{
  "lesson": "02-inference-platform-economics",
  "title": "Inference Platform Economics — Fireworks, Together, Baseten, Modal, Replicate, Anyscale",
  "questions": [
    {
      "stage": "pre",
      "question": "Which three market segments does the lesson use to organize 2026 inference vendors?",
      "options": [
        "Single-tenant, multi-tenant, on-prem",
        "Free, paid, enterprise",
        "Open-source, commercial, hybrid",
        "Custom silicon, GPU platforms, API-first marketplaces"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Around what sustained GPU utilization does per-minute billing (Baseten, Modal) start to beat per-token billing (Fireworks, Together)?",
      "options": [
        "5%",
        "90%",
        "60%",
        "30%"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which platform is described as Python-native serverless with per-second billing and 2-4s cold starts after pre-warming?",
      "options": [
        "Anyscale",
        "Baseten",
        "Fireworks",
        "Modal"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "What is Fireworks's notable LoRA pricing differentiator?",
      "options": [
        "LoRA requests require a separate dedicated GPU contract",
        "LoRA-served requests are charged at the base model's per-token rate",
        "LoRA-served requests cost more than base model",
        "LoRA is not supported at all"
      ],
      "correct": 1,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Which platform fits a regulated healthcare customer that needs SOC 2 Type II, HIPAA-ready posture, and dedicated GPUs?",
      "options": [
        "Together",
        "Anyscale",
        "Baseten",
        "Replicate"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Why does the lesson argue that custom-engine claims are mostly marketing shade at the platform layer?",
      "options": [
        "Custom engines never outperform vLLM",
        "Per-token pricing is the only real differentiator",
        "vLLM and SGLang represent roughly 80% of production open-source inference, so platform differentiation comes more from DX, attribution, and SLAs than engine",
        "All custom engines are forks of TensorRT-LLM"
      ],
      "correct": 2,
      "explanation": ""
    }
  ]
}