Top 1% of Consumer AI Users Spend $903 Monthly, a16z Study Shows

a16z研究显示：前1%的消费级AI用户月均花费达903美元 鉴于付费渗透率停滞在4.5%，且收入高度集中于技术极客，消费级AI必须摆脱订阅付费墙，转向广告与电商模式。 结合匿名消费者信用卡数据与流量排行榜，该讨论小组透露，仅有约4.5%的美国消费者会为AI订阅付费，且支出高度集中在技术重度用户群体中——其中排名前1%的用户月均支出高达903美元。主持人认为，循环订阅模式并不适合作为大众市场AI的长期变现模式，并指出OpenAI年化广告收入迅速达到10亿美元，以及高意图的对话式搜索发现，都是转向广告和电商模式的早期信号。他们探讨了Muse和Instinct等自主个人智能体的出现，评估了Anthropic的Claude在美付费订阅用户数上超越Google Gemini的原因，并强调了AI生产力应用如何通过自下而上的产品驱动增长迅速扩展至企业软件领域。最终讨论得出结论：尽管目前的AI工具几乎完全集中在搜索和编程等功利性“节省时间”的任务上，但消费级价值的下一个前沿将在于娱乐、购物、社交网络和约会等视觉化、互动性的“消磨时间”品类。 第7期a16z百大消费级AI应用榜单揭示出极端的幂律经济现象：仅约4.5%的美国消费者为AI订阅付费，而排名前1%的用户每月在个人信用卡上的平均支出高达903美元。 (14:39) 商业化正迅速从软件订阅拓展至广告领域，其标志是OpenAI低调达成了前所未有的10亿美元年化广告收入运行率。 (23:51) 流量排名显示出固化迹象，网页端和移动端合计仅有11个新上榜应用；但消费者支出数据却呈现出不同景象——收入排名前50的应用中有29个从未登上过流量榜单。 (0:56) 在基础模型实验室的竞争中，Claude在美国付费订阅用户数上已超越Gemini，部分得益于其严格的无广告模式和高端付费层级，其中7.5%的Claude订阅用户每月支付100美元或更多。 (29:21)

## Background 该视频由知名风险投资机构 a16z 于 2026 年 10 月 5 日发布在 YouTube 频道，视频时长约 50 分钟。节目由 a16z 编辑合伙人 Elena Burger 主持，邀请投资合伙人 Olivia Moore 和 Josh Elman 共同深入解读第七版《Top 100 消费级 AI 应用榜单》，重点剖析消费者的真实付费数据、个人智能体的崛起以及未来商业模式的演变。 ## TL;DR 本期节目揭示了消费级 AI 领域极其极化的商业现状：虽然半数美国人体验过 AI，但仅约 4.5% 的用户真正付费订阅，且前 1% 的极客核心用户以月均 903 美元的自费支出支撑了庞大的收入大头。对话深入剖析了个人 Agent 在易用性突破与隐私社交阻碍之间的张力，指出订阅制并非大众消费级 AI 的终极归宿，未来广告与交易分成将重塑商业图景。嘉宾们还探讨了 ChatGPT、Claude 与 Gemini 的商业化分化，并指出约会、原生社交、视觉电商与大众娱乐等“消磨时间”的蓝海赛道仍蕴藏着巨大的平台级创业机遇。 ## Atmosphere & style 整场对话洋溢着硅谷顶尖风投的专业洞察力与极客兴奋感，节奏紧凑而充满思辨。Elena 的敏锐提问与 Olivia 严谨的数据拆解、Josh 充满温度的产品人直觉形成互补，既有对万亿美元估值巨头商业闭环的理性推演，也有对深夜极客编程狂欢与普通人“空白文本框”困境的感同身受。氛围从对当前流量板结与高昂成本的清醒审视，逐步走向对下一代视觉交互和娱乐网络广阔蓝海的乐观展望，兼具前沿科技的颠覆感与商业底层逻辑的厚重质感。 ## Bulletin - [00:56](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=56s) a16z 发布第七版 Top 100 消费级 AI 应用榜单，首次引入消费者真实刷卡支出数据：全美仅约 4% 至 4.5% 的消费者为 AI 订阅买单，但付费群体内部呈现极端幂律分布，前 1% 的超级硬核用户每月自费高达 903 美元。 - [20:20](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1220s) 消费级 AI 商业模式面临结构性倒挂，Top 100 网页端产品中 85% 依赖订阅制，但因高昂推理成本被迫过早收费设卡，抑制了规模化；随着推理成本暴跌与小模型普及，商业模式料将重归经典互联网的广告与交易佣金分成。 - [23:51](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1431s) OpenAI 凭借 12 亿周活跃用户及即时交互意图，广告业务年化营收已迅速突破 10 亿美元；同时 Claude 付费订阅人数在美国样本中正式反超 Gemini，且 7.5% 付费用户选用 100 美元以上的 Max 套餐，高客单价策略成效显著。 - [09:12](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=552s) 个人代理（Personal Agents）正加速演进，但由于深度接管邮箱与私密数据，面临强效用与社交网络效应之间的天然张力；同时约会、招聘、视觉电商及大众娱乐等多人网络型赛道仍处大片蓝海。 ## Key takeaways - [00:56](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=56s) 消费端真实支付与流量分布严重脱节，按消费金额排名的前 50 款产品中，有 29 款从未出现在流量榜单上，展现出高价值工具极其隐蔽的吸金能力。 - [05:55](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=355s) 个人代理（如 Instinct 和 Muse）首月即展现极高客单与高绑定率，但高昂的单客推理与服务成本迫使其克制放量，无法复制传统社交产品凭借零边际成本极速裂变的路径。 - [09:12](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=552s) AI 助手的效用与其掌握的私人数据深度成正比，但因缺乏人际社交边界感和潜在隐私泄露顾虑，导致用户极力抗拒人际协同，难以自发形成人与人之间的网络效应。 - [11:40](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=700s) 早期硬核用户主要将消费级助手用于代码编写与技术自动化，而泛大众面对空白文本框则陷入无从下手的悖论，互动过程的私人化阻碍了使用玩法的病毒式模仿。 - [14:39](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=879s) AI 付费呈现极端幂律集中：排名前 10% 的付费用户贡献了超 50% 收入，前 1% 的用户月均支出达 903 美元，驱动了以开发、设计和变现为核心的新一代创造者经济。 - [18:30](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1110s) 大部分主流大众追求的是“消磨时间”而非“节省时间”，微弱的提效难以维系高频日常黏性，消费级 AI 破圈需激发将脑中构想变现的纯粹创造与娱乐之乐。 - [20:20](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1220s) 消费级 AI 依赖订阅制是受制于早期推理成本的暂态现象，未来随着小模型普及和单次成本下降，更符合C端习惯的广告与交易佣金模式将重新占据主导。 - [23:51](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1431s) 对话式 AI 掌握精准的实时情境与高商业意图（如出行规划与特色导购），在确保用户信任的前提下，其广告定向与直接闭环转化潜力远超传统搜索引擎与静态社媒画像。 - [29:21](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1761s) 三大模型商业路径显著分化：ChatGPT 兼顾规模与对话广告，Anthropic 坚定拒绝广告并全力推行每月 100 美元以上的高客单订阅，Gemini 则深度依赖 Google 原生生态绑定。 - [31:44](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1904s) 大模型巨头全力攻坚通用推理与代码而无暇顾及音乐版权等 IP 泥潭，为 Suno、ElevenLabs 等深耕特定模态工作流与交互界面的垂类创意工具留出了充裕护城河。 - [34:18](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2058s) Midjourney 虽遭通用大模型分流大众浅层流量，但凭借不可替代的审美控制力和专业调性在收入榜名列前茅；而视频生成则极度依赖多模态系统工程与海量数据训练。 - [37:16](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2236s) 大量消费级 AI 公司通过轻量化产品驱动增长（PLG）极速渗透进企业团队并建立 B 端收入，反衬出老牌巨头既不敢颠覆存量界面又困于单一聊天框的创新困境。 - [41:44](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2504s) 行业价值正重归软件层与交互体验，“套壳”偏见被打破；能够深度融合私有上下文、工作流模版并模仿个人语气的专用系统，沉淀了极难迁移的复合壁垒。 - [46:34](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2794s) 双向匹配网络（约会、招聘）、AI 原生社交、交易市场与大众娱乐（如 AI 微短剧）仍是消费级 AI 的广阔空白，从文本框走向视觉化探索与内容网络将是下一阶段质变的关键。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=0s) 个人代理崛起与消费级 AI 报告概览 对话开篇直指当前消费级 AI 领域最具颠覆性的趋势——个人代理（personal agents）。从 ChatGPT 首次让大众领略对话式界面的魅力，到如今智能代理直接嵌入即时通讯软件，软件的使用范式已经彻底改写。当下人们不仅是用代理来辅助写代码，而是在构建极其丰富完备的独立软件产品，它绝非仅仅套在底层大模型表面的一层“薄垫片”（shim）。与此同时，商业化层面的数据同样极为分化：尽管约有一半的美国人表示自己体验过 AI，但真正为 AI 产品支付订阅费用的美国消费者仅占约 4%；而在付费群体内部，消费呈现出极端的幂律分布——前 1% 的顶尖核心用户，仅凭个人信用卡每月在 AI 上的自费支出就高达 903 美元。此外，OpenAI 凭借广告业务迅速拉升的年化营收体量（run rate）也令人咋舌，这种增长速度在以往商业模式下即便上线广告产品也往往需要数年打磨。 ```mermaid flowchart LR accTitle: AI 用户层级与支出金字塔 accDescr: 展示从泛大众使用者到极客核心付费用户的层级与支出规模分布 subgraph UserBase["用户渗透率"] All["全美约 50% 用户<br/>使用过 AI"] Paid["约 4% 消费者<br/>付费订阅 AI 产品"] end subgraph Spend["付费集中度"] TopSpend["前 1% 核心高净值用户<br/>人均自费 $903/月"] end All --> Paid Paid --> TopSpend ``` - 全美约 50% 的人口报告使用过各类 AI 驱动的产品。 - 仅约 4% 的美国消费者目前在为 AI 工具支付常规订阅费。 - 付费顶端 1% 的个人高净值核心玩家，月均个人信用卡消费金额达到 903 美元。 --- ### [00:56](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=56s) 第七版 Top 100 榜单演变：流量固化与真实的消费支出 a16z 编辑合伙人 Elena Burger 与投资合伙人 Olivia Moore、Josh Elman 展开深入对话。Olivia 回顾了 a16z 追踪消费级 AI 应用以来的演化历程：在推出该系列榜单的最初阶段，评选仅依据全球原生 AI 网站的网页访问流量（traffic），因为彼时几乎所有用户都是通过浏览器里的一个 Prompt 提示框来调用 AI。然而随着应用形态快速演化，第七版榜单进行了史上最大幅度的扩充，首次引入了消费端真实支付数据。通过与 Earnest Analytics 合作提取消费者个人信用卡与借记卡流水（严格剔除了企业与中小商户公账支付），团队首次还原了普通消费者“真金白银”的流向。 从流量维度来看，现有格局已趋于固化板结。在合并网页与移动端的数据后，本次新上榜的产品仅有 11 款，创下历届榜单新人数量的历史新低。但一旦切入消费支出维度，战局立刻呈现出截然不同的全貌：在按消费金额排名的前 50 款产品中，竟有多达 29 款从未出现在两张流量榜单的任何一张上。正如 a16z 负责企业投资的 Sarah Wang 在《市场现状报告》（State of the Markets）中所点出的，无论在企业级还是消费级领域，“幂律法则”（Power Law）正在强力主导一切——绝大多数商业价值正在高度向极少数核心付费力量集中。 --- ### [03:31](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=211s) 从极客玩物到大众代劳：代理生态的代际跨越 尽管整体榜单由于统计周期看起来略显平淡甚至僵化，但最近数月底层势头却异常凶猛。整个范式正从单纯辅助生产力的“工具”（提效、加快搜索）跨向替人类跑腿办事的“代理”（get things done）。几个月前，开源工具 OpenClaw（后由相关团队并入更大平台）曾点燃极客圈层，人们为了本地跑 OpenClaw 甚至掀起抢购苹果 Mac mini 作为家用服务器的热潮。如今，随着 GrokBot、Town 以及 OpenAI 推出的全新 Dots 代理等工具的涌现，这一波浪潮正在加速进化。 最本质的变化发生于消费属性的蜕变：以往的产品必须具备极高的技术配置能力，而近三四周至一个月内涌现的消费级代理，已经达到了投资人愿意放心地推荐给家中长辈使用、且确信他们能安全上手的易用程度。即便这些新物种由于诞生时间极短未能完全反映在本次榜单排名的顶峰，其后台数据已经展现出惊人的蹿升态势。正如 Josh Elman 所言，虽然硅谷播客里成天在聊代理，但现实世界里 AI 正在更广泛地无孔不入；即便新面孔数量受限于统计门槛，老玩家的体量却全部在以指数级膨胀。 --- ### [05:55](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=355s) 爆款个人代理 Muse 与 Instinct 的实测数据与增长差异 在纯消费级助手阵营中，近期最为夺目的当属 Muse 和 Instinct 两款产品（早期先驱还包括 Poke 以及拥有庞大核心受众的 Tomo）。Instinct 公布的数据极为亮眼：用户体量达 10 万且正以日环比 10% 的惊人速度扩张；更为激进的是，约 40% 的用户在注册前三周内就主动绑定了信用卡，首月平均消费支出竟直接突破了 1000 美元大关。 作为科技圈瞩目焦点的 Muse 同样迎来现象级发布，在上线前 12 天内斩获了约 50 万次下载和 25 万日活用户。但 Elena Burger 随即抛出对比：若将 Muse 的爆发力与 Meta 此前推出的社交应用 Threads 相比，其在非科技圈普通大众中的渗透力依然相形见绌。Olivia 调取了可比维度的数据进行澄清：由于 Muse 当时仅在美加地区开放，在同样对比上线前 22 天的美加地区下载量时，Threads 狂揽约 1600 万次安装，而 Muse 维持在约 500 万次左右。 ```mermaid flowchart TD accTitle: Threads 与 Muse 早期冷启动机制对比 accDescr: 对比社交网络强网络效应与个人代理工具高服务成本之间的启动模式差异 subgraph ThreadsApp["Threads: 社交裂变模式"] T1["绑定既有 Instagram 账号"] --> T2["极低单客边际成本"] T2 --> T3["强迫性追求用户密度与网络效应<br/>前22天美加下载达 1600 万"] end subgraph MuseApp["Muse: 个人代理模式"] M1["独立端侧/代理形态安装"] --> M2["高昂的单客推理与服务成本"] M2 --> M3["克制放量，深耕单客高阶价值<br/>前22天美加下载约 500 万"] end ``` 这种增速差异的根源主要在于产品机制与商业逻辑的本质不同： 1. **获客通道与边际成本**：Threads 依托 Instagram 的底层账号图谱，几乎零门槛平移关系链，且单客边际服务器成本极低，天然需要迅速形成沟通的“临界规模”（Critical Mass）以激活网络效应；而 Muse 这类代理应用背后的模型推理与调用成本极高，团队主观上甚至不敢在极短时间内承接无节制的爆发流量，否则高昂的单客服务成本（cost to serve）会迅速压垮运营。 2. **产品形态与黏性壁垒**：Muse 早期主要作为深度私密、高度定制化的个人超级工具运行，并不极度依赖即时社交网络效应来实现留存。但 Josh Elman 强调，个人代理要想确立长期的商业护城河，未来依然必须演进出“平台级网络效应”与多边撮合市场（Marketplace）。 3. **生态连接的双刃剑**：代理正试图构建连接各类外部应用的生态，但已遭遇现实商业势力的防御阻击。例如亚马逊（Amazon）便明确设立禁令，严禁 Muse 代理在其电商网站上进行自动化的商品检索与静默代付；尽管如此，Muse 依然在迅速打通与数百家第三方商业服务商的直连通道，一场围绕消费入口与代理权限的控制权暗战已经全面打响。 ### [09:12](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=552s) 个人 Agent 的隐私高墙与网络效应困境 在探讨消费级 AI 助手时，Olivia Moore 指出，目前的助手类产品尚未真正解锁“人与人之间的网络效应”（person-to-person network effects）。核心阻碍在于，AI 软件比以往任何工具都更深入地洞悉用户的私人生活——用户不仅将其接入个人邮箱，还会向 AI 倾吐极其私密的想法。这就产生了一种天然的内在张力：产品对消费者越有用，它对个人的了解就越深；而一旦它知晓过多秘密，用户就会本能地抗拒将它引入与他人协作或社交的场景中。正如许多人会将 ChatGPT 的工作账号与个人账号严格分开一样，至今还没有团队彻底解决这一界限问题。 Josh Elman 补充道，在消费级 AI 领域，隐私、安全与信任正变得愈发关键。整体层面上，公众对“AI 是否有益于世界”的信任波动，已经在影响产品的普及和热度；而在微观层面，如果用户无法完全信任一个掌握了自己邮箱、日常生活乃至信用卡的软件，担心它做出不可控的出格行为或擅自泄露信息，这便会成为消费级 AI 应用规模化扩张的最大瓶颈。 人类天生懂得一套微妙的社交规范：我们本能地知道哪些信息可以公开、哪些从他人处听来的话应该守口如瓶。然而现在，人们却突然要求冷冰冰的软件也能遵循这种复杂的社交分寸，这带来了一整套全新的学习成本。 ```mermaid flowchart TD accTitle: AI 助手的隐私网络效应张力 accDescr: 解释用户私密数据输入与社交网络效应之间的结构性矛盾。 subgraph PrivacyTension["隐私与效用张力"] A["高频输入私密数据<br/>(邮箱/信用卡/个人心事)"] --> B["AI 深度个性化<br/>实用价值飙升"] B --> C{"跨人际传播抗拒"} end subgraph NetworkBarrier["社交网络受阻"] C -->|"缺乏人际边界边界感"| D["无法形成人与人网络效应"] C -->|"担忧失控或越权"| E["账号隔离<br/>(工作与个人分开)"] end ``` ### [11:40](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=700s) 早期用户的硬核画像：从高昂成本到“空白文本框”悖论 针对平台能力演进与大众接受度之间的博弈，Olivia 认为，平台与模型能力的进化速度远快于普通消费者的实际使用意愿。近期一些 AI 助手创始人公开表示，服务单个用户的月度推理成本甚至高达数百甚至数千美元，这令行业感到困惑。Olivia 引用了 Assistant Benchmark 负责人 David Polon（其追踪了大约 170 到 200 多个 Agent）的一项社区调研：在 1500 多名深度早期用户的社群中，即便是在面对 Muse 或 Instinct 这类被定位为消费级助手的工具时，排名第一的核心使用场景依然是**编写代码与技术自动化（coding and technical automation）**。正因为早期硬核用户把助手当成技术自动化管道在高频调用，才堆出了成百上千美元的账单；而对于面向泛大众主流消费者的助手，单月单客成本往往仅在十几美元左右。 Josh 剖析了大众用户面临的心理门槛：对于熟悉编程的人来说，AI 是无与伦比的杠杆；但对于普通人而言，打开这类产品就像面对一个“空白文本框”（blank box），根本不知道能拿它做什么。即便官方给出“预订餐厅”或“购买机票”的示例，如果用户当下并没有这些需求，便极难将其与日常生活联系起来。 人类通常是通过观摩他人的示范来学习新行为的。在过去的社交媒体时代，一个人发布了新玩法，其他人迅速模仿、二创，从而形成滚雪球般的网络爆款；但在 AI 时代，由于互动过程高度私人化，这种灵感裂变与病毒式传播几乎没有发生，如何将工具推向千千万万的主流大众仍是一个待解的难题。 - **Agent 追踪量级**：Assistant Benchmark 平台已收录并评测超过 170 至 200 余个 Agent。 - **早期社区规模**：David Polon 维护着 1,500+ 名核心早期采用者的互动社群，数据显示技术与自动化占绝对统治地位。 - **服务成本分化**：开发者高频调用下单月成本可达数千美元，而泛大众主流消费级助手的单月成本通常仅为几十美元。 ### [14:39](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=879s) 4.5% 付费群体与 $903 的超级创作者经济 讨论随后切入本期报告最核心的交易数据：当前约有 50% 的美国人表示自己使用过 AI，约 25% 的人认为自己几乎每天都在与 AI 打交道。然而，**消费者的实际付费行为极度向金字塔尖集中**。 根据 YipitData 面板监测，全美仅有约 4.5% 的消费者为 AI 产品订阅付费（部分数据源统计在 2% 至 2.5% 不等，尽管相比一年前已然翻倍）。更关键的是，在这 4.5% 的极少数付费群体内部，支出分布极度倾斜： 1. **排名前 10% 的付费用户**：贡献了全部 AI 消费收入的 50% 以上； 2. **排名前 1% 的超级硬核用户**：单月个人信用卡支出高达 **$903 美元**，贡献了 20% 的总收入； 3. **中位数付费用户**：月支出仅为 **$25 美元**； 4. **排名后 50% 的付费用户**：仅贡献了整体收入的约 16%。 ```mermaid flowchart TD accTitle: 美国家庭与个人 AI 付费结构分层 accDescr: 展示从全美人口到 AI 消费者的金字塔形支出倾斜。 subgraph PopTier["美国泛人群参与度"] P1["约 50% 美国人使用过 AI"] P2["约 25% 接近日常高频互动"] P3["仅 4.5% 真正为订阅付费"] end subgraph PayTier["4.5% 付费群体内部极化"] P3 --> T1["Top 1% 超级硬核用户<br/>人均 $903/月 | 占 20% 收入"] P3 --> T2["Top 10% 高端用户<br/>贡献 >50% 收入"] P3 --> T3["Median 中位线用户<br/>人均 $25/月"] P3 --> T4["Bottom 50% 长尾用户<br/>仅占 16% 收入"] end ``` 透过个人账单的底层数据，Olivia 指出，排名前列的超额支出工具大多集中在开发工具、生产力工具与创意工具上——即用于“构建、制作、变现”（build, make, sell）的软件，例如 n8n、Granola、Higgsfield、Manus 等。Josh 对此评价道，虽然这还没有覆盖所有人，但这标志着“全新一代创造者”（a whole new generation of makers）的崛起，许多人用个人信用卡自费购买工具，去写过去从未写过的代码、制作从未尝试过的艺术与视频。 Elena 总结指出，程序员和创作者天然具备一种思维模型，即“我如何借工具撬动时间杠杆”，但这种追求极致效率的视角目前仍然只属于少数人。Olivia 引述其投资组合公司 Wabi 的 CEO Eugenia 的金句，道出了消费级应用破圈的真正命门： > "大部分人并不是在寻找节省时间的方法，他们寻找的是如何消磨时间。"（"Most people aren't looking to save time. They're looking for ways to spend their time."） ### [18:30](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1110s) 从“节省时间”到“创造的纯粹快乐” 对话首先反思了消费级产品价值主张的演进。像 Netflix、TikTok 和 YouTube 这样的大型消费级娱乐与社交平台，占据了人们绝大部分的日常屏幕时间。相比之下，当前 AI 产品的典型宣发点通常停留在“如何把事情做得稍微快一点、容易一点、或者稍微更出色一点”。对于绝大多数普通大众而言，这种微弱的效率提升根本无法构成高频的、每小时甚至每天必用的核心价值驱动力。 拯救时间主要吸引的是那些极度追求效率（hyperproductivity）、专注于工作交付的人群。然而，一股全新的人群正在涌现：那些曾经想做视频却从未掌握 Pro Tools 或 Adobe Premiere 等复杂专业软件的人，或者曾经对编程望而却步的人，现在只需输入文字，就能将脑海中的想法直接落地变现。对他们而言，核心动力已经不再是“节省时间”，而是“将脑中的构想亲手变成现实的纯粹创造之乐（spending the joy of actually willing this thing into reality）”。这就解释了为什么在早期尝鲜阶段，会出现大家捧着笔记本电脑到处走、彻夜不眠用 Agent 疯狂写代码和做产品的狂热现象。尽管在当前数据中，这种高额消费主要集中在极少部分头部极客群体，但这正是大众赋能时代的先兆。 - 社交与流媒体平台（Netflix、TikTok、YouTube）构成了日常消费级应用的使用时长基石。 - 早期 AI 工具聚焦于“略微提效”，对非专业用户的高频黏性不足。 - 新兴门槛消除效应让未曾掌握 Adobe Premiere 或专业代码技能的用户获得了实现想法的能力。 - 头部 1% 的用户每月花费高达 900 美元，这被视为更多创造者获得赋能的先兆，而非孤立个例。 --- ### [20:20](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1220s) 消费级 AI 商业模式的倒挂：为什么订阅制不该是终点 Olivia Moore 提出了一个对消费级 AI 投资人来说略显激进的观点：她其实并不希望看到“直接为 AI 产品付费的用户比例仅有 4.5%”这一现象被简单粗暴地通过扩大直接订阅来解决。回顾整个消费级互联网的历史，当前的变现结构呈现出一种非常不自然的“倒挂（unnatural inversion）”。绝大多数成功的消费级科技巨头，其绝大部分收入均来自广告或交易佣金（transaction fees），而非向C端收取高昂的软件订阅费。 ```mermaid flowchart TD accTitle: 消费级 AI 商业模式现状与未来演进 accDescr: 对比当前依赖订阅的现状与未来转向广告及交易分成的演进路径 subgraph Current["当前消费级 AI 商业模式现状"] Sub["网页端 85% 采用订阅制"] Token["62% 采用积分/Token 计费"] AdLow["仅约 13% 采用广告或免费模式"] end subgraph Bottleneck["当前核心瓶颈"] COGS["高昂的推理成本 (COGS)"] Fear["过度增长导致成本失控的恐惧"] end subgraph Future["未来经典互联网模式回归"] CostDown["推理成本持续下降 / 低成本小模型普及"] AdsModel["高意图广告变现"] TxModel["交易分成 / 商业服务 (Wallet, Login)"] end Current --> Bottleneck Bottleneck --> CostDown CostDown --> Future ``` 榜单数据显示，Top 100 网页端产品中有约 85% 依靠订阅模式变现，62% 采用了代币（Token）或积分等额度充值机制，而仅有约 13% 采用了广告等“用户即产品”的免费变现路径。普通大众既没有为软件付费的充裕资金，也没有长期为软件买单的消费习惯。全球前 20 大消费级订阅服务中，除了上线仅三年半就奇迹般杀入榜单的 ChatGPT，其余几乎全是流媒体等内容媒体公司（如 Netflix）或高频履约平台（如 Amazon Prime、Uber）。 AI 领域之所以普遍依赖订阅，本质上是因为早期高昂的推理成本（COGS）。在传统的消费互联网创业中，前五年盈利甚至会被投资人视为异常警告信号——传统逻辑是先把边际服务成本近乎为零的用户密度做大，待网络效应建立后再通过广告和交易变现。然而，高昂的模型调用成本让 AI 创业团队害怕“过快增长”，如果缺乏计量计费限制，爆发式的免费流量很容易导致算力账单失控。但随着低成本小模型的成熟与推理单价的持续下跌，传统的广告和交易分成模式必将重归主导地位。 - 排名前 100 的消费级 AI 网页应用中，约 85% 支持订阅，约 62% 拥有 Token/积分充值通道，仅约 13% 试水了广告模式。 - 全球 Top 20 消费级订阅服务以媒体（Media）和高频综合平台（Amazon、Uber）为主，ChatGPT 是罕见的新入局软件服务。 - 传统消费级应用逻辑：零边际服务成本下先冲用户密度，后靠广告与交易分成变现。 - AI 初创公司当下面临两难：高推理成本迫使其过早收费设卡，抑制了规模化网络效应的积累。 --- ### [23:51](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1431s) 广告经济的重构：OpenAI 的 10 亿美元年化广告收入 Elena Burger 提到了最新 Top 100 报告中披露的一项惊人事实：OpenAI 的广告业务年化营收（run rate）已经达到了 10 亿美元（数据截至 8 月份，随着后续推进很可能更高）。这种以极快速度跨过 10 亿美元广告门槛的现象，在互联网历史上前所未见。 这一爆发主要归功于两点：其一是 OpenAI 已经具备了庞大的用户密度，其周活跃用户（WAU）最新公布数据已达到 12 亿；其二是转化率潜能。不同于传统搜索，对话式 AI 掌握了用户极其详细的个人语境。随着 "Login with ChatGPT" 以及未来可能内置的钱包（Wallet）等功能上线，广告主可以获得前所未有的精准触达与闭环转化能力。 | 维度 | 传统搜索引擎广告 | 消费级对话式 AI 广告 | |:--- |:--- |:--- | | **交互形式** | 关键词检索列表，单向展示 | 双向自然对话，针对个性化情境深度推荐 | | **用户意图** | 广泛的信息搜寻或单纯事实查询 | 从日常闲聊演进至旅游规划、深度购物决策等商业高意图 | | **植入难点** | 机械放置赞助链接，用户习惯跳过 | 极度私人化；生硬推销会严重伤害用户与 AI 之间的信任 | | **商业闭环** | 依靠外部外链跳转结算 | 结合原生登录（Login with ChatGPT）、内置钱包实现原生结算 | 探讨中特别强调了广告引入时的“信任危机（trust issue）”。用户把最私密的烦恼和疑问托付给 AI，如果在讨论健康问题时突然被强插减肥药广告，将会彻底毁掉产品的信任基石。因此，广告必须做得极其灵巧克制。目前 OpenAI 的试点广告以清晰的标识呈现，并自然融入对话逻辑。 用户对对话式 AI 的使用场景也正在经历关键转型：从最初查事实、改作业等低商业价值的查询，逐渐转向寻求人生建议、制定深度旅行路线乃至搜索特定商品。例如，Josh Elman 分享了自己近期的体验：在为纽约行程做旅行调研时，AI 自然推荐了当地即时活动的广告信息；在寻找一款专为代客泊车设计的钥匙扣（快速分离车钥匙与门禁卡）时，对话系统针对其具体痛点提供的商业推荐，远比传统电商搜索高效。由于此类高意图场景与传统推销迥异，广告在此刻转变成了高价值的直接答案，从而释放出惊人的商业变现潜力。 - OpenAI 广告业务已达 10 亿美元年化营收规模，背靠 12 亿周活跃用户（WAU）。 - 生态闭环动作：“Login with ChatGPT”已上线，钱包及交易功能预期将进一步打通交易链路。 - AI 广告成功的前提是维护交互的信任感，避免在敏感个人对话中机械粗暴地插入推销信息。 - 用户提问向旅行策划、特色硬件导购等高商业意图（commercial intent）场景迁移，极大抬高了广告转化的单价与点击价值。 ### [27:45](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1665s) 对话式搜索中的广告潜力与垂直类 AI 的变现 Josh Elman 分享了自己的一次真实购物经历。他在 Google 上怎么也搜不到心仪的东西，最后决定去问 ChatGPT。在多轮对话中，他向模型细致描述自己的需求，ChatGPT 随即给出了一些非常棒的参考方案。最后 Josh 补充道：“太棒了，我还希望它能是美国本土制造的。”结果 ChatGPT 竟从底特律一家专门从事金属加工的工厂里，为他挖到了一个极其独特、酷炫的钥匙配件。Josh 感慨，如果在传统搜索中没有这种来回推敲的对话体验，他根本不可能找到这个产品；尽管当时的对话里没有任何推广信息，但他意识到这背后存在巨大的广告植入空间——用户带着极强的真实购买意图进行多轮交互，广告商若能在此刻提供精准匹配，体验将极其自然顺畅，而他最终也确实去往了那家此前从未听闻过的工厂官网完成消费。 Elena Burger 对此深表认同，指出随着 OpenAI 逐步发力广告业务，其用户定向能力可能会达到前所未有的精准度。以往 Meta 凭借庞大的用户数据画像被视为广告定向的业界标杆，但 ChatGPT 拥有与用户当下正在进行的“实时活态对话”（live conversation），其意图捕捉能力甚至有望超越 Meta。Olivia Moore 补充说，广告不仅能为 ChatGPT 这样的平台带来巨大的变现收益，也会强力支撑那些更具垂直深度（Verticalized）的消费级 AI 应用。例如针对医生群体的 AI 产品 OpenEvidence，目前已经渗透了全美 50% 到 60% 的执业医师。当一个产品在极其高价值且身份明确的专业垂类中拥有如此高的用户密度时，定向广告的投放效果和商业转化力将会变得极其惊人。 - Josh 在 ChatGPT 的交互引导下，精准挖掘到底特律一家金属加工厂生产的定制配件，展现了对话式电商的闭环潜力。 - 区别于 Meta 基于历史数据的静态用户画像，生成式 AI 基于用户当下的实时交互意图进行广告定向，精准度上限更高。 - 面向医师的专业 AI 工具 OpenEvidence 在美国医生群体中已达到 50%–60% 的渗透率，展示了高价值垂直 AI 产品的广告变现壁垒。 --- ### [29:21](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1761s) 基础大模型三足鼎立：ChatGPT、Claude 与 Gemini 的用户分化与变现路线 Elena 将讨论引向了榜单常客——三大基础模型实验室 OpenAI、Anthropic 和 Google Gemini，探讨它们的用户群像与付费变现演变。Olivia 指出，从 a16z 首次发布报告至今，ChatGPT 在全球消费端无论在访问量还是营收规模上都牢牢占据统治地位，但 ChatGPT 身后的格局却在经历剧烈洗牌。在 Web 端流量上，ChatGPT 目前大约是 Claude 的 6 倍、Gemini 的 2 倍。然而，考虑到大部分 AI 消费主要由高净值核心用户支撑，其实际收入维度的参考价值正在显著超越单纯的访问量。 本期报告中最令人震惊的发现之一来自美国付费用户样本数据：Claude 的付费订阅人数已经正式超越了 Gemini。考虑到 Gemini 背靠 Google 强大的原生分发渠道和更为庞大的整体装机量，Anthropic 取得这一成绩极为不易。Olivia 分析，这一飞跃得益于 Anthropic 近期推出的一系列成功产品（如 Claude Design 等），以及自年初以来因多项高调事件（包括与美国国防部相关的合作舆论）带来的持续曝光。在变现模式的选择上，三大平台也展现出截然不同的路径分化： ```mermaid flowchart TD accTitle: 三大 AI 实验室消费端商业化路径分化 accDescr: 对比 ChatGPT、Claude 与 Gemini 在用户基数、广告态度及高客单价订阅上的差异 subgraph OpenAI["OpenAI (ChatGPT)"] A1["流量与营收绝对领先"] --> A2["美区付费用户规模约为 Claude/Gemini 的 3 倍"] A2 --> A3["探索对话广告与高阶订阅"] end subgraph Anthropic["Anthropic (Claude)"] B1["付费订阅量超越 Gemini"] --> B2["坚定拒绝广告 (No Ads)"] B2 --> B3["激进推行订阅：7.5% 付费用户选用 $100+/月 Max 计划"] end subgraph Google["Google (Gemini)"] C1["装机量与原生分发优势巨大"] --> C2["美区高阶付费占比仅约 1%"] C2 --> C3["深度融合 Google 搜索与生态系统"] end ``` - **付费用户基数**：在美区消费端，ChatGPT 的付费订阅用户数大约是 Claude 和 Gemini 各自的三倍左右。 - **高阶客单价分化**：Anthropic 旗帜鲜明地坚持“绝不做广告”（No Ads），因而对高客单价订阅的推进最为激进。其付费用户中有约 7.5% 购买了每月 100 美元以上的 Max 订阅计划；而 ChatGPT 与 Gemini 在该档位的订阅者占比均仅在 1% 左右，这反映出 Claude 聚集了一批对模型能力有极重度需求、愿意支付高溢价的专业核心用户。 --- ### [31:44](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1904s) 垂类创意工具的生存之道：音频壁垒与软件交互的专属价值 Elena 提出，各大模型实验室如今也都在涉足创意工具，甚至像 Anthropic 这类最初没有专门训练扩散图像模型（Diffusion Model）的实验室，也能凭借其卓越的代码生成能力让用户间接生成出高水准的图像与视频。在通用大模型无所不包的压力下，那些垂直类的独立创意工具究竟该如何立足？ Olivia 指出，垂直创意工具在特定模态中构筑了极高的防御壁垒，尤其是音频与音乐领域。OpenAI 和 Anthropic 等巨头当前正全力以赴投入到 Coding Agent 和 AGI 的研发攻坚中，根本无暇顾及音乐版权这一极其棘手的“IP 烂摊子”，更没有动力去单独开发一款竞品音乐模型来冒法律风险。这给了 Suno 和 ElevenLabs 充足的窗口期，使它们在流量榜和消费支出榜上均稳居前列。 Josh 赞同这一观点，并强调独立垂直工具的真正壁垒不仅在于底层模型，更在于其极致贴合用户工作流的软件交互层（Software Layer）。无论是 Suno 为专业创作或纯粹娱乐定制的音乐交互，ElevenLabs 独特的语音处理工作流，还是 Figma 为设计师量身定制的专业界面，通用模型都无法替代这种经过深度打磨的端到端产品体验。 - **通用巨头的研发侧重**：头部模型实验室优先调配算力与精力攻克代码代理和通用智能，无意陷入音乐领域的版权纠纷与合规泥潭。 - **音频/音乐垂直领跑**：Suno 在音乐生成领域拉开显著身位，ElevenLabs 在文本转语音（Text-to-Speech）和音频制作赛道确立了高黏性优势。 - **交互与工作流价值**：垂直创意工具的胜负手在于软件界面的深度设计、资产沉淀方式与交互质感，这些“面向特定人群定制”的体验构成了区别于通用聊天框的核心护城河。 --- ### [34:18](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2058s) 图像与视频前沿：Midjourney 的审美壁垒、中国视频模型与 Agent 驱动的创意演进 转到视觉与视频生成赛道，竞争格局则呈现出截然不同的态势。在图像领域，OpenAI（如 GPT-4o 图像生成能力的跃升）和 Google（通过 Nano、Imagen、Veo 等一系列迭代）已大幅侵蚀了早先独立图像生成网站的大众消费级流量。以 Midjourney 为例，在 a16z 最早的榜单中其流量名列前茅，如今在公开流量排名中已滑落出榜；然而一旦切换到“实际收入与付费榜单”，Midjourney 便强势回归。这证明普通小白用户可能会被通用模型的免费功能分流，但对审美质感（Taste）、艺术风格控制有严苛要求的高阶付费用户，依然会为 Midjourney 无法被替代的特有调性买单。 谈及视频生成时，Olivia 认为该领域目前是一片充满变数的“狂野前沿”（Wild Frontier）。视频生成需要将画面、音效以及视听时序同步完美结合，其系统复杂度极高，而这极度依赖海量训练数据的喂养。在这一背景下，中国厂商凭借能够高效调用海量数据进行模型训练的优势，在视频模型生成领域展现出了极具竞争力的表现。 最后，对话探讨了创意工具在智能体（Agent）时代的演变。Olivia 和 Elena 提到，目前通过向 Astra 等顶尖多模态模型发出指令，模型能够自动调用网络工具（Tool Calling）甚至自主安装环境来完成复杂的图形编辑与动画制作。这种由技术极客（Tinkerers）在推特（X）上分享的前沿玩法，通常在短短六个月内就会被封装成体验友好的独立消费级产品，推动 AI 从前沿技术真正沉淀为面向大众的可持续产品。 - **Midjourney 的分化表现**：大众浅层流量被大模型蚕食，但凭借独特的审美理解与精准可控性，在收入榜上稳居头部。 - **视频生成的复杂度**：视频涉及视觉、音效和帧级别的时间同步，训练门槛高昂；拥有海量数据训练优势的中国团队在这一赛道表现格外抢眼。 - **Agent 友好型工具的红利**：能够便捷响应智能体工具调用（Tool Calling）的创意软件将迎来全新增长机遇，极客探索正在加速向成熟消费级产品转化。 ### [37:16](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2236s) 从“大扩张”到巨头的界面困境：AI 产品的 PLG 突围 随着 Granola、Wisperflow、Superhuman 以及各类聊天即时通讯产品的爆发，消费级 AI 在分销模式与产品形态上正在经历激烈的演进。Olivia Moore 提到她在几个月前撰写的一篇文章《大扩张》（The Great Expansion），核心探讨了为什么大量消费级 AI 公司能够以惊人的速度转变为企业级公司。以往像 Canva 这样的公司，通常需要花费六年甚至更久才能建立起真正意义上的团队与企业级收入；而如今诸如 Krea、Replit、ElevenLabs、Gamma 以及 Whisperflow 和 Granola 等产品，依托产品驱动增长（PLG）模式迅速扩张，因为具备工作生产力属性，被员工自发引入企业内部，迫使创始人迅速补齐隐私、安全保障和团队订阅计划（Team Plan）。这演变成了一种极其高效、低成本的获客与分销路径。 与此同时，这一趋势也凸显了现有老牌巨头的核心软肋：他们既不愿、也无法自我蚕食已有的核心交互界面。最经典的代表就是 Google，在 AI 浪潮来袭后，外界并未看到它对 Docs、Gmail 或日历进行根本性的重塑。即便像 OpenAI 和 Anthropic 这样的大模型领头羊，除了局限在 ChatGPT、Codex 或 Claude、Claude Code 现有的对话与代码界面之内，在尝试构建其他全新形态时同样成效有限。尽管各大实验室私下都试图孵化针对初创公司的“杀手级产品”，但当团队精力被底层模型迭代大量分散时，想要跳出原有界面进行激进的产品形态创新，在工程和产品管理上都面临极大阻碍。 - **快速扩张的企业级跨界**：诸如 Granola、Whisperflow、ElevenLabs、Gamma 等产品，借助工作场景的天然延伸，以轻量门槛让团队迅速自发采纳，完成从个人向企业级订阅的无缝过渡。 - **现存巨头的界面包袱**：传统科技巨头不敢颠覆自家既有产品的既定 UI，而大模型实验室则困于“聊天框”单一维度的成功，给专精于垂直场景交互的独立应用留下了巨大的切入窗口。 ### [39:23](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2363s) 底层模型跃升与硬件初创公司的窗口期 过去一年到一年半之间，AI 产品的爆发式增长在很大程度上归功于底层基础模型能力的质变。在早期，许多开发者花费大量心血去打磨自然、易用、便于团队分享和协同的交互界面，但因为当时模型智商有限，产品很难完全兑现预期价值；而一旦底层模型能力迎来突破，那些早已备好优秀界面的产品，其实用性便呈指数级爆发，享受到了“水涨船高”（a rising tide lifts all boats）的红利。本次发布的报告直接打破了“全行业最终只剩一两家公司独揽所有体验层”的误区——在极其通用的搜索之外，无数垂直界面的重塑才刚刚开始，普通用户只在乎哪些工具切实好用并能融入自己的日常生活。 从初创公司与巨头的竞争格局来看，OpenAI 和 Anthropic 尽管在概念上依然算初创企业，但实质上已膨胀为数千人的庞然大物，俗话说“船大难掉头”。以本次营收榜单上的硬件公司 Plaud（AI 录音卡片硬件）为例，其模式为购买硬件并搭配软件订阅，先在东方市场立足随后攻入美国。OpenAI 理论上有能力且未来必然会推出硬件设备，但作为一家大型企业，产品立项往往需要繁重的跨产品集成、合规层层审批与利益制衡，耗时极长。因此，在垂直硬件与高度专注的软硬件结合领域，动作敏捷的初创公司拥有极其可观的时间窗口。 ```mermaid flowchart TD accTitle: 模型与体验层飞轮效应 accDescr: 展示基础模型能力的提升如何让专注交互体验的初创公司释放出几何级的产品价值 subgraph InfraLayer ["底层基础设施与模型层"] M1["基础模型能力突破 (OpenAI / Anthropic 等)"] M2["推理成本下降与速度提升"] end subgraph ProductLayer ["垂直产品与体验层"] P1["打磨极致的专用界面与交互流 (如 Granola, Plaud)"] P2["轻量化 PLG 获客与自发团队传播"] end subgraph Outcome ["最终商业闭环"] O1["水涨船高：实用价值呈指数级释放"] O2["形成上下文与网络效应的护城河"] end M1 --> P1 M2 --> P1 P1 --> P2 P2 --> O1 O1 --> O2 ``` ### [41:44](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2504s) 软件层的价值回归：告别“套壳”偏见与累积个人上下文 当前科技行业的核心价值正在重新回流至软件层。过去一段时间，行业普遍质疑创业公司只是大模型的“套壳”（wrapper），或者稍微好听一点的称呼“马鞍/支架”（harness）。然而事实证明，随着快速、低延迟、结构化输出的模型触手可及，真正的创意重回产品体验本身、网络效应的搭建，以及对用户上下文的深入沉淀。 能够穿越模型迭代周期的产品，其核心护城河在于“持续复合的个人价值”（compounding personal value）乃至未来的“多人协同价值”。以定位为专业消费级（prosumer）智能体助手的 Town 为例，该产品能够直接打通 Slack、电子邮件和 Google Drive 等全套生产力工具，深刻理解用户及其所在组织的深层背景并代为执行任务。这种系统沉淀了用户独一无二的语言风格与工作流程体系（playbooks）： - **99.9% 与 85% 的天壤之别**：一封模仿本人语气达到 99.9% 相似度的邮件，用户只需花费 10 秒确认即可发出；而如果相似度只有 85%，用户则需要耗费 10 分钟以上的时间去人工修正每一个细节，两者的使用意愿完全不可同日而语。 - **深层锁定的用户资产**：一旦 AI 学习并构建了详尽的个人行为模型，用户根本无法轻易迁移至第二天冒出的任何竞争工具。 - **超越“套壳”的完整软件生态**：产品不仅仅是模型的浅层调用接口，而是结合了自动化流、私有上下文记忆和完整端到端交付的丰富系统，模型只是其中负责推理的一环。 ### [45:13](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2713s) 消费级 AI 的广阔空白：从工具到网络型产品 目前纯粹的 C 端大众对于生成式 AI 的认知与使用，依然极度局限在替代传统搜索引擎（如 Google 检索）、辅助学生撰写论文、辅导作业，或是上班族撰写邮件等“办成某件事”（getting things done）的任务型范畴内。纵观当前的消费级 AI 图景，几乎所有跑出来的头部初创公司都集中在生产力工具、图像与视频设计编辑、搜索问答等单人效能场景。 然而，属于 Web 2.0 时代的浩瀚版图在 AI 领域依然是大片空白。大量高度依赖多人互动的网络型产品类别（network categories）几乎尚未被真正开发，其中约会交友（dating）便是一个典型赛道。虽然目前已有先锋初创企业在探索这一方向，且有望在未来跻身消费级百强榜单，但整体而言，从社交、娱乐到交易平台，消费级 AI 仍处在极其早期的开荒时代，蕴藏着巨大的创业机遇。 ### [46:34](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2794s) 消费级 AI 的空白领域：社交、匹配网络与电商平台 在 a16z 第七期 Top 100 消费级 AI 应用榜单中，依然存在大片尚未被充分开拓的蓝海市场。Olivia 指出，目前榜单上完全没有专注于约会（Dating）的产品，招聘（Recruiting）同样处于极早期的空白状态。从底层逻辑来看，约会与招聘在本质上高度相似，都是依赖精准双向匹配（Matching）的双边网络。与此同时，“原生社交 AI”目前也未见真正爆发的独立平台，绝大多数所谓的社交 AI 现象，依然只是用户将 AI 生成的图文和视频内容搬运分发到现有的成熟社交平台（如 X、Instagram 或 TikTok）上。 另一个极具想象力的大类是交易型平台与市场（Marketplaces），包括日常电商网购、房产交易（Home buying）和零售业。这类产品未来的形态仍充满悬念：它们究竟会内置在像 Instinct 或 Muse 这类通用个人 Agent 中，还是作为独立的 AI 原生产品存在？亦或是走向一种混合架构——即个人 Agent 充当调度层，在后台调用专门的 AI 原生电商网络。此外，在游戏与娱乐领域，底层模型的生成质量长期以来一直无法产出让大众持续喜爱的长内容，目前唯一的例外是正在迅速爆发的 AI 微短剧（AI micro-dramas）。但随着多模态技术的飞速迭代，这一内容瓶颈被打破只是时间问题。 - **匹配网络（Matching Networks）空白**：约会与招聘尚未诞生杀手级消费 AI 应用。 - **社交 AI 现状**：缺乏原生平台，目前仅充当传统社交媒体的内容生产工具。 - **交易与电商形态**：未来可能是独立垂直应用，也可能是个人 Agent 直接调用的后台基础设施。 - **娱乐内容拐点**：AI 微短剧率先跑通，更广泛的大众娱乐内容正在逼近临界点。 ```mermaid flowchart TD accTitle: 消费级 AI 空白领域与未来架构 accDescr: 展示消费级 AI 在匹配网络、内容娱乐以及交易市场中与底层智能体可能的交互形态 subgraph BlankSpaces["未完全开发的垂直场景"] Match["双向匹配网络<br/>(约会 / 招聘)"] Social["原生社交网络<br/>(非存量社媒搬运)"] Ent["大众娱乐与游戏<br/>(AI 微短剧 / 叙事)"] end subgraph CommerceStructure["交易与电商基础设施"] Agent["个人通用 Agent<br/>(如 Instinct / Muse)"] Marketplace["AI 原生交易网络<br/>(电商 / 房产 / 零售)"] Agent -->|"调度与匹配需求"| Marketplace end ``` --- ### [47:42](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2862s) 超越对话框：下一代视觉购物与叙事网络 Josh 对此表示高度赞同，并深入剖析了消费级交互形态的演进局限。ChatGPT 和各类即时通讯 Agent 将世界带入了“对话即交互”的全新范式，这种体验固然强大，但对话框绝非全能。以购物为例，消费者天生需要强烈的视觉探索与发散体验，而不是对着文字窗口做问答。用户需要看到直观的选项，体验量身定制的款式，甚至提出细致的个性化诉求——比如指着一件衣服说“我很喜欢这件，但如果衣领上有滚边设计就更好了”。AI 完全可以做到实时修改渲染，并协同供应链直接定制生产，这背后的商业重构空间极为广阔。 在娱乐与叙事领域，AI 目前仅仅揭开了冰山一角。Josh 认为，当年 TikTok 的巨大成功在于为全球用户提供了一个极低门槛的短视频表达网络，让无数原本默默无闻的人得以向世界展示创意。然而，大众脑海中依然封存着大量未被表达的绝妙想法与故事。AI 的核心价值在于赋能普通人，让他们具备一流的叙事与视觉表达能力，进而衍生出由大众创作、大众消费的全新内容网络。各大垂直领域正站在质变的前夜，而一旦底层大模型具备足够的生成与推理能力，商业价值最终依然会牢牢沉淀在专注于体验和交互的软件层（Software layer）。 --- ### [48:54](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2934s) “节省时间”与“消磨时间”：消费级 AI 的终极商业潜力 节目的最后，对话回归到一个根本性的产品哲学命题：AI 应用究竟是帮人“节省时间”（Save time），还是帮人“消磨时间”（Spend time）。纵观前七期 Top 100 消费级 AI 榜单，几乎所有涌现并获得商业化进展的产品（无论是编程工具、写作助手还是音视频编辑器）都在主打效率与工具属性，本质上都是在帮用户节省时间。 然而，回顾互联网和移动互联网的发展史，能够成长为体量最为庞大、吸金能力最强的超级巨头（如社交网络、流媒体娱乐、短视频和游戏），无一例外都是“消磨时间”型产品。如果消费级 AI 想要真正触达全球十亿级大众，就必须跨越纯生产力工具的局限，诞生属于自己的原生娱乐与消费级杀手应用。无数优秀的创业者正投身于这一领域进行探索。随着这一探讨的结束，主持人 Elena Burger 感谢了 Olivia Moore 与 Josh Elman 的深度剖析，并宣布本期第七版 Top 100 消费级 AI 盘点圆满收官，期待在第八版榜单发布时带来更多突破性进展。 ## Notable quotes > [14:39](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=879s) Most people aren't looking to save time. They're looking for ways to spend their time. — Eugenia > [18:30](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1110s) spending the joy of actually willing this thing into reality — Josh Elman > [20:20](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1220s) unnatural inversion — Olivia Moore > [11:40](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=700s) coding and technical automation — Olivia Moore > [11:40](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=700s) blank box — Josh Elman > [14:39](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=879s) a whole new generation of makers — Josh Elman > [29:21](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=1761s) No Ads — Olivia Moore > [34:18](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2058s) Wild Frontier — Olivia Moore > [37:16](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2236s) The Great Expansion — Olivia Moore > [39:23](https://www.youtube.com/watch?v=aCvrzhwUxg0&t=2363s) a rising tide lifts all boats — Olivia Moore ## Mentioned a16z, Elena Burger, Olivia Moore, Josh Elman, Sarah Wang, David Polon, Eugenia, Earnest Analytics, YipitData, OpenAI, ChatGPT, GPT-4o, Codex, Dots, Anthropic, Claude, Claude Design, Claude Code, Google, Gemini, Nano, Imagen, Veo, Google Docs, Gmail, Google Drive, Meta, Instagram, Threads, Amazon, Amazon Prime, Uber, Netflix, TikTok, YouTube, X, Slack, Adobe Premiere, Pro Tools, Figma, Canva, Replit, OpenClaw, GrokBot, Town, Muse, Instinct, Poke, Tomo, Wabi, n8n, Granola, Higgsfield, Manus, OpenEvidence, Suno, ElevenLabs, Midjourney, Astra, Krea, Gamma, Whisperflow, Superhuman, Plaud, Mac mini ## Video description a16z Editorial Partner Elena Burger sits down with investing partners Olivia Moore and Josh Elman to unpack the seventh edition of a16z’s Top 100 Consumer AI Apps, including a new dimension this time: what consumers are actually paying for. The data reveals a striking power-user economy. Only a small share of consumers currently pay for AI, but among those who do, spending is heavily concentrated at the top. Olivia and Josh discuss why developers, creators, and other power users dominate spending today, and why subscriptions may not be the business model that ultimately brings consumer AI to everyone. They also dig into the rise of personal agents, the different trajectories of ChatGPT, Claude, and Gemini, how ads could reshape AI economics, and the enormous amount of consumer white space still left to build, from shopping and entertainment to social, dating, and marketplaces. Timestamps: 00:00 - Intro 00:56 - Inside the 7th edition 05:55 - Muse and Instinct's launch numbers 09:14 - The privacy wall for personal agents 14:39 - 4.5% pay, $903/month at the top 20:20 - Why subscription is the wrong model 23:51 - OpenAI's $1B ad run rate 29:21 - ChatGPT vs Claude vs Gemini 32:30 - Where creative tools still win 45:12 - The white space: dating, shopping, entertainment Resources: Follow Olivia Moore: https://x.com/omooretweets Follow Josh Elman: https://x.com/joshelman Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures. [Source](https://www.youtube.com/watch?v=aCvrzhwUxg0)

[14:39] Eugenia: Most people aren't looking to save time. They're looking for ways to spend their time. [18:30] Josh Elman: spending the joy of actually willing this thing into reality [20:20] Olivia Moore: unnatural inversion [11:40] Olivia Moore: coding and technical automation [11:40] Josh Elman: blank box [14:39] Josh Elman: a whole new generation of makers [29:21] Olivia Moore: No Ads [34:18] Olivia Moore: Wild Frontier [37:16] Olivia Moore: The Great Expansion [39:23] Olivia Moore: a rising tide lifts all boats

[17:23] Olivia Moore (quoting Eugenia Kuyda): Most people aren't looking to save time. They're looking for ways to spend their time. [39:23] Olivia Moore: A larger ship is harder to kind of steer or to turn around [17:23] Elena Burger and Josh Elman: How do I get leverage over this thing? [47:42] Josh Elman: I really like that, but I wish it had some piping on the collar or something [41:44] Olivia Moore: 99.9% sounds like me [41:44] Olivia Moore: 85% sounds like me [13:17] Josh Elman: How did you make that? [37:16] Olivia Moore: The Great Expansion

[0:00] The most interesting big new trend is in [0:02] personal agents. [0:03] >> We're in this world where we're so [0:04] excited by what Chachi did with turning [0:06] things into conversation and now these [0:08] agents we can get in our messaging apps [0:10] to turn them into more conversation. [0:12] There's this incredible paradigm. [0:13] >> About half of Americans report using AI. [0:16] Around 4% of US consumers are paying a [0:20] subscription to an AI product. The top [0:23] 1% user is spending $93 per month [0:26] personally on their personal credit [0:28] cards on AI. The reality is people are [0:30] now building software. They're building [0:31] these really rich products. They might [0:33] be using their agents to write some of [0:34] that code, but when they're delivering [0:36] this whole experience to a user, it is [0:38] much more than just a shim on the model. [0:41] The [0:41] >> run rate that OpenAI has reached through [0:43] advertising. What does that tell us? [0:45] That would previously take a company [0:48] years and years and years to get to even [0:50] after launching an ads product. What I'm [0:52] fascinated to see in terms of the [0:54] evolution of the ads product is [0:58] Welcome to the A16Z podcast. I'm Elena [1:01] Burgerer and I'm joined by Olivia Moore [1:03] who's the author of our ongoing top 100 [1:05] consumer AI apps report and Josh Elman [1:08] on our consumer team. Today we're [1:11] digging into the seventh edition of this [1:13] report. Who's paying for AI? Why the big [1:16] assistants are heading in different [1:17] directions? and whether personal agents [1:19] can turn everyday errands into the next [1:22] big consumer business. Team, thanks so [1:24] much for joining. Um, [1:26] >> thanks so much for having us. [1:27] >> Olivia, let's just get right into it. I [1:29] would love to hear uh your top takeaways [1:32] from this report. [1:33] >> Yeah, so this is the seventh time we've [1:35] done this list. The very first time we [1:37] only ranked all global websites that [1:40] were AI native by just traffic, like [1:42] hits on the website. Uh, and that worked [1:44] then because the way that almost [1:46] everyone was using AI was like a prompt [1:48] box product in a browser. We've had to [1:51] evolve that a lot over the last few [1:53] years as AI expands. So, this list is [1:56] our most expansive yet. Um, we added [1:58] revenue data for the first time. So, we [2:00] worked with the EBIT data to get [2:02] specifically consumer card spend data. [2:04] So, not accounting for enterprises or [2:06] SMBs. Uh, and we ranked products by what [2:09] consumers are actually paying for. I [2:11] think for me the biggest takeaways from [2:13] this report is on a traffic basis, it [2:16] looks like a lot of these products are [2:18] kind of settling. Um, we only had 11 new [2:21] products across the web and mobile list [2:23] combined, which is the least we've had [2:26] in every uh in any of the prior editions [2:28] of the series. But the spend data really [2:30] introduced a lot of new variables [2:32] because of those 50 that ranked for [2:34] spend, 29 of them were not on either [2:37] traffic list. M [2:38] >> um and that also I think kind of works [2:40] towards or connects into the other big [2:42] conclusion of the report which is just [2:44] how much of a power user game consumer [2:46] AI is right now. [2:47] >> Yeah. [2:48] >> Yeah. That's something that's funny. We [2:50] we also recently released our state of [2:52] markets and something that Sarah Wang [2:55] actually commented directly on um is she [2:58] she said hey there there's such a power [3:01] law game happening now. It's happening [3:02] at the enterprise level and it's [3:04] happening at the consumer level. Um, so, [3:08] uh, one thing that's probably notable [3:11] is, yeah, we have the fewest number of [3:14] newcomers on this list and it seems like [3:16] the space is oifying, but I think the [3:18] past few months things have actually [3:20] been picking up a little bit more and I [3:22] think I think the most interesting big [3:24] new trend is in personal agents. Um, so [3:27] what are you seeing? What are the big [3:29] shifts that are happening? [3:31] >> Yeah, it's funny. The last report that [3:32] we did um it ended right before the cut [3:36] off of when openclaw traffic would come [3:38] in. So we had done kind of like a retro [3:40] before we published the report and we [3:41] had found that it would have been very [3:43] high up in the rankings. Um, and this [3:46] time actually openclaw is nowhere in the [3:48] rankings because the traffic has [3:50] >> completely declined as I'm I think that [3:53] the team was acquired by OpenAI. they're [3:55] probably working on, you know, the agent [3:56] assistant products there, but also it's [3:59] kind of been overtaken, I think, by some [4:00] of the both other attempts at proumer [4:03] agents like Grockbot, um, Town has some [4:07] fantastic agents. Dots now by OpenAI, [4:10] which is new. Um, but also on the [4:13] consumer side, we're finally seeing [4:14] things that I would feel comfortable [4:16] giving to a family member and telling [4:18] them that they can safely use. Most of [4:20] those also because to your point this [4:22] consumer assistant trend has literally [4:24] been the last 3 weeks to a month. A lot [4:27] of those didn't appear uh on the ranks [4:30] yet. But we did pull the data to see [4:32] what's happening with them. And it's [4:33] been um very very fast growth and yet [4:37] still not quite reaching the mainstream [4:40] user is what I would say. You know, I [4:42] think we're in this really interesting [4:43] era that did really change at the end of [4:46] last year, beginning of this year with [4:47] OpenClaw, that AI went from something [4:50] that we were using as a tool to enhance [4:52] our productivity to maybe search for us [4:54] faster, something that would actually [4:56] get things done for us. And I think [4:58] OpenClaw was an incredible pioneer to do [5:00] this and I know they're still doing a [5:01] lot of work through their open source [5:03] foundation to continue to advance that. [5:05] But it really turned a lot of heads on, [5:07] wait, what can these products actually [5:09] be? And we were seeing people buying Mac [5:11] minis. We were seeing people run all [5:13] their own data with their own local open [5:15] cloud. There was a huge run on that. But [5:17] it was really inspiring to kind of feel [5:20] this new energy of what is possible. And [5:22] you know like most trends you know that [5:25] spark happens. It takes a little while [5:27] for those first products to get built. [5:29] It takes a little while for the momentum [5:30] to then start getting those even [5:31] released in the market for users to see [5:34] them. And so, you know, in some ways [5:36] though, this report is showing like [5:38] that's what we're talking a lot about, [5:39] you know, on our podcast in Silicon [5:41] Valley. But this really does a great job [5:43] representing what's happening in the [5:44] broader world and and that, you know, I [5:46] do think the truth is AI really is [5:48] showing up everywhere. Even if there's [5:50] not a lot of new entrance, all the [5:52] entrance are all getting so much bigger. [5:54] >> Yeah. [5:54] >> Yeah. I think probably the biggest [5:56] things on when we talk about pure [5:58] consumer assistance uh the biggest [6:00] developments have been muse and instinct [6:02] I would say although there's many other [6:04] fantastic products that we kind of list [6:06] in the report that have also you know [6:07] poke was a very early pioneer here. [6:10] Exactly. Tommo is has a very big [6:12] audience and is doing incredibly well. [6:14] Um Instinct had announced like 100,000 [6:17] users growing 10% day over day. uh they [6:20] had announced like 40% of users [6:22] connected a credit card in the first [6:23] three weeks and they were spending you [6:25] know over $1,000 on average in their [6:27] first month. Um Muse is interesting [6:30] because uh in many ways very successful [6:34] launch especially within the tech [6:35] community. I think the number was like [6:37] 500,000 downloads and 250,000 active [6:40] users in the first 12 days or so. But if [6:43] you look at it versus Threads, which was [6:46] another, you know, somewhat recent big [6:48] splashy meta launch, um it just it still [6:51] pales in comparison in terms of appeal [6:54] to the average person who doesn't work [6:56] in tech. [6:57] >> Yeah. So, if you look at the the [6:59] downloads, um Muse is still only US and [7:02] Canada. So, we grabbed the threads US [7:04] and Canada downloads only. In the first [7:06] 22 days or so threads, it had about 16 [7:09] million downloads and Muse is still [7:10] sitting around 5 million. over that same [7:13] time period. [7:13] >> Is the distribution mechanism different? [7:16] Like I remember with threads, it was [7:19] already tied to your Instagram account [7:21] and there was kind of this like very [7:23] natural distribution. Is are they trying [7:26] a similar strategy with Muse or is it [7:30] just a different distribution strategy [7:32] too? I mean, from I'd be curious what [7:34] Josh thinks, but from my point of view, [7:35] it's not as aggressive for sure. Uh, and [7:38] also part of that is because there's a [7:40] much higher cost to serve these users. [7:41] So they don't want it to get too big too [7:43] fast. Well, and I also think threads [7:45] very much had a critical mass effect, [7:47] which is if they didn't get a lot of [7:48] people on it communicating quickly, you [7:50] don't actually get to that leverage of a [7:53] new network starting to form. Whereas [7:54] Muse is still such a personal tool that [7:57] you know, they're pushing it and there [7:58] it's very exciting to kind of even see [8:00] this early just adoption and people [8:02] having these experiences. And again, if [8:04] you were using OpenClaw in January, you [8:06] were having incredible experience using [8:07] Crockbot when that first launched [8:09] earlier this summer, spring, you were [8:10] having these incredible experiences. And [8:12] now we're seeing as you know I think [8:14] Olivia said you know our moms our [8:15] families you know people who don't just [8:18] live on the tech what's next in AI cycle [8:21] and I think that's what's really [8:22] exciting about this consumer. I I assume [8:24] when we do this again next time we're [8:26] going to see massive adoption. But I [8:27] think it hasn't needed to push as hard [8:29] to have that early network effect. But [8:32] then I also think about what's going to [8:33] make these things stick around for a [8:35] long time. And it is going to be things [8:36] like network effects and marketplaces [8:39] and all the businesses you can interact [8:41] with those agents. That's going to be [8:42] really transformative too. [8:44] >> Yeah, this is one of the things we talk [8:45] about a lot when it comes to consumer [8:47] agents and assistants which is I think [8:50] they are getting kind of the compounding [8:52] effects of building out a platform or at [8:53] least it seems like they are in terms of [8:56] the ecosystem of other apps that are [8:57] available and more easily usable by the [8:59] agent. We've seen that cut both ways. [9:01] Like Amazon is like, "We're not going to [9:03] allow Muse [9:04] >> to kind of browse and purchase products [9:06] on our site, but also, you know, Muse [9:08] has already hundreds of other [9:10] partnerships with different, [9:12] >> you know, to be able to buy things on [9:13] Shopify." [9:14] >> Exactly. Yes. Um I don't I still don't [9:17] feel like any of the assistant products [9:19] have unlocked persontoperson network [9:22] effects. And part of this for me is like [9:25] this software knows us more intimately [9:28] than any other product in the past [9:30] >> based on email, but also like I find [9:32] myself telling extremely personal things [9:35] to my AI. And so there's kind of this [9:37] natural um tension between the more [9:41] useful the product is to you as the [9:43] consumer, the more it knows about you [9:46] and like then do you really want to [9:47] introduce that into other contexts with [9:49] other people. It's the same thing as [9:51] like many people have a separate chatbt [9:53] work account and personal account and no [9:55] one has quite cracked that yet. [9:57] >> No. And I think you know as we talk [9:59] about consumers and AI, privacy, safety, [10:02] security are becoming more and more [10:03] important. I mean there's obviously the [10:05] overall trends of you know trust in AI [10:07] and and this kind of belief that AI is [10:09] good for the world that is in some ways [10:12] you know affecting adoption and [10:13] affecting some of the excitement around [10:15] these products. But on a very personal [10:16] level, if you don't trust the thing that [10:18] you're giving it your most intimate [10:20] access to, your email or your life or [10:22] your credit card and it might do rogue [10:23] things that you wouldn't expect or share [10:25] things with other people you wouldn't [10:26] expect, you know, that's going to be the [10:28] biggest impingement on these things [10:29] really growing is as massive consumer [10:32] apps. So, it's a we're in a fascinating [10:34] time where we're all learning these new [10:35] norms and these are norms that like as [10:37] people we're actually very good at it. [10:39] We're very good at knowing what [10:40] information we're comfortable sharing, [10:41] what information if somebody gives it to [10:43] us, what information that we're trusted [10:44] to be able to pass on to somebody or [10:46] should be protecting and keeping [10:48] ourselves. And we're just like all of a [10:49] sudden expecting software to do this, [10:50] too. It's a it's a very, you know, [10:53] interesting new set of things we're [10:55] learning. [10:56] >> Yeah. Well, you you both just brought up [10:58] like two interesting sort of facets of [11:01] this. One is like there's this [11:03] endogenous force which is all of the [11:06] platforms and how they allow agents to [11:08] interact with them and then there's more [11:10] like this exogenous force of like how [11:13] how comfortable am I using this and and [11:16] how maybe I mixed up indogenous and [11:19] exogenous one is on the outside one is [11:21] on the inside [11:23] what the platforms are doing and one is [11:25] how you feel about it or how you feel [11:27] about the agent and I I'm curious which [11:29] which do you think um is going to get [11:32] resolved sooner and and who's like whose [11:36] onus is that? Like whose whose [11:38] responsibility is resolving those kinds [11:40] of questions? I think what the platforms [11:43] are capable of is improving faster than [11:46] true consumer willingness to use them. [11:48] Um, which we've seen kind of throughout [11:51] AI so far. It's interesting because a [11:54] lot of people are like seeing these [11:56] numbers that founders of these assistant [11:58] products are posting that it costs like [11:59] hundreds or thousands of dollars a month [12:01] to serve a user and like how is that [12:03] possible? Um, and I was stumped as well. [12:07] But I uh have been working with David [12:08] Pollon who runs assistant benchmark [12:11] who's fantastic. So [12:12] >> there's a recent podcast with him. [12:14] >> So basically what he's done is he's [12:16] created a website where they track I [12:18] think over 170 agents. I checked it last [12:20] night. It's probably like 200 something [12:22] this morning. Um, and they he has users [12:25] kind of use them and then rank them on [12:27] success on different tasks. But the [12:29] other really interesting thing that he [12:31] does is he hosts kind of group chats and [12:33] communities of early adopters of these [12:35] agents. So he has a group of I think [12:37] it's 1500 plus users. And he pulled some [12:39] cool data for us which is like what are [12:42] they doing with agents? what are they [12:43] talking about doing with agents? And the [12:46] vast majority still the number one use [12:48] case was coding and technical automation [12:50] even for these consumer assistant [12:51] products like like Muse and Instinct. [12:53] And so under that framing it's like oh [12:56] of course it costs that much. This is [12:58] like when we talk to some agent or [13:00] assistant products that are uh aimed [13:02] towards the more mainstream consumer, [13:04] they're seeing costs in the tens of [13:06] dollars, not the thousands of dollars a [13:08] month. So um I think that is maybe a [13:11] representation of the fact that we have [13:12] a very long way to go in terms of [13:15] consumers getting comfortable with these [13:16] products [13:17] >> and I think you know I think people [13:18] who've been coding and have been working [13:21] with AI I mean it's been such an [13:22] incredible enhancer and that shows up in [13:24] a lot of the data you know on the on the [13:26] top 100 here like it's incredible and [13:28] then we've people who've been able to do [13:30] that have had these great ideas for all [13:31] the other things in their life but when [13:34] you just hand this to somebody new it [13:35] feels like a blank box and it feels like [13:37] what should I do with this and maybe [13:39] there's examples of getting a [13:40] reservation or booking a flight, but if [13:41] you're not actively seeking a [13:42] reservation or booking a flight, it's [13:44] hard to figure out how to personalize [13:46] it. And so much of what we do as people [13:48] is we learn from others and we get these [13:50] great ideas and we try the other things [13:52] people tell us. So much of how social [13:54] media has worked is people put things on [13:56] social media and other people see that [13:57] and they try the same thing and then it [13:59] kind of spawns these effects. I try they [14:01] make a little riff on it and then those [14:03] riffs become you know cascading [14:05] snowballs and that's how great ideas [14:07] flourish and with AI because so much of [14:09] it's so personal we haven't even seen [14:10] any of that flourish and yet you know [14:12] outside of areas like coding and [14:14] productivity you know all of the [14:16] creative work you know I I created this [14:18] incredible video and how'd you do that I [14:20] use AI um you know there's so many tools [14:23] that do that that that are you know been [14:25] growing and are here but it's so [14:27] interesting that you know now we have to [14:28] figure out how to mainstream this for [14:30] not the you know millions and millions [14:32] of people who are using these and paying [14:34] and and everything but for the you know [14:36] tens hundreds and you know hundreds and [14:38] hundreds of millions [14:39] >> if you look at broad usage data now [14:41] about half of Americans report using AI [14:45] and about 25% of Americans think they [14:48] are interacting with AI probably on [14:50] close to a daily basis uh but the [14:52] spending data is much more concentrated [14:55] so it depends on the source um our Yepit [14:58] panel shows probably sub 5% around 4 [15:02] 1/2% of US consumers are paying a [15:05] subscription to an AI product. Some of [15:07] their sources say as low as 2 or 2 and a [15:09] half or maybe a little bit higher. It is [15:12] expanding. So that's like doubled from a [15:14] year ago. Um but that spend is still [15:17] extremely concentrated. So even within [15:20] that top 4.5% of people who do spend the [15:25] top 10% of those are like more than half [15:28] of the revenue and the top 1% are 20% of [15:32] the revenue whereas the bottom 50% are [15:34] like 16%. Right? So like the top 10% of [15:37] users are driving this market. This was [15:40] the other interesting thing about [15:41] revenue data is we were able to look at [15:43] it kind of like on a almost person [15:45] byperson basis what it looks like. Who [15:48] are these people who are spending on AI? [15:50] >> The top 1% user is spending $93 per [15:55] month personally on their personal [15:56] credit cards on AI. [15:58] >> Wow. [15:58] >> And the median is spending $25. [16:01] >> And again, all of this is within the [16:03] only the 4.5% that are spending at all. [16:06] Yet also was helpful for us because we [16:08] could see what products are they [16:11] actually spending on like on a on on a [16:13] personal basis. And probably [16:15] unsurprisingly, it was mostly developer [16:18] tools, um, productivity tools, creative [16:22] tools, things used to like build, make, [16:25] sell. [16:26] >> Um, things like NADN, Granola, [16:29] Higsfield, Manis, those all were way [16:32] over represented amongst the top [16:34] spenders. I [16:35] >> mean, in a way, this is actually really [16:36] cool. I mean, even though these numbers [16:37] aren't everybody yet, like we're [16:39] enabling this whole new generation of [16:40] makers of people who are, you know, not [16:42] just using this for work, but actually [16:44] using their personal credit cards to [16:46] code things they they never coded before [16:48] to make art, make videos, make content, [16:51] and, you know, really have this power of [16:53] making. And so, it's actually in some [16:55] ways, you know, while it's all still so [16:57] early, it's actually really cool how [16:58] much making is going on that people are [17:00] actually so excited they're willing to [17:02] spend. If you're spending $900 a month [17:05] on a set of AI tools, you're really [17:08] using that to like hopefully make things [17:10] faster, better, have more ideas in your [17:12] head coming to life than ever before. So [17:14] that's kind of like when I when I was [17:15] looking at the categories between [17:17] creativity and coding like [17:20] >> you know that's actually hopefully [17:22] really really powerful. [17:23] >> Yeah. [17:23] >> Yeah. Well, I I I think this is a point [17:25] that's really important, which is I [17:27] think developers and coders are people [17:29] who are naturally attuned to sort of ask [17:32] how do I get leverage over this thing? [17:34] How do I build a product that allows me [17:36] to get leverage over my time? And AI [17:39] obviously supercharges that. But it's it [17:42] is a specific kind of mindset, I think, [17:44] and it's a specific kind of world view. [17:47] And I think the part of part of my [17:50] understanding of these numbers is it's [17:52] still a world view that gets reflected [17:55] only among a small smaller group of [17:58] people. And the question might be just [18:00] like how do you expand that world view? [18:02] How do you help people understand like [18:04] here are all the ways you can get [18:05] leverage over your time and over your [18:07] life? [18:08] >> We talked about this a little bit in [18:09] this report in terms of areas where we [18:11] have not yet seen AI native startups pop [18:14] up or categories where we haven't. Um, [18:17] >> Eugenia, who's one of our portfolio [18:19] CEOs, runs a company called Wabby, which [18:21] is fantastic. She has this great quote. [18:23] I'm probably going to botch it, but it's [18:24] something like people, most people [18:26] aren't looking to save time. They're [18:28] looking for ways to spend their time. [18:30] >> Um, like this is why social media, [18:32] entertainment, Netflix, Tik Tok, all of [18:34] these, YouTube are so are the most used [18:37] consumer products that we have. And I [18:40] think that much of what we've seen in AI [18:42] thus far to your point is like how do I [18:44] do this thing a little bit faster, a [18:46] little bit easier, a little bit more [18:48] impressively. And that's not an [18:51] incredibly compelling daily or hourly [18:53] active value proposition for most [18:55] people. [18:56] >> No, I think I think this, you know, [18:59] saving time is still very, you know, [19:02] hyperproductivity people who are focused [19:04] on, you know, work or being creative. [19:06] But I'm also seeing this next wave of [19:08] people who are maybe used to be coders. [19:11] They're no longer maybe people who, you [19:14] know, want to make videos but never [19:16] quite mastered the, you know, ProTools [19:18] and Adobe Premiere or something else or [19:20] they're all of a sudden able to sign up [19:22] for these things and take the ideas in [19:24] their head and make them happen. So, I [19:26] do think, you know, we're starting to [19:28] see these rise up. And again, as you [19:31] know, I look at all these numbers as [19:32] we're still so early. Yes. and and the [19:34] fact that we're even here and that we're [19:36] talking about how, you know, big some of [19:39] these companies are and that 1% of [19:40] people are spending $900 a month. Like [19:43] >> I see these as harbingers of what's to [19:45] come as more people feel empowered, as [19:47] more people feel like they can create [19:48] where it doesn't it's not about saving [19:51] time. It's about spending the joy of [19:53] like actually willing this thing into [19:56] reality. You know, we there there were [19:57] all these jokes of, you know, oh, when [20:00] people were in this like the peak phase [20:02] of everybody was using their open claw [20:03] for the first time, their agents, they'd [20:04] walk around with their computers and [20:05] their laptops. They had to keep them [20:06] open and they they were like, I'm not [20:08] sleeping. I'm coding and I'm making [20:10] things, you know, more than I ever have. [20:13] And I think that's starting to, you [20:15] know, again, shows up at the top of the [20:16] data, but I do think it's starting to [20:18] show up sort of everywhere this list. I [20:21] actually don't necessarily want to see [20:23] that 4.5% of people paying for AI [20:26] products directly expand, [20:28] >> which is maybe controversial because I'm [20:30] like invest in consumer AI products and [20:32] I'm a consumer AI maximalist. [20:34] >> But the reason I say that is because [20:37] >> substantially all of the consumer AI [20:39] revenue thus far has come from direct [20:41] subscriptions. [20:43] >> Um, like on this list, we looked and it [20:45] was like, you know, 85% or something of [20:47] our web list monetized via [20:48] subscriptions. Another 62% monetized via [20:52] kind of credits or token extra usage [20:54] payments. Only like 13% had ads or other [20:59] options where like you are the product [21:01] instead of paying for the product. And [21:03] that's like a if you look back at the [21:06] history of the consumer internet, that's [21:08] like a pretty unnatural inversion. Like [21:10] almost all of the really big consumer [21:13] technology companies that we have now [21:16] make the vast majority of their revenue [21:18] from ads or from transaction fees, not [21:21] from subscriptions. Because I think the [21:23] truth of it is most similar to the fact [21:26] that not everyone is looking for a [21:28] coding tool to help them, you know, [21:29] start a business or or create and sell a [21:32] product. Most people don't have the [21:35] funds or don't want to spend their funds [21:37] on software. If you look at the top 20 [21:40] consumer uh subscription products [21:43] globally, Chachi BT is already on there [21:46] which is crazy in like three and a half [21:48] years of growth. Almost all the other [21:49] ones are media companies [21:52] >> or big platforms like Amazon or even [21:55] Uber things like that uh where you have [21:57] a very frequent purchase behavior. So I [22:00] actually am a big believer that we need [22:02] to move. We have transcended the need [22:04] for everyone to buy a subscription to AI [22:06] and we need to see these other business [22:08] models come back. [22:09] >> You know, I think we're still at the [22:10] early days of inference costs coming [22:11] down. We're still at the early days of [22:14] people finding ways to run lowerc cost [22:16] models to build great products. And as [22:18] they do that, you know, a lot of the [22:20] other business models that have always [22:21] been on the internet from, you know, [22:22] transaction fees and ads will finally [22:25] start working. But we we need that to [22:27] happen. And that's I think where we're [22:29] still, you know, there's so much [22:30] learning and and so so much that like, [22:33] well, if it's going to cost me a lot to [22:34] serve, at least let me just charge money [22:36] so that people pay for it became sort of [22:38] an early default. [22:39] >> Yes. [22:40] >> I mean, you can probably speak to this [22:41] from seeing many prior eras of the [22:43] consumer internet, but like it used to [22:45] be, at least for me as a consumer [22:47] investor, if a company was making money [22:49] in the first 5 years, it was like, whoa, [22:52] what's happening? The the rule was like [22:53] let's build density of users and because [22:55] they cost near zero to serve, we can [22:59] then wait and make money via ads or [23:01] start charging for transactions once we [23:03] have liquidity and density. And that [23:04] just isn't the case given how high COGS [23:07] are. [23:07] >> No, I think I think this cost thing is a [23:09] is a really interesting challenge and [23:11] companies get afraid of growing too [23:13] fast. Yeah. like they actually if they [23:15] grow too fast and they don't have meters [23:18] on how they're charging for it, you can [23:20] actually have out of control spend that [23:21] can be very very hard to manage. So, you [23:23] know, it's created this this very [23:25] different shape of what we see in the [23:27] top apps and the top 100, you know, [23:30] traffic and everything right now. But [23:32] again, this is again we're all still so [23:34] early. There's so much to build and [23:36] learn and bring cost down and bring [23:38] value up and introduce advertising and [23:41] introduce other ways to to spend money. [23:43] And by the way, if if it is we subscribe [23:45] to these services we love and that's [23:47] actually what we get continual value [23:49] from, that's amazing, too. [23:51] >> I think this is a a great transition to [23:53] talk about ads. Yeah, [23:55] >> your uh consumer uh top 100 report had a [23:59] really really interesting fact in there [24:01] which was the run rate that OpenAI has [24:03] reached through advertising. They're at [24:06] a billion dollars now in in annual run [24:08] rate on on advertising and [24:11] >> what does that tell us? You know, [24:13] >> again, the number was from August and so [24:16] there's probably a chance that it's [24:17] reasonably higher even a month or so [24:19] later. But they have been very slowly, I [24:21] would say, rolling out ads um with kind [24:24] of a network of partners and they're [24:26] already at, you know, a billion billion [24:29] in annualized run rate. Like that would [24:32] previously take a company years and [24:34] years and years to get to even after [24:36] launching an ads product. I think [24:38] there's two interesting things there. [24:39] One, they have real density now. So I [24:42] think the last reported number a few [24:43] days ago was 1.2 billion weekly active [24:46] users, which is like a real density. And [24:49] then the second thing and what I'm [24:51] fascinated to see in terms of the [24:53] evolution of the ads product is [24:56] hypothetically they should be able to [24:58] charge more and have higher conversion [24:59] rates because they just know so much [25:01] more about you as a user and chatbt has [25:04] launched things like login with chatbt. [25:06] I think a wallet is coming, things like [25:08] that. And so you could imagine that um [25:12] it becomes a much more compelling value [25:16] prop for the advertisers and then that [25:19] just like juices the ad business even [25:21] more. Uh but we'll see. It's very early [25:23] days there. [25:24] >> It's early days and I think you know [25:25] there have been some really funny [25:26] conversations about ads and AI. if you [25:28] are talking to this incredibly personal [25:30] thing and then it's like let me tell you [25:32] about health and let me introduce you to [25:33] this like crazy product that can help [25:35] you you know lose weight fast like [25:38] that's not building trust with AI and so [25:40] I think the other really interesting [25:41] thing is like how do we as a ecosystem [25:43] these products are so different they're [25:44] so much more personal than ever before [25:46] they're giving you answers that are [25:48] customtailored to what you're doing and [25:50] what you're asking about and what it [25:51] knows about you that actually [25:53] introducing ads has been done very [25:54] definitely I've actually been really [25:55] impressed with the open AI roll out of [25:57] ads what I've ever seen an ad? It's it's [26:00] very clearly labeled and it actually [26:01] feels like part a natural ad to the [26:04] conversation that I'm having. And I [26:06] think we're also, you know, one of the [26:07] other shifts is people are starting to [26:09] rely on, you know, chat with a lot more [26:11] for a lot more queries. It might have [26:12] used to been like look up a fact or tell [26:14] me the history of this thing or help me [26:16] with my homework, not necessarily [26:18] problem solving and things we're [26:19] advertising, but now people are going it [26:21] to it for real advice or shopping or [26:23] trying to figure out exactly, you know, [26:26] where they want to go or what they want [26:27] to do. And those are great opportunities [26:29] when you're in a discovery mode where [26:31] ads are actually often a great answer to [26:33] help you solve that. And so a lot more [26:34] of these commercial queries and these [26:36] commercial intent creates that too. But [26:38] again, it has to be done very very [26:39] deafly, not this weird interruptive very [26:42] personal experience I'm having that ads [26:44] just get shoved into the conversation. [26:46] >> Um, I'd be I'd be super curious to how [26:49] the unit economics on ads work. like [26:51] presumably it's a it's a lower cost [26:54] question or lower cost compared to you [26:56] know a coding or or something like that [26:58] or image gen. Um, so, so for OpenAI, you [27:03] know, what what kinds of margins are [27:06] they seeing on ads is a super [27:08] interesting question. [27:09] >> Um, I curious also, Josh, what what kind [27:12] of ads do you see? Because I I I don't [27:14] get any ads. I guess I I I guess I pro [27:18] enterprise so we don't Yeah. [27:21] You know, I I was doing a bunch of [27:22] travel research recently and um was was [27:25] going to be in New York for a few days [27:26] and I just started seeing some ads for a [27:28] couple things that were going on in New [27:30] York that I might have wanted to do and [27:32] I just found that really useful and like [27:34] I've started using it more for [27:36] commercial shopping. I was looking for a [27:38] a special um keychain that would have a [27:41] really easy separation for a valet key [27:43] and I was having trouble finding this [27:45] actually on Google and I finally was [27:46] like, "Let me just go ask ChachiBT." And [27:48] I went back and forth and I was [27:49] describing what I wanted and started [27:51] giving me some really good examples. And [27:52] at the end I was like, great. And I'd [27:53] also love one that's made in America. [27:55] And it actually found this incredible [27:57] like option from a factory in Detroit [27:59] that often makes a lot of other metal [28:01] work and other stuff that had this like [28:02] really cool key for me. And like it was [28:05] one of those things that if I had never [28:06] had that chance to really go back and [28:08] forth and actually have that [28:09] conversation and if there had been other [28:12] ads that got no no ads were presented in [28:14] that conversation, but it was entirely [28:16] open for ads that they would have helped [28:18] me find stuff and eventually I went to [28:20] this website. I'd never heard of that [28:21] company before. Um and it was a great [28:24] experience. [28:24] >> So interesting. I feel like we may see [28:27] as they lean into it, OpenAI uh being [28:29] able to do crazy good targeting uh on [28:33] consu like even better than I think Meta [28:35] which previously was like best in class [28:36] for this just because they knew so much [28:38] about you but like you're having this [28:39] live conversation with them. [28:41] >> I do think that it'll help chatbt but [28:44] also I think it will also support a lot [28:47] of these more focused almost [28:49] verticalized to use the enterprise term [28:51] consumer AI products. So another company [28:53] that has seen early success in ads is [28:55] open evidence which is um an AI product [28:59] for doctors and it already has 50 I [29:02] think the most recent numbers were 50 to [29:03] 60% density of US all of all US [29:06] physicians on the product and so when [29:08] you have that kind of density of a [29:10] really valuable audience and it's like a [29:12] targeted audience because you know [29:14] they're physicians you can I think [29:16] advertise really effectively and there [29:18] will be and should be more of those I [29:20] think. [29:21] >> Yeah. Is it is it worth talking and [29:23] taking a step back and talking about uh [29:26] OpenAI, [29:27] Anthropic and Gemini and and sort of [29:30] these have these are companies, these [29:32] are labs that have been on the list [29:34] pretty much since the beginning. Um are [29:37] we seeing any kind of evolution in in [29:39] the way people are paying for these [29:41] products? Um what can we surmise about [29:44] the different kinds of users of all [29:46] three? Uh super super curious what [29:48] you're what you're seeing. Yeah, [29:50] absolutely. Um, from the very first time [29:53] we did this report, the main conclusion [29:55] was like chatbt is the dominant global [29:58] user by both usage and revenue on the [30:00] consumer side. That continues to be [30:02] true. Everything below chatbt has seen, [30:05] I think, more movements and more [30:06] shakeups. So, chatbt is now ahead of [30:09] Claude um around 6x on the web, around [30:14] 2x uh ahead of Gemini on the web. [30:17] probably going forward given a lot of [30:20] the usage is concentrated in these power [30:22] users who are paying um the revenue [30:25] numbers could be even more useful here. [30:27] Um, and probably that's one of the [30:29] biggest surprises to me in this report [30:31] was we got a panel of US spenders and [30:35] Claude has actually passed Gemini in [30:37] terms of number of paid subscribers [30:39] which is kind of crazy given Gemini has [30:41] a much bigger install base and user base [30:44] overall and it also has like the natural [30:46] distribution mechanism of something like [30:48] a Google. I think that's a big, you [30:50] know, accomplishment of course for [30:51] Anthropic. I think it's come on the wave [30:53] of a bunch of very successful products [30:55] that they've launched over the last few [30:57] months like claw design. Uh it's also [31:00] come from a lot more press that they've [31:02] had starting with the department of war [31:04] stuff all the way back in I think it was [31:05] February. Um the other main difference I [31:09] would say in how these products are [31:10] monetizing and again chatbt still far in [31:13] the lead by monetization. They have [31:15] around three times more paid consumer [31:17] subscribers in the US than both Gemini [31:19] and Claude. Um, but Anthropic has said [31:22] very famously, no ads. So, they're much [31:25] more aggressive about subscription. They [31:28] have around seven and a half% of [31:30] subscribers on the $100 plus per month, [31:33] like the max plan. [31:34] >> And that's like 1% for both Chachi and [31:37] Gemini. And so those users naturally [31:39] look different and are doing different [31:40] things on the product even in that paid [31:42] user base. [31:44] >> Should we talk about creative tools? And [31:47] um I think one thing that's interesting [31:49] is actually all of the big labs have [31:51] gotten really good at creative tools. [31:53] And one thing that that actually like [31:56] Anthropic kind of famously abstained [31:59] from but then somehow achieved anyway [32:01] was like creating an image model or [32:03] video model but now people are making [32:06] images and videos because coding is so [32:09] good at creating like images and videos. [32:12] Um so it's interesting that you know [32:15] they didn't build the diffusion model [32:17] yet they're still capable of doing uh [32:20] image gen but but about the the pure [32:24] creative tools the the companies that [32:26] have kind of like verticalized in that [32:28] category what are we what are we seeing? [32:30] >> Creative tools is very interestingly [32:32] distributed in terms of who is [32:34] succeeding where. So there's a few [32:36] categories of creator tools or maybe [32:38] modalities is the better term where the [32:41] labs have not focused energy and [32:43] attention for many reasons we could [32:45] discuss. Uh a lot of those are around [32:47] audio. So like text to speech music 11 [32:51] labs inso are now very reliably near the [32:55] top of our traffic list and are also [32:57] ranked very highly in our spend list. [32:59] The labs have so many things to do. [33:01] They're building coding agents. They're [33:03] building AGI. Does it make sense at this [33:05] point for them to deal with all of the [33:07] IP headaches that Suno has had to go [33:09] through and launch a competing music [33:11] model? I don't know. And in the [33:13] meantime, Sunno has been able to kind of [33:14] run away with a really strong lead. [33:16] >> Yeah. And you know, I'm a big believer [33:18] that, you know, the labs are going to do [33:19] what they're going to do and they're [33:20] going to keep growing and they obviously [33:22] are great baselines for everything. But [33:24] in so many of these markets, you're [33:25] seeing really specialized products that [33:27] are really focused knowing who their [33:29] audience is and what they're doing [33:30] that's able to build really custom [33:32] things. You know, if you want to make [33:33] music just for fun or for something that [33:35] you're really producing, Sunno is an [33:37] incredible tool and it speaks to you [33:39] from the moment you start using it [33:40] through the end. You know, 11 Labs, if [33:42] you need voice and audio within the [33:44] product that you're making, it just has [33:46] a way of working and and interacting [33:48] with it that feels completely different. [33:50] when you look at a bunch of the video [33:51] tools, um, you know, when you look at [33:53] design tools, even like Figma, like [33:55] they're so designed to to use deeply for [33:58] the thing you're trying to express [34:00] yourself and it speaks to designers, not [34:02] to people who, you know, want to just [34:05] skip that step. And I think we're going [34:07] to see so much more value from these [34:08] really bespoke interfaces where the [34:10] software layer of all the things you can [34:12] do and the things you save and the way [34:14] you interact with it become really where [34:16] the value sits, not just the models. [34:18] >> Yes. So this is the second category I [34:20] think of creative tools which is um [34:22] images and video that looks different [34:24] from audio. On images I think to your [34:27] point like OpenAI has really leaned into [34:29] images quite successfully like images [34:32] 2.0 is very very good both for image [34:34] creation and editing and then of course [34:36] Google with the various iterations of [34:38] Nano Banana and VO. I think they have [34:41] candidly taken away a lot of the [34:43] highlevel consumer traffic that was [34:45] previously going to these standalone [34:47] image generators. Um so like Midjourney [34:49] was fairly high on our first ever [34:51] rankings. It has fallen off the traffic [34:53] rankings completely but to your point [34:56] empower users if you look at the revenue [34:58] rankings midjourney is back. So people [35:00] the power users still want the model [35:02] that they feel has taste has an [35:05] aesthetic sense and that they're able to [35:07] tune in a really specific way. Um, and [35:10] then that brings us to video, which to [35:12] me is kind of like a wild frontier right [35:14] now because [35:16] uh, Chinese companies that can train on [35:19] any data and do uh, have a real [35:21] advantage. Video is just so complex to [35:24] generate with like the visual, the [35:27] sound, the visual and the sound synced [35:29] and looking a specific way. Uh, that I [35:32] think like the more data you have to [35:33] train the better. They've been [35:34] successful there thus far. Um, I think [35:37] more broadly in creative tools, I'd be [35:39] curious if either of you have [35:40] experienced this, but I have recently [35:42] been asking Astra and other really [35:44] powerful models to generate or create [35:47] uh, generate or edit creative for me and [35:50] it does a very very good job and it will [35:52] go tool call to elsewhere on the [35:53] internet. And so I think that creative [35:56] tools that are agentfriendly [35:58] uh, and agent accessible are going to [36:00] increasingly see really compelling [36:02] tailwinds. [36:02] >> Yeah. Yeah. It's actually crazy what you [36:05] can do if you just like empower an agent [36:07] and just say like, "Yeah, install [36:09] whatever you want to install." Um, I had [36:12] it I actually had uh Astra and 5.5 go [36:17] like headtohead and animating some stuff [36:19] for me and like [36:20] >> it it was just like it knew exactly what [36:22] to do. It's like yeah. So, [36:24] >> and and what I love about this is, you [36:26] know, this is the tinkerers, the people [36:27] at the edge are figuring out these [36:29] capabilities and we're going to build [36:31] products over the next year that are [36:33] going to show up on these lists because [36:35] they're going to take all of this [36:36] complexity of what you can do and [36:38] package it up in a way that makes it [36:40] really easy to use. And like that's [36:41] this, you think of this gap of like all [36:43] AI versus consumer AI, it's exactly this [36:47] this movement of packaging these up and [36:49] building great product experiences. [36:51] Yeah. that really turns it into into [36:53] these things that sustain. [36:55] >> Yeah. Yeah. And I guess it it has to [36:57] start with the tinkerers and then they [36:59] post on X and then you know six months [37:02] later we get a real like product and [37:04] startup out of it or like you know 12 [37:06] different startups all competing. [37:08] Outside of the web window there have [37:12] been a big proliferation of of mature [37:16] products. So there's Granola, uh there's [37:19] Whisper Flow, there's Superhum. [37:21] Um what what are you kind of seeing [37:24] there? And then obviously the ones that [37:26] exist in chat and messaging, so in [37:28] instinct. Um what are you seeing when it [37:32] comes to distribution and what are you [37:34] seeing just when it comes to the kinds [37:36] of products and users that they're [37:37] approaching? [37:38] >> Yeah, I think there's two maybe [37:39] interesting threads to pull on here. [37:41] when I I wrote this piece called maybe [37:44] six nine months ago called the great [37:45] expansion which is basically like why [37:48] are consumer companies becoming [37:49] enterprise companies so quickly uh like [37:52] things like uh crea you know replet um [37:55] 11 labs gamma these products previously [37:59] like Canva took six plus years to have [38:02] real team and enterprise revenue and [38:04] what's happening is that they're very [38:05] successful and growing via PLG [38:07] whisperflow and granola another great [38:09] examples there and And because they are [38:12] kind of work productivity related, they [38:15] get pulled into enterprises and then the [38:17] founders are like, "Oh, we got to add [38:19] privacy and security and a team plan." [38:21] But it's very effective, very lowcost [38:24] distribution. And again, it works really [38:25] well for these tools that are can be [38:28] used personally but are like work [38:29] adjacent. Um, I think things like [38:32] Whisper Flow and Granola and Superhum [38:35] are another good example of this trend [38:38] we've seen in general, which is [38:40] incumbents being unwilling or unable to [38:42] cannibalize their existing interfaces. [38:44] The classic example of this is Google. [38:46] Like we haven't really seen them [38:47] reinvent Docs or Gmail or calendar for [38:51] the AI age. I think it's even happening [38:53] with open AAI and anthropic where they [38:56] have been much less successful in [38:58] anything that is not a launch within [39:01] their existing chatbt or codeex or [39:04] claude or claude code interfaces. [39:06] >> Um so and I know many of them have you [39:09] know attempted to incubate killers for [39:11] these many startups but I think it is um [39:15] difficult from a product perspective for [39:17] them to innovate beyond that especially [39:19] when they have so many other things on [39:20] their plates. I think all that's true [39:23] and I think you we now as we get used to [39:25] some interfaces we're just still in such [39:27] a period of incredible reinvention like [39:29] a lot of the reason we've seen so much [39:31] growth over the past year year and a [39:32] half is the models got so much better [39:34] and the models got so much better so the [39:36] people who had been spending time [39:38] creating these incredible interfaces [39:39] that were easy to use that were that [39:41] were natural that were you know easy to [39:43] share and get a bunch of your team on [39:45] and you know whisper flow and granola [39:47] were just so easy to get started with [39:49] you then would say oh my gosh now [39:50] they've gotten so much better as the [39:52] models have gotten better that the [39:53] products utility gets so much better [39:55] too. And so we're kind of this this [39:57] incredible you build something great, [39:59] but it may not do everything you want [40:00] because the model capabilities weren't [40:02] quite there. Then the models get better [40:03] and your product gets to explode in [40:06] demand and utility and this kind of [40:08] rising tide lifts everything. And so to [40:10] think that you know we only have one or [40:11] two companies that can build that [40:12] experience layer I think is what this [40:14] sort of report sort of debunks. Now [40:16] those are certainly some of the top [40:17] companies and these very general use [40:19] cases but it's you know the reason that [40:21] so many other things are growing is [40:22] there's so many interfaces that get to [40:24] be reinvented and and the sort of rising [40:26] tide list everything and then you know [40:28] the infra value may still go to the [40:31] model providers or may go to the other [40:33] things that are serving inference and [40:34] there's a whole different conversation [40:36] there but what consumers are adopting [40:38] are things that work for them and are [40:39] easy to to use and bring into their [40:41] lives. [40:42] >> Yeah. Maybe to also make the to continue [40:44] to make the case for startups here. Open [40:47] AAI and Anthropic in many ways are [40:50] startups, but they have been so [40:52] successful that they have become big [40:54] companies with thousands of people and [40:58] there are just so many extra [40:59] consideration like a a a larger ship is [41:02] harder to kind of steer or to turn [41:04] around. Yeah. [41:04] >> Um and so I think that's like one of the [41:07] companies on our revenue list this time [41:08] is Plaude which is an AI notetaker [41:11] device. um that started kind of in China [41:14] in the east and now has come to the US [41:15] and you buy the device and you can also [41:17] buy a subscription. Um OpenAI can and [41:20] will do a hardware device but as we've [41:23] seen with them it just everything takes [41:24] longer when you're a big company. Uh you [41:27] have to integrate with maybe other [41:29] existing products. There's more checks [41:30] and balances it has to go through. So I [41:33] do think that the big labs will continue [41:34] to be very very successful and there's a [41:37] lot more for them to do and I think that [41:39] startups will continue to be able to [41:41] find these quite big windows of [41:43] opportunity. [41:44] >> Yeah. And well it also sounds like the [41:47] the question for startups is how do you [41:49] build a product that as you said Josh [41:51] gets better as models get better but [41:54] doesn't get cannibalized as models get [41:56] better. um which is you know what like [41:59] like an interesting kind of sweet spot [42:02] to exist in. I mean, part of the reason [42:04] that I was so excited to join Andre and [42:06] Horowitz a couple months ago was this [42:08] belief that the value really now is [42:10] moving back to the software layer that [42:12] now that we have these incredible models [42:13] at our, you know, capabilities and we [42:15] can access them, you can build really [42:17] rich products that leverage the models [42:20] to do it, you know, and there's new [42:21] innovations like, you know, Jeb from [42:23] Typesafe that allows even developers to [42:25] build things in even new ways. is and we [42:26] haven't even seen all the incredible new [42:29] ideas that can be done when you have [42:31] access to a model like that that's fast [42:33] that's easy to integrate that you know [42:35] gives you code you know are very [42:37] structured answers back um and so [42:39] because of that now the creativity is [42:41] back to the experience the network you [42:43] can build the ways that you can serve a [42:45] lot of people and bring them together [42:47] using the models to be something [42:48] powerful but the context and the [42:51] community really becomes the asset [42:54] there's so much in invention and I I [42:56] think almost every category has a chance [42:58] to be reinvented and you know some big [43:00] companies are going to be able to make [43:01] the leap and become the relevant company [43:03] for their sector in the AI era but not [43:05] all of them will and that's where [43:07] startups come in to kind of beat them to [43:09] that punch and really be the best new [43:11] possible experience. [43:12] >> Yeah. Yeah. I think to that point I [43:14] think a lot about compounding personal [43:16] value and hopefully eventually [43:18] compounding multiplayer value that [43:20] becomes like a lockin to some of these [43:22] startup products. One example I would [43:24] give is town which is more on the [43:25] proumer end of the agent assistant. It [43:28] can connect into all your systems, [43:30] slack, email, Google drive, understand [43:34] kind of the context of you and your org [43:36] and then do things for you. Like I use [43:38] it to write a lot of my emails, helped [43:40] with a lot of this report. And I know [43:43] it's easy for people to dismiss things [43:45] like this as like, oh, it's just a [43:47] harness, which is now the slightly less [43:49] derogatory term for rapper, I guess. [43:51] >> But Town has built these playbooks of [43:55] who I am, what I sound like, that no [43:58] other product has, and that I cannot [44:00] easily take and migrate over to [44:02] something that comes along tomorrow. [44:04] Because the difference between an email [44:06] that is 99.9% [44:09] sounds like me and even an email that's [44:11] 85% sounds like me is the difference [44:14] between spending like 10 seconds to fix [44:16] it and like 10 plus minutes. Uh and so I [44:19] think that we'll see more products like [44:21] Town that do a really good job of [44:22] collecting context and acting on that [44:24] context to build playbooks around you [44:26] continue to be successful. [44:28] >> No, I think that's totally right. And I [44:29] try to avoid using words like harness [44:30] and rapper because because I think the [44:33] reality is people are now building [44:34] software and they're building these [44:36] really rich products and they're you [44:37] know they might be using their agents to [44:39] write some of that code but they're [44:40] actually when they're delivering this [44:42] whole experience to a user. It is much [44:44] more than just a a shim on the model or [44:47] a harness that brings some context to [44:49] the model. It's actually a really rich [44:50] product and an experience that happens [44:52] to now use a model as part of what makes [44:55] them able to deliver that. And I think [44:57] that's really the excitement of where [45:00] that's going to, you know, what we're [45:01] going to see on this list. You know, [45:02] many of these already are doing that. [45:04] And we think really the next wave, they [45:06] are all going to be these companies that [45:08] have built so much more value. And the [45:10] model is something that is, you know, [45:11] just one piece. [45:12] >> Yeah. [45:13] >> Um Olivia, I think you referenced this [45:15] this chart that you made, but maybe we [45:17] can bring it up. is all of the white [45:19] space that still exists for, you know, [45:23] just all the product categories that [45:25] exist in web 2 and the pre-ai age and [45:29] maybe you can just talk through some of [45:31] these categories and um where are the [45:34] big opportunities? [45:35] >> Yeah, the vast majority of pure [45:37] consumers are basically using AI still [45:40] as a replacement for search products, [45:42] things like Google. Um maybe especially [45:44] as you get into education use cases, [45:46] kids are using it to write essays. A [45:49] little bit more of an expansive use [45:50] case, homework help. Exactly. Yes. Or [45:54] people in the workplace to write emails. [45:56] But if you look at this this chart, [45:57] which will show almost all of the kind [46:00] of areas where we've seen startups [46:03] succeed thus far have been in and around [46:05] that theme of getting things done. Uh so [46:08] like productivity, photo and video [46:10] design and editing, search and answers. [46:12] Um, so many categories are wide open. [46:16] Uh, and many of those categories are I [46:20] would say I would call them network [46:21] categories where you need a multiplayer [46:24] product. Um, you know, dating is one. We [46:28] haven't seen we there are certainly [46:30] startups there that I hope and believe [46:32] will make future additions of the list, [46:34] but there's no one on the list doing [46:35] dating now. Recruiting is another one. [46:37] And I think of dating and recruiting as [46:39] somewhat similar networks actually in [46:41] terms of trying to make matches. [46:43] >> Um, social AI, we really haven't seen [46:46] anything take off. Most of [46:49] >> most of social AI is people posting AI [46:51] generated content on existing social [46:53] platforms. Um, and then I think there's [46:56] all of these other categories that are [46:58] look maybe more like marketplaces did [47:01] like shopping, uh, home buying, [47:04] um, you know, retail. We'll have to see [47:08] if those end up living in agents like [47:11] Instinct or Muse or living as standalone [47:15] products or the next ones which there [47:17] will be many uh or living as standalone [47:20] products or perhaps a mix of both. [47:21] Perhaps your agent is calling the AI [47:23] native marketplace. [47:25] >> Um so I'm very excited about those [47:27] things. is I think also categories like [47:29] gaming and entertainment, the models [47:31] just have not been good enough to [47:33] produce content that the average person [47:35] likes to watch, which maybe the [47:37] exception of AI micro dramas, which are [47:38] blowing up. [47:39] >> Um, but I think that is coming soon, [47:41] too. [47:42] >> No, I think that's exactly right. And I [47:43] just think we're in this world where, [47:45] you know, we're kind of like so excited [47:48] by what Chach did with turning things [47:49] into conversation and now these agents [47:51] we can get in our messaging apps to turn [47:53] them into more conversation. It's [47:55] incredible paradigm, but when you're [47:57] shopping, you want to visually see [47:59] options and explore and have things [48:02] customized for you and be able to say, [48:03] "I really like that, but I wish it had [48:05] some piping on the collar or something [48:07] and actually have AI helped you both [48:09] realize that and maybe even get it made [48:10] for you." And like shopping and have so [48:12] much invention there. entertainment. I [48:14] think we've barely started scratching [48:16] the surface of the ability to tell [48:18] better stories, of the way to empower [48:20] people to be better storytellers that [48:22] then creates these incredible networks [48:23] of people consuming stories. You know, [48:25] Tik Tok was this incredible place where [48:27] people came together and just create [48:28] little videos. You see people around the [48:30] world that you're now watching, but [48:32] there's still so much capture in [48:33] people's heads that if only they could [48:34] express their ideas even better. I is [48:36] going to help that and we're going to [48:37] have incredible new network. So I feel [48:40] like we are just at the cusp of all [48:42] these categories and we've needed the [48:43] models to get better so that you could [48:45] start to dabble in some of these ideas [48:46] and invent them. But as that happens [48:48] over the coming years and with the new [48:50] model capabilities, I still think the [48:52] value is going to end up in this [48:53] software layer in the process. [48:54] >> Yeah, to me this gets back to what we [48:56] were talking about of products that help [48:58] people save time versus spend time. [49:00] Almost everything we've seen in consumer [49:02] AI is save time [49:04] >> and that spend time those end up often [49:06] actually being amongst the biggest [49:08] companies if not the biggest ones [49:10] >> and so we need that to to exist and many [49:13] many builders I'm sure are tinkering [49:14] away on that right now. [49:15] >> Absolutely. [49:16] >> Absolutely. I think that's a great note [49:18] to end on. Thank you both so much for [49:20] joining and uh absolutely make sure to [49:24] check out the top 100 consumer AI apps [49:27] seventh edition. [49:28] >> Thank you. Thanks for having us. We will [49:30] see you for edition 8. [49:32] >> Yes. Yes.