but we're looking for a more systematic approach. What cost control strategies have actually moved the needle for teams running LLMs at enterprise scale?

— VP Engineering, B2B SaaS, 500+ enterprise clients