LLM infrastructure (orchestration, retrieval augmented generation (RAG), vector storage) is leaking 5x faster than core model providers.