{
 "about": "VC / 投资人 / 投研机构写的端侧 AI、小模型、端侧语音、边缘推理文章（投资备忘、赛道论点、市场地图、年度预测、行业报告），外加专业媒体的分析文章；每篇打开原文核对过标题、日期、作者（没核到的在 cav 写明）。vs / why / route / top / region 是我们的归类，不是文章原话：vs = 与图谱判断「小模型和端侧语音已经够用，越来越多 AI 会在设备本地跑」一致 / 部分 / 相反 / 无关，route = 文章看好的方向（端侧 / 云端 / 混合 / 不涉及），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": "2026: The State of Enterprise AI",
   "org": "Menlo Ventures",
   "au": "Tim Tully、Joff Redfern、Deedy Das、Derek Xiao",
   "date": "2026-11-03",
   "url": "https://menlovc.com/perspective/2026-the-state-of-enterprise-ai/",
   "kind": "行业报告",
   "seg": [
    "m1"
   ],
   "co": [
    "Anthropic",
    "OpenAI Whisper"
   ],
   "gist": "小闭源模型（Haiku、Luna）在企业 token 中的占比从 29%（2025Q3）降到 13%（2026Q3），开源从 11% 升到 19% 但只占约 4% 开支；全文不讨论端侧或本地部署。",
   "cav": "口径是企业 API 采购，不含本地运行，不能直接当端侧的反证；全文无端侧内容。",
   "stance": "小模型（企业 API）：偏空；端侧：未涉及。",
   "vs": "相反",
   "why": "企业 API 里小闭源模型占比从 29% 降到 13%；全文不谈本地运行",
   "route": "云端",
   "region": "美国"
  },
  {
   "t": "a16z bets on local AI assistant Underdog",
   "org": "a16z（据 BlockBeats 动察快讯）",
   "au": "署名未显示",
   "date": "2026-10-03",
   "url": "https://en.theblockbeats.news/flash/370139",
   "kind": "分析文章",
   "seg": [
    "m2",
    "s2"
   ],
   "co": [
    "Conway Research",
    "Underdog",
    "Khosla Ventures"
   ],
   "gist": "a16z 确认领投 Conway Research 首轮（Khosla 等也在投资方里）；Underdog 在 Mac 本地跑 4B 参数的 Woof 模型（4 位量化约 2.37GB），邮箱和日历数据不交给云，可离线；自研 Husky 推理引擎，在 M5 Max 上同模型比 MLX 最多快 4.5 倍；创始人称「前沿 AI 约半年后能进入个人设备」。",
   "cav": "是媒体快讯，不是 a16z 的文章；速度等数字是公司自测、快讯转述；a16z 自己的页面没打开。日期取快讯页显示的 10-03（R1 记 10-02）。",
   "stance": "本地个人助理：看多；理由：隐私、离线、小模型够用。",
   "vs": "一致",
   "why": "a16z 领投本地跑小模型的个人助理，数据不出设备",
   "route": "端侧",
   "region": "美国"
  },
  {
   "t": "AI Infrastructure Roadmap: Five frontiers for 2026",
   "org": "Bessemer Venture Partners",
   "au": "Janelle Teng Wade、Lance Co Ting Keh、Talia Goldberg、David Cowan、Grace Ma、Bhavik Nagda、Brandon Nydick、Bar Weiner",
   "date": "2026-03-30",
   "url": "https://www.bvp.com/atlas/ai-infrastructure-roadmap-five-frontiers-for-2026",
   "kind": "赛道论点",
   "seg": [
    "s2",
    "h2",
    "m2"
   ],
   "co": [
    "WebAI",
    "femtoAI",
    "PolarGrid",
    "Aizip",
    "Mirai",
    "OpenInfer",
    "Perceptron",
    "TurbineOne",
    "Dominion Dynamics",
    "Picogrid",
    "Breaker"
   ],
   "gist": "第四个前沿「推理拐点」里把边缘和端侧列为推理创新的「一个主要例子」：AI 要贴近用户所在的地方，不总在云上；点名 WebAI、FemtoAI、PolarGrid、Aizip、Mirai、OpenInfer 在消费设备端侧部署上做探索，模型厂商 Perceptron 的端侧创新对物理 AI 重要；国防场景通信被干扰时也要靠边缘 AI。",
   "cav": "原文已在 bvp.com 打开核过（R1 当时没拿到原文，用的是微信译文）。端侧只占五大前沿中第四条里的一两段，没有讨论小模型本身；原文列表写作「Aizip Mirai」，疑为 Aizip、Mirai 两家并列。",
   "stance": "端侧推理：看多（作为推理创新的一例）；形态：混合；理由：机器人、消费设备、国防都不能只靠云。",
   "vs": "部分",
   "why": "把端侧列为推理创新的一个例子并点名公司，但全文主线是云端推理和智能体基础设施",
   "route": "混合",
   "region": "美国",
   "top": true
  },
  {
   "t": "Miraiへの投資（投资 Mirai）",
   "org": "Z Venture Capital（ZVC）",
   "au": "Hyung Kim",
   "date": "2026-02-19",
   "url": "https://zvc.vc/content/investing_in_mirai/",
   "kind": "投资备忘",
   "seg": [
    "s2"
   ],
   "co": [
    "Mirai",
    "Apple Core ML / MLX"
   ],
   "gist": "Mirai 做的是此前缺失的「端侧推理基础设施层」；未来不是全云也不是全端而是混合，但在实时交互场景里端侧会占到出乎意料的大比例；引用 Mirai 内部基准（比 MLX 生成快 37%、预填充最多快 59%）。",
   "cav": "页面为日文；页面不显示日期，2026-02-19 取自融资新闻（档案站标 02-20）；基准是 Mirai 自测。",
   "stance": "端侧推理：看多（混合）；理由：成本、延迟、隐私、离线。",
   "vs": "一致",
   "why": "把端侧推理基础设施当作缺失的一层，实时场景里端侧占比会超预期",
   "route": "混合",
   "region": "日本"
  },
  {
   "t": "To the edge (again!) — welcoming Mirai to the Uncork portfolio",
   "org": "Uncork Capital",
   "au": "Andy McLoughlin",
   "date": "2026-02-18",
   "url": "https://uncorkcapital.com/blog/to-the-edge-again-welcoming-mirai-to-the-uncork-portfolio",
   "kind": "投资备忘",
   "seg": [
    "s2"
   ],
   "co": [
    "Mirai",
    "Fritz",
    "Apple Core ML / MLX",
    "Spotify"
   ],
   "gist": "领投 Mirai 1,000 万美元种子轮；论点：云推理太贵，口袋里的设备已能跑像样的模型，最佳形态是设备与云之间的智能路由；Mirai 14 人，自研 Rust 推理栈，称生成速度比 MLX 快 37%、预填充最多快 59%。作者承认 2018 年投 Fritz「早了八年」，当时没有足够高价值的场景。",
   "cav": "页面不显示日期，2026-02-18 按融资新闻日期推断；速度数字是 Mirai 的内部基准。",
   "stance": "端侧推理基础设施：看多（附自我警示）；需求：降云成本、延迟、隐私；形态：混合。",
   "vs": "一致",
   "why": "投资方明说设备已能跑像样的模型，云推理太贵，且自承 2018 年投端侧早了八年",
   "route": "混合",
   "region": "美国",
   "top": true
  },
  {
   "t": "Qualcomm Ventures' 2025 Year in Review",
   "org": "Qualcomm Ventures",
   "au": "署名为公司",
   "date": "2026-02-12",
   "url": "https://www.qualcommventures.com/insights/blog/qualcomm-ventures-2025-year-in-review/",
   "kind": "分析文章",
   "seg": [
    "all"
   ],
   "co": [
    "Figure AI",
    "Nothing",
    "Auradine"
   ],
   "gist": "2025 年投资 16 家公司，覆盖 AI、数据中心、边缘计算、机器人、互联智能；愿景是让端侧智能「即时、个人、安全」；称云与边缘结合的混合架构已成必需，「AI 是新的 UI」。",
   "cav": "页面已读，没有任何端侧模型公司的投资案例，只有愿景句；芯片厂商自己的 CVC，立场不中立；页面不显示日期（R1 记 2026-02-12）。",
   "stance": "端侧智能：看多；形态：混合；立场：芯片厂商。",
   "vs": "一致",
   "why": "愿景是端侧智能，称云加边缘的混合架构成为必需；但自己是芯片厂商",
   "route": "混合",
   "region": "美国"
  },
  {
   "t": "Marc Andreessen's 2026 Outlook: AI Timelines, US vs. China, and The Price of AI",
   "org": "a16z",
   "au": "Marc Andreessen（主持 Erik Torenberg）",
   "date": "2026-01",
   "url": "https://bidclub.ai/e/marc-andreessen-s-2026-outlook-ai-timelines-us-v",
   "kind": "分析文章",
   "seg": [
    "all"
   ],
   "co": [
    "Moonshot AI（Kimi）"
   ],
   "gist": "他设想一个金字塔：数据中心里少数「上帝模型」，往下是逐级更小的模型，最底层是嵌入式芯片里的微小模型；大模型的能力 6–12 个月后会被压进小模型，最新例子是 Kimi 推理模型据称能在 1–2 台 MacBook 上跑；最聪明的在顶端，数量最多的是小模型。",
   "cav": "读的是第三方站点的节目文字稿与摘要，a16z 原视频没看，所以只当二手；「1–2 台 MacBook」是他引用的早期基准，原话带「据基准」的限定。日期只到月（R1 记逐字稿 2026-01-06）。",
   "stance": "端侧小模型：看多（分层共存）；理由：能力压缩快、体量上小模型最多；顶端仍是云端大模型。",
   "vs": "部分",
   "why": "认为小模型会按金字塔扩散到本地和嵌入式，但最聪明的模型仍在数据中心",
   "route": "混合",
   "region": "美国"
  },
  {
   "t": "New Investment: Gradium",
   "org": "FirstMark Capital",
   "au": "Matt Turck",
   "date": "2025-12-02",
   "url": "https://mattturck.com/gradium/",
   "kind": "投资备忘",
   "seg": [
    "m3",
    "m4",
    "m5"
   ],
   "co": [
    "Gradium",
    "Kyutai"
   ],
   "gist": "与 Eurazeo 共同领投 Gradium 7,000 万美元种子轮；认为语音是最被低估的界面，瓶颈是延迟、轮流说话和部署成本；团队来自 Kyutai，那里发布过首个实时对话模型。",
   "cav": "补查：全文没有讲端侧或本地运行，谈的是云端 API 的质量-延迟-成本；Gradium 2026-07 把种子轮追加到约 1 亿美元（TechCrunch 转述，未打开原文）；Kyutai 本身是非营利实验室，没有投资方文章。",
   "stance": "端侧语音：未涉及；理由：低延迟、低成本的语音基础模型。",
   "vs": "无关",
   "why": "领投方谈的是云端语音基础模型的质量、延迟与成本",
   "route": "不涉及",
   "region": "美国"
  },
  {
   "t": "State of AI: An Empirical 100 Trillion Token Study（与 OpenRouter 合作）",
   "org": "a16z",
   "au": "Malika Aubakirova、Anjney Midha（OpenRouter：Alex Atallah 等）",
   "date": "2025-12",
   "url": "https://www.a16z.news/p/the-state-of-ai",
   "kind": "行业报告",
   "seg": [
    "m1",
    "m2"
   ],
   "co": [
    "OpenRouter",
    "阿里 Qwen 小尺寸",
    "DeepSeek",
    "Moonshot AI（Kimi）"
   ],
   "gist": "OpenRouter 全文显示：开源模型里「小模型」（<150 亿参数）的使用份额持续下降，15–70B 的中型模型成为新类别、用户要么选中型要么集中到最强的大模型；a16z 自己的摘要页只讲推理走向智能体化。",
   "cav": "数据只来自 OpenRouter 的云端 API 流量，报告自己写明本地部署、企业内部系统不在统计范围内，所以不能直接当作端侧小模型的反证。a16z 页面本身没有小模型结论，该结论出自 openrouter.ai/state-of-ai 全文（已读）。日期只核到「2025 年 12 月」。",
   "stance": "云 API 小模型：份额下降；对端侧：不涉及；与同一机构 Jennifer Li 的看多是两种声音。",
   "vs": "相反",
   "why": "云 API 里小模型份额在降，中型模型走强；但报告自己说本地部署不在统计内",
   "route": "云端",
   "region": "美国",
   "top": true
  },
  {
   "t": "Why We Invested in EdgeCortix — Building the Future of Thick-Edge AI",
   "org": "TDK Ventures",
   "au": "Ankur Saxena",
   "date": "2025-11-14",
   "url": "https://tdk-ventures.com/news/insights/why-we-invested-in-edgecortix-building-the-future-of-thick-edge-ai/",
   "kind": "投资备忘",
   "seg": [
    "h2"
   ],
   "co": [
    "EdgeCortix"
   ],
   "gist": "「厚边缘」指 5–30 W 的算力段，介于 TinyML 与云/手机芯片之间，2029 年约 200 亿美元；最耐久的壁垒是编译器和运行时软件；列出的风险是硬件优势能撑多久、市场碎片化、巨头竞争。",
   "cav": "投资方自己写；200 亿美元是文中预测，第三方来源未核；日期 R1 记 11-14（搜索显示 11-13）。",
   "stance": "边缘 NPU：看多（列出风险）；需求：5–30 W 的机器人、工业、车载；理由：软件栈是壁垒。",
   "vs": "部分",
   "why": "看好 5–30 W 的厚边缘，但对象是机器人和工业盒子，且列了硬件寿命与碎片化风险",
   "route": "端侧",
   "region": "美国",
   "top": true
  },
  {
   "t": "Plumerai: Bringing vision AI to millions of devices",
   "org": "OTB Ventures",
   "au": "署名为公司",
   "date": "2025-09-15",
   "url": "https://otb.vc/article/plumerai-bringing-vision-ai-to-millions-of-devices/",
   "kind": "投资备忘",
   "seg": [
    "s4"
   ],
   "co": [
    "Plumerai",
    "Partech",
    "Chamberlain Group"
   ],
   "gist": "与 Partech 联合领投 Plumerai 870 万美元 A 轮；路线是设备上约 5MB 的 Tiny AI（可跑在 1–4 美元的低成本芯片上）加云端视觉大模型的混合方案，称延迟从 7 秒降到 0.7 秒、云推理成本最多降 135 倍；已进入 Chamberlain（myQ、LiftMaster）设备，覆盖超 5,000 万户。",
   "cav": "页面已读；数字都是 Plumerai 与客户基准的口径；页面不显示日期，沿用 R1 的 2025-09-15。",
   "stance": "端侧视觉小模型：看多（混合）；理由：成本、功耗、隐私。",
   "vs": "部分",
   "why": "押设备上的 Tiny AI 加云端视觉大模型，不是纯本地",
   "route": "混合",
   "region": "英国"
  },
  {
   "t": "Moonshine AI Secures Funding to Advance On-Device Voice AI",
   "org": "Moonshine AI（Wing VC、IQT 投资）",
   "au": "署名未显示（引述 Pete Warden、Jake Flomenberg、Gareth Keane）",
   "date": "2025-07-30",
   "url": "https://pulse2.com/moonshine-ai-raises-funding-to-advance-on-device-voice-ai/",
   "kind": "投资备忘",
   "seg": [
    "m3",
    "m5"
   ],
   "co": [
    "Moonshine AI（原 Useful Sensors）",
    "Wing VC",
    "IQT"
   ],
   "gist": "语音命令在设备本地处理，不再经云，换来隐私和低延迟，也能离线用；Wing 合伙人称语音是 AI 的关键界面，IQT 合伙人称安全、低延迟的端侧语音在多个行业有价值。种子轮金额未披露。",
   "cav": "是公司新闻稿的转载，投资人只有一句引述，没有投资方自己的论点文章。",
   "stance": "端侧语音：看多；理由：隐私、延迟、离线。",
   "vs": "一致",
   "why": "投资的就是本地处理语音，换隐私和低延迟",
   "route": "端侧",
   "region": "美国"
  },
  {
   "t": "Why We Invested in Rhizome",
   "org": "Critical Ventures",
   "au": "署名未显示",
   "date": "2025-06-24",
   "url": "https://critical-ventures.com/news/why-we-invested-rhizome",
   "kind": "投资备忘",
   "seg": [
    "all"
   ],
   "co": [
    "Rhizome"
   ],
   "gist": "AI 浪潮不会缩小边缘，而是淹没它；实时推理和端上微调让低延迟基础设施更值钱。",
   "cav": "边缘在这里指网络边缘/边缘计算，不一定是手机或 PC 上的本地运行。",
   "stance": "边缘基础设施：看多；理由：低延迟。",
   "vs": "部分",
   "why": "看好边缘低延迟基础设施，但指网络边缘，不一定是设备本地",
   "route": "端侧",
   "region": "未核"
  },
  {
   "t": "Multiverse Computing: Why We Invested",
   "org": "Forgepoint Capital",
   "au": "署名未显示（第一人称）",
   "date": "2025-06-11",
   "url": "https://forgepointcap.com/perspectives/multiverse-computing-why-we-invested/",
   "kind": "投资备忘",
   "seg": [
    "m2",
    "s2"
   ],
   "co": [
    "Multiverse Computing",
    "Bullhound Capital"
   ],
   "gist": "参与 Multiverse 1.89 亿欧元（2.15 亿美元）B 轮；CompactifAI 把开源 LLM 压小最多 95%，称推理成本和功耗降 50–80%、速度快 4–12 倍、精度只损失 2–3%；小节标题把企业、IoT、边缘并列为部署场景。",
   "cav": "补查：论点主要是云端成本和能耗，端侧只在一个小节标题里出现一次；数字是公司口径。Multiverse 的新闻稿提到超压缩模型可直接上手机、树莓派。",
   "stance": "模型压缩：看多；端侧：顺带。",
   "vs": "部分",
   "why": "把压小的模型用在企业、IoT、边缘，但论点主要是云成本与能耗",
   "route": "混合",
   "region": "美国"
  },
  {
   "t": "Why edge AI changes everything about how we build products for Samsung's 2 billion customers",
   "org": "Samsung Next",
   "au": "Raymond Liao",
   "date": "2025-06",
   "url": "https://www.samsungnext.com/blog",
   "kind": "赛道论点",
   "seg": [
    "all"
   ],
   "co": [
    "Samsung Galaxy AI"
   ],
   "gist": "导语：从云中心到边缘优先不只是技术演进，而是重新思考网络里谁拥有智能，对三星这样的硬件厂商意味着什么。",
   "cav": "只核到博客列表页上的标题、作者、月份和导语；单篇正文链接没拿到，正文没读。同作者 2024-09《Investing at the edge: how Samsung's device footprint changes the AI deployment equation》也只见标题与导语。",
   "stance": "端侧/边缘优先：看多；理由：硬件厂商的设备规模（仅见导语）。",
   "vs": "一致",
   "why": "标题和导语主张边缘优先（正文未读）",
   "route": "端侧",
   "region": "美国"
  },
  {
   "t": "Cartesia: Pioneering real-time voice AI",
   "org": "Kleiner Perkins",
   "au": "署名未显示",
   "date": "2025-03-11",
   "url": "https://www.kleinerperkins.com/perspectives/cartesia-pioneering-real-time-voice-ai/",
   "kind": "投资备忘",
   "seg": [
    "m4"
   ],
   "co": [
    "Cartesia"
   ],
   "gist": "领投 Cartesia A 轮；认为实时 AI 是下一次计算范式转移，SSM 让低延迟、低成本的实时语音成为可能；Sonic 已有 1 万多家客户。",
   "cav": "补查：全文没有提设备端或本地运行，只讲实时性与成本；A 轮金额页面没写。",
   "stance": "端侧语音：未涉及；理由：实时、延迟、成本。",
   "vs": "无关",
   "why": "领投方只谈实时与成本，全文没提设备端",
   "route": "不涉及",
   "region": "美国"
  },
  {
   "t": "Cartesia: AI comes for Voice",
   "org": "Dell Technologies Capital",
   "au": "Radhika Malik",
   "date": "2025-03-10",
   "url": "https://www.delltechnologiescapital.com/resources/cartesia-ai-comes-for-voice",
   "kind": "投资备忘",
   "seg": [
    "m4"
   ],
   "co": [
    "Cartesia"
   ],
   "gist": "参与 Cartesia 6,400 万美元 A 轮（与 Kleiner Perkins、Index、Lightspeed、Factory、Samsung Ventures 等一起）；SSM 适合音频这类高维带噪信号，Cartesia 的推理栈云端和设备上都能跑。",
   "cav": "投资方自己写；端侧只一句话。日期取页面与作者领英帖显示的 03-10（R1 记 03-11，Kleiner 的稿 03-11 发）。",
   "stance": "端侧语音：看多（顺带）；理由：模型架构省算力，部署位置灵活。",
   "vs": "一致",
   "why": "推理栈云端和设备上都能跑，但只是一笔带过",
   "route": "混合",
   "region": "美国"
  },
  {
   "t": "The Race for AI—Search, National Infrastructure, & On-Device AI（The a16z Show 第 834 期）",
   "org": "a16z",
   "au": "Anjney Midha、Jennifer Li、Alex Immerman",
   "date": "2024-12-25",
   "url": "https://a16z.simplecast.com/episodes/the-race-for-aisearch-national-infrastructure-on-device-ai-TnX6QVkd",
   "kind": "分析文章",
   "seg": [
    "all"
   ],
   "co": [],
   "gist": "节目介绍把「更小的端侧模型带来隐私、性能和新消费应用」列为 2025 年三大主题之一，另两个是 AI 搜索和国家级基础设施。",
   "cav": "只读到节目介绍页（42 分 45 秒的音频没听），端侧部分的具体论点未核；日期取页面显示的 2024-12-25（R1 记 12-24）。",
   "stance": "端侧小模型：看多；理由：隐私、性能、新应用（仅见节目简介）。",
   "vs": "一致",
   "why": "节目简介把更小的端侧模型当作 2025 年的主题之一（音频未听）",
   "route": "端侧",
   "region": "美国"
  },
  {
   "t": "Building the Next Generation of Real-Time AI Models: Our Investment in Cartesia",
   "org": "Index Ventures",
   "au": "署名为公司（标签 Shardul Shah）",
   "date": "2024-12-12",
   "url": "https://www.indexventures.com/perspectives/building-the-next-generation-of-real-time-ai-models-our-investment-in-cartesia/",
   "kind": "投资备忘",
   "seg": [
    "m4"
   ],
   "co": [
    "Cartesia"
   ],
   "gist": "领投 Cartesia 2,200 万美元新一轮（累计 2,700 万美元）；状态空间模型比 Transformer 更省资源、延迟更低；首款产品 Sonic TTS 称是最快的，可以在任何设备上不联网本地运行。",
   "cav": "Index 是投资方；「最快」是公司口径，没见第三方基准；日期沿用 R1（搜索结果显示 12-11，差一天）。",
   "stance": "端侧语音合成：看多（作为能力之一）；理由：SSM 省算力、延迟低。",
   "vs": "一致",
   "why": "投资方强调 Sonic 可在任何设备上离线运行",
   "route": "端侧",
   "region": "英国"
  },
  {
   "t": "We raised $250M to scale capable and efficient general-purpose AI",
   "org": "Liquid AI（AMD Ventures 战略领投）",
   "au": "署名未显示（引述 CEO Ramin Hasani、AMD 的 Mathew Hein）",
   "date": "2024-12-12",
   "url": "https://www.liquid.ai/blog/we-raised-250m-to-scale-capable-and-efficient-general-purpose-ai",
   "kind": "投资备忘",
   "seg": [
    "m2"
   ],
   "co": [
    "Liquid AI",
    "AMD XDNA（Ryzen AI）"
   ],
   "gist": "2.5 亿美元 A 轮用于扩算力、加快 LFM 在边缘和本地（on-premise）部署的产品就绪，并进入消费电子、电信、金融、电商、生物技术等行业；AMD 的引述讲的是在 AMD Instinct GPU 上训练和部署。",
   "cav": "公司自己的公告，不是投资方文章；「边缘」只有一句；AMD 的话讲的是 GPU 训练，不是端侧。",
   "stance": "端侧小模型：看多（顺带）；主战场：企业与算力。",
   "vs": "部分",
   "why": "只一句提到边缘和本地部署，主战场是企业与算力",
   "route": "端侧",
   "region": "美国"
  },
  {
   "t": "Small, But Mighty: On-Device AI",
   "org": "a16z",
   "au": "Jennifer Li",
   "date": "2024-12",
   "url": "https://a16z.com/big-ideas-in-tech-2025/",
   "kind": "赛道论点",
   "seg": [
    "all"
   ],
   "co": [],
   "gist": "设备上运行的小模型会在使用量和部署量上领先，大模型也许仍赚更多收入；实时处理、软件框架和定制芯片会抬高用户对响应速度的预期。",
   "cav": "页面不显示发文日期，12 月是按「Big Ideas in Tech 2025」系列推断的；文中没有点名具体公司。",
   "stance": "端侧小模型：看多；理由：延迟与体验预期；但承认大模型可能仍占收入大头。",
   "vs": "一致",
   "why": "直接预测设备上的小模型会在使用量和部署量上领先",
   "route": "端侧",
   "region": "美国",
   "top": true
  },
  {
   "t": "Argmax CEO Atila Orhon on Developing On-Device AI Alongside Salesforce Ventures",
   "org": "Salesforce Ventures",
   "au": "Caroline Fiegel、Jason Spinell（访谈 Argmax CEO）",
   "date": "2024-11-26",
   "url": "https://salesforceventures.com/perspectives/argmax-ceo-atila-orhon-on-developing-on-device-ai-alongside-salesforce-ventures/",
   "kind": "分析文章",
   "seg": [
    "s2",
    "m3"
   ],
   "co": [
    "Argmax"
   ],
   "gist": "是访谈，不直接陈述投资人论点；创始人称端侧推理的市场窗口 2025 年开启，风险是端侧推理若没有强玩家突破局限就长不成大市场，早期也曾过度追求前沿模型的压缩。",
   "cav": "论点是创始人的，不是投资机构的。",
   "stance": "端侧推理：中性偏多；风险：没有强玩家就长不大。",
   "vs": "部分",
   "why": "创始人说窗口 2025 年开启，也承认没有强玩家端侧推理会长不大",
   "route": "端侧",
   "region": "美国"
  },
  {
   "t": "Roadmap: Voice AI",
   "org": "Bessemer Venture Partners",
   "au": "Mike Droesch、Libbie Frost、Aia Sarycheva",
   "date": "2024-11-11",
   "url": "https://www.bvp.com/atlas/roadmap-voice-ai",
   "kind": "赛道论点",
   "seg": [
    "m3",
    "m4",
    "m5"
   ],
   "co": [],
   "gist": "对「更小的模型」特别兴奋，但理由是延迟和成本，未讲本地运行。",
   "cav": "地图图标注 as of 2024-12-09，文章发布日取 2024-11-11。",
   "stance": "小模型：看多（作为成本/延迟手段）；端侧：未涉及。",
   "vs": "部分",
   "why": "同意小模型有用，但理由是延迟和成本，没讲本地运行",
   "route": "云端",
   "region": "美国"
  },
  {
   "t": "Arcee Is Going Small To Win Big in the Long Run",
   "org": "Emergence Capital",
   "au": "署名未显示",
   "date": "2024-07-16",
   "url": "https://www.emcap.com/thoughts/arcee-is-going-small-to-win-big-in-the-long-run",
   "kind": "投资备忘",
   "seg": [
    "m2"
   ],
   "co": [
    "Arcee AI"
   ],
   "gist": "小语言模型会成为企业 AI 的支柱，因为性能、成本和隐私；Arcee 用模型合并和 Spectrum 训练让企业在自己的数据上训小模型，文中客户案例称成本降 89–96%。",
   "cav": "补查：讲的是企业私有化部署的小模型，不是手机或 PC 本地运行；客户案例数字是文中口径，未核第三方。",
   "stance": "小模型：看多；部署：企业私有环境；端侧：未涉及。",
   "vs": "部分",
   "why": "同意小模型够用，但讲的是企业私有环境，不是设备本地",
   "route": "不涉及",
   "region": "美国"
  },
  {
   "t": "Silicon Synapses: AI hardware in data centers and on the edge",
   "org": "Activant Capital",
   "au": "Simphiwe Msibi、Marc Wu",
   "date": "2024-03-12",
   "url": "https://activantcapital.com/research/ai-hardware",
   "kind": "行业报告",
   "seg": [
    "h2"
   ],
   "co": [
    "Armada",
    "NXP eIQ Neutron NPU",
    "Mythic",
    "Etched"
   ],
   "gist": "边缘推理是新进入者唯一现实可竞争的部分，因为数据产生处就地推理省传输、降成本。",
   "cav": "Kinara 后被 NXP 收购（见收购页）。",
   "stance": "边缘推理硬件：看多。",
   "vs": "一致",
   "why": "认为边缘推理是新进入者唯一能打的部分",
   "route": "端侧",
   "region": "美国"
  },
  {
   "t": "Investing in Arcee.ai — The future of Specialized Models",
   "org": "Flybridge",
   "au": "Jesse Middleton（据其领英帖署名，页面未署名）",
   "date": "2024-01-23",
   "url": "https://www.flybridge.com/ideas/the-bow/investing-in-arceeai-the-future-of-specialized-models-n8ram",
   "kind": "投资备忘",
   "seg": [
    "m2"
   ],
   "co": [
    "Arcee AI"
   ],
   "gist": "相信未来是多模型世界，小型专用语言模型（SSLM）会起关键作用，因为企业要精度、隐私和定制；Arcee 让企业在自己的 VPC 里训练、部署、监控小模型，称训练成本最多低 50%。",
   "cav": "补查：同样是企业私有云场景，不涉及设备端；页面不显示日期，取作者领英帖日期；「最多低 50%」是文中口径。",
   "stance": "小模型：看多；部署：企业 VPC；端侧：未涉及。",
   "vs": "部分",
   "why": "同意小型专用模型会起关键作用，部署在企业 VPC，不涉及设备端",
   "route": "不涉及",
   "region": "美国"
  }
 ]
}