Prompt para gestão de custos de LLM em apps de IA de grande escala

You are an AI architecture consultant tasked with designing cost-efficient, scalable LLM-powered AI applications. Given that most projects rely heavily on external LLM API calls, and rough calculations suggest self-hosting a 10B-parameter LLM for 10k users making ~50 calls/day would cost around $90k/month (~$9/user), which is not practical at scale. Many apps serve 1M+ users with thousands of daily active users. Your job is to propose a comprehensive, actionable strategy to manage AI infrastructure costs while maintaining profitability. Deliver a multi-part analysis:

1) Identify the top cost drivers for large-scale LLM workloads (compute, memory, egress, embeddings, persistence, orchestration, and API usage).
2) Propose a hybrid architecture that mixes API-based calls with self-hosted or quantized/inference-optimized models where appropriate, including decision criteria for when to call an API vs when to serve locally.
3) Outline caching and data-structuring strategies beyond prompt or query caching, such as response caching, historical memory/state caching, embedding caches, and memoization of repeated intents or classifications.
4) Suggest model-tiering and model-selection strategies (when to use smaller/quantized models, task-specific fine-tuned models, or external APIs) to balance quality and cost.
5) Provide a cost model with sample calculations for 1M+ users and thousands of daily active users, including monthly run-rate estimates for different architectures.
6) Propose engineering, monitoring, and governance plans: metrics to track (per-user cost, cache hit rate, latency, reliability), alerting, SLAs, and rollback strategies.
7) Deliver a practical 8-week rollout plan with milestones and minimal viable features to achieve significant cost reductions while preserving user experience.

Include concrete examples, pseudo-code for a cache layer, and a checklist of trade-offs and risks. End with a short executive summary.