A16z Data Shows Top 1% of AI Users Drive 20% of Spending

a16z数据：前1%的AI用户贡献20%支出 消费级AI应用面临不自然的订阅瓶颈，因为只有高阶用户愿意为省时工具买单，迫使初创公司转向广告和休闲体验以触达大众市场。 分析显示，尽管近半数美国人尝试过AI，但仅有4.5%的用户保持付费订阅，且支出高度集中在技术高阶用户与创作者群体中，他们每月为高杠杆工作流支付高达900美元。高昂的推理成本过去迫使AI产品采用预付订阅和积分制度，但随着OpenAI等公司扩展广告业务，以及专业垂直应用在音频和桌面集成领域开辟出持久的细分市场，商业模式正在发生转变。展望未来，节目主持人认为消费级AI领域最大的未开发机会将超越功利性的“省时”生产力工具，转向游戏、娱乐、约会和智能体电商等广阔的“消磨时间”品类。 第7版a16z消费级AI应用Top 100榜单结合了YipitData信用卡面板数据，显示尽管约50%的美国人使用AI，但仅有约4.5%的用户支付周期性订阅费用。 (0:00) 消费级AI支出呈现极端的幂律分布：前1%的消费者每月支付超过900美元，占所有订阅收入的20%，而前10%的用户则贡献了超过一半的总支出。 (15:00) 尽管Anthropic拒绝采用广告支持模式，但在高端深度用户（7.5%每月支付100美元以上）的推动下，Anthropic的Claude在美国付费个人订阅用户数上超越了Google Gemini。 (30:00) 在12亿周活跃用户和高意图对话定向能力的推动下，OpenAI截至8月的年化广告收入运行率达到10亿美元。 (24:10) 顶级消费级AI应用的流量格局正在固化，只有11个新晋应用闯入网页端和移动端综合排名前100位，创下该报告历史上的最低更替率。 (0:00)

According to the seventh edition of the a16z Top 100 Consumer AI Apps report, consumer AI adoption is characterized by a stark monetization divide, where general consumer penetration is broad but recurring revenue is concentrated among power users funding high-leverage workflows. The Power-User Economy and Monetization Concentration (0:00) Analysis of consumer credit card transactions via YipitData reveals a sharp drop-off between trial and recurring payment. While about 50% of Americans report using AI and 25% interact with it daily, only roughly 4.5% maintain an active paid subscription (with other industry estimates placing the figure between 2.0% and 2.5%). Elena Berger and Olivia Moore note that spending obeys an aggressive power law: the top 1% of spenders account for 20% of subscription revenue—spending over $900 monthly on their tool suites—and the top 10% generate more than half of all revenue. The bottom 50% of paying users contributes just 16% of total spend. Rather than mainstream consumers paying for novelties, the top spenders are makers, developers, and creators swiping personal cards for high-utility productivity and automation tools such as n8n, Granola, Higgsfield, and Manus. Organic turnover across top rankings has also slowed, with only 11 newcomers entering the combined web and mobile lists. Elena points out this mirrors observations by a16z's Sarah Wang that winner-take-most dynamics are solidifying across both enterprise and consumer categories. High COGS and the Subscription Inversion (20:00) Olivia Moore characterizes the current prevalence of direct subscriptions as an unnatural inversion of traditional consumer internet models. Historically, consumer tech scaled on free access, subsidizing operations through advertising or transaction fees. In consumer AI, high inference costs and cost of goods sold (COGS) forced founders to charge upfront, fearing that unmonetized viral growth would trigger unsustainable compute bills. Currently, 85% of top consumer AI web apps monetize via subscriptions, 62% use credit or token tiers, and only 13% rely on advertising or free-to-use models. Unit economics vary widely by user persona: David Paolon of Assistant Benchmark found that power users utilizing agents for heavy coding and technical pipelines drive serving costs into hundreds or thousands of dollars per seat each month, while mainstream consumer assistants run costs in the tens of dollars. As inference costs fall, traditional high-margin internet models are emerging. For instance, OpenAI achieved an annualized advertising run rate of $1 billion by August, leveraging its 1.2 billion weekly active users and the commercial intent inherent in conversational search. Software Defensibility Beyond the Foundation Model (30:00) Josh Elman and Olivia Moore argue that durable enterprise and consumer value resides in the software layer rather than bare foundation models. While frontier labs dominate broad conversational queries, vertical tools capture defensible niches by tailoring workflows to practitioner needs. Suno and ElevenLabs, for example, have built commanding leads in audio and text-to-speech—areas frontier labs deprioritize due to licensing, complex music royalties, and focus on core AGI infrastructure. Similarly, Midjourney dropped out of top web traffic tiers but remains near the top of paid panels due to user loyalty to its parameters and aesthetic control. Startups also leverage bottom-up product-led growth to expand from personal consumer utilities into enterprise powerhouses—a trend Olivia Moore terms 'The Great Expansion.' Desktop utilities like Granola and Wispr Flow acquire users individually, who then bring them into workplace environments, forcing founders to deploy enterprise billing, security, and workspaces. As Josh Elman notes, bespoke context integration across email, documents, and workflows creates high switching costs, debunking the idea that applications are merely disposable wrappers around model APIs. Agent Bottlenecks and the Consumer White Space (45:26) Autonomous agents are shifting the paradigm from browser-based question-and-answer prompts toward multi-step task execution, as seen in platforms like Crockbot, Town, and Dots by OpenAI. However, consumer assistants such as Muse and Instinct lack the person-to-person network effects that drove viral adoption on social platforms like Threads. Consumer interactions are private and solitary, leading to the 'blank box' interface problem where users struggle to find open-ended use cases outside immediate utility. Olivia Moore, citing Eugenia Kuyda, emphasizes that consumer AI remains heavily skewed toward tools designed to save time—such as search, writing, and administrative drafting. However, the largest historical consumer tech platforms (such as TikTok, YouTube, and Netflix) monetize leisure by giving people ways to spend time. Untapped white space remains in categories like dating, recruitment networks, collaborative social platforms, and interactive entertainment, which have yet to produce breakout AI leaders. Consumer AI businesses are presently tethered to utilitarian productivity and power-user subscriptions to offset steep compute costs. The panel concludes that long-term consumer market expansion depends on transitioning past efficiency tools into expressive, leisure-oriented categories that capture human attention, alongside falling inference costs that unlock traditional consumer business models.

根据 a16z 发布的第七期消费级 AI 应用 Top 100 榜单与讨论，消费级 AI 领域呈现出广泛尝鲜与极度集中的付费意愿之间的尖锐背离，整个市场的商业化目前仍由极少数高净值重度用户与生产力工具主导。 付费转化与极端幂律分布 (0:00) 本期报告首次通过与 YipitData 合作排查美国消费者信用卡真实交易，剥离了企业与中小企业账户。数据显示，尽管约 50% 的美国消费者表示使用过 AI，约 25% 几乎每天交互，但实际维持付费订阅的消费者仅占约 4.5%（另有行业追踪数据估计在 2.0% 至 2.5% 之间）。在付费群体中呈现出极其极端的幂律分布：前 1% 的付费者贡献了 20% 的总订阅收入，其每月在 AI 工具链上的支出超过 900 美元；前 10% 的付费者贡献了超过 50% 的收入；而排名后 50% 的付费用户仅贡献了 16% 的收入池。榜单更替率降至历史最低，仅有 11 款新产品进入网页与移动端综合百强，显示出头部通吃的沉淀趋势。 商业模式倒置与高昂边际成本 (20:00) 传统消费级互联网大多依靠免费使用、沉淀受众密度后通过广告或交易佣金变现，但消费级 AI 出现了商业模式的“非自然倒置”：百强网页端 AI 产品中有 85% 依赖订阅制，62% 采用积分或 Token 计费模式，仅 13% 采用广告或完全免费模式。这一现象的核心在于推理算力带来的销售成本（COGS）。若缺乏计费或门槛限制，病毒式用户增长可能直接引发难以承受的服务器账单。例如，David Paolon 的 Assistant Benchmark 追踪数据显示，重度极客将代理工具用于代码编写等复杂任务时，每月单座算力成本可达数百乃至数千美元；普通大众日常使用成本则仅为几十美元。不过，随着模型成本降低，商业模式正在拓展，如 OpenAI 凭借 1.2 亿周活跃用户及富上下文对话数据，在 8 月份广告年化营收已达 10 亿美元；专注医生群体的 Open Evidence 亦通过高价值专业受众变现。 从浅层包装到深层体验层软件 (3:45) 与会者驳斥了将 AI 应用贬低为底层大模型“浅层包装（wrapper）”或“线束（harness）”的论调，认为行业正从最初的浏览器提示词输入框，演进为自主执行多步骤任务的端到端软件。虽然类似 OpenClaw 等早期本地代理依赖极客在 Mac mini 上运行，但如今的 Tau、Wispr Flow 与 Granola 等产品，正通过深度绑定用户在 Slack、邮件和云盘中的私域上下文构建高昂的转移成本。在底层基座模型方面，ChatGPT 在网页流量上分别约 6 倍和 2 倍于 Claude 与 Gemini，但在付费订阅方面，Claude 依靠更偏技术的极客受众，在美国付费用户数上已超越 Gemini，且有 7.5% 的订阅者购买每月超 100 美元的高端档位（远高于 ChatGPT 和 Gemini 的约 1%）。在音频与音乐赛道，Suno 和 ElevenLabs 因垂直工作流设计及规避了基座大厂不愿触碰的版权与版税泥潭，稳居高位；而在图像领域，Midjourney 虽在网页流量层掉出榜单前列，但在真实信用卡付费榜单中依然名列前茅。 从“节省时间”走向“消磨时间”的空白 (45:26) 消费级 AI 的另一大瓶颈在于缺乏人与人之间的网络效应。与 Meta 旗下 Threads 凭借 Instagram 社交图谱在 22 天内获得 1600 万北美下载量相比，Muse 同期仅获 500 万下载；且由于 AI 交互多为高度私密的单人体验，无法像传统社交软件那样通过公开内容引发裂变与模仿，普通用户往往面临不知道该用它做什么的“空白输入框”困境。目前主流 AI 依然拥挤在搜索、作业解答和邮件撰写等“节省时间”的工具属性领域。然而，正如 Replika 首席执行官 Eugenia Kuyda 所指出的，大多数普通消费者并非寻求节省时间，而是在寻找消磨时间的方式。在约会匹配、招聘网络、互动娱乐、社交空间以及基于视觉生成的定制化电商等领域，消费级 AI 仍存在巨大的未开发市场空间。 消费级 AI 目前已有效验证了技术开发者与生产力极客群体的超高付费能力，但其高昂的算力成本、单机私密性导致的网络效应缺失，以及过度偏向“节省时间”的功能定位，阻碍了它向大众消费市场的进一步渗透。未来消费级 AI 的破局，取决于模型推理成本能否下降到支撑广告与交易佣金模式，以及产品能否从纯粹的效率工具拓展至社交、娱乐等“消磨时间”的大众场景。

## Background 本期内容由 a16z 编辑合伙人 Elena Burger 与投资合伙人 Olivia Moore、Josh Elman 于 2026 年 10 月 5 日在播客平台联合发布。节目围绕 a16z 发布的第七版《Top 100 消费级 AI 应用报告》展开，重点剖析了本次报告新增的真实消费者信用卡付费维度，以及用户究竟在为哪些 AI 产品买单。 ## TL;DR 节目深入揭示了消费级 AI 领域极端的“超级用户经济”：虽然约半数美国人接触过 AI，但仅有约 4.5% 的消费者持续付费，且收入高度向排名前 1% 的开发者与专业创作者集中。嘉宾们指出，当前 85% 产品依赖订阅制的模式受制于高昂算力成本，并非消费互联网的终局，未来广告、交易抽成等商业模式将重塑格局。同时，围绕 ChatGPT、Claude 与 Gemini 的付费分化、个人智能体的隐私与成本瓶颈，以及音频、视频工具的突围展开了研讨。投资人最终强调，AI 正在从单纯“帮人节省时间”的生产力工具，向社交、约会、娱乐与电商等能让用户“消磨时间”的万亿级消费蓝海跃迁。 ## Atmosphere & style 对谈洋溢着硅谷顶尖一线创投的敏锐嗅觉与理性思辨，兼具极客探索的兴奋感与审视商业逻辑的冷静克制。主持人节奏把控清晰，两位投资合伙人在真实数据与生动日常案例间切换自如，既有对 OpenClaw 极客狂热的怀旧会心，也有对 AI 广告与隐私张力的犀利拆解。整场对话既打破了单纯看流量增长的行业浮躁，又充满着坚信应用层重构消费生态的坚定乐观。 ## Bulletin - `00:00` a16z 发布第七期 Top 100 消费级 AI 应用榜单，首次引入剔除企业/SMB 干扰的纯个人信用卡消费数据；数据显示全美仅约 4.5% 消费者为 AI 付费，且付费呈现极端幂律分布，头部 1% 用户每月自费超过 900 美元。 - `03:00` 流量格局进一步固化，本期 Web 与移动端合并后仅出现 11 款全新产品入围，创历届新低；与此同时，早期极客型智能体（如 OpenClaw）热度回落，消费级个人智能体（如 Muse、Instinct）开始向大众日常场景渗透。 - `20:00` 商业模式面临倒挂挑战：Top 100 榜单中 85% 的产品仍依赖高昂推理成本驱动的订阅制，而 OpenAI 广告业务已达 10 亿美元年化运行率，证明高意图对话驱动的精准广告有望重塑 AI 商业生态。 - `30:00` 大模型底层生态出现分化，Claude 在美国付费订阅规模上反超 Google Gemini，且其百美元以上的高客单价订阅比例达到 7.5%，远超 ChatGPT 与 Gemini 的 1%。 ## Key takeaways - `00:00` 公众对 AI 的“认知使用”与“真金白银掏钱包”存在严重断层，全美约 50% 的人声称使用过 AI，但实际付费订阅率仅在 4.5% 左右（部分统计低至 2%~2.5%）。 - `15:00` 消费级 AI 的收入大盘完全由极客与超级用户支撑：头部 10% 的付费者贡献了超 50% 收入，前 1% 的顶级用户贡献了 20% 收入，其支出高度集中在 n8n、Granola、Manus 等开发与生产力工具。 - `10:00` 极客重度自动化与大众日常消费的单客推理成本悬殊：极客进行高强度开发导致单客月服务成本破千美元，而面向大众的日常助手月成本仅需数十美元。 - `10:00` 人际社交与个性化智能体之间存在隐私张力：由于缺乏类似社交裂变的公开传播机制，掌握用户最私密日程与邮件的智能体目前仍停留在“单人对机器”的孤岛模式，尚未解锁人与人的网络效应。 - `18:00` 生成式 AI 应用过度聚焦于“帮用户节省时间”，但消费级互联网真正诞生的万亿级巨头往往依靠“帮用户打发时间”（如流媒体、社交平台与泛娱乐内容）。 - `20:00` 消费级 AI 商业模式目前呈现反直觉的倒挂：85% 依赖订阅制、62% 依赖 Token 计费，这主要是受高昂推理成本（COGS）所迫，未来广告与交易佣金抽成必将重新回归大众主流模式。 - `23:45` OpenAI 凭借 12 亿周活跃用户（WAU）和多轮实时高意图对话，在极低打扰的测试下已实现广告年化运行率突破 10 亿美元，精准定位潜力甚至有望超越 Meta。 - `27:15` 多轮自然语言交互能够完美处理带有复杂多重约束的长尾商业需求，相较于传统关键词搜索引擎展现出显著更强的导购与转化价值。 - `30:00` 大模型底层生态格局中，Claude 凭借 Artifacts 等核心产品力在美付费人数反超 Gemini，其 100 美元以上月度订阅方案占比达 7.5%，聚拢了极高粘性的重度核心用户。 - `32:15` 大模型实验室将精力锁定于 AGI 与通用编程，版权风险与专注度限制使其忽略音频等细分赛道，为 ElevenLabs 和 Suno 构筑了深厚的垂直工作流壁垒。 - `37:28` “大扩张”现象表明消费级生产力产品（如 Granola、Whisperflow）正通过个人自发使用的 PLG（产品主导增长）路径自下而上反向渗透进企业 IT 采购，形成极低 CAC 的获客优势。 - `42:47` 应用层绝非简单的“套壳”或“线束”，而是深度接入个人与组织上下文（如 Slack、邮件、网盘），构建出高度拟真、不可迁移的高复利工作流护城河。 - `45:00` 消费级 AI 在约会交友、社交网络、双边交易市场及互动娱乐等强网络效应赛道依然处于巨大的应用空白期，未来蕴含着由“度过时光”驱动的百亿级商业机会。 ## Detailed walkthrough ### `00:00` 消费级 AI 的真实付费图景与第七期榜单发布 播客开篇点出了当前 AI 领域的两大核心现实：一方面，对话式交互正在向个人智能体（Personal Agents）深度演进，AI 正在从一个被动的问答框变成能够帮用户实际办事、深度融入日常即时通讯应用的主动助理；另一方面，公众对 AI 的“认知使用”与“真金白银掏钱包”之间存在着巨大的断层。大约一半的美国人表示自己在日常中使用过 AI，但实际上仅有约 4.5% 的美国消费者正在为 AI 订阅产品付费。更引人注目的是，在这些极少数的付费群体中，支出呈现出极度陡峭的幂律分布（Power-law dynamic）。高额支出的主力军是开发者、创作者等重度用户（Power Users），其中排名前 1% 的消费者每月在各类 AI 工具上的支出甚至超过了 900 美元。此外，OpenAI 等头部公司通过广告等新型变现方式所达到的年化营收增速（Run rate），放在以往的互联网公司身上往往需要历经数年甚至更久的时间才能实现。 本期节目中，a16z 编辑合伙人 Elena Berger 邀请了投资合伙人 Olivia Moore（负责主导编制《Top 100 消费级 AI 应用报告》）和消费级投资合伙人 Josh Elman。Olivia 详细解释了这期（第七版）报告与往期截然不同的重大迭代：在最早的第一版中，团队仅仅通过 Web 端的流量数据（如网站访问量）来给原生 AI 产品排名。当时这种方法非常有效，因为几乎所有用户与 AI 的交互都局限在网页端的提示词对话框（Prompt box）中。但随着消费级 AI 生态的演化与膨胀，单看流量已经无法反映真实商业价值。因此，本次第七期报告首次引入了真实的消费者信用卡消费数据（Consumer card spend data），通过与 EBIT 数据合作，精确剔除了企业（Enterprise）及中小企业（SMB）的采购支出，只针对纯个人消费者的实际扣款进行排名。 分析数据后，Olivia 强调了两个极为鲜明的趋势：首先，从纯流量角度来看，整个 AI 榜单的格局正在沉淀和固化。在 Web 端和移动端榜单合并后，本期仅仅出现了 11 款新产品，这是历次发布报告以来新面孔最少的一期。这种现象与 a16z 市场状态报告中所反映的企业端与消费端“强权律分布”不谋而合。然而，在榜单看似停滞固化的表象之下，底层的技术浪潮实际上在过去几个月里剧烈翻涌，尤其是在个人智能体（Personal Agents）赛道，一系列更具颠覆性的消费级落地正在孕育。 - 约 50% 的美国人声称使用过 AI，但仅有约 4.5% 的消费者为 AI 产品持续支付订阅费。 - 付费呈现极端集中的头部效应，Top 1% 的重度付费用户每月 AI 账单超过 900 美元。 - 第七期 a16z Top 100 消费级 AI 应用榜单首次引入去除了企业/SMB 干扰的纯个人信用卡消费数据。 - Web 与移动端榜单合计仅有 11 款全新产品入围，创下历届榜单新人数量新低，显示出头部流量格局正在沉淀。 --- ### `03:00` 智能体热潮崛起与早期技术探索者的演进 在讨论为何流量榜单出现“固化”时，Elena 引用了投资人 Sarah Wang 此前关于强权律游戏的判断，指出头部巨头的统治力日益加强，但底层的产品范式正在发生突变。Olivia 回顾了上期榜单的情况：上一次统计截止时，以 OpenClaw 为代表的极客型智能体流量正如日中天，高居榜单前列；而到了这一期，OpenClaw 的流量已经大幅滑落，甚至完全退出了榜单。其核心原因一方面在于该团队被 OpenAI 招募/收购，成员很可能转去研发更官方的智能体助理产品；另一方面，市场涌现出了更多面向专业消费者（Prosumer）与普通消费者的全新智能体工具，例如 Crockbot、Town 旗下的智能体产品，以及 OpenAI 新推出的 Dots。 针对这种从底层极客走向普通大众的过渡，Josh 充满感触地回忆了最初的惊艳与震撼。在智能体探索的最初阶段，AI 从“帮我们搜索信息、提升效率的被动工具”，飞跃为“真正能替我们动手解决问题并完成闭环操作的自主软件”。Josh 描述了早期的极客生态：“OpenClaw 确实是一个不可思议的先驱……人们甚至为此成批购买 Mac mini，把所有的私有数据部署在本地运行的 OpenClaw 上，在科技圈引发了一场真正的狂热。”这种自主性彻底打开了人们对 AI 软件边界的想象力。 尽管 OpenClaw 这类产品为技术世界注入了火花，但从技术极客的狂欢到真正被大众接纳，必然经历一段漫长的产品化磨合。极客们愿意为了尝鲜去折腾环境配置、忍受系统崩溃和安全隐患；但要把一个智能体产品交给自己的父母或普通亲友，必须确保它足够安全、简单且具备极高的确定性。Josh 指出，目前硅谷播客里热烈讨论的 AI 智能体革命，大多还停留在早期的原型和前瞻实验阶段，而 a16z 的榜单则忠实地反映了更广泛的大众现实——即便新产品层出不穷，真正实现规模化落地的依然是那些体量越做越大的巨头平台。 ```mermaid flowchart TD accTitle: 智能体演化路径图 accDescr: 展示从被动问答工具到本地极客智能体再到大众消费级智能体的转变历程 subgraph Phase1 ["第一代交互 (2022-2023)"] A["网页端提示词输入框<br/>(Prompt Box)"] B["被动问答与信息检索<br/>(提升个人生产力)"] A --> B end subgraph Phase2 ["极客探索期 (2023-2024)"] C["本地部署与开源探索<br/>(例如 OpenClaw / Mac mini 狂热)"] D["自主执行复杂任务<br/>(但配置复杂、伴随安全顾虑)"] C --> D end subgraph Phase3 ["大众化消费级阶段 (当前演进)"] E["消费级安全封装<br/>(如 Muse / Instinct)"] F["融入即时通讯与日常杂务<br/>(家人与非技术用户无门槛使用)"] E --> F end B --> C D --> E ``` --- ### `05:00` 消费级 AI 助理的爆发：Muse、Instinct 与 Threads 式分发的对比 进入纯粹的消费级 AI 助理（Consumer Assistants）领域后，话题转向了近期增长最迅猛的产品。Olivia 提到，这一波消费级助理的真正爆发实际上就发生在最近这三到四周之内，虽然很多产品的绝对数据尚未冲进历史榜单的总榜，但增长曲线极其陡峭。除了早期先驱 Polk 以及拥有庞大活跃受众且表现极为稳健的 Tomo 之外，目前最亮眼的新秀当属 Muse 和 Instinct。以 Instinct 为例，它在上线短短三周内就宣布积累了 10 万名用户，保持着每日 10% 的惊人环比增长（10% day over day），且这些新用户在第一个月的平均花费就超过了 1000 美元；而 Muse 在上线前 12 天左右就拿下了约 50 万次下载和 25 万活跃用户。 然而，当拿 Muse 与 Meta 此前推出的现象级社交产品 Threads 对比时，数据揭示了消费级 AI 助理在破圈广度上的局限性。Elena 提问这种差距是否源自天然的分发机制（Distribution mechanism）差异——毕竟 Threads 直接绑定了 Instagram 现成的关系链账号体系。Olivia 补充了详细的对比数据：由于 Muse 目前仅在美加地区上线，对比两款产品在美加地区前 22 天的下载量，Threads 斩获了约 1600 万次下载，而 Muse 在同样的时间窗口内仅徘徊在 500 万次左右。 Josh 深度剖析了这种分发差异背后的本质原因与商业考量。首先是单位服务成本（Cost to serve）的悬殊区别：Threads 作为传统的社交网络，新增一个纯文字用户的服务器成本极低，天然依赖“关键规模效应（Critical mass effect）”——如果不能在极短时间内拉入海量用户互相沟通，新的社交关系链就无法沉淀，网络效应便会失效；相反，对于 Muse 这样的生成式智能体助理，每一个活跃用户的日常调用都伴随着极为高昂的模型推理计算成本，团队根本不敢也不希望无节制地进行病毒式激进化获客，否则成本会迅速失控。 其次，两者的定位有着根本不同。Muse 目前依然是一个高度私人化（Personal tool）的工具，其早期价值在于给用户带来革命性的单点体验。正如 Josh 所言： > “Because Muse is still such a personal tool that they're pushing it and it's very exciting this early just adoption and people having these experiences.”（“因为 Muse 仍然是一款如此个人化的工具，他们正在全力推进它，而看到早期的这种采用率以及人们获得的惊艳体验，是非常令人兴奋的。”） Josh 和 Olivia 进一步探讨了让智能体助理具备长期留存壁垒（Stickiness）的核心引擎。光靠单向的私人助理工具属性远远不够，未来真正能够形成护城河的，是**网络效应（Network effects）以及能够打通外部商业体系的双边/多边市场（Marketplaces）**。虽然像亚马逊（Amazon）等平台出于数据护城河的考虑，明确表示不允许 Muse 自主在其电商网站上进行商品比价与直接采购，但 Muse 已经迅速与 Shopify 等生态系统内的数以百计的产品和平台达成了深度伙伴关系，实现了在 Shopify 商家网络内替用户完成消费决策与直接下单。 对话的最后，Olivia 提出了一个深层的结构性思考：尽管消费级智能体正在以惊人的速度演化，但迄今为止，市面上**没有任何一款助理类产品真正解锁了“人与人之间的网络效应（Person-to-person network effects）”**。现在的智能体软件通过深入读取用户的私人电子邮件、日程和通讯记录，比以往任何历史软件都更亲密、更透彻地了解一个人。但这种极其强大的认知和连接能力，目前仍完全停留在“单人对机器”的孤岛模式中，尚未演进出人与人借由智能体媒介进行社交、协作与价值交换的全新网络范式。 ### `10:00` 个人智能体的隐私困境与“超级用户”的高昂服务成本 随着 AI 越来越深入人们的生活，产品对个人了解得越多就越实用，但这种高度个性化也带来了天然的张力：当一个 AI 掌握了你所有的秘密，你真的愿意把它带入与其他人的社交或工作场景中吗？Josh Elman 指出，这就像许多人至今必须将 ChatGPT 的工作账户与个人账户严格分开一样，业内还没有人真正彻底解决这一问题。AI 的采用不仅受制于公众对“AI 是否造福世界”的宏观信任度，更取决于极其微观的个人安全感。Josh Elman 直言：“But on a very personal level, if you don't trust the thing that you're giving it your most intimate access to your email or your life or your credit card, and it might do rogue things that you wouldn't expect or share things with other people you wouldn't expect, you know, that's going to be the biggest impingement on these things really growing as massive consumer apps.”人类在数万年演化中已经非常擅长把握人际社交的分寸，清楚什么话可以分享、什么秘密必须守护，而如今人们却突然要求一串软件代码同样具备这种细腻的分寸感。 Elena Burger 进而将这个问题拆解为双重力量：一种是平台如何允许智能体（Agent）相互调用的内生力量；另一种是消费者自身在使用时是否安心的外在心理防线。Olivia Moore 认为平台技术能力的进化速度远快于普通消费者的心理接纳度。针对外界盛传“许多 AI 助手产品服务单个用户的月度计算成本高达数百甚至数千美元”的谜团，Olivia Moore 分享了她与 Assistant Benchmark 创始人 David Paolan 的合作发现。David Paolan 追踪了超过 170 到 200 个智能体，并在一个拥有 1500 多名早期核心用户的社群中进行了数据调研。结果显示，即便是面对 Muse 或 Instinct 这类消费级 AI 助手，这群早期狂热用户的第一大使用场景依然是代码编写与技术自动化。面对如此繁重的技术计算任务，单客成本飙升到几千美元自然不足为奇；而面向真正主流大众消费者的助手产品，单客月度成本其实仅在几十美元左右。 Josh Elman 补充道，这种认知错位凸显了 AI 迈向大众市场的巨大鸿沟。对于程序员而言，AI 是一剂惊人的加速器，但将同样的产品交到一个普通人手里时，它往往只像一个让人不知所措的“空白搜索框”。现实中大部分人的学习机制是社会化的：在社交网络上看到别人发了有趣的视频或玩法，自己才会模仿并二次创作，从而形成滚雪球般的网络效应。然而在 AI 领域，除了编程和创意生产领域已经出现“你用什么做的？我用了 AI”的交流外，绝大多数日常消费场景由于过度私密，尚未形成任何自发繁荣的社交传播飞轮。 ```mermaid flowchart TD accTitle: AI Agent 采纳层级与成本结构差异 accDescr: 展示 AI Agent 在专业开发者/早期极客群体与大众主流用户群体之间的使用场景、成本以及交互模式差异。 subgraph PowerUsers ["早期极客 / 开发者群体"] direction TB A1["高强度任务 (编程、技术自动化)"] --> B1["单客服务成本: 数百至数千美元/月"] B1 --> C1["明确的需求驱动与高杠杆回报"] end subgraph Mainstream ["主流大众消费者"] direction TB A2["日常消费任务 (订票、日常闲聊)"] --> B2["单客服务成本: 仅数十美元/月"] B2 --> C2["面对'空白对话框'的探索门槛与隐私顾虑"] end ``` - **智能体追踪规模**：David Paolan 的 Assistant Benchmark 持续评测超过 170 至 200 个智能体，覆盖 1500 多名深度早期极客用户。 - **成本反差**：技术极客进行重度自动化导致部分智能体单客月成本破千美元；而大众消费场景的真实月成本仅数十美元。 - **冷启动阻碍**：缺乏公开的社交裂变机制让个性化 AI 停留在私密工具层面，难以自发形成流行趋势。 --- ### `15:00` 谁在真正为消费级 AI 买单？极度集中的“超级用户经济” 讨论转向全美范围的 AI 普及与真实消费数据。尽管宏观渗透率看起来十分繁荣——大约 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 互动——但背后的真实支付意愿却呈现出令人咋舌的极端集中。根据 Yipit 消费面板的数据，全美仅有不到 5%（约 4.5%）的消费者会为 AI 产品的订阅付费，其他部分统计甚至显示该比例低至 2% 到 2.5%。虽然这一付费比例在过去一年中已经翻倍，但资金来源的高度集中化极为惊人。 Olivia Moore 详细剖析了这 4.5% 付费群体内部的金字塔结构：在所有付费用户中，前 10% 的超级用户贡献了超过一半的行业总收入；前 1% 的顶级用户贡献了整整 20% 的收入；而处于后半段的 50% 用户加起来，贡献的收入份额仅有可怜的 16%。换句话说，整个消费级 AI 的付费大盘完全是由那前 10% 的超级拥趸一手托起的。更耐人寻味的是用户的单人月度开销分布，通过穿透至个人信用卡的扣款记录可以发现，排名前 1% 的个人用户每月自费投入在 AI 上的金额极其庞大，而中位数付费额则相对平缓。这群人自掏腰包购买的产品名单高度一致：n8n（自动化工作流）、Granola（AI 会议笔记）、Higgsfield（视频生成）、Manus（通用自动化智能体）等工具在头部消费账单中占据了极高的比例，清一色偏向开发者、生产力和专业创意创作领域。 Josh Elman 对这种现象持乐观态度，他认为这代表着新一代“创作者/制造者（Makers）”的崛起。这些人不仅把 AI 用于公司业务，更主动掏出个人信用卡，去学习自己此前从未掌握的编程技术，去制作前所未有的艺术作品、视频和数字内容。Josh Elman 强调，当你愿意每个月从个人账户中支出数百美元购买一系列 AI 工具时，你就是在为思维的具象化加速买单。然而 Elena Burger 指出，这种现象也暴露了一种思维局限：程序员和开发者天然具有寻求“杠杆（Leverage）”的世界观，他们本能地思考如何撬动时间、提升效率；但放眼普通大众，这种“追求时间杠杆”的心态其实只属于极少数精英圈层。 - **宏观使用率**：约 50% 的美国人声称使用过 AI，约 25% 的人接近每日高频互动。 - **真实付费率**：全美仅约 4.5%（部分统计仅 2%~2.5%）的消费者为 AI 产品订阅付费。 - **收入二八定律**：在 4.5% 的付费人群内部，头部 10% 的用户贡献了超过 50% 的收入，前 1% 的用户贡献了 20% 的收入，后 50% 的用户仅贡献 16%。 - **头部支出偏好**：个人信用卡付费高度集中在 n8n、Granola、Higgsfield、Manus 等重度提升生产力与自动化水平的工具。 --- ### `18:00` 生产力工具的瓶颈：人类更想打发时间，而非节省时间 关于如何将 AI 从“生产力杠杆”推向更广泛的大众市场，讨论切入到了人性本质的底层逻辑。Olivia Moore 引用了其投资组合公司 Wabi 的 CEO Eugenia 的精辟见解，指出当前 AI 创业者的盲区：“Eugenia who is one of our portfolio CEOs runs a company called Wabi, which is fantastic, she had this great quote—I'm probably going to botch it—but it was something like, 'Most people aren't looking to save time. They're looking for ways to spend their time.'” Olivia Moore 进一步阐述，这就是为什么全球最主流、用户时长最长的消费级互联网产品永远是社交媒体、流媒体娱乐、Netflix、TikTok 和 YouTube。而当前几乎所有的生成式 AI 应用都在围绕“如何更快、更容易、更惊艳地完成某件事”展开宣传。对于普通大众而言，“帮你把事情做得快一点”并不能构成强烈的、每小时或每天都必须打开的长期核心价值。 Josh Elman 对此深表赞同，他认为“节省时间”本质上仍是属于超高生产力人群、严肃工作者或硬核创作者的语境。但与此同时，他也观察到门槛降低带来的涟漪效应正在扩散：曾经那些因为无法掌握 Pro Tools 或 Adobe Premiere 等复杂专业软件而放弃制作视频的人，如今可以通过 AI 界面轻松将脑海中的画面落地；曾经不懂代码的人也可以一键生成专属程序。在两位投资人看来，尽管当前消费级 AI 的付费大盘仍然极度集中于生产力与开发工具，但这并不意味着未来受限，而是说明整个行业依然处于极早期阶段。大众对 AI 的需求绝不会止步于“更高效地干活”，如何切入情感、娱乐与“消磨时间”的广阔蓝海，才是消费级 AI 爆发的真正胜负手。 ### `20:00` 订阅制的“不自然反转”与高昂推理成本（COGS）的束缚 嘉宾们深入探讨了消费级 AI 领域极为极端的权力定律（Power Law）：在付费用户中，前 1% 的超级用户每月花费高达 900 美元。Josh Elman 认为这是未来趋势的先兆，代表着人们不再仅仅是为了“节省时间”而使用工具，而是体会到“将想法亲手化为现实”的纯粹喜悦。他回忆起当大家第一次狂热地使用 OpenClaw 等个人代理（Agents）时的搞笑场景：人们随时随地抱着不合盖的笔记本电脑到处走，甚至兴奋地喊着“我不睡觉了，我正在写代码，正在创造前所未有的东西”。然而，这种由开发者和硬核创作者驱动的支出集中现象，也揭示了当前消费级 AI 商业模式的畸形。 ```mermaid flowchart LR accTitle: 消费级 AI 与传统互联网变现模式对比 accDescr: 对比传统互联网依赖广告与交易费，而当前消费级 AI 严重依赖订阅和 Token 计费 subgraph Traditional ["传统消费级互联网"] T1["免费获取用户/高密度"] --> T2["广告变现"] T1 --> T3["交易佣金"] end subgraph CurrentAI ["当前消费级 AI"] A1["高昂算力成本 (COGS)"] --> A2["85% 依赖订阅制"] A1 --> A3["62% 依赖 Token/点数计费"] end ``` Olivia Moore 提出了一个反直觉且颇具争议的观点：尽管她是消费级 AI 的坚定拥护者和投资人，但她并不认为直接让消费者为 AI 产品买单会成为唯一的未来。数据显示，Top 100 Web 榜单中约 85% 的产品依赖订阅制变现，另有 62% 采用积分或 Token 超额使用计费，而仅有约 13% 依赖广告或将“用户作为产品”的模式。回顾消费级互联网的历史，这是一种“相当不自然的反转”，因为绝大多数万亿级消费科技巨头都是依靠广告或交易抽成崛起的。ChatGPT 在短短三年半内就冲进了全球前 20 大消费级订阅服务行列，而榜单上其余席位几乎全被流媒体或 Amazon、Uber 这种高频刚需平台占据。 - **榜单变现模式分布**：约 85% 的 Web AI 产品依赖订阅制，约 62% 使用积分/Token 计费，仅约 13% 采用广告等免费变现模式。 - **高昂算力成本（COGS）困境**：以往消费级创投遵循“前五年专注于增长与流动性，之后再靠广告赚钱”的逻辑；如今由于高昂的推理成本，公司若不设门槛直接免费狂飙，很容易因失控的算力账单而陷入破产恐慌。 - **商业模式演进预测**：随着更低成本的模型出现，交易费（Transaction fees）和广告（Ads）等经典互联网盈利模式必须回归，从而解放大众市场。 > “如果回顾消费互联网的历史，这可以说是一个相当不自然的反转。”（Olivia Moore：“If you look back at the history of the consumer internet, that's like a pretty unnatural inversion.”） --- ### `23:45` OpenAI 的广告破局：10 亿美元年化与极度个性化意图 讨论随后转向 OpenAI 正在展开的商业化大动作。Top 100 报告中披露了一个令人震惊的数据：截至 8 月，OpenAI 通过广告业务达到的年化运行率（Annual Run Rate）已经突破 10 亿美元，且这一数字目前很可能已经更高。在极少打扰用户、仅低调联合少数合作伙伴测试的情况下，OpenAI 实现这一体量所用的时间，在传统互联网历史上几乎不可想象。推动这一惊人爆发的核心因素在于其强大的网络密度与前所未有的用户数据上下文。 ```mermaid flowchart TD accTitle: OpenAI 广告转化的双飞轮 accDescr: 展现密集活跃用户与深度上下文对话如何赋能高精准广告 U["12 亿周活跃用户 (WAU)"] --> D["网络密度与用户规模"] C["日常深度对话与意图"] --> P["超高精准个性化画像"] D --> ADV["广告变现高转化率"] P --> ADV ADV --> ECO["极高商业毛利与商业闭环 (如内置钱包)"] end ``` 与传统的搜索广告不同，AI 广告面临着极高的信任门槛。如果用户正在与 AI 进行极其私密的情感或健康倾诉，突然插播一条“快速减肥产品”的广告，便会彻底摧毁用户信任。因此，AI 广告的切入必须极度巧妙。嘉宾们指出，ChatGPT 现在的周活跃用户（WAU）已达 12 亿，并且用户使用它的习惯正从单纯的“查资料、补历史、做作业”向“寻求现实生活建议、商业购物与旅行规划”演进。这种“商业意图”（Commercial Intent）的自然涌现，让广告不再是骚扰，而成为极具价值的解决方案。 - **商业化增速**：OpenAI 广告业务的年化运行率已达 10 亿美元，且处于极早期、轻量上线的阶段。 - **用户密度基石**：官方披露的最新数据已达 12 亿周活跃用户（1.2B WAU）。 - **极高转化潜力**：由于掌握长上下文、未来可能推出的“使用 ChatGPT 登录”（Log in with ChatGPT）以及内置钱包，广告转化的精准度和变现价值有望显著超越 Meta 曾经建立的最佳水平。 > “我真的对 OpenAI 推出广告的方式感到印象深刻，在少数几次我真正看到广告的场景中，它都与我正在进行的对话极度相关且高度匹配。”（Josh Elman：“I've actually been really impressed with the OpenAI rollout of ads when I've actually ever seen an ad. It's very integrated and very relevant to the conversation that I'm having.”） --- ### `27:15` 对话式购物场景与垂直 AI 广告标杆 Josh Elman 分享了自己真实的购物经历来佐证对话式搜索带来的颠覆。他曾在 Google 上苦苦寻找一种“带有一键快拆功能的代客泊车钥匙扣”，但传统搜索引擎给出的结果令人沮丧。于是他转向 ChatGPT 展开了连续的多轮对话，逐步明确细节，甚至在最后追加了一个严苛条件——“最好是美国本土制造”。最终，ChatGPT 为他精准挖掘出底特律一家专门从事传统金属加工的工厂所制造的精美钥匙扣。Josh 指出，在整个探索和发现的过程中，用户的心态是完全开放的，如果当时有相关厂商投放上下文赞助广告，那将是一次完美的体验，不仅能帮用户解决问题，还能让毫无知名度的独立工坊直接触达最精准的买家。 这种基于专业人群和极高密度的商业化逻辑，同样在垂直消费/企业级 AI 领域得到了印证。Olivia Moore 举了面向医生群体的 AI 工具 Open Evidence 的例子。该产品在美国所有执业医师中的渗透率/密度达到了惊人的 50% 至 60%。面对这群极难触达且价值极高的高净值专业人群，精准的赞助与广告变现展示出了无比恐怖的转化率。这不仅验证了通用聊天助手的广告前景，也为更多垂直领域的垂类 AI 应用指明了一条摆脱单一高价订阅、走向多元变现的康庄大道。 - **传统搜索与 AI 对话的对比**：传统关键词检索难以捕捉多层约束（如“快拆代客泊车钥匙扣”+“底特律本土金属工厂制造”），而多轮对话能够完美处理长尾、复杂的商业意图。 - **垂直 AI 广告标杆 Open Evidence**：覆盖全美 50% 至 60% 的执业医师，凭借极高的人群密度与专业价值，在早期广告变现上取得了显著成功。 > “在用户与其实时对话的当下，OpenAI 有能力进行极其惊人的精准定位，甚至可能超越此前被公认为业内最佳的 Meta。”（Olivia Moore：“I feel like we may see as they lean into it, OpenAI being able to do crazy good targeting on like even better than I think Meta, which previously was like best in class, because they have live data of you in a live conversation with them.”） ### `30:00` 基础大模型格局演变：ChatGPT、Claude 与 Gemini 的付费用户差异 在讨论三大多模态 AI 巨头时，Elena Burger 提问大家是否看到了消费者付费行为的变化，以及用户群体特征的差异。Olivia Moore 回顾了历次排名的变化，指出虽然 ChatGPT 从一开始就在全球消费者层面的使用量和收入上占据绝对主导地位，但下层的竞争格局发生了剧烈震荡。目前在网页端，ChatGPT 的流量大约是 Claude 的 6 倍，是 Google Gemini 的 2 倍。 ```mermaid flowchart TD accTitle: 三大主流大模型付费层级与商业模式对比 accDescr: 对比 ChatGPT、Claude 和 Gemini 在美国付费用户规模及订阅定价策略上的差异。 subgraph MarketShare["美国付费订阅规模对比"] direction TB GPT["ChatGPT (绝对领先)"] CL["Claude (高客单价黑马)"] GEM["Google Gemini (受庞大装机量支撑)"] GPT -->|"付费订阅用户数约 3 倍于"| CL GPT -->|"付费订阅用户数约 3 倍于"| GEM CL -->|"付费订阅用户数反超"| GEM end subgraph TierStrategy["高客单价方案占比 (100美元+/月)"] direction TB CL_Max["Claude: 约 7.5% 用户选择"] GPT_GEM["ChatGPT & Gemini: 仅约 1% 用户选择"] end ``` 然而，针对美国付费用户群体的样本数据显示出了令人震惊的逆转：Claude 的付费订阅人数已经超越了 Gemini。考虑到 Google 拥有无可比拟的庞大装机量、基础用户池以及天然的分发渠道，Anthropic 取得这一成绩极为不易。这一突破主要归功于 Anthropic 近期推出的一系列极具吸引力的产品（例如 Claude Artifacts / Claude Projects），以及密集的媒体曝光度。从总体付费盘子看，ChatGPT 依然遥遥领先，其美国消费者付费用户数大约是 Claude 和 Gemini 各自的三倍。但 Anthropic 旗帜鲜明地坚持“绝不做广告”的路线，因此在订阅商业化上表现得极其激进。数据显示，Claude 的付费用户中有约 7.5% 购买了每月 100 美元以上的顶配（Max）方案，而 ChatGPT 和 Gemini 对应百美元级别档位的用户占比仅有约 1%。这充分表明，即使在付费用户内部，Claude 聚拢的重度核心用户在产品使用方式和支出意愿上也截然不同。 - ChatGPT 网页端流量优势：约 6 倍于 Claude，约 2 倍于 Gemini。 - 美国付费订阅用户体量：ChatGPT 约是 Claude 和 Gemini 的 3 倍；Claude 付费人数反超 Gemini。 - 100美元+/月高客单价订阅比例：Claude 约为 7.5%，而 ChatGPT 和 Gemini 仅约为 1%。 --- ### `32:15` 垂直创意工具的破局：音频领域的 ElevenLabs 与 Suno 随后对话转向了创意工具赛道。Elena Burger 指出，尽管各大前沿研究室在创意工具方面能力突飞猛进，但例如 Anthropic 早期并未直接训练 Diffusion（扩散）或视频模型，而是通过其强大的代码生成能力间接实现了图像与视频生成。而在垂直创意工具领域，市场分布展现出了极为独特的动态。 各大模型实验室将战略重心死死锁定在构建 AGI（通用人工智能）和 Coding Agents（编程智能体）上，这导致它们在某些创意模态上投入甚少——最为典型的就是音频领域（文本转语音、音乐生成）。ElevenLabs 和 Suno 因此不仅常年稳居流量榜前列，在支出收入榜上也名列前茅。对于大模型巨头而言，冒着巨大版权诉讼风险去涉足音乐领域并不划算，这就给了 Suno 迅速拉开身位、建立绝对壁垒的绝佳窗口。Josh Elman 强调，尽管基础大模型设立了极高底线，但那些深刻理解受众的专业化垂直工具依然无可替代：“如果你做音乐只是为了好玩，或是为了真正的生产创作，Suno 都是一个令人惊叹的工具，从你开始使用的那一刻到最后，它都懂你。” ElevenLabs 在语音集成上的体验，以及 Figma 等设计工具在深层工作流上的沉淀，都证明了软件层交互和保存管理资产的价值，而非仅仅依靠底层模型。 > “在这么多市场中，你看到的都是真正专业化的产品，它们高度聚焦，清楚自己的受众是谁以及他们在做什么，从而能够打造出高度定制化的东西。” —— Josh Elman --- ### `34:42` 图像与视频分化：Midjourney 的韧性、Sora 与中国视频 AI 的竞争 在图像和视频领域，竞争态势与音频大不相同。Olivia Moore 分析指出，OpenAI 的图像能力（如内置于模型的编辑与生成）非常强大，Google 凭借 Imagen 等多次迭代，分流了大量原本属于独立图像生成器的普通消费者泛流量。在他们最初的榜单中名列前茅的 Midjourney，如今在整体流量榜上已经跌落，但在付费收入榜上却卷土重来。这说明 Midjourney 牢牢锁住了专业创作者和有审美调优需求的重度用户（Power Users）。 ```mermaid flowchart LR accTitle: 图像与视频 AI 工具分化路径 accDescr: 展现泛消费者流量向大模型聚合，专业创作者向具备审美和深层工作流的独立工具沉淀的过程。 subgraph InputDemand["创作者需求分层"] Casual["泛消费者日常需求"] Pro["专业制作与重度用户"] end subgraph ToolEcosystem["工具流向与产品"] GenAI["巨头内置图像/视频工具<br/>(OpenAI / Google Nano & Veo)"] NicheTool["垂直审美/工作流工具<br/>(Midjourney / 专属视频产品)"] end Casual --> GenAI Pro --> NicheTool Pro -->|"Agent 自动化调用外部工具"| GenAI ``` 视频领域则被视为狂野西部（Wild Frontier）。一方面，中国公司展现出极强的竞争力，其在丰富数据训练和多模态视音频同步对齐方面进展飞速；另一方面，视频生成涉及画面、声音、音画同步与风格一致性等多重复杂维度，越是敢于深度投入训练并解决复杂性的团队，优势越明显。 此外，智能体（Agents）调用外部工具的趋势正在重塑创意工作。Olivia Moore 提到，她近期会让强大的推理模型或多模态 Agent（如 Astra 等前沿模型）去生成或修改创意内容，智能体能够自动进行网络工具调用（Tool Calling），甚至自主决定安装运行所需的工具环境。Josh Elman 也兴奋地分享了他让模型同时对比制作动画的极客体验。两位投资人一致认为，边缘探索者（Tinkerers）正在挖掘这些极致可能，而未来的杀手级消费级应用正是要将这种复杂的智能体链路“打包”成普通人极易上手的产品体验。 > “当我们思考‘全部 AI’与‘消费级 AI’之间的鸿沟时，正是这种把复杂能力打包、打造出色产品体验的运动，才真正把它们变成了可持续沉淀的事物。” —— Josh Elman --- ### `37:28` 大扩张时代：Granola、WhisperFlow 与企业端“被动”出海 Elena Burger 随后把视角拉向了浏览器网页窗口之外的成熟桌面端与本地端产品，例如会议记录与笔记工具 Granola、WhisperFlow、Superhuman 以及各类植入聊天界面的智能体，询问消费级产品向企业渗透的动态。 Olivia Moore 提到她在六到九个月前撰写的文章《大扩张》（The Great Expansion），核心探讨了为何像 Krea、Replit、ElevenLabs、Gamma 等大量纯消费级产品以惊人的速度转变为企业级产品： 1. **PLG（产品主导增长）天然驱动**：个人用户在日常或办公场景中自发使用这些高生产力属性的工具，并在组织内部自发安利。 2. **被动拉入 Enterprise 阶段**：随着在工作环境中的泛滥，企业 IT 部门被惊动，创始团队随即不得不顺应需求，火速开发企业安全合规、隐私保护以及团队管理（Team Plan）功能。这构成了一种极其高效、低获客成本（Low CAC）的分发范式。 与此同时，这种现象也映射出传统巨头与底层模型实验室的困局。老牌巨头（如 Google）由于害怕颠覆自身原有的利益格局，迟迟无法也不愿为 AI 时代彻底重构 Google Docs、Gmail 或 Calendar 的界面。令人意外的是，这一窘境同样发生在 OpenAI 和 Anthropic 身上：两家顶尖实验室至今没有打造出颠覆性的办公套件产品，也没有成功孵化出能够杀死外部创业公司的生产力杀手应用，其精力完全被底层算力、通用模型训练和核心界面（如 ChatGPT、Claude Code）所牢牢牵制，从而为专注于新形态界面的创业公司留出了广阔的创新土壤。 > “这是我们总体上看到的另一个趋势的极好范例，即在位巨头不愿意或无法蚕食他们现有的交互界面。” —— Olivia Moore ### `40:00` 水涨船高：大模型进化与初创公司的应用层窗口 基础大模型的演进带来了显著的协同效应。Josh Elman 指出，随着底层模型能力的大幅跃升，此前专注于打磨易用界面、自然交互和团队分享体验的产品（如 Whisperflow 和 Granola）迎来了爆发。最初受限于模型能力的软件，在模型进化后不仅没有被淘汰，反而在实用性和市场需求上全面爆发。Josh 强调：“这是一种水涨船高（rising tide lifts everything）的态势；认为只有一两家公司能构建体验层的想法，正是这份报告所打破的。” 在初创公司与大厂的竞争格局上，Olivia Moore 补充分析称，尽管 OpenAI 和 Anthropic 等机构本质上也是初创公司，但它们如今已成长为拥有数千名员工的庞大组织。如同“大船难调头”，大公司在推进硬件或新产品时面临更多流程制衡与现有生态协同的顾虑。例如本次营收榜上出现的硬件记录仪 PLAUD，最初发轫于中国等东方市场，如今已走向全球。大模型实验室虽然实力雄厚，但行动周期较长，这为灵活敏捷的垂直初创公司留下了广阔的发展窗口。 Elena Burger 总结道，初创企业的最佳生态位在于找到一种“能随着底层模型变强而同步增强、但又不会被模型升级轻易吞噬”的产品形态。Josh 进一步透露，这正是他选择加入 Andreessen Horowitz 的核心动因——他坚信价值正在重回软件应用层。得益于像 TypeSafe 的 Jeff 所带来的创新工具，开发者能以结构化、极快的方式调用模型，未来的核心竞争力将彻底回归到用户体验、网络效应、组织上下文与社区壁垒上。 - 底层模型能力的快速迭代使 Granola、Whisperflow 等早期体验优秀的应用实现价值倍增。 - PLAUD 作为软硬件结合的 AI 录音备忘设备，展示了初创企业跨区域落地与快速执行的敏捷性优势。 - 大模型实验室具备通用能力，但面临繁琐的企业审核与整合链条，初创公司因而具备速度与细分场景的破局机会。 --- ### `42:47` 摆脱“套壳”偏见：个人与组织上下文带来的复利壁垒 对于市场上轻视应用层软件、称其为“套壳”（wrapper）甚至新代名词“线束/支架”（harness）的论调，Olivia Moore 表达了明确的异议。她以自身高频使用的企业级/专业消费者（prosumer）智能体助手 Town 为例：Town 通过接入 Slack、电子邮件和 Google Drive，全面汲取用户及组织的上下文背景，进而接管大量邮件撰写和数据整理工作（包括参与整理本次榜单报告）。 Olivia 深入分析了这种个性化 context 所构建的护城河：“一封听起来 99.9% 像我本人语气的邮件，与一封听起来 85% 像我语气的邮件，二者之间的差距，是花 10 秒微调与花 10 分钟重写的本质区别。”这种深度融入工作流的个人知识体系是无法被轻易迁移或替代的，因此这类精准捕获上下文并构建个性化操作手册（playbooks）的应用具有极其稳固的留存壁垒。 Josh 强烈赞同应摒弃 harness 和 wrapper 这类带有贬义的词汇。他指出，现代 AI 应用绝非大模型外层薄薄的垫片，而是高度复杂的完整产品架构。开发者即便在开发过程中使用编码智能体，最终交付给用户的依然是极其丰富立体的全套体验。在这类产品中，大模型已经转变为核心功能组件之一，而不再是产品的全部定义。 ```mermaid flowchart LR accTitle: 上下文数据转化为产品壁垒的闭环机制 accDescr: 展示组织数据如何经过智能体处理，转化为带有个人风格的成果，最终沉淀为不可迁移的护城河。 subgraph Inputs["多源工作流数据接入"] S["Slack 实时沟通"] E["Email 往来邮件"] D["Google Drive 文档"] end subgraph Engine["个性化上下文引擎 (Town 等)"] P["构建个人与组织 Playbook"] M["大模型推理能力支持"] end subgraph Output["高复利输出与壁垒"] R["99.9% 拟真人输出<br/>(仅需10秒微调)"] W["不可迁移的<br/>工作流护城河"] end S --> P E --> P D --> P M <--> P P --> R R --> W ``` --- ### `45:00` 消费级 AI 的巨大留白：从“省时间”走向“花时间” 随后，Olivia Moore 结合消费级 AI 现状图表，剖析了 Web2 与 AI 时代交替之间依然存在的巨大空白领域。目前绝大多数纯消费级 AI 的应用逻辑高度集中在“替代搜索”或“提升效率”上，如学生用其辅导作业、写论文，职场人群用其撰写邮件等，整体均围绕“完成任务/节约时间”（save time）展开。 然而，在涉及多人协同、强网络效应以及更复杂的社交场景中，原生 AI 应用几乎仍是一片荒原： 1. **约会交友（Dating）与人才招聘（Recruiting）**：这两类本质上高度依赖网络匹配算法的赛道，在第七版榜单中尚无任何一家公司上榜。 2. **社交网络（Social AI）**：尚未出现真正规模化爆发的原生社交平台。 3. **交易平台与电商零售（Marketplaces & Shopping）**：无论是购房、电商还是零售领域，未来究竟会作为类似 Instinct 或 Muse 这样的独立智能体存在，还是诞生全新的 AI 原生交易平台，仍处于激烈的探索期。 4. **游戏与娱乐（Gaming & Entertainment）**：此前模型在生成长篇内容上能力不足，但目前 AI 微短剧（AI micro dramas）已呈现爆火势头，预示着拐点临近。 Josh 强调，对话框交互并非零售体验的终局。理想的 AI 购物应当兼具高度视觉化、自由探索与个性化定制能力——用户可以圈出衣服细节提出“我喜欢这个款式，但希望领口加一道滚边”，AI 甚至可以直接对接柔性供应链制作出来。此外，在内容创作领域，过去的 TikTok 让大众通过短视频表达自我，未来的多模态生成工具将抹平构思与表达之间的鸿沟，催生全新的叙事网络与互动社区。 Olivia 总结了消费级产品的终极商业逻辑：“我们在消费级 AI 中看到的绝大部分产品都在帮人节省时间。但那些帮人们‘花时间’（spend time）的产品，往往最终会成为规模最庞大、市值最高的一批巨头。”随着底层多模态能力的就位，下一代消费级杀手级应用必将由围绕“度过时光”与“内容沉浸”的开发者所创造。 - 现阶段大众消费者对 AI 的认知与使用习惯仍停留在“高效版 Google 搜索”与“作业/邮件辅助工具”。 - 匹配类双边网络（如约会交友、猎头招聘）存在巨大的 AI 重构机会，但目前榜单中尚无代表企业。 - “节省时间”型产品奠定实用基石，而“消耗时间”型产品（娱乐、社交、个性化消费）将催生市值上限最高的商业帝国。 --- ### `49:33` 结语与投资免责声明 在对消费级 AI 广阔蓝海的展望中，Elena Burger 代表 a16z 团队向两位投资合伙人 Olivia Moore 和 Josh Elman 致谢，并宣布第七版《Top 100 Consumer AI Apps》播客对谈圆满结束，期待在第八版榜单发布时带来更多维度的产业洞察。 播客末尾给出了标准合规与投资免责声明：本期节目内容仅供信息参考，不构成任何法律、商业、税务或投资建议，亦不得用于评估任何证券或投资决策，不可作为针对 a16z 旗下任何基金现有或潜在投资者的要约。同时，Andreessen Horowitz 及其关联公司可能在节目所提及的部分企业中持有投资权益，相关完整披露信息请参阅 a16z 官方网站。 ## Notable quotes > `45:00` 我们在消费级 AI 中看到的绝大部分产品都在帮人节省时间。但那些帮人们‘花时间’（spend time）的产品，往往最终会成为规模最庞大、市值最高的一批巨头。 — Olivia Moore > `18:00` Most people aren't looking to save time. They're looking for ways to spend their time. — Eugenia > `42:47` 一封听起来 99.9% 像我本人语气的邮件，与一封听起来 85% 像我语气的邮件，二者之间的差距，是花 10 秒微调与花 10 分钟重写的本质区别。 — Olivia Moore > `32:15` 在这么多市场中，你看到的都是真正专业化的产品，它们高度聚焦，清楚自己的受众是谁以及他们在做什么，从而能够打造出高度定制化的东西。 — Josh Elman > `34:42` 当我们思考‘全部 AI’与‘消费级 AI’之间的鸿沟时，正是这种把复杂能力打包、打造出色产品体验的运动，才真正把它们变成了可持续沉淀的事物。 — Josh Elman > `37:28` 在位巨头不愿意或无法蚕食他们现有的交互界面。 — Olivia Moore > `27:15` OpenAI being able to do crazy good targeting on like even better than I think Meta, which previously was like best in class, because they have live data of you in a live conversation with them. — Olivia Moore ## Mentioned Elena Burger, Olivia Moore, Josh Elman, Sarah Wang, David Paolan, Eugenia, OpenAI, ChatGPT, Claude, Gemini, Google, Meta, Threads, Instagram, Anthropic, Amazon, Shopify, Netflix, TikTok, YouTube, OpenClaw, Crockbot, Town, Dots, Muse, Instinct, Polk, Tomo, Wabi, n8n, Granola, Higgsfield, Manus, Open Evidence, ElevenLabs, Suno, Midjourney, Sora, Figma, WhisperFlow, Superhuman, Krea, Replit, Gamma, PLAUD, Andreessen Horowitz (a16z), Yipit ## 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. [Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/15cd7eea-99bc-4db9-b027-1fdb2500162c/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=15cd7eea-99bc-4db9-b027-1fdb2500162c&feed=JGE3yC0V)

[45:00] Olivia Moore: 我们在消费级 AI 中看到的绝大部分产品都在帮人节省时间。但那些帮人们‘花时间’（spend time）的产品，往往最终会成为规模最庞大、市值最高的一批巨头。 [18:00] Eugenia: Most people aren't looking to save time. They're looking for ways to spend their time. [42:47] Olivia Moore: 一封听起来 99.9% 像我本人语气的邮件，与一封听起来 85% 像我语气的邮件，二者之间的差距，是花 10 秒微调与花 10 分钟重写的本质区别。 [32:15] Josh Elman: 在这么多市场中，你看到的都是真正专业化的产品，它们高度聚焦，清楚自己的受众是谁以及他们在做什么，从而能够打造出高度定制化的东西。 [34:42] Josh Elman: 当我们思考‘全部 AI’与‘消费级 AI’之间的鸿沟时，正是这种把复杂能力打包、打造出色产品体验的运动，才真正把它们变成了可持续沉淀的事物。 [37:28] Olivia Moore: 在位巨头不愿意或无法蚕食他们现有的交互界面。 [27:15] Olivia Moore: OpenAI being able to do crazy good targeting on like even better than I think Meta, which previously was like best in class, because they have live data of you in a live conversation with them.

[15:00] Eugenia Kuyda: Most people aren't looking to save time. They're looking for ways to spend their time. [5:40] Olivia Moore: I still don't feel like any of the assistant products have unlocked person to person network effects.

[0:00] The most interesting big new trend is in personal agents. We're in this world where we're so excited by what Chats VT did with turning things into conversation. And now these agents we can get in our messaging apps to turn them into more conversation. It's an incredible paradigm. About half of Americans report using AI. Around four and a half percent of U.S. consumers are paying a subscription to an AI product. credit cards on AI. The reality is people are now building software. They're building these really rich products. They may be using their agents to write some of that code, but when they're delivering this whole experience to a user, it is much more than just a shim on the model. The run rate that OpenAI has reached through advertising, what does that tell us? That would previously take a company years and years and years to get to, even after launching an ads product. Half of Americans now report using AI, but only a small fraction are actually paying for it. And the people who do are spending a lot. In this episode, A16Z's Elena Berger sits down with partners Olivia Moore and Josh Ellman to unpack the 7th edition of our Top 100 Consumer AI Ads Report. For the first time, the report includes consumer spending data, revealing a very different picture from traffic alone. The biggest spenders are developers, creators, and other power users, with the top 1% spending more than $900 a month on AI tools. Why Advertising could unlock new consumer AI business models. and the categories that remain surprisingly wide open, from shopping and entertainment to social and dating. Welcome to the A16Z podcast. I'm Elena Berger, and I'm joined by Olivia Moore, who's the author of our ongoing Top 100 Consumer AI Apps report, and Josh Ellman on our consumer team. Today, we're digging into the seventh edition of this report, who's paying for AI, turn everyday errands into the next big consumer business. Team, thanks so much for joining. Thanks so much for having us. Olivia, let's just get right into it. I would love to hear your top takeaways from this report. Yeah. So this is the seventh time we've done this list. The very first time we only ranked all global websites that were AI native by just traffic, like hits on the website. And that worked then because the way that almost everyone was using AI was like a prompt box product in a browser. We've had to evolve that a lot over the last few years as AI expands. So this list is our most new data for the first time. So we worked with the EBIT data to get specifically consumer card spend data. So not accounting for enterprises or SMBs. And we ranked products by what consumers are actually paying for. I think for me, the biggest takeaways from this report is on a traffic basis. It looks like a lot of these products are kind of settling. We only had 11 new products across the web and mobile list combined, which is the least we've had in any of the prior editions Yeah. Yeah, that's something that's funny. We also recently released our state of markets and something that Sarah Wang actually commented directly on is she said, hey, there is such a power law game happening now. It's happening at the enterprise level and it's happening at the consumer level. So one thing that's interesting. probably notable is, yeah, we have the fewest number of newcomers on this list, and it seems like the space is ossifying. But I think the past few months, things have actually been picking up a little bit more. And I think the most interesting big new trend is in personal agents. So what are you seeing? What are the big shifts that are happening? Yeah, it's funny. The last report that we did, it ended right before the cutoff of when open-claw traffic would come in. So we had been very high up in the rankings. And this time, actually, OpenClaw is nowhere in the rankings because the traffic has completely declined. I think that the team was acquired by OpenAI, so they're probably working on the agent assistant products there. But also, it's kind of been overtaken, I think, by some of the both other attempts at prosumer agents like Crockbot, Town has some fantastic agents, Dots now by OpenAI, which is new. But also on the consumer side, giving to a family member and telling them that they can safely use it. [5:00] Most of those also, because to your point, this consumer-assistant trend has literally been the last three weeks to a month, a lot of those didn't appear on the ranks yet. But we did pull the data to see what's happening with them. And it's been very, very fast growth and yet still not quite reaching the mainstream users, what I would say. using as a tool to enhance our productivity, to maybe search for us faster, something that would actually get things done for us. And I think OpenClaw was an incredible pioneer to do this. And I know they're still doing a lot of work through their open source foundation to continue to advance that. But it really turned a lot of heads on, wait, what can these products actually be? And we were seeing people buying Mac minis. We were seeing people run all their own data with their own local OpenClaw. And there was a huge run on that. But it was really inspiring to kind of feel this new energy of what is possible. And like most trends, that spark happens. It takes a little while for those first products to get built. It takes a little while for the momentum to then start getting those even released in the market for users to see them. And so in some ways, though, this report is showing like that's what we're talking a lot about on our podcast in Silicon Valley. But this really does a great job representing what's happening in the broader world. And I do think the truth is AI really is showing up everywhere, even if there's not a lot All the entrants are all getting so much bigger. Yeah. I think probably the biggest things on, when we talk about pure consumer assistance, the biggest developments have been Muse and Instinct, I would say. Although there's many other fantastic products that we kind of list in the report that have also, you know, Polk was a very early pioneer here. They were a pioneer of this, yeah. Exactly. Tomo has a very big audience and is doing incredibly well. Instinct had announced like 100,000 users, growing 10% day over day. in the first three weeks and they were spending over $1,000 on average in their first month. Muse is interesting because in many ways, very successful launch, especially within the tech community. I think the number was like 500,000 downloads and 250,000 active users in the first 12 days or so. But if you look at it versus Threads, which was another somewhat recent big splashy meta launch, it still pales in comparison in terms of appeal to the average person who Yeah. So if you look at the downloads, Muse is still in the U.S. and Canada. So we grabbed the Threads U.S. and Canada downloads only. In the first 22 days or so, Threads had about 16 million downloads and Muse is still sitting around 5 million over that same time period. Is the distribution mechanism different? Like I remember with Threads, it was already tied to your Instagram account and there was kind of this very natural distribution. or is it just a different distribution strategy too? I'd be curious what Josh thinks, but from my point of view, it's not as aggressive for sure. And also part of that is because there's a much higher cost to serve these users, so they don't want it to get too big too fast. And I also think threads very much had a critical mass effect, which is if they didn't get a lot of people on it communicating quickly, you don't actually get to that leverage of a new network starting to form. Because Muse is still such a personal tool that they're pushing it and it's very exciting this early just adoption and people having these experiences. And again, if you're using OpenClaw in January, you were having incredible experience of using CrockBot. When that first launched earlier this summer, it was spring, you were having these incredible experiences. And now we're seeing, as I think Olivia said, our moms, our families, people who don't just live on the tech, what's next in AI cycle. And I think that's what's really exciting about this consumer. I assume when we do this again next time, we're going to see massive adoption. But I think it hasn't needed to push as But then I also think about what's going to make these things stick around for a long time. And it is going to be things like network effects and marketplaces and all the businesses you can interact with those agents. That's going to be really transformative, too. Like Amazon is like, we're not going to allow Muse to kind of browse and purchase products on our site, but also Muse has a already hundreds of other partnerships with different products. Shopify, you know, to be able to buy things on Shopify. Exactly. Yes. I still don't feel like any of the assistant products have unlocked person to person network effects. And part of this for me is this software knows us more intimately than any other product in the past based on email. But also I find myself telling you [10:00] Extremely personal thanks to my AI. And so there's kind of this natural tension between the more useful the product is to you as the consumer, the more it knows about you. And then do you really want to introduce that into other contexts with other people? It's the same thing as many people have a separate ChatGVT work account and personal account. And no one has quite cracked that yet. No, and I think as we talk about consumers and AI, privacy, safety, security are becoming more and more important. I mean, there's obviously the overall trends of, you know, trust in AI and this kind of belief that AI is good for the world that is in some ways, you know, affecting adoption and affecting some of the excitement around these products. But on a very personal level, if you don't trust the thing that you're giving it your most intimate access to your email or your life or your credit card, and it might do rogue things that you wouldn't expect or share things with other people you wouldn't expect, you know, that's going to be the biggest impingement on these things really growing as massive consumer apps. So we're in a fascinating time where we're all learning these new norms. And these are norms that as people, we're actually very good at it. We're very good at knowing what information we're comfortable sharing, what information if somebody gives it to us, what information that we're trusted to be able to pass on to somebody or should be protecting and keeping ourselves And we just all of a sudden expecting software to do this too It a very you know interesting new set of things we learning Yeah Well you both just brought up like two interesting sort of facets of this One is like there's this endogenous force, which is all of the platforms and how they allow agents to interact with them. And then there's more like this exogenous force of like how comfortable am I using this and how maybe I mixed up endogenous. One is on the outside, one is on the inside. We knew what you meant. One is what the platforms are doing and one is how you feel about it or how you feel about the agent. And I'm curious, which do you think is going to get resolved sooner? And whose onus is that? Whose responsibility is resolving those kinds of questions? I think what the platforms are capable of is improving faster than true consumer willingness to use them, which we've seen kind of throughout AI so far. It interesting because a lot of people are like seeing these numbers that founders of these assistant products are posting that it costs like hundreds or thousands of dollars a month to serve a user And like how is that possible And I was stumped as well but I have been working with David Paolon who runs Assistant Benchmark who fantastic I think there's a recent podcast with him. Yes, he's great. So basically what he's done is he's created a website where they track, I think, over 170 agents. I checked it last night. It's probably like 200 something this morning. And he has users kind of use them and then rank them on success on different tasks. But the other really interesting thing that he does is he hosts kind of group chats and communities of early adopters of these agents. So he has a group of, I think it's 1500 plus users. And he pulled some cool data for us, which is like, what are they doing with agents? What are they talking about doing with agents? And the vast majority still, the number one use case was coding and technical automation, even for these consumer assistant products like Muse and Instinct. And so under that framing, it's like, oh, of course it costs that much. This is like when we talk to some agent or assistant products that are aimed towards the more mainstream consumer, they're seeing costs in the tens of dollars, not the thousands of dollars a month. So I think that is maybe a representation of the fact that we have a very long way to go in terms of consumers getting comfortable with these products And I think you know I think people who been coding and have been working with AI I mean it been such an incredible enhancer And that shows up in a lot of the data you know on the top 100 here Like it incredible And then we people who've been able to do that have had these great ideas for all the other things in their life. But when you just hand this to somebody new, it feels like a blank box. And it feels like, what should I do with this? And maybe there's examples of getting a reservation or booking a flight. But if you're not actively seeking a reservation or booking a flight, it's hard to figure out how to personalize it. And so much of what we do as people is we learn from others and we get these great ideas and we try the other things people tell us. So much of how social media has worked is people put things on social media and other people see that and they try the same thing. And then it kind of spawns these effects. So I try, they make a little riff on it and then those riffs become, you know, cascading snowballs. And that's how great ideas flourish. And with AI, because so much of it's so personal, we haven't even seen any of that flourishing yet, you know, outside of areas like coding and productivity, you know, all of the creative work, you know, I, I created this incredible video and how'd you do that? I use AI. Um, you know, and there's so many tools that do that, that, that are, you know, been growing and are here, but it's so interesting that, you know, now we have to figure out how to mainstream this for not the, you know, millions and millions of people who are using these and paying and everything, but for the. [15:00] You know, tens, hundreds and, you know, hundreds and hundreds of millions. If you look at broad usage data now, about half of Americans report using AI and about 25 percent of Americans think they are interacting with AI probably on close to a daily basis. But the spending data is much more concentrated. So it depends on the source. our YIPIT panel shows probably sub 5% around four and a half percent of U.S. consumers are paying a subscription to an AI product. Some other sources say as low as two or two and a half or maybe a little bit higher. It is expanding. So that's like doubled from a year ago. But that spend is still extremely concentrated. So even within that top 4.5% of people who do spend, the top 10% of those are like more than half of the revenue. And the top 1% are 20% of the revenue, whereas the bottom 50% are like 16%, right? So like the top 10% of users are driving this market. This was the other interesting thing about revenue data is we were able to look at it kind of like on a almost person by person basis what it looks like who are these people who are spending on AI The top 1 user is spending per month personally on their personal credit cards on AI And the median is spending And again, all of this is within the only the 4.5% that are spending at all. Yibbit also was helpful for us because we could see what products are they actually spending on, like on a personal basis. And probably unsurprisingly, it was mostly developer tools, productivity tools, creative tools, things used to like build, make, sell. Things like N8N, Granola, Higgsfield, Manus, those all were way overrepresented amongst the top spenders. I mean, in a way, this is actually really cool. I mean, even though these numbers aren't everybody yet, like we're enabling this whole new generation of makers of people who are not just using this for work, but actually using their personal credit cards to code things they never coded before, to make art, make videos, make content, and really have this power of making. And so it's actually, in some ways, while it's all still so early, it's actually really cool how much making is going on that people are actually so excited they willing to spend I mean if you spending a month on a set of AI tools you really using that to like hopefully make things faster better have more ideas in your head coming to life than ever before. Yeah. That's kind of like when I when I was looking at these categories between creativity and coding, like, you know, that's actually hopefully really, really powerful. Yeah. Well, I think this is a point that's really important, which is I think developers and coders are people who are naturally attuned to sort of ask, how do I get leverage over this thing? How do I build a product that allows me to get leverage over my time? And AI obviously supercharges that. But it is a specific kind of mindset, I think. And it's a specific kind of worldview. And I think part of my understanding with these numbers is it's still a worldview that It gets reflected only among a smaller group of people. And the question might be just like, how do you expand that world view? How do you help people understand, like, here are all the ways you can get leverage over your time and over your life? We talked about this a little bit in this report in terms of areas where we have not yet seen AI native startups pop up or categories where we haven't. Eugenia who one of our portfolio CEOs runs a company called Wabi which is fantastic She this great quote I probably going to botch it but it something like people most people aren looking to save time. They're looking for ways to spend their time. Like this is why social media, entertainment, Netflix, TikTok, all of these YouTubers are the most used consumer products that we have. And I think that much of what we've seen in AI thus far, to your point, is like, how do I do this thing a little bit faster, a little bit easier, a little bit more impressively. And that's not an incredibly compelling daily or hourly active value proposition for most people. No, I think this saving time is still very hyper productivity, people who are focused on work or being creative. But I'm also seeing this next wave of people who are maybe used to be coders. They're no longer maybe people who want to make videos but never quite mastered the Pro Tools and Adobe Premiere or something else, where they're all of a sudden able to sign up for these things and take the ideas in their head and make them happen. So I do think we're starting to see these rise up. And again, I look at all these numbers as we're still so early and the fact that we're [20:00] that we're talking about. how big some of these companies are, and that 1% of people are spending $900 a month. I see these as harbingers of what's to come as more people feel empowered, as more people feel like they can create. Where it doesn't, it's not necessarily about saving time, it's about spending the joy of actually spending, wheeling this thing into reality. There were all these jokes of when people were in the peak phase of everybody was using their open claw for the first time, their agents, they'd walk around with their computers and their laptops, and they had to keep them open. And they were like, I'm not sleeping. I'm coding, and I'm making things more than I ever have. And I think that's starting to, again, shows up at the top of the data, but I do think it's starting to show up paying for AI products directly expand, which is maybe controversial because I'm like invested in consumer AI products and I'm a consumer AI maximalist. But the reason I say that is because substantially all of the consumer AI revenue thus far has come from direct subscriptions. Like on this list, we looked and it was like, you know, 85% or something of our web list monetized via subscriptions, another 62% monetized via kind of credits or token extra usage payments. only like 13% had ads or other options where like you are the product instead of paying for the product. And that's like a... If you look back at the history of the consumer internet, that's like a pretty unnatural inversion. Like almost all of the really big consumer technology companies that we have now make the vast majority of their revenue from ads or from transaction fees, not from subscriptions. Because I think the truth of it is most similar to the fact that not everyone is looking for a coding tool to help them start a business or create and sell a product. or don't want to spend their funds on software. If you look at the top 20 consumer subscription products globally. ChatGPT is already on there, which is crazy in like three and a half years of growth. Almost all the other ones are media companies or big platforms like Amazon or even Uber, things like that, where you have a very frequent purchase behavior. So I actually am a big believer that we need to move. We have transcended the need for everyone to buy a subscription to AI and we need to see these other business models come back. I think we're still at the early days We're still at the early days of people finding ways to run lower cost models to build great products. And as they do that, you know, a lot of the other business models have always been on the Internet from, you know, transaction fees and ads will finally start working. But we need that to happen. And That's, I think, where we're still, you know, there's so much learning and so much that like, well, if it's going to cost me a lot to serve, at least let me just charge money so that people pay for it became sort of an early default. Yes. I mean, you can probably speak to this from seeing many prior eras of the consumer Internet. But like it used to be, at least for me as a consumer investor, if a company was making money in the first five years, it was like, whoa, what's happening? can then wait and make money via ads or start charging for transactions once we have liquidity and density. And that just isn't the case given how high COGS are. No, I think this cost thing is a really interesting challenge and companies get afraid of growing too fast. Yeah. Like they actually, if they grow too fast and they don't have, meters on how they're charging for it. You can actually have out of control spend that can be very, very hard to manage. So, you know, it's created this very different shape of what we see in the top apps in the top 100, you know, traffic and everything right now. But again, this is, again, we're all still so early. There's so much to build and learn and bring costs down and bring value up and introduce advertising and introduce other ways to spend money. And by the way, if it It's actually what we get continual value from. That's amazing, too. I think this is a great transition to talk about ads. Yeah. Your consumer top 100 report had a really, really interesting fact in there, which was the run rate that OpenAI has reached through advertising. They're at a billion dollars now in annual run rate on advertising. And what does that tell us? Again, the number was from August. And so there's probably a chance that it's reasonably higher, even a month or so later. they have been very slowly, I would say, rolling out ads with kind of a network of partners. And they're already at you know, a billion, billion in annualized run rate. Like that would previously take a company years and years and years. [25:00] to get to even after launching an ads product. I think there's two interesting things there. One, they have real density now. So I think the last reported number a few days ago was 1.2 billion weekly active users, which is like a real density. And then the second thing, and what I'm fascinated to see in terms of the evolution of the ads product is hypothetically, they should be able to charge more and have higher conversion rates because they just know so much more about you as like log in with ChatGPT, I think a wallet is coming, things like that. And so you could imagine that it becomes a much more compelling value prop for the advertisers. And then that just like juices the ad business even more. But we'll see. It's very early days there. It's early days. And I think, you know, there've been some really funny conversations about ads and AI. If you are talking to this incredibly personal thing, and then it's like, product that can help you, you know, lose weight fast. Like that's not building trust with AI. And so I think the other really interesting thing is like, how do we as an ecosystem, these products are so different. They're so much more personal than ever before. They're giving you answers that are custom tailored to what you're doing and what you're asking about and what it knows about you. That actually introducing ads has been done very deftly. I've actually been really impressed with the open AI rollout of ads when I've actually ever seen an ad. It's very conversation that I'm having. I think we're also, you know, one of the other shifts is people are starting to rely on chat with T a lot more for a lot more queries. It might have used to have been like, look up a fact or tell me the history of this thing or help me with my homework, not necessarily problem solving and things we're advertising. But now people are going to it for real advice or shopping or trying to figure out exactly, you know, where they want to go or what they want to do. And those are great opportunities when you're in a discovery mode where ads are actually often a great answer to help you solve that. And so a lot more of these commercial queries and these commercial intent creates that too. But again, it has to be done very, very deftly, not this weird, interruptive, very personal experience I'm having that ads just get shoved into the conversation. Yeah. I'd be, I'd be super curious to how the unit economics on ads work. Like presumably it's a, it's a lower cost question or lower cost compared to, you know, the coding or, or something like that or image gen. So, so for open AI, you know, what, what kinds of margins are they seeing on ads is a super interesting question. Yeah. I curious also, the ads i guess i i guess you're on a pro plan yeah enterprise so we don't yeah you know i was doing a bunch of travel research recently and um was was going to be in new york for a few days and i just started seeing some ads for a couple things that were going on in new york that i might have wanted to do and i just found that really useful and like i've started using it more for commercial shopping i was looking for a special um key chain that would have a really easy separation for a valet key. And I was having trouble finding this actually on Google. And I finally was like, let me just go ask ChachiBT. And I went back and forth and I was describing what I wanted and start giving me some really good examples. And at the end, I was like, great. And I'd also love one that's made in America. And I actually found this incredible option from a factory in Detroit that often makes a lot of other metal work and other stuff that had this really cool key for me. And it was one of those things that if I had never that conversation with disability. And if there had been other ads, no ads were presented in that conversation, but it was entirely open for ads that they would have helped me find stuff. And eventually I went to this website. I'd never heard of that company before. And it was a great experience. So interesting. I feel like we may see as they lean into it, OpenAI being able to do crazy good targeting on like even better than I think Meta, which previously was like best in live conversation with them. I do think that it'll help ChatGPT, but also I think it will also support a lot of these more focused, almost verticalized, to use the enterprise term, consumer AI products. So another company that has seen early success in ads is Open Evidence, which is an AI product for doctors. And it already has 50, I think the most recent numbers were 50 to 60% density of all US physicians on the product. And so when you have that kind of of a really valuable audience, and it's like a targeted audience because you know they're physicians, you can, I think, advertise really effectively. And there will be and should be more of those, I think. Is it worth talking and taking a step back and talking about open AI and... Anthropic and Gemini and sort of these are companies, these are labs that have been on the list pretty much since [30:00] the beginning. Are we seeing any kind of evolution in the way people are paying for these products? What can we surmise about the different kinds of users of all three? Super, super curious what you're seeing. Yeah, absolutely. From the very first time we did this report, the main conclusion was like ChatGPT is the dominant global user by both usage and revenue on the consumer side. has seen I think more movements and more shakeups so Chachi BD is now ahead of Claude um around 6x on the web, around 2x ahead of Gemini on the web. Probably going forward, given a lot of the usage is concentrated in these power users who are paying, the revenue numbers could be even more useful here. And probably that's one of the biggest surprises to me in this report was we got a panel of U.S. spenders and Claude has actually passed Gemini in terms of number of paid subscribers, has a much bigger install base and user base overall. And it also has like the natural distribution mechanism of something like a Google. I think that's a big accomplishment, of course, for Anthropic. I think it's come on the wave of a bunch of very successful products that they've launched over the last few months, like Claude Design. It's also come from a lot more press that they've had, starting with the Department of War stuff all the way back in, products are monetizing. And again, ChatGBT is still far in the lead by monetization. They have around three times more paid consumer subscribers in the U.S. than both Gemini and Claude. But Anthropic has said very famously, no ads. So they're much more aggressive about subscription. They have around seven and a half percent of subscribers on the hundred dollar plus per month, like the max plan. And that's like one percent for both ChatGBT and Gemini. And so those and are doing different things on the product, even in that paid user base. Should we talk about creative tools? And I think one thing that's interesting is actually all of the big labs have gotten really good at creative tools. And one thing that actually like Anthropik kind of famously abstained from, but then somehow achieves anyway, was like creating an image model or video model. But now people are making... images and videos because coding is so good at creating like images and videos. So it's interesting that, you know, they didn't build a diffusion model. Yeah, they're still capable of doing image gen. But about the pure creative tools, the companies that have kind of like verticalized in that category, what are we seeing? Creative tools is very interestingly distributed Few categories of creator tools or maybe modalities is the better term where the labs have not focused energy and attention for many reasons we could discuss. A lot of those are around audio. So like text to speech, music, 11 labs in Suno are now very reliably near the top of our traffic lists and are also ranked very highly in our spend list. The labs have so many things to do. They're building coding agents. They're building AGI. to deal with all of the IP headaches that Suno has had to go through and launch a competing music model. I don't know. And in the meantime, Suno has been able to kind of run away with a really strong lead. And, you know, I'm a big believer that, you know, the labs are going to do with they're going to do and they're going to keep growing and they obviously are great baselines for everything. But in so many of these markets, you're seeing really specialized products that are really focused, knowing who their audience is and what they're doing, that's able to build really custom things. If you want to make music just for fun or for something that you're really producing, Suno is an incredible tool and it speaks to you from the moment you start using it through the end. Eleven Labs, if you need voice and audio within the product that you're making, When you look at a bunch of the video tools, you know, when you look at design tools, even like Figma, like they're so designed to to use deeply for the thing you're trying to express yourself. And it speaks to designers, not to people who, you know, want to just skip that step. And I think we're going to see so much more value from these really bespoke interfaces where the software layer of all the things you can do and the things you save and the way you interact with it become really where the value sits, not just the models. Yes. I think of creative tools, which is images and video that looks different from audio. On images, I think to your point, OpenAI has really leaned into images quite successfully. Images 2.0 is very [35:00] very good both for image creation and editing. And then of course, Google with the various iterations of Nano Banana and Beo, I think they have candidly taken away a lot of the high level consumer traffic that was previously going to these standalone image generators. So like MidJourney was fairly high on our first ever rankings. It has fallen off the traffic rankings completely. But to your point in power users, if you look at the revenue rankings, MidJourney is back. has taste, has an aesthetic sense, and that they're able to tune in a really specific way. And then that brings us to video, which to me is kind of like a wild frontier right now, because Chinese companies that can train on any data and do have a real advantage. Video is just so complex to generate with like the visual, the sound, the visual and the sound synced, and looking a specific have to train the better. They've been successful there thus far. I think more broadly in creative tools, I'd be curious if either of you have experienced this, but I have recently been asking Astra and other really powerful models to generate or create, generate or edit creative for me. And it does a very, very good job and it will go tool call to elsewhere on the internet. And so I think that creative tools that are agent friendly and agent accessible are going to increasingly see Yeah, it's actually crazy what you can do if you just like empower an agent and just say like, yeah, install whatever you want to install. I had it. I actually had Astra and 5.5 go like head to head and animating some stuff for me. And like it was just like it knew exactly what to do. It's like, yeah. So. And what I love about this is this is the tinkerers, the people at the edge are figuring out these capabilities. that are going to show up on these lists because they're going to take all of this complexity of what you can do and package it up in a way that makes it really easy to use. And like that's this, you know, when I think of this gap of like all AI versus consumer AI, it's exactly this movement of packaging these up and building great product experiences that really turns it into these things that sustain. a real product and startup out of it, or 12 different startups all competing. Outside of the web window, there have been a big proliferation of mature products. So there's granola, there's whisper flow, there's superhuman. What are you kind of seeing there? And then obviously the ones that exist in chat and messaging, so in instinct. What are you seeing when it comes to and what are you seeing just when it comes to the kinds of products and users that they're approaching? Yeah, I think there's two maybe interesting threads to pull on here. One, I wrote this piece called maybe six, nine months ago called The Great Expansion, which is basically like, why are consumer companies becoming enterprise companies so quickly? Like things like CREA, you know, Replit, Eleven Labs, Gamma. revenue. And what's happening is that they're very successful in growing via PLG. Whisperflow and Granola are another great examples there. And then because they are kind of work productivity related, they get pulled into enterprises and then the founders are like, oh, we got to add privacy and security and team plan. But it's very effective, very low cost distribution. And again, it works really well for these tools that are can be used personally, but are like work adjacent. and superhuman are another good example of this trend we've seen in general, which is incumbents being unwilling or unable to cannibalize their existing interfaces. The classic example of this is Google. Like we haven't really seen them reinvent docs or Gmail or calendar for the AI age. I think it's even happening with OpenAI and Anthropic where they have been much less successful or codex or cloud or cloud code interfaces. So, and I know many of them have, you know, attempted to incubate killers for these many startups, but I think it is... from a product perspective for them to innovate beyond that, especially when they have so many other things on their plates. I mean, I think all that's true. And I think we now, as we get used to some interfaces, we're just still in such a period of incredible reinvention. Like a lot of the reason we've seen so much growth over the past year, year and a half, [40:00] is the models got so much better. And the models got so much better. So the people who had been spending time creating these incredible interfaces that were easy to use, that were natural, that were easy to share and get a bunch of your team on and Whisperflow and Granola were just so easy to get started with. You then would say, oh my gosh, now they've gotten so much better as the models have gotten better that the products, the utility gets so much better too. And so we're kind of this, it's incredible, you build something great, but it may not do everything you want because the model capabilities weren't quite there. the models get better and your product gets to explode in demand and utility. And this kind of rising tide lifts everything. And so to think that, you know, we only have one or two companies that can build that experience layer, I think is what this sort of report sort of debunks. Now, those are certainly some of the top companies in these very general use cases. But it's, you know, the reason that so many other things are growing is there's so many interfaces that get to be reinvented. And the sort of rising tide lifts everything. And then, you know, the infra value that are serving inference. And there's a whole different conversation there. But what consumers are adopting are things that work for them and are easy to use and bring into their lives. Yeah. Maybe to also make the, to continue to make the case for startups here, OpenAI and Anthropic in many ways are, startups, but they have been so successful that they have become big companies with thousands of people. And there are just so many extra considerate, like a larger ship is harder to kind of steer to turn around. And so I think that's like one of the companies on our revenue list this time is Plod, which is an AI note taker device that started kind of in China in the East and now has can and will do a hardware device. But as we've seen with them, it just everything takes longer. When you're a big company, you have to integrate with maybe other existing products. There's more checks and balances it has to go through. So I do think that the big labs will continue to be very, very successful. And there's a lot more for them to do. And I think that startups will continue to be able to find these quite big windows of opportunity. Yeah, and it also sounds like the question for startups That, as you said, Josh, gets better as models get better, but doesn't get cannibalized as models get better. Yes. Which is, you know, what is like an interesting kind of sweet spot to exist in. I mean, part of the reason that I was so excited to join Andreessen Horowitz a couple months ago was this belief that the value really now is moving back to the software layer. do it. And there's new innovations like you know, Jeff from TypeSafe that allows even developers to build things in even new ways. And we haven't even seen all the incredible new ideas that can be done when you have access to a model like that that's fast, that's easy to integrate, that, you know, gives you code, you know, or very structured answers back. And so because of that, now the creativity is back to the experience, the network you can build, the ways that you can serve a lot of people and bring them together using the models to be something powerful, but the context and the community really becomes the asset. And there's so much invention. And I think almost every category has a chance to be reinvented. And, you know, some big companies are going to be able to make the leap and become the relevant company for their sector in the AI era, but not all of them will. And that's where startups come in to kind of beat them to that punch and really be the best new possible experience. Yeah. Yeah. I think to that point, I think a lot about compounding personal value and hopefully eventually compounding multiplayer value that becomes like a lock-in to One example I would give is town, which is more on the prosumer end of the agent assistant. Connecting to all your systems, Slack, email, Google Drive, understand kind of the context of you and your org and then do things for you. Like I use it to write a lot of my emails, helped with a lot of this report. And I know it's easy for people to dismiss things like this as like, oh, it's just a harness, which is now the slightly less derogatory term for rapper, I guess. and that I cannot easily take and migrate over to something that comes along tomorrow. Because the difference between an email that is 99.9% sounds like me and even an email that's 85% sounds like me is the difference between spending like 10 seconds to fix it and like 10 plus minutes. And so I think that we'll see more products like Town that do a really good job of collecting context and acting on that context to build playbooks around you continue to be successful. No, I think that's totally right. to avoid using words like harness and wrap. Yes. Because I think [45:00] Reality is people are now building software and they're building these really rich products and they're, you know, they may be using their agents to write some of that code, but they're actually when they're delivering this whole experience to a user, it is much more than just a shim on the model or a harness that brings some context to the model. It's actually a really rich product and an experience that happens to now use a model as part of what makes them able to deliver that. And I think that's really the excitement of where that's going to, you know, what we're going to see on this list. these already are doing that. And we think really the next wave, they're all going to be these companies that have built so much more value. And the model is something that is, you know, just one piece. Olivia, I think you referenced this chart that you made, but maybe we can bring it up, is all of the white space that still exists for, you know, all of the, just all the product categories that exist in Web 2 and the pre-AI age. And maybe you can just talk through some of these categories and where are the big opportunities. Yeah. The vast majority of pure consumers are basically using AI still as a replacement for search products, things like Google. Maybe especially as you get into education use cases, kids are using it to write essays. A little bit more of an expansive use case. Homework help. Yes. Or people in the workplace write emails. But if you look at this this chart, which will show almost all of the kind of areas where we've seen startups succeed thus far have been in and around that theme of getting things done. You need a multiplayer product. You know, dating is one we haven't seen. There are certainly startups there that I hope and believe will make future additions of the list, but there's no one on the list doing dating now. Recruiting is another one. I think of dating and recruiting as somewhat similar networks, actually, in terms of trying to make matches. Social AI, we really haven't seen anything take off. social platforms. And then I think there's all of these other categories that are look maybe more like marketplaces did, like shopping, home buying, you know, retail. We'll have to see if those end up living in agents like Instinct or Muse or living as standalone products or the next ones, which there will be many, or living as standalone products or perhaps a mix of both. Perhaps your AI native marketplace. So I'm very excited about those things. I think also categories like gaming and entertainment, the models just have not been good enough to produce content that the average person likes to watch, which may be the exception of AI micro dramas, which are blowing up. But I think that is coming soon, too. No, I think that's exactly right. And I just think we're in this world where, you know, we're kind of like so excited by what Chachi VT did with turning things into conversation. And now these agents, we can get in our messaging apps to turn them into more conversation. There's this incredible paradigm. But when you're shopping, you want to visually see options and explore and have things customized for you and be able to say, I really like that, but I wish it had some piping on the collar or something and actually have AI helped you both realize that maybe even get it made for you. I like shopping and have so much invention there. Entertainment, the surface of the ability to tell better stories, of the way to empower people to be better storytellers that then creates these incredible networks of people consuming stories. You know, TikTok was this incredible place where people came together and just create little videos. You see people around the world that you're now watching, but there's still so much capturing people's heads that if only they could express their ideas even better, it's going to help that. And we're going to have an incredible new network. So I feel like we are just at the cusp of all these categories and we've needed the models these ideas and invent them. But as that happens over the coming years and with the new model capabilities, I still think the value is going to end up in this software layer in the products. Yeah. To me, this gets back to what we were talking about of products that help people save time versus spend time. Almost everything we've seen in consumer AI is save time. And that spend time, those end up often actually being amongst the biggest companies, if not the biggest ones. And so we need that to exist. And many, many builders, I'm sure, are tinkering away on that Absolutely. I think that's a great note to end on. Thank you both so much for joining and absolutely make sure to check out the top 100 consumer AI apps. Seventh edition. Thank you. Thanks for having us. We will see you for edition eight. Yeah. [50:00] thanks for listening to this episode of the a16z podcast if you like this episode be sure to like comment subscribe leave us a rating or review and share it with your friends and family for more episodes go to youtube apple podcast and spotify follow us on x a16z and subscribe to our sub stack at a16z.substack.com thanks again for listening and i'll see you in the next episode As a reminder, the content here is for informational purposes only. It 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. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see A16Z.com forward slash disclosures. Thank you.