{
 "about": "海外 VC / 投资人自己写的 AI 云文章（投资备忘、赛道论点、市场地图、年度预测），外加研究机构报告；每篇打开原文核对过标题、日期、作者。stance = 对「neocloud 是不是好生意」「推理会不会商品化」「芯片 vs GPU」等争议的立场。vs / why / rail / top / region 是我们的归类，不是文章原话：vs = 与图谱判断「租卡越来越像重资产基建，差异化往推理、Serverless、沙箱这些软件层走」一致 / 部分 / 相反 / 无关，rail = 立场标签，top = 最值得读。gist 是核心观点的中文转述，不是原文；seg 对应 layers.json 的 code，all = 赛道总论。",
 "articles": [
  {
   "t": "The Four Questions of AI",
   "org": "Theory Ventures（合伙人博客）",
   "au": "Tomasz Tunguz",
   "date": "2026-11-02",
   "url": "https://tomtunguz.com/the-four-questions-of-ai/",
   "kind": "赛道论点",
   "seg": [
    "all",
    "a2"
   ],
   "co": [
    "CoreWeave"
   ],
   "gist": "AI 分四层：应用、harness、模型、数据中心。数据中心层未来 5 年约 4 万亿美元债务、CoreWeave 5 年 CDS 1090bp、B300 指数 12.29 美元/卡时，卡能否用满 5 年、卡时价涨跌都未定；另称约 83% 的 agent 任务路由到小模型。",
   "cav": "提问框架而非结论；数字多引自第三方。",
   "stance": "软件层（harness）价值在升；数据中心层风险在债务与卡时价。",
   "vs": "一致",
   "why": "价值放在 harness 与上下文，数据中心层只列债务风险，同向",
   "rail": "看空 neocloud",
   "region": "美",
   "top": true
  },
  {
   "t": "The GPU Revenue Ladder",
   "org": "Theory Ventures（合伙人博客）",
   "au": "Tomasz Tunguz",
   "date": "2026-10-21",
   "url": "https://tomtunguz.com/the-gpu-revenue-ladder/",
   "kind": "赛道论点",
   "seg": [
    "a2",
    "a4",
    "b1"
   ],
   "co": [
    "SpaceX"
   ],
   "gist": "同一张卡一小时的收入有四档：理论 38 美元、B200 现货 8.77、SpaceX 合同价 5.27、实际 2.44；SpaceX CFO 称回本不到一年，作者按脚注算合同价约 1.21 年；结论是大规模推理在现价下健康、价格有上行空间，风险在需求、降价和端侧推理。",
   "cav": "全是推算；回本口径含不含税是作者假设。",
   "stance": "租卡跑推理在现价下赚钱，GPU 价格看涨。",
   "vs": "部分",
   "why": "认为卡时本身有利润、价格还能涨，与「钱往软件层走」不同",
   "rail": "看好 neocloud",
   "region": "美",
   "top": true
  },
  {
   "t": "Our Investment in Modal",
   "org": "Accel",
   "au": "",
   "date": "2026-10-14",
   "url": "https://www.accel.com/news/our-investment-in-modal",
   "kind": "投资备忘",
   "seg": [
    "b3",
    "d1"
   ],
   "co": [
    "Modal"
   ],
   "gist": "Modal 一年增长 15 倍，2026-09 年化收入突破 10 亿美元，其中累计超 30 亿个沙箱贡献 3 亿美元年化；Accel 领投 7.5 亿美元 D 轮。",
   "cav": "数据为公司披露；Accel 是领投方。",
   "stance": "Serverless 加沙箱是 AI 云里增长最快的软件层。",
   "vs": "一致",
   "why": "Serverless 与沙箱收入快速放大，正是图谱说的软件层",
   "rail": "推理有护城河",
   "region": "美",
   "top": true
  },
  {
   "t": "Price War in the Chip Aisle",
   "org": "Theory Ventures（合伙人博客）",
   "au": "Tomasz Tunguz",
   "date": "2026-10-13",
   "url": "https://tomtunguz.com/intelligence-per-dollar-2026/",
   "kind": "赛道论点",
   "seg": [
    "b2",
    "a1",
    "b1"
   ],
   "co": [
    "Google Cloud（TPU / GPU）",
    "Microsoft Azure（ND GPU 系列）",
    "OpenAI Hosted Sandbox",
    "Nvidia"
   ],
   "gist": "五个前沿模型智能分差在 5 分内，但每任务成本差 3.8 倍，Google 的 TPU 解释了差距，微软 Maia 300 和 OpenAI 自研芯片在跟进；价格持续通缩，毛利仍在 40–70%。",
   "cav": "TPU 数据是 Google 自己的、且为 2023 年（v4 对 A100）；OpenAI 芯片对比对象是 GB200/GB300 而非 Rubin。",
   "stance": "芯片 vs GPU：自研芯片拉低成本，通用 GPU 云溢价会被压缩（证据偏薄）。",
   "vs": "部分",
   "why": "自研芯片压低 token 成本，但讲的主要是模型厂商而非 AI 云",
   "rail": "推理商品化",
   "region": "美"
  },
  {
   "t": "（X 长帖）Inference is a permanent reverse auction",
   "org": "Foundation Capital（合伙人 X 长帖）",
   "au": "Jaya Gupta",
   "date": "2026-10-10",
   "url": "https://x.com/JayaGup10/status/2109414688924856810",
   "kind": "赛道论点",
   "seg": [
    "b1",
    "a2"
   ],
   "co": [],
   "gist": "推理是「永久的反向拍卖」：切换成本只是「一行代码」，新实验室又在出租闲置算力，所以推理是栈里护城河较弱的一层，却仍拿着大估值。",
   "cav": "X 长帖无标题（标题为概括）；无数据；帖下有人回「这是规模游戏」，作者身份未核实，不计入。",
   "stance": "推理会商品化：低粘性加持续降价。",
   "vs": "相反",
   "why": "认为推理层是最难防守的一层，与「差异化往推理走」相反",
   "rail": "推理商品化",
   "region": "美",
   "top": true
  },
  {
   "t": "How GPU Prices Can Double While AI Gets Cheaper",
   "org": "Theory Ventures（合伙人博客）",
   "au": "Tomasz Tunguz",
   "date": "2026-09-28",
   "url": "https://tomtunguz.com/how-gpu-prices-double-while-ai-gets-cheaper/",
   "kind": "赛道论点",
   "seg": [
    "a2",
    "b1"
   ],
   "co": [
    "Microsoft Azure（ND GPU 系列）"
   ],
   "gist": "六个月内 GPU 租金从每卡时 4.40 美元涨到 8.08 美元，而 token 价格仍在降，两者靠效率调和（微软称每 GPU 年产 token 增 90%），关键指标是「每卡时毛利」。",
   "cav": "数字多引自第三方指数。",
   "stance": "算力涨价与 token 降价并存，效率暂时追平。",
   "vs": "部分",
   "why": "卡贵 token 贱，夹在中间的是效率软件层",
   "rail": "推理商品化",
   "region": "美"
  },
  {
   "t": "Concrete, Silicon, & Leverage",
   "org": "Theory Ventures（合伙人博客）",
   "au": "Tomasz Tunguz",
   "date": "2026-09-04",
   "url": "https://tomtunguz.com/the-4-trillion-dollar-ai-data-center-debt-wave/",
   "kind": "赛道论点",
   "seg": [
    "a2"
   ],
   "co": [],
   "gist": "未来 5 年 AI 数据中心约 4 万亿美元债务（相当于美国市政债的 91%）；要还本付息，2030 年 AI 年收入需达 1.2–1.5 万亿美元，而现在约 1500 亿，年复合增速约 55%；作者称这是宏观信用事件。",
   "cav": "数字是作者模型推算；没点名具体 neocloud。",
   "stance": "融资是重资产 AI 基建的头号风险。",
   "vs": "一致",
   "why": "租卡/数据中心是债务驱动的重资产，与图谱判断同向",
   "rail": "看空 neocloud",
   "region": "美"
  },
  {
   "t": "Stripe to Acquire OpenRouter: Why Everyone Is Obsessed With Model Routing",
   "org": "Menlo Ventures",
   "au": "Matt Murphy, Deedy Das",
   "date": "2026-08-18",
   "url": "https://menlovc.com/perspective/stripe-to-acquire-openrouter-why-everyone-is-obsessed-with-model-routing/",
   "kind": "投资备忘",
   "seg": [
    "b5"
   ],
   "co": [
    "OpenRouter",
    "Stripe",
    "Fireworks AI",
    "Modal"
   ],
   "gist": "Stripe 收购 OpenRouter；Menlo 认为护城河是规模与双边市场，年化 token 超 4.5 千万亿；只靠提示词做路由是「傻子的差事」；文中把 Fireworks、Modal 列为 Menlo 组合公司。",
   "cav": "Menlo 是早期投资方，也持有 Fireworks、Modal。",
   "stance": "模型层可互换，网关与数据成为价值点。",
   "vs": "一致",
   "why": "入口、数据与规模效应在聚合层，不在卡",
   "rail": "推理商品化",
   "region": "美",
   "top": true
  },
  {
   "t": "Own Your Intelligence: A How-To Guide",
   "org": "Sequoia Capital（合伙人 X 长文）",
   "au": "Sonya Huang",
   "date": "2026-08-16",
   "url": "https://x.com/sonyatweetybird/status/2089390902162895350",
   "kind": "赛道论点",
   "seg": [
    "b1",
    "b4"
   ],
   "co": [
    "Fireworks AI",
    "Mercor"
   ],
   "gist": "应用公司应「拥有自己的智能」：理由是成本、速度、专有数据和控制权；独立的后训练栈（Fireworks、Mercor）已成熟；「不是你的权重，就不是你的产品」。",
   "cav": "日期取自检索结果（X 页未显示发布日）；Sequoia 是 Fireworks 的投资方。",
   "stance": "后训练与推理的软件栈有价值。",
   "vs": "一致",
   "why": "价值在后训练与推理软件栈，而非自己买卡",
   "rail": "推理有护城河",
   "region": "美",
   "top": true
  },
  {
   "t": "Charts of the Week: Head In The Neoclouds",
   "org": "a16z（a16z.news 栏目）",
   "au": "",
   "date": "2026-08-13",
   "url": "https://www.a16z.news/p/charts-of-the-week-head-in-the-neoclouds",
   "kind": "赛道论点",
   "seg": [
    "a2"
   ],
   "co": [
    "CoreWeave",
    "Nebius（AI Cloud）"
   ],
   "gist": "neocloud 增长极快：CoreWeave 约 25 个季度到 26 亿美元季收入，AWS 用了约 40 个季度；但 CoreWeave 过去一年股价约 -16%，芯片折旧超过收入一半，利息支出在涨；栏目声明「不对此表态」。",
   "cav": "页面未写作者和发布日，日期取自检索结果；媒体转载常改标题，原标题以此为准。",
   "stance": "增长快但重资产，折旧和利息吃利润。",
   "vs": "一致",
   "why": "租卡是重资产，折旧与利息是主问题，与图谱判断同向",
   "rail": "看空 neocloud",
   "region": "美",
   "top": true
  },
  {
   "t": "Candid with Index: In Conversation with Fireworks' Lin Qiao",
   "org": "Index Ventures",
   "au": "Sahir Azam",
   "date": "2026-07-16",
   "url": "https://www.indexventures.com/perspectives/candid-with-index-in-conversation-with-fireworks-lin-qiao/",
   "kind": "投资备忘",
   "seg": [
    "b1"
   ],
   "co": [
    "Fireworks AI"
   ],
   "gist": "Fireworks 年化收入从 2.8 亿美元增至 10 亿美元以上，日 token 从 15 万亿增至 43 万亿，约三分之二来自专用模型；Index 共同领投 15 亿美元 D 轮。",
   "cav": "数字为公司披露，只核对到导语；Index 是投资方。",
   "stance": "专用模型加开源推理是企业默认方向。",
   "vs": "一致",
   "why": "推理平台靠专用模型提高粘性，属软件层",
   "rail": "推理有护城河",
   "region": "欧"
  },
  {
   "t": "AI's $1.5T Question",
   "org": "Sequoia Capital（合伙人 Substack）",
   "au": "David Cahn",
   "date": "2026-07-08",
   "url": "https://dcahn.substack.com/p/ais-15t-question",
   "kind": "赛道论点",
   "seg": [
    "all"
   ],
   "co": [
    "Nvidia",
    "Anthropic"
   ],
   "gist": "按「英伟达数据中心年化收入 ×2 ×2」或「2026 年超大云资本开支约 7500 亿美元 ×2」估算，自 ChatGPT 以来累计需要约 3 万亿美元终端收入才能回本；Anthropic 传闻年化收入 600 亿美元让变现路径比三年前清晰，但内存等瓶颈涨价会让纯 GPU 口径低估所需收入。",
   "cav": "餐巾纸算法，不是预测；Anthropic 600 亿是文中「传闻」。",
   "stance": "AI 投入的收入缺口仍在，但变现路径比以前清楚；不单独评 neocloud。",
   "vs": "部分",
   "why": "同意钱先砸在重资产基建上，但谈的是总量缺口，没谈钱往软件层走",
   "rail": "不涉及",
   "region": "美",
   "top": true
  },
  {
   "t": "Together AI: why we invested",
   "org": "Geodesic Capital",
   "au": "Arvind Ayyala",
   "date": "2026-06-30",
   "url": "https://geodesiccap.com/insight/together-ai-why-we-invested/",
   "kind": "投资备忘",
   "seg": [
    "b1",
    "a2"
   ],
   "co": [
    "Together AI"
   ],
   "gist": "称 Together 是技术差异化最强的独立 AI 云；GPU 云只转售裸算力、没有软件差异；每 GPU 多约 15% token，约 90% token 是开源基座的后训练衍生，收入增长 6 倍以上。",
   "cav": "效率数字是公司自述；Geodesic 是投资方。",
   "stance": "裸卖算力没有差异化，软件加自有算力才有。",
   "vs": "一致",
   "why": "明确说 GPU 云卖裸算力缺软件差异",
   "rail": "推理有护城河",
   "region": "美",
   "top": true
  },
  {
   "t": "Baseten: the whole game",
   "org": "Blackbird",
   "au": "",
   "date": "2026-06-23",
   "url": "https://www.blackbird.vc/blog/baseten-the-whole-game",
   "kind": "投资备忘",
   "seg": [
    "b1",
    "b4"
   ],
   "co": [
    "Baseten"
   ],
   "gist": "推理支出到 2030 年将从 1.3 万亿增至逾 3 万亿美元；后训练的开源模型成本低约 70%；Baseten 收入一年约 20 倍；前沿模型公司在与自己的客户竞争；这是 Blackbird 最大一笔投资。",
   "cav": "市场数字未给出处；Blackbird 是投资方。",
   "stance": "推理加后训练形成闭环才有护城河。",
   "vs": "一致",
   "why": "推理平台向后训练延伸，是软件层加价值",
   "rail": "推理有护城河",
   "region": "其他"
  },
  {
   "t": "Our Investment in RadixArk: Building the Open Infrastructure for AI",
   "org": "Accel",
   "au": "Ivan Zhou, Josh Fang",
   "date": "2026-05-05",
   "url": "https://www.accel.com/news/our-investment-in-radixark-building-the-open-infrastructure-for-ai",
   "kind": "投资备忘",
   "seg": [
    "c2",
    "b1"
   ],
   "co": [
    "RadixArk",
    "SGLang",
    "Decagon"
   ],
   "gist": "瓶颈已从造模型转向运营模型；Decagon 80% 的工作负载已跑在自有模型上。",
   "cav": "Accel 是投资方。",
   "stance": "推理引擎是新的基础层。",
   "vs": "一致",
   "why": "投资方认为瓶颈在运营模型的软件",
   "rail": "推理有护城河",
   "region": "美"
  },
  {
   "t": "Why We Invested in RadixArk",
   "org": "HOF Capital",
   "au": "",
   "date": "2026-05-04",
   "url": "https://hofcapital.substack.com/p/why-we-invested-in-radixark",
   "kind": "投资备忘",
   "seg": [
    "c2",
    "b1"
   ],
   "co": [
    "RadixArk",
    "SGLang"
   ],
   "gist": "RadixArk（SGLang）种子轮 1 亿美元；趋势是从「通过别人的 API 租 token」走向「拥有并运营自己的系统」。",
   "cav": "页面未署名；HOF 是投资方。",
   "stance": "推理引擎开源加商业化运维是新层。",
   "vs": "一致",
   "why": "新增的推理引擎商业层，在 GPU 云之上",
   "rail": "推理有护城河",
   "region": "美"
  },
  {
   "t": "Our Investment in Modal",
   "org": "General Catalyst",
   "au": "",
   "date": "2026-05",
   "url": "https://www.generalcatalyst.com/stories/our-investment-in-modal",
   "kind": "投资备忘",
   "seg": [
    "b3",
    "d1"
   ],
   "co": [
    "Modal"
   ],
   "gist": "Modal 是能跑任何 GPU 工作负载的 serverless 执行层，GC 称之为「AI 时代的 AWS」；GPU 供给在超大云之外仍是硬约束。",
   "cav": "页面无日期和作者，日期按 C 轮时间推断，需复核；GC 是投资方。",
   "stance": "软件调度层是中枢。",
   "vs": "一致",
   "why": "软件调度层比卡更有价值，同向",
   "rail": "推理有护城河",
   "region": "美"
  },
  {
   "t": "AI Infrastructure Roadmap: Five frontiers for 2026",
   "org": "Bessemer Venture Partners",
   "au": "Janelle Teng Wade 等",
   "date": "2026-03-30",
   "url": "https://www.bvp.com/atlas/ai-infrastructure-roadmap-five-frontiers-for-2026",
   "kind": "年度预测",
   "seg": [
    "all",
    "b1",
    "c2"
   ],
   "co": [
    "fal",
    "Together AI",
    "Baseten",
    "Fireworks AI",
    "Inferact",
    "RadixArk"
   ],
   "gist": "2026 年 AI 基础设施五个前沿，其中「推理拐点」：推理需求与经济重要性已追平甚至超过训练；点名 Fal、Together、Baseten、Fireworks 以及 vLLM 与 SGLang 背后的 Inferact、RadixArk。",
   "cav": "只核对到总体结构与推理一节；Bessemer 是部分公司的投资方。",
   "stance": "价值向推理优化软件转移。",
   "vs": "一致",
   "why": "把新一代基础设施放在推理优化与软件层，同向",
   "rail": "推理有护城河",
   "region": "美",
   "top": true
  },
  {
   "t": "ClusterMAX 2.0: The Industry Standard GPU Cloud Rating System",
   "org": "SemiAnalysis",
   "au": "SemiAnalysis",
   "date": "2025-11-06",
   "url": "https://newsletter.semianalysis.com/p/clustermax-20-the-industry-standard",
   "kind": "行业报告",
   "seg": [
    "a2",
    "a3",
    "a1"
   ],
   "co": [
    "CoreWeave",
    "Nebius（AI Cloud）",
    "Crusoe",
    "Fluidstack",
    "Lambda"
   ],
   "gist": "评了 84 家 GPU 云（跟踪 209 家，访谈 140+ 用户）：CoreWeave 是唯一 Platinum 且唯一能卖溢价；Nebius、Oracle、Azure、Crusoe、Fluidstack 为 Gold；Google、AWS、together.ai、Lambda 为 Silver；顶部云自 3 月以来新增约 4000 亿美元 RPO。",
   "cav": "评级是 SemiAnalysis 自家标准；本条只核对到摘要与评级结论。",
   "stance": "neocloud 内部分化，只有顶部能卖溢价。",
   "vs": "部分",
   "why": "GPU 云内部分化、顶部有溢价，说明租卡不全是商品；未谈软件层",
   "rail": "看好 neocloud",
   "region": "美",
   "top": true
  },
  {
   "t": "Our Investment in Fireworks AI",
   "org": "Lightspeed Venture Partners",
   "au": "",
   "date": "2025-10-28",
   "url": "https://lsvp.com/stories/our-investment-in-fireworks-ai-the-inference-platform-aiming-to-power-every-genai-application/",
   "kind": "投资备忘",
   "seg": [
    "b1"
   ],
   "co": [
    "Fireworks AI"
   ],
   "gist": "企业想「拥有自己的 AI 栈」；Fireworks 日处理超 10 万亿 token、服务约 1 万客户。",
   "cav": "数字为公司自述；Lightspeed 是投资方。",
   "stance": "推理平台受益于自有栈趋势。",
   "vs": "一致",
   "why": "推理平台屏蔽算力采购复杂度，价值在软件",
   "rail": "推理有护城河",
   "region": "美"
  },
  {
   "t": "Inference pricing is a race to the bottom. Why Baseten is different",
   "org": "Premji Invest",
   "au": "",
   "date": "2025-09-07",
   "url": "https://www.premjiinvest.com/blog/inference-pricing-is-a-race-to-the-bottom-why-baseten-is-different",
   "kind": "投资备忘",
   "seg": [
    "b1"
   ],
   "co": [
    "Baseten"
   ],
   "gist": "推理定价对单点方案是向下竞速，但 Baseten 卖的是以可靠性（4–5 个 9）为核心的系统，加工作流粘性；认为在这一层超大云只剩算力与存储。",
   "cav": "页面未署名，日期取自检索结果；Premji 是投资方。",
   "stance": "推理有向下竞速风险，但系统化可避开。",
   "vs": "一致",
   "why": "重点是软件系统与可靠性而非卡",
   "rail": "推理有护城河",
   "region": "亚",
   "top": true
  },
  {
   "t": "Why We Doubled Down: fal's $125M Series C",
   "org": "Notable Capital",
   "au": "Glenn Solomon, Dan Cahana",
   "date": "2025-07-31",
   "url": "https://www.notablecap.com/blog/why-we-doubled-down-fals-125m-series-c",
   "kind": "投资备忘",
   "seg": [
    "b1",
    "b3"
   ],
   "co": [
    "fal"
   ],
   "gist": "二月以来年化收入翻倍；定位「生成媒体时代的 AWS」，全栈体验是壁垒。",
   "cav": "数字为公司披露；Notable 是追加投资方。",
   "stance": "全栈推理平台能形成速度和体验壁垒。",
   "vs": "一致",
   "why": "全栈软件体验是护城河",
   "rail": "推理有护城河",
   "region": "美"
  },
  {
   "t": "Investing in OpenRouter, the One API for All AI",
   "org": "Menlo Ventures",
   "au": "Matt Murphy, Deedy Das",
   "date": "2025-06-24",
   "url": "https://menlovc.com/perspective/investing-in-openrouter-the-one-api-for-all-ai/",
   "kind": "投资备忘",
   "seg": [
    "b5"
   ],
   "co": [
    "OpenRouter"
   ],
   "gist": "Menlo 领投 OpenRouter 4000 万美元 A 轮：一个 API 调用多家模型与提供商，开发者在成本、延迟、性能间取舍。",
   "cav": "作者取自页面署名；Menlo 是投资方。",
   "stance": "模型和提供商可互换，入口层有价值。",
   "vs": "一致",
   "why": "价值落在聚合入口，不在单个提供商",
   "rail": "推理商品化",
   "region": "美"
  },
  {
   "t": "Why We Invested In fal.ai",
   "org": "Notable Capital",
   "au": "Glenn Solomon, Dan Cahana",
   "date": "2025-02-12",
   "url": "https://www.notablecap.com/blog/why-we-invested-in-fal-ai",
   "kind": "投资备忘",
   "seg": [
    "b1",
    "b3"
   ],
   "co": [
    "fal"
   ],
   "gist": "B 轮：延迟最高低 10 倍，日推理请求超 1 亿次，A 轮后收入涨 20 倍。",
   "cav": "数字为公司披露；Notable 是投资方。",
   "stance": "生成媒体推理需要专用优化层。",
   "vs": "一致",
   "why": "fal 卖优化后的推理 API 而非卡",
   "rail": "推理有护城河",
   "region": "美"
  },
  {
   "t": "Our Investment in Together AI",
   "org": "General Catalyst",
   "au": "",
   "date": "2025-02",
   "url": "https://www.generalcatalyst.com/stories/our-investment-in-together-ai",
   "kind": "投资备忘",
   "seg": [
    "b1",
    "a2"
   ],
   "co": [
    "Together AI"
   ],
   "gist": "Together 是覆盖 200 多个模型的全栈开放 AI 云。",
   "cav": "页面既无日期也无作者；日期按 B 轮时间推断，需复核。",
   "stance": "开源模型加全栈 AI 云。",
   "vs": "一致",
   "why": "卖的是模型平台，不只是卡",
   "rail": "不涉及",
   "region": "美"
  },
  {
   "t": "Welcome to LLMflation: LLM inference cost is going down fast",
   "org": "a16z",
   "au": "Guido Appenzeller",
   "date": "2024-11-11",
   "url": "https://a16z.com/llmflation-llm-inference-cost/",
   "kind": "赛道论点",
   "seg": [
    "b1",
    "all"
   ],
   "co": [
    "Together AI"
   ],
   "gist": "同等性能的 LLM 推理价格每年降为十分之一：GPT-3 水平每百万 token 60 美元，三年后 Together 上的 Llama 3.2 3B 只要 0.06 美元，降了约 1000 倍。",
   "cav": "以模型基准分作为「同等性能」，口径有限。",
   "stance": "推理价格持续通缩，是个商品式曲线。",
   "vs": "部分",
   "why": "token 价格暴跌说明单卖推理难靠价格差异化",
   "rail": "推理商品化",
   "region": "美",
   "top": true
  },
  {
   "t": "AI Neocloud Playbook and Anatomy",
   "org": "SemiAnalysis",
   "au": "SemiAnalysis",
   "date": "2024-10-02",
   "url": "https://newsletter.semianalysis.com/p/ai-neocloud-playbook-and-anatomy",
   "kind": "行业报告",
   "seg": [
    "a2"
   ],
   "co": [
    "CoreWeave"
   ],
   "gist": "系统拆解 neocloud 的成本与集群经济，把玩家分为超大云、neocloud 巨头、新兴 neocloud 和经纪平台四类，并讨论 H100 租价下跌。",
   "cav": "仅依据检索摘要，正文未完整打开，细节未核对。",
   "stance": "GPU 租价下行，运营品质决定谁能活。",
   "vs": "一致",
   "why": "早在 2024 就把 GPU 云当重资产生意拆解，租价下行",
   "rail": "不涉及",
   "region": "美"
  },
  {
   "t": "AI's $600B Question",
   "org": "Sequoia Capital",
   "au": "David Cahn",
   "date": "2024-06-19",
   "url": "https://sequoiacap.com/article/ais-600b-question",
   "kind": "赛道论点",
   "seg": [
    "all",
    "a2"
   ],
   "co": [
    "Nvidia"
   ],
   "gist": "把 2023 年的 2000 亿美元问题按英伟达最新数据中心收入重算，缺口扩大到 6000 亿美元；AI 收入与基建投入之间的差距比去年更大。",
   "cav": "只核对到标题、日期和核心结论，正文细节未逐段核对。",
   "stance": "缺口在扩大，不是缩小。",
   "vs": "部分",
   "why": "同样是总量缺口论，不区分租卡与软件层",
   "rail": "不涉及",
   "region": "美"
  },
  {
   "t": "Our investment in Together AI & the new AI infra",
   "org": "Emergence Capital",
   "au": "",
   "date": "2023-11-30",
   "url": "https://emcap.substack.com/p/our-investment-in-together-ai-and",
   "kind": "投资备忘",
   "seg": [
    "b1"
   ],
   "co": [
    "Together AI"
   ],
   "gist": "Together 完成 1.025 亿美元 B 轮；企业会同时运行几十上百个模型，闭源模型是「没有护城河的黑盒」。",
   "cav": "Surfacing 通讯里的短公告；Emergence 是投资方。",
   "stance": "多模型、开源是企业方向。",
   "vs": "一致",
   "why": "新基础设施栈是软件平台",
   "rail": "不涉及",
   "region": "美"
  },
  {
   "t": "AI's $200B Question",
   "org": "Sequoia Capital",
   "au": "David Cahn",
   "date": "2023-09-20",
   "url": "https://www.sequoiacap.com/article/follow-the-gpus-perspective/",
   "kind": "赛道论点",
   "seg": [
    "all",
    "a2"
   ],
   "co": [
    "Nvidia"
   ],
   "gist": "GPU 产能被建过头：每年资本开支对应的终端收入缺口在 1250 亿美元以上；短期会乱，长期供给过剩让 AI 开发成本大降；建议创业者把精力放在终端客户价值上。",
   "cav": "页面正文未显示日期，日期取自检索结果。",
   "stance": "过度建设对算力价格是利空、对应用是利好。",
   "vs": "一致",
   "why": "算力供给过剩、价值上移，与「租卡像基建、差异化在软件层」同向",
   "rail": "不涉及",
   "region": "美"
  },
  {
   "t": "Navigating the High Cost of AI Compute",
   "org": "a16z",
   "au": "Guido Appenzeller, Matt Bornstein, Martin Casado",
   "date": "2023-04-27",
   "url": "https://a16z.com/navigating-the-high-cost-of-ai-compute/",
   "kind": "赛道论点",
   "seg": [
    "a1",
    "a2",
    "a3",
    "b3"
   ],
   "co": [
    "CoreWeave",
    "Lambda"
   ],
   "gist": "生成式 AI 受算力约束，不少公司把融到的钱约 80% 花在算力上；同一块 A100 在不同云上价差约 4 倍，专用 AI 云（CoreWeave 自建专用机房、Lambda 套利别家云）更便宜更容易拿到货，大云多只提供专用整卡；多数创业公司应直接上云。",
   "cav": "2023 年文章，价格与货源早已变化。",
   "stance": "算力同硬件不同价，专用 AI 云在 2023 年是好选择。",
   "vs": "部分",
   "why": "早期看好专用 AI 云，同时把算力看成可比价的商品",
   "rail": "看好 neocloud",
   "region": "美"
  }
 ]
}