Hyperscalers Lift AI Spend to $780B as Profit Growth Anchors Multiples

盈利增长支撑估值，超大规模云厂商AI支出升至7800亿美元 尽管生成式工作负载开始挤占传统企业软件预算，但企业利润的大幅增长与持续的算力紧缺，支撑着超大规模云厂商向年基础设施支出1万亿美元的目标迈进。 a16z成长投资团队指出，尽管市场经历了快速上涨，但当前科技周期是由两位数的盈利增长与估值倍数收缩所推动，而非投机性泡沫机制。预计超大规模云厂商的年度资本支出将突破1万亿美元，带动一个以能源、芯片和数据中心地产巨额投资为特征的工业“原子时代”。在应用层，企业表现出明显的落地差距——虽然69%的企业已开展实际部署，但仅有2%系统性地追踪投资回报率（ROI）；同时，软件供应商面临严重分化，一边是被动压缩利润率的防守型企业，另一边则是重新加速收入增长的进攻型平台。与此同时，顶尖的非上市超级巨头继续通过高流动性的二级市场要约收购推迟IPO，风险投资亦日益转向实体自主性、机器人、数字生物学和现代国防等领域。 在Alphabet、Amazon和Microsoft总计1.7万亿美元云业务积压订单的支撑下，超大规模云厂商的资本支出预计将在2026年达到7800亿美元，并从2027年起每年突破1万亿美元。 (6:28) 尽管股指创下历史新高，但市场市盈率倍数却收缩了约20%，这表明市场回报是由两位数的盈利增长驱动，而非投机性的估值倍数扩张。 (4:57) 企业在成效追踪方面仍存在巨大差距：尽管标普500指数成分股中有69%的公司已上线运行AI应用，但只有2%在持续时间周期内系统性追踪业务影响指标。 (15:50) 后期非上市科技公司的规模已达历史性水平，包括OpenAI、Anthropic、Databricks和Stripe在内的前六大非上市领军企业估值约达2.4万亿美元，超过了过去十年所有科技类IPO市值的总和。 (42:23)

The a16z growth team argues that artificial intelligence has initiated an economy-wide capital cycle where software demand directly compels historic investments in physical infrastructure. Contrary to dot-com bubble comparisons, they contend this buildout is anchored in corporate earnings expansion, even as enterprise workflow diffusion and consumer monetization remain in their early stages. Valuation Fundamentals and Hyperscaler Capital Expenditures (4:57) Despite the S&P 500 gaining roughly 90% in the four years following the launch of ChatGPT, the team highlights that average forward price-to-earnings multiples contracted by approximately 20%. The broader index trades below 20 times forward earnings alongside corporate profit growth of roughly 15%. They contrast this disciplined profile with the 100x multiples of the dot-com peak, noting that critical hardware suppliers like memory manufacturers frequently trade at single-digit multiples between 6x and 7x forward earnings. Simultaneously, hyperscaler capital expenditures across Alphabet, Amazon, Meta, Microsoft, and Oracle are projected to increase from $416 billion in 2025 to $780 billion in 2026, topping $1 trillion annually starting in 2027. While upfront data center shells and chip procurement depress near-term free cash flows, the team argues that an aggregate cloud backlog of roughly $1.7 trillion across Microsoft, Alphabet, and Amazon reflects durable customer demand that consensus models expect will inflect into significant cash flow recovery by 2028. Physical Infrastructure Bottlenecks and Grid Economics (12:12) The hosts assert that AI demand has transitioned technology from pure software into a physical "age of atoms," with annual compute and data center expenditures outpacing the 19th-century American railroad buildout relative to GDP. They point to an estimated $90 trillion in worldwide infrastructure capital requirements through 2040 across power generation, utilities, transit networks, and facilities, mirrored by large-scale physical manufacturing footprints from portfolio companies like Anduril, Waymo, and SpaceX. Addressing concerns regarding the impact of data centers on municipal infrastructure, the team cites US electrical grid analyses indicating that a 10% increase in local data center capacity correlates with an average 0.40% decline in residential power rates. Because regional electrical grids operate as shared fixed-cost assets, high-volume baseload consumption from data centers allows utilities to distribute capital overhead across greater aggregate volume, which can suppress unit costs for local households. The Enterprise Diffusion Gap and Software Bifurcation (15:50) Enterprise AI adoption exhibits a severe drop-off between experimentation and operational measurement: while 69% of S&P 500 firms have live AI implementations, only 30% demonstrate quantifiable business impact, and just 2% track recurring AI performance metrics. A stark user divide has also emerged; the top 1% of enterprise AI spenders deploy roughly eight times more capital than the top 10%, while AI-native startups spend as much as 10% of their headcount-equivalent expenses on AI infrastructure. This shift is cannibalizing traditional enterprise IT budgets. Public software metrics show an aggressive pivot toward profitability (roughly 75% of tracked SaaS firms are profitable), but only 30% maintain annual revenue growth at or above 20%. To prove platform durability against automation, the team establishes a 10% or greater revenue re-acceleration target. Concurrently, specialized vertical software startups like Harvey, Abridge, and EliseAI are outpacing historical SaaS growth rates by integrating domain-specific orchestration around low-cost inference models. Consumer Agent Disruption and Marketplace Tensions (28:40) Consumer AI subscription adoption is still nascent, with just over 2% of US households paying for an AI subscription, compared to over 70 million households for Netflix. However, early consumer cohorts exhibit rare upward-sloping "smiling" retention curves, as regular foundation model upgrades enhance product capability without requiring active consumer software updates. The partners contend that traditional attention metrics, such as 30 to 60 minutes of daily screen time, fail to capture the value of proactive, asynchronous agents executing complex background tasks. This transition to zero-click agentic execution is reshaping marketplace discovery. The team contrasts Amazon’s decision to block third-party agent Muse to protect its $70 billion-plus high-margin retail ad business with Instacart’s willingness to integrate autonomous agents. Because grocery e-commerce penetration remains low, under-penetrated platforms are incentivized to aggregate net-new gross merchandise value through agents, whereas ad-dependent discovery engines must defend sponsored query margins. Private Market Scale and Next Frontier Allocations (42:23) Elite technology companies are remaining private across extended cycles. The top six private market leaders—including OpenAI, Anthropic, Databricks, Stripe, and Revolut—command a combined valuation of roughly $2.4 trillion, surpassing the $1.7 trillion aggregate market capitalization of all tech IPOs over the past decade. Secondary share market discounts have largely compressed to parity with primary round valuations, while Carta data shows employee tender offer participation dropping to 58%, reflecting internal workforce conviction in ongoing equity appreciation. At the macro level, AI-related companies captured 86% of total US venture capital deal activity in 2026, up from 65% in 2025. Looking ahead, the partners anticipate that physical robotics could match or exceed software LLMs in economic scale over a three- to five-year lag. They also project that autonomous vehicle fleets—which operate roughly 14 times safer than human drivers—will radically expand their share of vehicle miles traveled, while modern defense contractors modernizing military procurement will scale far beyond their current footprint of under 5% of defense spend. While the a16z team concludes that the current cycle is anchored in real earnings growth rather than speculative multiple expansion, long-term returns remain contingent on resolving acute physical supply bottlenecks in power and compute, closing the enterprise measurement gap, and demonstrating durable software monetization as workflows shift to autonomous agents.

## Background 该视频由知名风险投资机构 a16z 于 2026 年 9 月 30 日在 YouTube 官方频道发布。节目中，a16z 成长投资团队合伙人 David George、Sarah Wang、Alex Immerman 与 Santiago Rodriguez 围绕其最新发布的《市场现状》（State of Markets）报告展开拆解，深入剖析了近万亿美元规模的 AI 算力基建竞赛、企业端落地真实现状以及下一个科技投资周期的演进脉络。 ## TL;DR 本期节目通过 25 张核心图表深入复盘了当前 AI 驱动的资本超级周期。四位合伙人指出，美股大盘近年的强劲涨势主要由坚实的底层利润驱动，而非估值倍数泡沫；与此同时，超大规模云厂商的 CapEx 军备竞赛正向万亿美元狂飙，并在“原子的新时代”全面引爆电力、物理工厂等底层基建需求。在应用与企业端，虽然 69% 的标普 500 公司已部署 AI，但真正长期量化追踪成效的仅有 2%，超级用户与普通用户的支出差距已被拉大到 20 倍以上；面对传统 SaaS 的严重分化与挤压，未来兼具自主执行力的 Agent、具身智能机器人、全自动驾驶以及 AI 与生物结合的交叉领域，将孕育出颠覆历史的超大体量科技巨头。 ## Atmosphere & style 整场对谈呈现出专业投资人闭门会议式的冷静、敏锐与高密度推演风格。四位合伙人不仅没有被市场当下的概念狂热所裹挟，反而用大量详实的一级与二级市场财务数据、客户调用指标进行严密的逻辑穿透，在宏大愿景与残酷现实之间构建起清晰的坐标系。对话节奏干练紧凑，合伙人们在观点交锋与内部案例互抛之间穿插着恰到好处的幽默，既散发着身处技术剧变风暴中心的兴奋感，又始终保持着风险投资老手独有的严谨与清醒。 ## Bulletin - [06:28](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=388s) 超大规模云厂商 2026 年资本开支预计达 7800 亿美元，2027 年起将破 1 万亿美元，同时微软、谷歌与亚马逊合计握有 1.7 万亿美元云积压订单，支撑自由现金流走出 J 曲线。 - [17:55](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1075s) 标普 500 企业中虽有 69% 部署了 AI 应用，但能够量化实际成效的仅占 30%，真正纳入长期 KPI 持续考核的企业仅有 2%，企业级渗透仍处于极早期阶段。 - [36:14](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2174s) 美股公开市场 SaaS 板块呈现低增长与高盈利特征，年营收增速达 30% 以上的上市软件公司已不足 5 家，企业 AI 预算直接挤压了传统软件采购与席位扩展。 - [42:23](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2543s) OpenAI、Anthropic、Databricks、Stripe、Waymo 与 Revolut 六大未上市超级独角兽估值总和达 2.4 万亿美元，已超越过去十年除 SpaceX 外美股全部 IPO 公司的总市值。 - [48:00](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2880s) AI 相关公司在全美风投交易活动中的占比已攀升至 86%，资本投资全面从前沿大模型向电力、半导体、防务、具身智能及全自动驾驶等物理基础设施扩散。 ## Key takeaways - [02:56](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=176s) 科技已成为拉动全美宏观经济投资的引擎，高科技设备、软件及研发投入占全美 CapEx 的 55%，本轮基础设施建设占 GDP 比例已超越 19 世纪美国铁路繁荣期。 - [04:56](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=296s) 当前牛市上涨是由坚实的基本面盈利支撑而非虚假估值倍数扩张驱动，在标普 500 上涨期间其远期市盈率反而下行，并未重现 2000 年互联网泡沫的结构性风险。 - [06:28](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=388s) 算力使用范式正由一问一答的偶发搜索转向全天候长程自主 Agent，持续并行计算让算力需求全面击穿预测模型，促使超大规模云服务商持续调高 CapEx 预算。 - [13:15](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=795s) 到 2040 年全球基础设施投资需求预估达 90 万亿美元，数据中心作为稳定基荷用户接入电网能够有效摊薄固定资产维护成本，全美数据中心容量每增加 10% 居民电价反而下降 0.4%。 - [15:50](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=950s) 推理成本剧降两个数量级引发杰文斯悖论，催生出大量具备经济可行性的新用例，推动 OpenAI 与 Anthropic 的年化经常性收入（ARR）以远超历史上任何软件公司的速度狂飙。 - [20:28](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1228s) 企业 AI 消费呈现极端马太效应，AI 原生初创公司超级用户月支出可达中位数的 20 倍以上，AI 预算占人力成本比重在激进初创中达 10%，而传统大企业仅在 1% 左右徘徊。 - [22:43](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1363s) AI 对企业损益表的影响开始显现确定性验证，Chime 单客服务成本连续四年复合下降逾 10%，Shopify 商家冷启动转化率提升 8%，增收型投资的商业天花板显著高于单纯降本。 - [25:51](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1551s) 为平衡任务 SLA 与推理开销，企业正广泛采用 Prompt 缓存、智能模型路由与专用小模型微调策略，使软件公司的核心商业度量重构为“完成端到端任务的全包成本”。 - [28:40](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1720s) 全美仅略高于 2% 的家庭拥有 AI 付费订阅，其留存呈现独特的微笑曲线；同时由于 Agent 转向后台无感运行，以使用时长为核心的传统消费级尽调框架正在失效。 - [31:34](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1894s) 具备高变现意图的 AI 流量引发传统平台博弈分化，年广告收入超 700 亿美元的亚马逊筑墙拦截第三方 Agent 以捍卫搜索广告现金流，而 Instacart 则选择拥抱接入以抢占增量 GMV。 - [39:18](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2358s) 软件资产估值正在剧烈分化，通用横向应用面临工作流重构压力，而受益于 Agent 复杂调用的网络安全、可观测性以及深嵌垂直流程的行业 AI 软件迎来了爆发式逆势增长。 - [43:48](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2628s) 创业公司延长私有化周期使得创始人能承担更长周期、更具进攻性的战略押注，而定期二级市场员工要约回购以趋近于零的折价率成为维系人才竞争力和定价公信力的核心机制。 - [48:43](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2923s) 未来 3 到 10 年的投资罗盘全面辐射硬科技与垂直落地，具身智能、全自动驾驶网络重构、AI 制药、个性化健康管理以及新兴国防供应链将迎来资金的大规模迁徙。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=0s) 引言与“市场现状”报告概览 在节目开场，David George、Sarah Wang、Alex Immerman 和 Santiago Rodriguez 抛出了一组勾勒当前科技与资本周期的核心数据：全球市值排名前十的公司中，有八家是美国科技巨头；自 ChatGPT 推出近四年来，美股大盘已上涨约 90%，年化复合增长率达 17%；全球基础设施到 2040 年的投资需求预估高达 90 万亿美元，涵盖范围远超数据中心本身，蔓延至电力、水务、道路与交通等物理层；而在标普 500 企业中，虽然已有 69% 部署了 AI 实际应用，但若按真正持续追踪成效的指标来衡量，比例仅有 2%。 David George 宣布，a16z 成长投资团队（Growth team）正式发布年度《市场现状》（State of Markets）报告。本期节目将精选其中 25 张核心图表展开拆解，主要划分为两大主线： - **宏观层面**：研判算力资本支出（CapEx）、数据中心建设规模以及加速超预期的宏观需求。 - **产业实操层面**：剖析模型进展、应用落地、垂直行业实况，以及在物理基建承压背景下的软硬件变革。 ```mermaid flowchart TD accTitle: 基础设施投资向上扩散路径 accDescr: 展示从电力冷却等物理层到超大规模云厂商，再到模型与应用的价值扩散过程。 subgraph Physical["物理层基建 (Atoms)"] A["电力、水务与电网"] B["冷却、厂房与施工"] C["先进制程芯片与服务器"] end subgraph Cloud["云平台与资本 (Capital)"] D["超大规模厂商 CapEx"] end subgraph Intelligence["软件与应用层 (Bits)"] E["前沿大模型 (OpenAI / Anthropic 等)"] F["企业级 Agent 与垂直行业落地"] end A --> D B --> D C --> D D --> E E --> F ``` ### [02:56](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=176s) 科技即一切：历史级资本支出周期 David George 翻开第一张幻灯片——“科技是包罗一切的投资周期（Tech is the everything cycle）”。高科技设备、软件与研发（R&D）目前已占美国总资本支出的约 55%，科技板块市值占美国股市整体总市值的近 40%。科技不再只是单一垂直行业，而是彻底成了拉动全美乃至全球经济投资周期的主引擎，直接拉动了从底层电力、土建工程、工业产能到上层模型开发的全产业链。 这一轮由 AI 驱动的基础设施投入，在占 GDP 的比重上已经正式超越了 19 世纪美国铁路大扩建时期的峰值。各大前沿模型研发公司累计融资金额已突破 3500 亿美元。Alex Immerman 引用 Marc Andreessen 著名的“软件正在吞噬世界（Software is eating the world）”指出，过去 15 年软件虽重构了各行各业，但当时真正掌握软件构建能力的人凤毛麟角；如今借助 AI 与各类自动化脚本，团队内部普通人每晚都能自动跑数个自动化任务，软件生产门槛被彻底击碎，进而激化了对算力、芯片、电力与物理建设近乎无限的渴求。团队预测，如果站在 5 到 10 年后的维度回望，当前的累计资本开支可能还会再翻 20 倍。 - 高科技设备、软件与研发投入占全美 CapEx 比例：约 55% - 科技股市值占美股大盘比例：近 40% - 全球市值前 10 大企业中美国科技公司数量：8 家 - 大模型研发公司累计融资金额：超 3500 亿美元 - 历史横向对比：AI 基础设施建设规模占 GDP 比例已正式超越当年的“铁路繁荣期” ### [04:56](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=296s) 究竟是不是泡沫？盈利驱动 vs 估值倍数扩张 面对外界最核心的质疑——“这到底是不是泡沫（Is it a bubble）？”，四位合伙人展开了深入讨论。美股指数虽然持续刷新历史高点，但背后的驱动逻辑与 2000 年互联网泡沫或 2021 年流动性泛滥时截然不同： 股价整体上涨约 20% 的同时，美股的整体估值倍数反而下降了约 20%。标普 500 指数的远期市盈率（P/E）目前低于 20 倍，且伴随着 15% 左右的坚实底层利润增长。这表明行情上涨完全是由实打实的底层基本面盈利（Fundamental earnings）支撑，而非概念炒作带来的估值泡沫。2000 年互联网泡沫破裂前夕，诸多尚未盈利甚至营收微薄的领跑公司交易在 100 倍以上的市盈率，而当前即使是身处高景气度的周期性存储芯片龙头，其远期市盈率也仅在 6 到 7 倍左右。 Alex Immerman 坦言，ChatGPT 问世以来市场 4 年大涨 90%（年化 17%），投资者的第一直觉往往是“市场过热、势必回调”。但深入分析底层数据就会发现，高增长与合理估值倍数罕见地并存，这与以往纯粹依靠估值乘数扩张（Multiple expansion）推动的虚假繁荣存在本质区别。 ### [06:28](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=388s) 7800 亿美元：超大规模云厂商的算力军备竞赛 超大规模云服务商（Alphabet、Amazon、Meta、Microsoft 和 Oracle）的资本开支增长令人咋舌：五大巨头在 2026 年的 CapEx 预估合计达到 7800 亿美元，相比 2025 年的 4160 亿美元近乎翻倍；业界一致预期到 2027 年起，五大厂的年度 CapEx 总额将正式突破 1 万亿美元大关。 不同于个人电脑或智能手机早期从零普及硬件网络的缓慢周期，本轮 AI 浪潮是“直接架设在互联网、成熟云计算与数十亿存量智能手机”之上。现成的全球分发渠道，让任何算力与应用落地都能在极短时间内触达海量受众，导致需求持续压过供给。 Sarah Wang 补充指出，算力需求的暴增还因为使用范式已从单纯的“一问一答偶发式搜索（Occasional queries）”进化为“自主智能体（Agents）”。Agent 涉及大量并行任务（Parallel tasks）与超长运行任务（Long running tasks），即使在人类休息的夜间，后台仍有无数 Agent 在自主编写代码、检索信息、清洗数据并推进工作流，这种工作机制带来了前所未有的算力饥渴。 巨头们的资本开支预测在过去几个季度内被反复大幅上调，过去被视作不可逾越的天花板数字，在短短几个月后就变成了近在眼前的底线。Santiago Rodriguez 将这种算力需求形容为“彻底击溃预测模型的狂潮（Model buster）”。他提到一个生动案例：一两年前 Sam Altman 与 Sarah Friar（OpenAI 首席财务官）因为大举签下巨额算力长协合同而遭遇外界严厉抨击，被指责行事鲁莽激进；但从现在的视角来看，这一决策堪称极具远见。即便如此，OpenAI 此前仍因算力承载达到极限，被迫一度暂停了高端 Pro 订阅计划的注册，哪怕用户揣着每个月 2000 美元的支票想要付费也被拒之门外。整个数据中心供应链上下游，从最底层的冷却设备到核心算力硬件，全部处于供不应求的缺货状态。 - 五大超大规模云厂商（Alphabet、Amazon、Meta、Microsoft、Oracle）2026 年 CapEx 预期：约 7800 亿美元 - 2025 年对应 CapEx：约 4160 亿美元 - 2027 年及之后预期：每年超 1 万亿美元 - 算力瓶颈现实写照：OpenAI 曾因算力告急暂停 Pro 订阅方案扩容，需求严重倒挂于供给 ### [09:50](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=590s) 1.7 万亿美元云积压订单与 GPU 的真实经济寿命 超大规模云服务商（Hyperscalers）的高额资本支出（CapEx）一直备受市场审视，甚至有巨头因发债扩建算力而遭到华尔街的惩罚。然而，在需求侧存在着极其坚实的业务基本面：微软、谷歌和亚马逊三家科技巨头目前合计拥有约 1.7 万亿美元的云计算积压订单（Cloud Backlog）。当这些巨头大举投资基础设施以满足订单履约时，短期内自由现金流固然承压，但市场共识预测其现金流将在 2028 年触底反弹并开启强劲增长。 在最近的财报电话会议中，亚马逊详细拆解了底层资本投资的“J 曲线”动态（J-Curve dynamic）。数据中心的机房外壳（shell）建设确实耗时漫长，芯片采购也耗资巨大，但无论是 GPU 还是 TPU，其实际有效经济寿命远比外界最初预期的要长得多。此外，GPU 现货市场的价格持续坚挺，每一张能够点亮上线的计算卡都能以极具吸引力的高价格对外出租并产生丰厚回报。在这场算力竞赛中，谁若稍有迟疑或缩减开支，立刻就会陷入被动和后悔。 在整个技术扩散过程中，AI 正在为终端消费者和企业创造巨大的“消费者剩余”（Consumer Surplus）。用户因获得了前所未有的生产力价值而疯狂调用模型；提供上层服务的供应商借此赚取可观利润；即使追溯到最底层的芯片层级——尽管当前的算力采购成本远高于一两年前——全产业链各层依然维持着极高的利润率。 - **三巨头云积压订单**：微软（Microsoft）、谷歌（Google）与亚马逊（Amazon）合计订单储备达 1.7 万亿美元。 - **现金流转折点**：自由现金流预计于 2028 年走出 J 曲线低谷，随后迎来实质性反弹增长。 - **核心逻辑**：超大规模云服务商的 CapEx，本质上就是上游整个硬件供应链的销售订单簿（Order Book）。 ```mermaid flowchart TD accTitle: 基础设施投资J曲线与产业链价值传导 accDescr: 展示超大规模云厂商的资本开支如何流向物理基建，并通过长寿命算力转化为云订单积压与消费者剩余。 subgraph CapEx["资本开支投放（J曲线前期）"] A["1.7万亿美元云积压订单"] --> B["超大规模服务商举债/现金投入"] B --> C["土地、厂房外壳与电力基建"] B --> D["GPU/TPU 加速芯片采购"] end subgraph Operations["运营与经济寿命"] C --> E["数据中心长期并网运营"] D --> E E --> F["超长经济寿命 + 现货高租金回报"] end subgraph ValueCapture["价值释放（2028+ 拐点）"] F --> G["自由现金流强劲复苏"] F --> H["创造巨大消费者剩余与全栈利润"] end ``` --- ### [13:15](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=795s) “原子的新时代”：从软件代码走向万亿美元物理世界 随着科技巨头将天量利润倾倒回物理现实，这一轮科技繁荣已经彻底演变成为一场波澜壮阔的“工业繁荣”。风投机构过去的重心是投资具备零边际成本特性的纯软件（即“一次构建，无限分发”），而当下他们正把大量时间投入到数据中心供应链以及物理实体设施中。这类重资产业务要求团队具备截然不同的能力：精通大规模项目融资、复杂的供应商生态管理，以及精准预测电力与容量需求。 人类正迈入“原子的新时代”（The New Age of Atoms）。据测算，到 2040 年全球基础设施投资缺口及需求高达 90 万亿美元，涵盖的不单单是 AI 和数据中心，还包括电网、供水系统、公路和公共轨道交通。在投资组合中，这种物理扩张已经处处可见：国防科技独角兽 Anduril 正在建设占地相当于 87 个美式足球场大小的超级制造工厂；自动驾驶先驱 Waymo 正激进扩建其运营车队枢纽（Depots）；而航天巨头 SpaceX 也在推进上千亿美元级别的基地布局。正如埃隆·马斯克（Elon Musk）当年的名言“工厂本身就是产品”（The factory is the product），制造与物理落地能力已成为科技初创企业的终极竞争壁垒。 与大众流行舆论截然相反的是，数据中心不仅没有“耗尽全美电力资源”，反而在经济机制上降低了平民电费。全美最新研究表明，**数据中心容量每增加 10%，居民电价反而下降了 40 个基点（0.4%）**。电网本质上属于重资产的共享固定成本设施（包括电线杆、高压线缆与变电站），当数据中心作为体量巨大、用电稳定的基荷客户接入时，固定维护成本被摊薄至更庞大的总用电量上。例如 Meta 高管在播客中分享其路易斯安那州基地案例时就指出，科技公司与当地社区携手扩建电网，最终切实拉低了周边居民的日常电价。 - **全球基建需求规模**：至 2040 年，全球包括能源、水利与交通在内的基础设施需求约 90 万亿美元。 - **物理扩展典型范例**：Anduril 占地相当于 87 个足球场的先进制造基地；Waymo 运营车场网络扩张；SpaceX 的巨额实体投入。 - **电价经济学**：电网共享固定成本被数据中心大规模平摊，打破了“AI 抢电抬价”的舆论误区。 --- ### [15:50](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=950s) 杰文斯悖论、模型降价与前所未见的营收狂飙 将视线移至模型层与应用层，整个生态正在经历成本断崖式下跌与推理能力爆发的交汇点。随着推理成本以数量级（例如降幅达两个数量级）的剧烈速度压缩，AI Agent 在更加广泛的真实业务场景中具备了充分的经济可行性。这正是典型的“杰文斯悖论”（Jevons Paradox）：当一种资源的使用效率大幅提高、单位消耗成本骤降时，其总消费需求非但不会减少，反而会因为门槛降低催生出指数级爆发的新用例，为底层模型商和上层应用软件公司打开了庞大的增长空间。 在未上市资产领域，头部大模型实验室的变现速度完全改写了科技史记录。合伙人们每次在内部合伙人大会（GP Offsite）上展示 OpenAI 与 Anthropic 的年度经常性收入（ARR）合并走势图时，每个月都必须重新刷新数据，因为数字更新得太快。将这两家前沿实验室的年度新增收入轨迹与历史上最优秀的几家 SaaS 巨头进行对比，前沿模型实验室无论是在扩张规模还是在冲刺速度上，都实现了对传统顶级软件公司的全面碾压。 - **杰文斯悖论显现**：模型推理成本下降两个数量级，直接引爆了企业级 Agent 的大规模经济性可行方案。 - **颠覆软件增长历史**：OpenAI 与 Anthropic 的年化合并营收增长曲线，以极其夸张的陡峭角度超越了历史上任何传奇软件公司。 --- ### [17:55](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1075s) 早期采纳之谜：从 69% 概念验证到仅 2% 的长期量化指标 尽管模型实验室的营收数据亮眼，但从全社会的采用广度与深度来看，AI 变革才刚刚露出冰山一角。与 2008 年智能手机浪潮（只需简单统计持有一台 iPhone 的人口百分比）不同，AI 转型属于一个复合型的平台级变迁：投资人不仅要评估接触并使用 AI 产品的绝对人数，更要衡量 AI 在用户整体工作流和企业“钱包份额”（Share of Wallet）中的渗透程度。 标普 500（S&P 500）企业在落地 AI 上的真实转化漏斗揭示了一个极其悬殊的阶梯落差： 1. **试点部署（Live Deployments）**：高达 **69%** 的标普 500 企业已经上线部署了某种形式的 AI 应用； 2. **产生可量化影响（Quantifiable Impact）**：仅有 **30%** 的企业能够明确证明该部署带来了切实的商业价值或节约； 3. **长期追踪考核（Tracked Over Time）**：真正将 AI 的业务产出纳入长期关键绩效指标（KPI）并持续跟踪的标普 500 企业，仅仅只有 **2%**。 这种巨大的落差表明，虽然外界舆论已陷入狂热，但整个企业界对 AI 的实际利用还停留在表面。AI 创造实际成效的潜力是真实存在的，但企业在组织内部真正深化使用场景、重构工作流程并将价值长期量化沉淀的进程中，依然存在着无法估量的广阔上升空间。 ### [19:06](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1146s) 应用层机遇：填补模型能力与企业专有数据的鸿沟 对话从企业落地的现状切入。目前绝大多数企业对 AI 的探索仍处于单点工具部署的极早期阶段，距离将其深度嵌入具备高影响力的核心工作流还有很长的路要走。合伙人们坦言，与他们交流的大多数传统企业，目前对 AI 的接触几乎仅限于微软的 Microsoft Copilot，这直观地表明整个市场尚处于何等稚嫩的初期阶段。 然而，这种基础大模型“无所不知”与“对具体企业业务一无所知”之间的巨大断层，恰恰构成了 AI 应用层（Application Layer）最令人兴奋的创业空间。Sarah 回忆起与 Databricks 首席执行官 Ali Ghodsi 以及 Martin 的对话，Ali 明确指出通用 AI 了解广阔的世界，却对企业自身的私有上下文茫然无知。应用层初创公司的核心价值，正在于构建可靠工程框架去缝合这一鸿沟。以英国数字银行巨头 Revolut 为例，即便其自身拥有极其顶尖精悍的工程团队，在构建安全、自动化的银行客服支持时，仍选择与 ElevenLabs 合作，将其顶尖的语音模型安全接入客户账户与金融银行业务流中，确保客户拨打电话时能以极高效率、可靠且合规地解决实际问题。眼下的核心商业机遇，正在于把通用的前沿模型能力驯化为稳定、高度可依赖的企业级服务。 ```mermaid flowchart LR accTitle: 大模型通识与企业专有工作流的结合流程 accDescr: 展示通用大模型如何通过 AI 应用层框架注入私有数据，驱动端到端业务工作流。 subgraph S1["能力供给"] M["通用大模型能力<br/>(通识丰富 / 缺业务上下文)"] end subgraph S2["AI 应用层与工程化"] A["应用架构与安全桥接<br/>(Decagon / ElevenLabs)"] D["企业私有数据与上下文<br/>(Databricks / 核心数据)"] A --- D end subgraph S3["业务交付"] W["端到端生产工作流<br/>(金融客服 / 商家服务)"] end M --> A A --> W ``` ### [20:28](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1228s) 极端分化：超级用户支出狂飙 20 倍与渗透率指标 伴随着企业整体采纳仍处初期的现实，另一个惊人的结构性分化现象正在显现——超级用户（Power Users）与普通企业用户之间的支出差距被急剧拉大。来自 Yepit 的支出数据显示，在 AI 采购上排名前 1% 的超级企业客户，其支出中位数大约是前 10% 客户的 8 倍之多；前 1% 群体的消费量甚至几乎追平了第 2% 到第 10% 企业支出的总和。 这种分化在顶级 VC 的被投企业内部表现得更加淋漓尽致： - **AI 原生被投企业的内部支出**：那些彻底拥抱 AI、高度“AI-pilled”的科技初创企业中，顶级超级用户的月度 AI 账单高达 7,500 至 9,000 美元；而中位数用户通常只订阅一两款月费 200 至 400 美元的 SaaS 辅助工具，头部开销达到了中位数的 20 倍以上。 - **AI 支出占人力成本比重**：团队常以“AI 工具采购支出占公司总人力成本（Headcount Spend）的百分比”作为衡量组织 AI 化程度的标尺。当前相对前沿、敢于尝鲜的 Fortune 500 大企业，该比例仅在 1% 左右徘徊；而在 a16z 投资的最激进的 AI 原生初创公司中，这一比例已高达 10%。 - **采购与落地的巨大鸿沟**：大量企业在首次采购 Cursor 或 Cognition（Devin）时激动不已，但这往往只是马拉松的第一步。从最初的尝鲜采购到全员全流程的深度渗透，中间横亘着漫长而艰巨的落地跑道。 ### [22:43](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1363s) 确定性成果验证：从降本增效到业务增量 市场上正开始涌现出第一批能够被严密量化的上市公司公开业绩案例，清晰印证了 AI 对企业损益表的积极冲击。在降本方面，数字银行 Chime 在被投企业 Decagon 的技术支持下，连续四年将单客服务成本（Cost to Serve）以每年超过 10% 的复合速度持续压缩，累计降幅接近 50%。单客服务成本的断崖式下降为企业提供了更充足的竞争腾挪空间，使其不仅能为客户提供更具竞争力的定价和更全面的服务，还能将省下来的利润再次注回业务扩张。 在增收与核心业务驱动方面，Shopify 推出的 AI 助手 Sidekick 显著缩短了新商家的冷启动周期：商家入驻后 15 天内达成 5 笔订单的转化率提升了 8%，而“5 单”正是 Shopify 内部判定商家是否会长期留存的关键生命周期指标。与此同时，传统软件巨头并未坐以待毙，ServiceNow 不仅斩获了超过 10 亿美元的 AI 年度合同价值（ACV），其 Agent 部署量更激增了 9 倍。 合伙人们随之抛出了一个与 Stripe 等前沿公司深入探讨过的终极试金石问题：企业正在将增量 AI 投资资金引向何方？ - **增收（Revenue Growth） vs 降本（Cost Efficiency）**：增加收入的想象空间在理论上是上不封顶的，而节约成本虽然重要，却是一项始终存在的滞后机会。如果一家公司的管理层认为今天资本回报率最高的用途仅仅是微调内部成本结构，这本身就反衬出其产品在收入增长空间上的窘迫。 - **避免内部过度工程化**：降本也分高下，降低单客履约成本能直接转化为竞争壁垒；但半年前不少硅谷团队耗费数名核心工程师的精力，去内部重构一套系统记录（System of Record）仅仅为了省下区区几十万美元，在当前技术浪潮下无疑是对工程资源的巨大浪费。 ### [25:51](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1551s) Agent 爆发与推理成本断崖：智能路由与模型微调的双重红利 ServiceNow 展现出的 Agent 规模只是冰山一角。伴随自主 Agent 开始接管复杂的端到端任务，单个任务往往需要拆解为多步骤、调用多次底层模型，这直接推动了 OpenRouter 上 Agent Token 消耗量实现了惊人的 14 倍爆炸式增长。然而多步调用的经济性正被快速下沉的技术红利所拯救：通过 Prompt 缓存（Prompt Caching）技术，系统能够直接复用背景信息与上下文，无需从头重算。例如 Hebia 在其金融文档分析与聊天场景中，借助缓存技术将运行成本骤降了 10 倍，使得相同预算下能够支撑十倍于以往的任务吞吐。推理成本与延迟的大幅下降，使得 Agent 自主进行自我校验、试错反思（Try-Check-Retry）在商业上成为可能。 ```mermaid flowchart TD accTitle: 企业级 AI 成本与延迟优化策略 accDescr: 展现智能路由与垂直微调如何降低推理成本并保障任务 SLA。 subgraph TaskRouter["策略一：智能路由 (以 Databricks 为例)"] Req["用户任务输入"] --> Router["Databricks 智能路由器"] Router -->|"简单任务"| SmallM["高性价比模型"] Router -->|"复杂逻辑"| FrontierM["顶级前沿模型"] Router --> Out1["解决率提升，成本降低 35%"] end subgraph FineTune["策略二：专用小模型微调 (以 Elise AI / Harvey 为例)"] Domain["垂直领域数据"] --> FT["专用轻量模型微调"] FT --> Perf["成本降低 60% / 超低延迟"] Perf --> Out2["实时语音交互落地"] end ``` 在工程落地层面，企业正在系统性地探索性能、延迟与开销的平衡点： - **模型智能路由（Smart Routing）**：Databricks 开发的智能路由机制能够针对具体任务的难易程度分发给最合适的模型。实测显示，该路由器在解决问题成功率超越单一大模型的同时，整体推理成本降低了 35%。 - **垂直领域专有小模型微调（Fine-Tuning）**：Harvey 与 Elise AI 展现了小模型路线的巨大威力。Elise AI 通过微调轻量化专用模型，不仅让推理成本直降 60%，更实现了极低的交互延迟，让对响应时间要求极严苛的实时语音交互场景完全具备了商业可行性。 归根结底，对于 AI 应用层企业而言，衡量商业模型的最核心单位变成了“为客户彻底搞定一项工作的全包成本（Cost of Getting the Job Done）”。当应用软件能够以工作完成的结果向客户计费，同时在后端利用智能模型路由和轻量微调极度压低执行成本，软件公司的单位经济模型将迎来本质性的蜕变。 ### [28:40](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1720s) 消费级 AI 的初期渗透与“微笑”留存曲线 谈及企业端 AI 部署与应用软件毛利率的改善之后，讨论自然延伸到了消费级市场的现状。尽管消费级市场历来诞生了体量最为庞大的超级应用，也是投资人重度倾注精力的领域，但目前的渗透率依然处于极早期阶段。最新调研数据显示，全美仅有略高于 2% 的家庭真正为 AI 付费订阅。虽然长期来看，订阅制未必是主流家庭的最佳变现模式，但就当前而言，订阅收入展现出了消费软件历史上罕见的韧性——它的留存不仅能走平，甚至会随着产品与模型的持续迭代演化成昂头向上的“微笑曲线（Smile Curve）”，用户流失后又重新回流，这极为直观地证明了产品所交付的真实价值。 然而，如果横向对比传统消费级巨头的订阅量，现有 AI 订阅的基数小得惊人。以美国家庭规模对比，Amazon Prime 拥有超过 2 亿家庭订户，Netflix 也有约 7000 万家庭用户，AI 的付费渗透才刚刚拉开序幕。正如 a16z 合伙人 Josh Elman 所界定的那样，“消费级 AI 是你日常生活中频繁使用的工具，而不一定是你拿去报销的东西”，因此高达 97% 使用 AI 的家庭目前尚未掏钱付费完全在情理之中。 与此同时，传统的消费级平台评估维度正在失效。过去衡量 Facebook、Instagram、TikTok 或 Snap 等巨头的核心指标是用户使用时长（日均 30 到 60 分钟），而新一代具备持久性、全天候主动响应的 AI Agent 助理，其目标恰恰是替用户在后台把事情办妥，而不是争夺前台屏幕时间。投资团队调侃道，如果用户在开着特斯拉去旧金山的路上随手抛出一个“深度研究（Deep Research）”任务让 Agent 在后台默默运行，传统依赖“外部屏幕使用时长”的分析框架就彻底失效了，这无疑让风险投资人的尽调工作变得更具挑战性。 - **全美付费渗透率**：仅略超 2% 的美国家庭目前拥有 AI 付费订阅。 - **成熟订阅体量参照**：Amazon Prime 拥有超 2 亿家庭订户；Netflix 在美国家庭用户达 7000 万。 - **用户行为模式变迁**：传统社交媒体占据每日 30 至 60 分钟高粘性前台时长，而 Agent 转向“后台无感运行”，前台屏幕时长不再是衡量参与度的唯一有效指标。 --- ### [31:34](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1894s) Agent 冲击流量生态：亚马逊的筑墙与 Instacart 的拥抱 以 Muse、Instinct 及 ChatGPT 为代表的新一代 AI 工具正迅猛抢占传统搜索的查询份额，迫使所有传统消费级发现平台与交易市场（Marketplaces）重新审视自己的防御与进攻策略。摆在每个平台面前的有两个核心考量：第一，AI Agent 究竟能为平台带来多少增量订单需求？第二，自身的利润池中到底有多大比例依赖于掌控“客户发现流程与用户关系”？ 上周行业内上演了截然相反的两幕决策：亚马逊对 Muse 坚决说“不”，选择筑墙拦截；而 Instacart 则大方敞开大门，张开双臂欢迎对接。这一分歧背后的经济学逻辑极为直观：除去云计算 AWS，亚马逊与 Instacart 的广告收入规模均超过了其核心电商业务的全部营业利润。对年广告收入超过 700 亿美元的亚马逊而言，广告依赖的是海量用户在站内反复浏览、点击赞助广告位的超高毛利现金流，而接入 Muse 几乎无法为其带来有意义的增量新客户或新订单；但对 Instacart 来说，生鲜杂货的线上化渗透率尚处早期，借助 Agent 消除摩擦、抢占增量订单与 GMV（商品交易总额）的吸引力，远大于短期内被削弱广告入口的风险。 ```mermaid flowchart TD accTitle: 平台应对 AI Agent 的权衡决策 accDescr: 对比高广告依赖成熟巨头与高增长渗透型平台对 Agent 流量的不同考量与策略 subgraph Evaluation["核心评估维度"] Q1["增量订单/GMV 增长潜力"] Q2["广告业务与流量入口依赖度"] end subgraph Strategies["平台策略分化"] AMZN["亚马逊 (Amazon)<br/>• 广告收入 >700 亿美元<br/>• 增量用户空间极小<br/>• 结果：拦截 Agent (拒绝 Muse)"] INST["Instacart<br/>• 生鲜电商渗透率尚早<br/>• 消除点击转化摩擦<br/>• 结果：开放接入 (欢迎 Muse)"] end Q1 --> AMZN Q2 --> AMZN Q1 --> INST Q2 --> INST ``` 关于 Agent 对商业生态的长期影响，市场上存在“零和博弈”与“正向飞轮”两种争论。悲观论调认为 Agent 将吞噬点击，摧毁平台高毛利的广告利润池；但更乐观的宏观视角指出，过去许多因为繁琐对比和点击链条而放弃的消费意图——无论是想要预订的度假行程，还是全家今晚的晚餐订单——都将被低门槛的 Agent 顺畅完成，从而带动整体 GMV 与全社会消费的扩张。广告主投入的 700 亿美元并不会凭空消失，而是会迁移至更高效的 Agent 交互载体或精准匹配通道。尽管利润池在微观层面会面临重组，但宏观生产力提升将推动总体经济效益扩大；资本市场的即时反应也验证了这一点：Meta 股价上涨带来的市值增量，已经远远覆盖了电商交易市场遭受的短期估值折损。 - **广告利润核心比重**：亚马逊广告业务年营收超 700 亿美元，且属于极高毛利率的流转收益。 - **单用户广告变现水平**：在欧美成熟市场，Meta 与 Google 从每位活跃用户身上获取的年化广告收入均已突破 200 美元。 - **高变现意图搜索防线**：Google 搜索业务此前展现出极强韧性，原因在于类似“购买保险”、“预订城市酒店”等高商业变现意图的关键词较难被早期对话式 AI 直接完成履约闭环；一旦 Agent 全面具备代客行动与履约能力，传统搜索商业模式将迎来实质性冲击。 --- ### [36:14](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2174s) 软件市场的结构性分化：SaaS 大繁荣的终结与防御战 话题回到软件生态时，团队将视线转向公开市场。当前美股公开上市的软件公司正在发生深刻的结构性逆转：整个软件板块的资产组合已全面向“低增长、高盈利”倾斜。在被追踪的样本企业中，约 75% 的上市软件公司已实现盈利，但维持 20% 以上年化收入增速的公司仅剩下 30%。更令人震惊的是，如果将增速基准线拉高到 30%，公开市场中符合条件的企业甚至不足 5 家；与之一级私募市场中几乎每家被投企业都以远超 30% 速度狂飙的景象形成了鲜明割裂。 这种增长骤降背后的驱动力在企业一线 IT 采购决策中显而易见。企业首席信息官（CIO）和 IT 负责人的年度预算盘子是有限的，在生成式 AI 浪潮下，为了给昂贵的 AI 试点与算力采购挪出空间，最顺理成章的做法就是冻结增量预算，不再立项采购新的传统 SaaS 工具或增添新席位。 面对市场上沸沸扬扬的“SaaS 末日浩劫（SaaS Apocalypse）”论调，现存 SaaS 软件公司与创始人的重心聚焦于如何守住自身坚固的现有分发渠道。这些老牌软件企业掌握着客户核心工作流与极高壁垒的“锁定客群”，其当务之急是将具有高度吸引力的新一代 AI 产品嵌套进成熟的分发体系中，以此拉动客单价和二次增长。团队认为，要彻底驱散市场对软件商业模式崩塌的深层恐惧，公开市场的软件巨头必须在接下来的财报周期中持续拿出 98% 净金额留存率（Gross Dollar Retention）这类硬核防守数据，向市场证明其客户基盘的不可替代性。 - **上市公司盈利与增速倒挂**：约 75% 的样本上市软件公司处于盈利状态，但仅有 30% 的企业营收增速达到 20% 及以上。 - **超高成长软件股极度稀缺**：公开市场中年增速超过 30% 的纯软件公司已缩减至 5 家以内。 - **企业预算挤出效应**：企业对 AI 项目的预算扩张并非来源于增量 IT 资金注入，而是直接通过压缩传统 SaaS 新项目与席位预算进行挪移。 ### [37:58](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2278s) 软件公司的两条出路与 10%+ 增长再加速目标 对话延续了对软件公司投资逻辑的讨论。投资者过去青睐软件资产的关键在于其极高的资本效率和防御性壁垒，但在当前环境下，单纯依赖提效已经远远不够，关键在于能否实现“营收增速的再加速”（revenue growth acceleration）。这不仅证明企业具备足够的防御韧性，更说明它们在进攻性方向上的投资正在切实转化成商业成果。 David George 提及他在数月前撰写的《软件只剩两条路》（There Are Only Two Paths Left for Software）博客文章。他指出，在与众多创始人交流后发现，几乎每个人都将驱动营收增长置于首要位置。这与一年前的市场恐慌形成了鲜明反差——当时人们普遍担心所有人都会用 AI“随性编码”（vibe code）从而颠覆现有软件系统，但实际现实并未朝这一方向发展。不过，软件企业面临的业绩压力依然沉重。David 提出以“10% 以上的增长再加速”作为参考基准：尽管门槛极高，但鉴于当前流向 AI 的庞大企业预算，如果能打造出真正具有魔力的产品，这一目标完全可行，预计未来 12 到 18 个月内就能在部分密切跟踪的公司中看到这种拐点。 ### [39:18](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2358s) 公开市场与一级市场的 SaaS 严重分化 Sarah Wang 指出，不能将整个软件行业一概而论为估值倍数下调或增长停滞，市场实际上呈现出剧烈的分化态势。根据演示数据，网络安全（Cybersecurity）和系统可观测性（Observability）表现尤为突出，垂直软件（Vertical Software）的整体抗跌能力也远强于通用横向应用（Horizontal Applications）。判断软件类别能否坚挺甚至走强的核心分析框架，在于审视 **AI 究竟如何重构客户对该产品的底层需求**： ```mermaid flowchart TD accTitle: 软件受 AI 冲击的分化逻辑 accDescr: 对比网络安全与垂直软件等受益板块与通用横向应用软件受到的冲击 subgraph Winners["受 AI 顺风推动的板块"] A["AI 普及与 Agent 大量调用系统"] --> B["产生全新安全防护与监控需求"] B --> C["网络安全与可观测性（如 CrowdStrike）需求激增"] D["垂直领域专有业务工作流"] --> E["深入行业特定编排与交互"] E --> F["垂直 AI 软件（如 Harvey、Abridge）超速成长"] end subgraph Losers["面临变革压力的板块"] G["通用横向应用软件"] --> H["遭遇 AI 工作流重构"] H --> I["被推上 AI 的'断头台'（Chopping Block）"] end ``` 随着 AI Agent 访问更多系统并执行自动化操作，安全与监控需求水涨船高，不仅利好 CrowdStrike 这类传统网络安全在位巨头（近期 OpenAI 与 Hugging Face 相关的安全事件更是给全行业敲响了警钟），也推动了细分垂直 AI 公司的爆发。Alex Immerman 补充道，在私募市场中，Harvey（法律）、Abridge（医疗）、EliseAI（物业管理）等垂直 AI 公司的增长速度刷新了所属行业的所有历史先例。这证明客户越来越意识到，核心竞争力绝不仅仅是底层模型，而是如何将模型有效编排进特定业务流程之中。 Alex 和 Santiago Rodriguez 进一步总结了公开市场全年的演变轨迹：年初市场普遍弥漫着“SaaS 启示录”（SaaS Apocalypse）的恐慌情绪，估值遭遇剧烈抛售；但截至目前，软件指数已基本修复至年初点位。然而指数持平的表象下隐藏着赢家与输家的巨大断层——那些被视作赢家的公司不仅优化了自身成本结构，更切实抢占了 AI 带来的全新增量预算。 ### [42:23](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2543s) Stripe 的“复兴”数据与超级未上市巨头的崛起 Sarah 引入了来自一级市场运营的真实数据，指出 Stripe 的 SaaS 客户数据打破了市场悲观的刻板叙事：数据显示，无论是初创企业还是成熟企业，其营收增速在迈向 2026 年时均在重新加速。Stripe 将这一趋势形象地命名为“复兴”（Renaissance），这一强劲反弹有力驳斥了软件业终局将至的论调。 与此同时，创业公司选择在更长时间内保持私有化（staying private for longer）的趋势愈发显著。Alex 指出，当前全球估值最高的六家私有科技企业——Anthropic、OpenAI、Databricks、Stripe、Waymo、Revolut——在最近一轮估值中总和已达约 2.4 万亿美元。这一数字令人惊叹： - 过去 10 年美股所有 IPO 公司（剔除 SpaceX 后）的当前总市值约为 1.7 万亿美元，这 6 家超级未上市公司市值总和已远超该规模。 - 该体量已逼近聚集了数百名基金经理日常重仓的罗素 2000 指数（Russell 2000）总规模。 这意味着最具爆发力的后期增长红利大量保留在了一级市场内部，使这些龙头能够不受短期季度业绩掣肘，持续向长期增长下注。 ### [43:48](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2628s) IPO 决策权衡与二级市场员工要约回购 David George 从投资人视角阐述了他们如何辅导被投企业权衡 IPO 时机。他提出一个核心观点：“在当今阶段，创始人本身就是一种独立资产类别（the founder is the asset class）”。未上市状态的最大红利在于创始人可以承受更长的投资回报周期，敢于进行更大胆的战略押注（在公开市场上，唯有扎克伯格或马斯克等极少数人拥有这种任性）。例如 Databricks 和 Stripe 能够在一级市场环境下持续砸重金布局全新产品线，并由此驱动了新的营收加速；而在公开市场，这种大笔资本开支极易引致公开质疑——正如 Meta 曾因大举投入 AR/VR 导致股价跌破 100 美元低谷一样。 在团队看来，IPO 归根到底只是一种“融资事件”，必须综合权衡其战略利弊： - **上市的收益**：像 Navan（原 TripActions）上市后，公开披露的透明财务数据显著增强了大型企业和消费者的信任度，反而推动了业务的二次加速；同时公开市场在面对动辄不可思议的超大规模资本需求时，依然具备一级市场无法匹敌的资金承载力。 - **留在一级市场的机制**：对于选择保持私有化的企业，流动性问题主要通过员工要约收购（Tender Offers）解决。Carta 平台公布的数据显示，在员工要约收购中，实际参与变现的比例仅为 58%，意味着有超过四成的员工主动放弃变现。这充分说明员工对公司未来发展抱有极强的信心，直接推翻了外界所谓的“员工急于套现跑路”的假想。最后大家还开玩笑感叹：“如果这帮员工当初 100% 全套现了，旧金山湾区的房价不知又得被推高成什么样。” ### [46:35](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2795s) 一级半与二级市场常态化：员工流动性与估值重置 二级市场交易在当下的非上市公司中变得越来越高频，且已经成为一级市场不可或缺的良性机制。这一操作主要服务于两大核心目的： 1. **与上市公司争夺顶尖人才**：对于员工和候选人而言，非上市公司往往很难匹敌公开市场股票的变现诱惑——上市公司每个季度在扣税后准时到账的 RSU（受限股票单位）具有极强的确定性。因此，提供定期的二级流动性，是非上市公司在人才保留与招聘上对抗上市巨头的关键武器。 2. **及时重置并刷新公司估值**：更频繁地在市场上确认股权价值，能使非公开市场的股票定价保持“新鲜”。这不仅让公司在与员工及潜在雇员沟通薪酬方案时更具说服力，也为企业在发起并购（M&A）时提供了具备公信力且最新定价的“股权货币”。 从折价率的走势来看，过去五年二级市场的动态发生了根本性逆转。在 2021 年之后的一段时间（2022 至 2024 年），二级市场相较于上一轮融资估值的中位折价率非常显著，主要原因是那一时期许多早期融资轮次定价过高且已严重过时。然而在当前环境下，二级市场交易对上一轮价格的折价率几乎已经收窄至零。这表明当下一级公司的融资估值非常贴近实际市场，且源源不断有全新投资者愿意以同等定价入场接盘。 --- ### [48:00](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2880s) “万物皆 AI”：86% 风险投资流向与全产业链扩散 根据 2026 年初的最新市场数据切片，AI 相关公司在美国风投交易活动中的占比已经高达 86%，较 2025 年的 65% 实现了进一步的跃升。毫无疑问，AI 已成为全球风险投资最核心的目的地。 ```mermaid flowchart TD accTitle: AI投资版图扩散与辐射领域 accDescr: 展示风险投资不仅流向基础模型层，还广泛渗透到底层物理设施与顶层应用服务。 subgraph Core["底层大模型核心"] M1["OpenAI"] M2["SpaceX / xAI"] M3["Anthropic"] end subgraph Infra["硬件与物理基础设施"] I1["半导体 (Semiconductors)"] I2["电力与能源 (Power)"] I3["国防军工 (Defense)"] end subgraph Apps["场景落地与服务"] A1["企业级应用 (Enterprise Apps)"] A2["消费级应用 (Consumer Apps)"] A3["专业服务 (Services)"] end Core --> Infra Core --> Apps ``` 虽然公众视野和媒体聚光灯主要聚焦在头部前沿实验室（如 OpenAI、Anthropic、xAI 等），但在“AI”这个统一的标签之下，实际的投资机会已经全面泛化。从半导体、电力能源、军工国防，到企业级软件、消费级应用以及外包服务，投资团队所面对的投资机会纵深与广度达到了前所未有的水平。 --- ### [48:43](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2923s) 未来投资罗盘：机器人、全自动驾驶、生物与企业落地 结合 AI 浪潮的推演，团队详细梳理了未来 3 到 10 年内最值得重注的几大核心前沿赛道： * **长程自主 Agent 与消费级应用（Long-running / Autonomous Agents）**：消费者代理能够接管繁琐无趣的杂务，为个人创造以往不愿投入时间或金钱才能获得的价值，这类应用的扩散渗透速度将远超预期。 * **具身智能与机器人（Robotics）**：该领域的潜在市场规模未来甚至可能超越 LLM（大型语言模型）。尽管其发展阶段可能比当前 LLM 晚 3 到 5 年，但接下来的 5 年将迎来爆发式的资金涌入与突破。 * **完全自主驾驶（Autonomous Driving）**：自动驾驶时代已经真正到来，技术已被验证成熟。出行与汽车是全球最大的产业之一，目前 Uber 和 Lyft 仅占全美总行驶里程的约 1%；随着比人类驾驶安全 14 倍的自动驾驶网络铺开，这一里程渗透率至少将提升一个数量级。此外，全美每年新售汽车约 1700 万辆，预计未来 10 年内这些新车都将全面迈向自主化。 * **AI + 生物医疗（AI × Bio）**：实验室前沿正在被大模型重塑，从全新靶点与药物发现（Drug Discovery）到攻克人类最顽固的绝症，这一跨界方向承载着极高的现实期望。 * **个性化健康管理（Personal Health）**：作为“健康狂热者（Health Maxer）”的一大痛点，长期以来一直缺乏一个能够汇总个人所有身体指征数据、并输出超个性化医疗/健康建议的智能中枢。 * **AI 在企业端纵深渗透（Enterprise Diffusion Beyond Coding）**：目前除了辅助编程之外，AI 技术真正在传统企业内部的业务级渗透依然处于极早期，未来的普及空间巨大。 * **美国活力体系重构（American Dynamism）**：在纯 AI 之外，重构国防与硬科技供应链也是极其广阔的领域。目前诸如 Anduril、Saronic、Castelion 等新兴防务承包商在全美整体军事支出中的占比尚不足 5%，随着全球局势演变与防务需求切换，新一代科技供应商的份额将迎来爆发式增长。 纵观整体宏观环境，尽管当前阶段围绕算力与物理设施的基建支出（Buildout）堪称天文数字，但其对全美经济与生产力的长远拉动将是极其深远的。 ## Notable quotes > [02:56](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=176s) Tech is the everything cycle — David George > [02:56](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=176s) Software is eating the world — Alex Immerman > [13:15](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=795s) The factory is the product — Elon Musk > [28:40](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=1720s) 消费级 AI 是你日常生活中频繁使用的工具，而不一定是你拿去报销的东西 — Josh Elman > [37:58](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2278s) There are only two paths left for software — David George > [43:48](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2628s) the founder is the asset class — David George > [43:48](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2628s) 如果这帮员工当初 100% 全套现了，旧金山湾区的房价不知又得被推高成什么样。 — David George > [36:14](https://www.youtube.com/watch?v=lr3hNhA0IfQ&t=2174s) SaaS Apocalypse ## Mentioned David George, Sarah Wang, Alex Immerman, Santiago Rodriguez, Marc Andreessen, Sam Altman, Sarah Friar, Elon Musk, Ali Ghodsi, Josh Elman, OpenAI, Anthropic, Microsoft, Alphabet, Google, Amazon, Meta, Oracle, Anduril, Waymo, SpaceX, Databricks, Revolut, ElevenLabs, Decagon, Chime, Shopify, ServiceNow, Stripe, Hebia, EliseAI, Harvey, Abridge, CrowdStrike, Hugging Face, Instacart, Muse, Carta, Saronic, Castelion, Uber, Lyft, Netflix, TikTok, Instagram, Facebook, Snap, Navan, ChatGPT, Microsoft Copilot, Cursor, Cognition (Devin), Shopify Sidekick, OpenRouter, Yepit ## Video description a16z’s David George, Sarah Wang, Alex Immerman, and Santiago Rodriguez unpack 25 key charts from the latest State of Markets presentation, from the scale of the AI infrastructure buildout to what adoption looks like inside companies today. They examine why rising markets have so far been supported by earnings rather than multiple expansion, why hyperscaler CapEx is approaching $1 trillion annually, and why demand for compute continues to outrun supply. They also look at the downstream effects of that spending across chips, power, construction, and physical infrastructure. State of Markets Then they move up the stack: OpenAI and Anthropic’s revenue growth, the gap between AI deployment and measurable enterprise impact, the rise of agents, falling inference costs, and what all of this means for SaaS. They close with where the team is spending time next, including consumer agents, robotics, autonomy, AI and biology, personal health, defense, and the continued diffusion of AI across the enterprise. State of Markets Timestamps: 00:00 - Intro 00:59 - Is it a bubble? 06:28 - The $780B hyperscaler CapEx race 13:15 - The new age of atoms 18:21 - Only 2% of enterprise AI is truly tracked 20:28 - Power users spend 20x the median 31:34 - Amazon blocks Muse, Instacart opens up 36:14 - The SaaS bifurcation 42:23 - Stripe's Renaissance data 48:43 - Robotics, autonomy, bio: what's next Resources: Follow David George on X: https://x.com/DavidGeorge83 Follow Sarah Wang on X: https://x.com/sarahdingwang Follow Alex Immerman on X: https://x.com/aleximm Follow Santiago Rodriguez on X: https://x.com/santiago__rdz Read David’s piece ‘There are only two paths left for software’: https://a16z.com/there-are-only-two-paths-left-for-software/ 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=lr3hNhA0IfQ)

[2:56] David George: Tech is the everything cycle [2:56] Alex Immerman: Software is eating the world [13:15] Elon Musk: The factory is the product [28:40] Josh Elman: 消费级 AI 是你日常生活中频繁使用的工具，而不一定是你拿去报销的东西 [37:58] David George: There are only two paths left for software [43:48] David George: the founder is the asset class [43:48] David George: 如果这帮员工当初 100% 全套现了，旧金山湾区的房价不知又得被推高成什么样。 [36:14] SaaS Apocalypse

[2:56] Alex Immerman: software is eating the world [4:57] David George: That's all great, but is it a bubble? [6:28] Santiago Rodriguez: model buster [6:28] Santiago Rodriguez: reckless and aggressive [12:12] Santiago Rodriguez: the factory is the product [24:35] Alex Immerman: the opportunity to drive revenue growth is unbounded to the upside [28:40] Alex Immerman: smiling [37:58] David George: There are only two paths left for software [37:58] David George: vibe coding [39:18] Ali Ghodsi: chopping block of AI [42:23] David George: Renaissance [44:48] David George: if that was 100% [46:35] Alex Immerman: weapon of competition [48:43] David George: health maxers

[0:00] Eight of the top 10 valued companies in [0:02] the world are US tech companies. [0:04] >> Since JBT came out almost 4 years ago, [0:06] the market's up 90% which is 17% [0:09] annualized. The natural instinct is well [0:11] that's got to come down. We're [0:12] definitely in a hot period. [0:13] >> This puts us in a new age of atoms. [0:15] Global infrastructure investment needs [0:17] are estimated at $90 trillion through [0:20] 2040. This goes way beyond AI and data [0:24] centers. It includes power, water, [0:26] roads, transit, [0:28] >> live deployments at S&P 500 companies, [0:31] that's at 69%. Now, if you go to the [0:33] ultimate barometer, which is a metric [0:35] tracked over time, that's actually only [0:37] at 2%. AI is generating major revenue [0:40] and savings. On the other hand, adoption [0:42] is still extremely early. [0:43] >> Today's opportunity is so much around [0:46] taking these capabilities and harnessing [0:49] them to build reliable services. That's [0:51] all great, but is it a bubble? Um [0:54] >> um um. [0:58] >> Welcome back to the A16Z podcast. I'm [1:01] David George. I'm here with my [1:03] colleagues Sarah Wang, Alex Emerman, and [1:05] Santiago Rodriguez. Today we are walking [1:09] through 25 key slides from our latest [1:12] state of markets presentation. the [1:14] earnings behind the market's rise, the [1:15] scale of the AI buildout, the evidence [1:18] of growing adoption, and what this cycle [1:20] means for hardware, software, and the [1:22] next generation of private companies. [1:24] We'll explain what the charts show and [1:26] discuss what we're seeing inside [1:28] businesses along the way. So, you can [1:30] follow along whether you're watching or [1:32] listening. So, Sarah, Alex, Santi, [1:35] thanks for joining me. [1:36] >> Of course, [1:36] >> the day this podcast goes live, we will [1:39] be releasing our state of markets [1:41] presentation. And so this is a now [1:43] yearly tradition from our growth team [1:45] where we synthesize the biggest trends [1:47] in tech, AI, infra and markets. Uh today [1:50] we're just picking out a subset of [1:52] interesting slides and having a [1:53] discussion about them. So I would direct [1:55] you to uh look at the whole version [1:57] which has which is filled with a lot of [1:59] nuggets. Um the areas that we're going [2:02] to discuss today are um macro so where [2:06] we'll talk about capex data centers [2:09] accelerating demand at a high level and [2:11] then on the ground takeaways where we [2:13] will discuss models apps uh and some [2:16] vertical deep dives. Technology is [2:18] driving an economywide investment boom [2:21] with rising earnings supporting market [2:22] gains and AI demand pushing [2:24] infrastructure spending far beyond [2:25] earlier forecasts. Hyperscalers are [2:28] investing all of their near-term [2:29] operating cash flow to build out this [2:31] capacity for demand that continues to [2:33] outstrip supply in almost every case [2:35] that we see. They are channeling this [2:37] investment into chips, power, cooling, [2:39] construction, skilled labor, and much [2:42] more. This expansion coincides with a [2:44] broader need to modernize [2:46] physical infrastructure, creating [2:49] opportunities across industries and [2:50] potentially lowering shared costs for [2:52] businesses and households. So with that, [2:56] let's jump in. First slide, tech is the [2:59] everything cycle. So um this some of [3:02] this may be a little bit obvious, but um [3:04] the numbers are are somewhat striking at [3:06] this point, right? So high-tech [3:08] equipment, software, and R&D now account [3:10] for roughly 55% of US capital spending, [3:14] which is just a staggering number. Um [3:16] tech is driving the investment cycle [3:18] obviously across all areas of the [3:19] economy. uh from software and models to [3:22] power construction and industrial [3:24] capacity. Tech is almost 40% of the [3:28] aggregate value of the whole stock [3:29] market in the US. Um and eight of the [3:33] top 10 valued companies in the world are [3:36] US tech companies. So this is a broad [3:39] story. Uh you're seeing it in capex. [3:41] This is heavily covered in in the data [3:43] center buildout. Um it's obviously [3:44] heavily covered in the amount of capital [3:46] that's been going into the to fund the [3:48] model development models model companies [3:50] together raising I think over $350 [3:52] billion um at this point. Um and just to [3:56] put it into into very deep historical [3:58] context because this is probably the [3:59] analogy that we've seen the most of uh [4:01] this buildout just surpassed railroads [4:04] as a percentage of GDP. Um, so you know, [4:07] exciting times, massive buildout. Uh, we [4:11] all happen to think that if you fast [4:13] forward five to seven, maybe 10 years [4:16] from now, um, you know, we'll be looking [4:18] at these numbers and they'll probably be [4:20] 20x higher cumulatively. So [4:23] >> yeah, this this does feel like the next [4:25] chapter of Mark's software is eating the [4:28] world. You know, over the last 15 years, [4:30] software has transformed all these [4:32] industries. uh but only a small fraction [4:35] of the population could build it. None [4:37] of us could. Uh but today, you know, [4:40] look at us. We all have like, you know, [4:41] a handful of automations running every [4:43] night. And so if you think about the [4:45] software demand, that means a lot more [4:47] compute, chips, power, construction. Um [4:50] and so it's it's no surprise to see the [4:52] majority of investment now now covered [4:54] in tech. So one of the questions that we [4:57] get all the time from various audiences [4:59] is okay that's all great but is it a [5:03] bubble? So the market has reached new [5:05] highs um and at the same time that it [5:07] has reached new highs the trading [5:09] multiples of the market are actually [5:11] down. So stocks are up about 20% while [5:14] multiples are down about 20%. So what [5:17] that means is the performance is is [5:18] driven by fundamental earnings right not [5:20] not increased multiples. um the S&P 500 [5:23] earnings multiple is below 20 times. So [5:27] um you know if you just start with that [5:29] these are many in most cases very high [5:31] quality businesses. It's nothing like [5:33] the dot boom in that way where some of [5:36] the highest market cap companies in the [5:37] world um had their massive stock runups [5:40] based on increases in their trading [5:42] multiples and would would trade in many [5:43] cases for like a 100 times PE. That's [5:45] not what's happening here. Um in [5:48] contrast uh you know some of the memory [5:50] companies which again are very cyclical [5:52] uh are trading you know for call it six [5:55] times seven times forward earnings. So [5:57] very very different [5:58] >> and I think there's an important double [5:59] click here where you know since JVD came [6:01] out almost four years ago the market's [6:03] up 90%. which is 17% annualized. And I [6:06] think anytime there's been a 17% [6:07] annualized growth to four years, the [6:09] natural instinct is well that's got to [6:11] come down, right? Like we're we're [6:12] definitely in a hot period. But when you [6:14] compare that to, as you said, the market [6:16] trading for below 20 times earnings [6:18] growing 15%. This definitely feels a [6:21] little different than maybe the 2021 [6:22] period that was recent or the 2000 [6:24] period when multiples in growth were not [6:26] really going together. [6:28] >> Yeah, totally agreed. So I mentioned the [6:31] scale of the capex buildout um you know [6:34] in the context of of the railroads you [6:37] know if you just look at the [6:39] hyperscalers Alphabet, Amazon, Meta, [6:41] Microsoft and Oracle their capex in 2026 [6:45] is about $780 billion. That's up from [6:48] $416 billion in 2025 and uh all [6:52] expectations point to them spending over [6:54] a trillion dollars annually from 2027. [6:58] Um so you know the pattern recognition [7:00] is is each of these computing platforms [7:02] has supported a much larger population [7:04] of users and uses. In this case um this [7:08] buildout can happen so quickly and [7:10] demand can still outstrip supply because [7:13] the amount of users is driven by the [7:16] fact that there's already existing [7:17] distribution. This trend is built on top [7:19] of the internet and cloud computing [7:20] obviously and mobile phones. Um and so [7:22] you know sort of immediately could reach [7:24] billions of users um you know in [7:26] contrast to previous technology cycles. [7:29] Yeah, and I think it's um an important [7:30] point to double click on just because [7:32] we've moved well beyond this model of [7:34] occasional queries, right? If you think [7:36] about the rise of agents, you have [7:38] parallel tasks, you have long running [7:40] tasks. And so, I mean, Alex, you [7:42] mentioned this previously, right? You [7:43] have tasks going into the evening, [7:45] during the day when you're doing other [7:47] things. Um, and so if you think about [7:49] agents autonomously writing code, [7:51] searching, carrying out these tasks, um, [7:53] it just gives these platforms a much [7:55] larger compute requirement than ever [7:57] thought before. Yeah, absolutely. [7:59] Totally agree. Um, this one is another [8:02] one that just points to what is [8:04] happening happening in the capex side of [8:06] things. Successive forecasts for the [8:09] five largest hyperscalers capex have [8:12] moved sharply higher um in pretty short [8:15] succession. So spending that once looked [8:17] like a ceiling, you know, a number of [8:19] quarters or years out has become [8:21] near-term um as the demand for the [8:23] compute keeps keeps expanding. Yeah, [8:25] look, if you look at the chart here, I [8:27] think at any point in time, I think [8:29] natural instinct is just to say like the [8:31] investment will kind of flatline from [8:32] here. Yet for the last four years, like [8:36] as an economy, we keep underestimate [8:38] just the the strength of the trend. And [8:40] as Sarah said, you know, with model [8:42] developments, usage of of our installed [8:44] capacity keeps keeps being taken up. So [8:47] right now, the numbers that David [8:48] pointed to or the current estimates. [8:50] >> Yeah. I mean, maybe to use DG's [8:52] language, I think we'd all call the [8:54] demand for compute a model buster at [8:56] this point. Um, I think one anecdote [8:59] that's telling that that we've all [9:01] experienced is, you know, Sam Alman, [9:03] Sarah Frier, they got a lot of flak, you [9:06] know, a year or so ago for their massive [9:08] compute uh commitments. Uh, they were [9:10] being reckless and aggressive. And I [9:12] think at this point, like everyone would [9:13] say they're incredibly preient with that [9:15] decision. Um, and even so, two weeks [9:18] ago, we all saw that they had to pause [9:20] new subscriptions on their pro plans. [9:22] Like showing up with a $2,000, you know, [9:25] uh, you know, service. No, no thanks. [9:28] Um, pretty amazing, uh, insatiable [9:30] demand that we're experiencing here. [9:32] >> Yeah. Absolutely. Yeah. Pretty much [9:34] every, you know, everyone we talk to at [9:37] every stage of the supply chain uh, is [9:39] telling us the same thing, some version [9:41] of the same thing that demand, you know, [9:43] outstrips supply. uh there are certain [9:45] elements in the data center supply chain [9:48] um you know where you can't get access [9:50] to materials or products until 2028 um [9:54] and so this is not softened so um you [9:57] know I I would say at the same time you [9:59] can look to the hyperscalers and see [10:02] some evidence of highquality you know [10:04] business on the demand side that you can [10:06] hang your hat on right so Microsoft [10:08] Google and Amazon have uh about 1.7 [10:11] trillion of combined cloud backlog [10:14] together. Um and those customer [10:16] commitments are are building rapidly [10:18] while the platforms invest heavily in [10:19] the capacity to serve them. So um while [10:22] right now free cash flow is depressed [10:24] during this buildout um you know [10:26] consensus forecasts are showing a [10:27] recovery from 2028 and substantial [10:29] growth thereafter. Um, you know, you [10:32] could look to Amazon's actually Amazon's [10:34] latest earnings call where they did a [10:36] really good job of explaining this sort [10:38] of J-Curve dynamic where the useful life [10:40] of GPUs is actually pretty or TPUs is [10:44] actually pretty long. Um, and so, you [10:46] know, you have to build out the shell, [10:47] the data center, you know, that's a [10:49] certain amount of time. You have to buy [10:50] the the chips. Um, but you know, those [10:52] will have a very useful economic life [10:54] for a long period of time. And it's and [10:56] it's been longer than I think any of us [10:58] expected. [10:59] >> Yeah. Yeah, and I know some of the [11:00] hyperscalers have frankly gotten dinged [11:02] for raising uh debt for capex, but um [11:04] you know, similar to the model lab [11:06] dynamic, I think the ones who have [11:08] blinked and been less aggressive have [11:09] regretted it. Um I know on the podcast [11:11] recently that you did with Gavin, uh [11:13] Microsoft came up. Um but I think this [11:16] is uh you know an issue across the [11:17] board. [11:17] >> Yeah. The other dynamic that's been [11:18] spoken about a lot is pricing on the [11:21] spot markets for existing GPUs which is [11:23] just another signal that like any GPU [11:25] that you can bring online is being [11:27] priced at an attractive rate where like [11:28] the hyperscalers earning an attractive [11:30] return. Yeah, it's a key point. Um, you [11:32] know, this is one of the things that we [11:33] talk about all the time is like each one [11:35] of these successive waves just creates a [11:37] tremendous amount of user or consumer [11:39] surplus. And so what's actually [11:42] happening right now, at least as far as [11:43] we can tell, um is, you know, consumers [11:46] and users get a tremendous amount of [11:47] value out of this. Like that's why [11:49] they're using it so much. Um you know, [11:51] the the the vendors who are serving [11:54] those users are making very good money. [11:57] Um and then, you know, you go all the [11:59] way down every level of the stack to the [12:01] chips where, you know, you have to pay [12:03] much higher than you did 12, 18, 24 [12:05] months ago to get access to them. And [12:07] yet you could still make very high [12:08] margins and create a tremendous amount [12:10] of surplus out of the users. [12:12] >> And as we uh you know talk about this [12:14] capex uh for the hyperscalers uh we [12:17] should think of it as someone else's [12:18] order books. So these big platforms with [12:20] historically you know the largest [12:22] profits are pouring their cash back into [12:25] AI infrastructure. Um that's putting [12:27] pressure on their free cash flow [12:28] shortterm as we just talked about. Um [12:30] but it's been a boon uh for chip orders, [12:34] for power, for construction and that's [12:36] why this broader technology boom has [12:38] become an industrial boom and we I mean [12:41] we've been spending more of our time [12:43] looking at businesses serving this [12:44] industrial boom and everywhere along the [12:47] data center supply chain. I think it's [12:48] interesting because you know [12:50] traditionally our world was more about [12:52] monetizing existing IP you know build [12:54] once and then sell infinitely but with [12:56] these businesses there's many more [12:58] complexities that the companies need to [12:59] manage like you know financing managing [13:02] vendor relationships predicting capacity [13:04] and predicting demand and we've seen you [13:07] know there are some teams that excel at [13:08] that and some teams that have struggled [13:10] or kind of taken a pause and it's a [13:12] little bit of a different expertise than [13:13] we typically spend our time with. This [13:15] puts us in a new age of atoms. Uh big [13:19] numbers on this page. Global [13:20] infrastructure investment needs are [13:22] estimated at 90 trillion dollars through [13:25] 2040. Um importantly though, this goes [13:28] way beyond AI and data centers. It [13:31] includes power, water, roads, transit. [13:34] Um we're seeing this across our [13:36] portfolio. You can see this uh at [13:38] Anderoll with their massive [13:40] manufacturing facility. Um it, you know, [13:42] is it's the the scale of it is 8 87 [13:46] football fields. Uh Whimo, uh they're [13:49] you know, aggressively expanding their [13:51] depots. And of course, you know, SpaceX [13:53] with their hundred billion dollar [13:54] investment in Louisiana. Um betting on [13:57] the future today from our perspective [13:59] isn't, you know, what it exactly was in [14:01] the as it was in the past. As Santi [14:03] said, it means building factories, [14:05] expanding infrastructure, uh more [14:08] skilled jobs, uh and across the country [14:10] at large. [14:11] >> And I mean, similar to what we just said [14:13] on data centers, I think investing in [14:16] physical infrastructure is very [14:18] different than like, you know, investing [14:19] in software. And Elon coined the term, [14:22] you know, the factory is the product [14:23] many years ago. It's no surprise that a [14:25] lot of the great founders that we've [14:27] backed have come out of Tesla or SpaceX [14:29] because they just learned that, you [14:31] know, executing on the factory ends up [14:33] being a massive competitive advantage um [14:35] as they scale. So, there are many myths [14:37] that we hear all the time about data [14:40] centers. They're draining all of [14:41] America's water. Rich people don't want [14:44] to live near them. Uh and then, you [14:46] know, one of our favorites, uh my [14:47] electricity bill is going to go through [14:49] the roof. It may seem counterintuitive [14:52] uh but a data center can help lower your [14:55] electricity rate. A recent US study [14:57] showed that for every 10% increase in [15:00] data center capacity uh residential [15:02] rates went down by 40 bips. Um, you [15:05] should think about it as a power grid is [15:08] a shared fixed cost space, poles, wires, [15:13] substations. And so a large stable [15:16] customer like a data center can help [15:18] spread those costs across more units of [15:20] electricity. Simply put, more demand [15:22] across a shared system is a positive. [15:24] >> Yeah. Yeah. I mean, yeah, totally agree [15:26] with what you're saying. Investment in [15:28] shared infrastructure can also improve [15:30] the economics for the households and [15:31] businesses connected to it. [15:32] >> Yeah. And I don't think that point gets [15:34] enough media attention. Um Dina Powell [15:36] went on a podcast recently from from [15:38] Meta and talked about the Louisiana site [15:41] that they did where they worked with the [15:42] community to lower um electricity costs. [15:45] And so there's real examples of that [15:47] today. It's not just lip service. [15:49] >> Totally agreed. [15:50] >> So now we're going to talk about the [15:52] trends that we're seeing at the model [15:54] and application layer. Um on the one [15:56] hand, uh as we've sort of previewed, AI [15:58] is generating major revenue and savings. [16:02] On the other hand, adoption is still [16:04] extremely early. Um, and then, you know, [16:06] I think on top of this, the the trend is [16:08] very much so that costs are plummeting. [16:11] And, you know, we previewed agents [16:13] already, but agents are actually [16:15] economical now for a much wider range of [16:17] work. And of course, this is changing [16:19] even as we speak, right? New innovations [16:21] like Jev by newer labs like Typesafe are [16:24] taking this to a greater extreme. you're [16:26] starting to see things like two orders [16:29] of magnitude uh cost differences really [16:32] impact the number of use cases. Um it's [16:34] sort of classic Jevans paradox which is [16:37] why it's very aptly named. Um and you [16:40] know I'd say all in all this is an [16:41] awesome setup for both the model layer [16:43] and the app layer. Um these improvements [16:45] are creating real opportunities for both [16:48] growth and profitability among software [16:50] companies. And that's not to say that [16:52] everyone will win, but it's no surprise [16:55] that AI is attracting venture dollars [16:57] into an very much expanding range of [17:00] industries. Um, and we're also seeing [17:02] more companies just reach enormous scale [17:04] in the private markets. So, this is not [17:06] a new chart, but it is one of my [17:08] favorite charts, which is the combined [17:11] scale of revenue for OpenAI and [17:13] anthropic. Um, and this is one that we [17:16] we've had sort of in every GP offsite as [17:18] far as I can remember the last few [17:20] years. And we have to keep updating it [17:21] every month because it it just sort of [17:22] gets out of date that quickly. I think [17:24] the point here, you can see in the [17:25] numbers, OpenAI, Anthropic, their [17:27] combined annualized revenue has climbed [17:29] to an extraordinary degree. Um, we [17:33] love to show this relative to the [17:35] greatest software companies in history. [17:37] Um and you can see that in the right [17:38] that um if you look at the estimates for [17:40] revenue added um for the the best [17:43] software companies ever built um the [17:46] estimates for net new for the leading [17:48] labs has s well well surpassed that um [17:50] so it's really a striking combination of [17:52] not only scale but also speed to get [17:54] there [17:55] >> but we're still pretty early like I [17:56] think like relative to other platform [17:58] shifts the interesting thing about the [18:00] AI shift is we're not only early in [18:02] terms of percentage of population or [18:05] percentage of users that use AI But [18:07] still very early on in kind of the share [18:08] of wallet gains within those users. So [18:10] it's a little bit different than like [18:11] you know in 2008 we would model [18:13] percentage of people with an iPhone. [18:15] Here we need to think about percentage [18:17] of people that will use AI products and [18:19] then to what extent they're using the [18:20] products. [18:21] >> Yeah, I think that point that we're [18:23] still so early is such an important one [18:25] and there's a bunch of metrics in here [18:26] that I think elucidate that. Um, one is, [18:29] you know, if you think about live [18:31] deployments at S&P 500 companies, that's [18:35] at 69%. So, you know, if you're AI [18:38] pilled as we all are, u maybe something [18:39] close to what we would all expect. Um, [18:41] but then if you move toward quantifiable [18:43] impact, which is probably a good metric [18:45] of how far along they are in their [18:46] deployment, that's 30%. Um, now if you [18:49] go to the ultimate barometer, which is a [18:51] metric tracked over time, that's [18:53] actually only at 2%. So we believe that [18:57] AI is delivering results but there is a [18:59] huge amount of room to actually deepen [19:02] its use inside the organization and [19:04] track those results over time. Um and so [19:06] moving from these individual deployments [19:08] that you're seeing to really deeply [19:11] influential recurring workflows is is [19:13] the next stage frankly. Um, and uh, I I [19:16] would just say that most the enterprises [19:17] that we speak with anecdotally, their [19:20] exposure to AI is still mostly with [19:22] Microsoft Copilot, which just shows you [19:24] how far they have to go. [19:25] >> Yeah. This gap between what the models [19:28] can do and how they're being used is [19:31] what makes me so excited about the [19:32] application layer. Sarah, you did this [19:34] great, you know, conversation with Ali [19:36] Godsy at data bricks with Martin and he [19:39] talked about how, you know, the AI can [19:41] know a lot about the world but know very [19:43] little about your company. And we're [19:45] seeing this uh in our application layer [19:47] companies where they're bridging this [19:48] gap. Um, and like one example that has [19:51] stuck out to me over the last few months [19:53] is if you were to take Revolute, who has [19:55] a really sophisticated engineering team, [19:58] they still had to partner with 11 Labs, [20:01] uh, and take 11's, you know, [20:03] best-in-class voice model and harness it [20:05] to connect securely with customer [20:08] accounts and banking workflows so that [20:10] when I, as a customer, call, I can have [20:12] my, you know, problem resolved smoothly, [20:15] efficiently, and, you know, securely. um [20:18] today's opportunity is so much around [20:21] taking these capabilities and harnessing [20:24] them to build uh reliable services. [20:28] >> So in addition to the the fact that um [20:31] enterprise adoption is still quite early [20:33] for AI, uh the other really remarkable [20:36] trend that we wanted to um put some [20:38] stats to is just that the power users [20:40] are totally pulling away. Um so you can [20:42] think of AI spending rising generally [20:44] across companies but the power users as [20:47] we talked about the most intensive users [20:49] are spending much much more. Um so we we [20:51] looked at uh some Yepit data for this [20:53] and if you if you look at sort of median [20:55] AI vendor spending in the top 1% that's [20:59] roughly eight times the level of the top [21:01] 10%. [21:02] >> What's impressive about that is it's [21:04] almost as much as the 2 to 10% combined [21:06] is basically the 1%. Yeah. So just you [21:09] got these power users that are are [21:11] clearly early adopters. [21:12] >> Yeah, absolutely. And I mean [21:14] anecdotally, even inside our own [21:15] portfolio companies, which you could say [21:17] are um pretty much fully AI native or AI [21:20] pilled, um you're seeing the top users [21:22] spend anywhere from call it 7 and a half [21:25] to 9,000 a month. And the median users [21:28] again for the most AI native companies [21:30] probably closer to the cost of like a [21:32] monthly subscription of you know 200 [21:34] maybe you're stacking two subscriptions [21:36] on top. So 200 to 400. Um so over 20 [21:39] times the spend if you think about [21:40] median versus um sort of top users. [21:43] >> Yeah. the you know we we talk to our own [21:45] portfolio companies and you know one of [21:47] the questions that we discuss is like [21:49] you know how how how much adoption do [21:50] they have and how do you measure that [21:52] and you know if you just look at like [21:54] the percentage of money that they're [21:56] spending on AI tools compared to [21:58] headcount as an example um you know I [22:01] think like relatively forwardleaning [22:03] large Fortune 500 type companies are [22:06] probably today at like 1%. M [22:08] >> um and you know the most forwardleaning [22:11] >> you know AI build companies in our [22:12] portfolio can be as high as like 10%. [22:15] And so all of these are just questions [22:17] of like how early are we into diffusion [22:20] and how deep will that diffusion go. [22:22] >> Yeah. Another example we've seen within [22:24] our portfolio um you know many [22:26] enterprises are really excited when they [22:30] you know buy cursor or cognition for the [22:33] first time but the reality is that's [22:34] just like the beginning of the journey. [22:36] the runway uh from just procuring to and [22:39] trying these tools to like full adoption [22:42] is is massive. [22:43] >> And I mean Sarah, you touched on this [22:45] early, but we're starting to see [22:47] quantifiable case studies. And you know, [22:49] we've highlighted a couple here of [22:50] public companies that actually report on [22:52] the metrics where they've seen [22:54] meaningful improvements with AI. You [22:56] know, two two that we call out. One on [22:58] the cost side. For example, Chime has [23:01] reported that they've reduced their cost [23:02] to serve by over 10% a year for the last [23:05] four years. you know, compounded. We're [23:06] talking about almost a 50% reduction in [23:09] cost to serve. You know, happily [23:11] supported by one of our portfolio [23:12] companies, Decagon. Uh, but just in [23:14] general, many initiatives to lower the [23:16] cost to serve. You know, in the revenue [23:17] side, an interesting one that I found [23:19] was Shopify here, who, you know, they [23:22] they launched their AI sidekick. And [23:24] this AI sidekick helps merchants get up [23:26] to speed much much faster. So, you know, [23:29] the percentage of customers that reach [23:31] five orders within 15 days after [23:33] onboarding has grown 8%. And that is [23:36] kind of a metric that Shopify's tracks [23:38] is once you reach five orders, you know, [23:40] you're going to stick around and you're [23:41] going to retain on Shopify. So, it's [23:43] been a meaningful tailwind to their [23:44] business as well. [23:46] >> Yeah, absolutely. And I think to your [23:48] point on um just the cost to serve [23:50] coming down, right, if you think about [23:51] the advantages that lower cost to serve [23:54] gives you, you just have more room to [23:55] compete. So, you have better pricing. [23:57] you know, maybe broader service, you can [23:58] reinvest in growth. Um, and so I think [24:01] the benefits are reaching the customers [24:04] um the not only the customers but also [24:06] the the margins and then they rotate [24:08] that back in. Um, and then I think the [24:10] other thing that's been really [24:11] fascinating to see is um, of course they [24:14] vary so you can't throw everything into [24:15] one bucket, but the incumbents have done [24:17] a pretty nice job on u monetizing this [24:20] as well. And um, service now is a good [24:22] example, right? They've reported more [24:23] than a billion in um in a in AI ACV um [24:27] and actually a 9x increase in agentic [24:30] deployments. Um so you're kind of seeing [24:32] it across the stack incumbents to newer [24:35] companies. [24:35] >> Yeah, one of the interesting things that [24:37] we've talked about a lot and you know [24:39] again some of our most forwardleaning [24:40] companies like Stripe have discussed [24:42] with us is you know where are they [24:44] actually putting their incremental AI [24:46] investment dollars? Are they putting it [24:48] toward things like building new products [24:50] for customers that could drive higher [24:51] revenue or are they putting it toward uh [24:54] optimization or efficiency gains on the [24:56] cost side? Uh and this is sort of a [24:58] litmus test for us of like where are the [25:01] you know founders or CEOs of these [25:03] companies seeing the most amount of [25:05] opportunity like the opportunity to [25:07] drive revenue growth is unbounded to the [25:09] upside [25:10] >> whereas the opportunity for cost [25:12] improvement um you know yes you could [25:14] take that and reinvest it but that's [25:16] sort of a latent opportunity that will [25:18] continue to exist and if you think that [25:21] the best and highest use of your dollars [25:23] today is to optimize your cost structure [25:25] what does that say about the revenue new [25:27] opportunity for you [25:28] >> and even within cost optimizations [25:29] there's different flavors right like one [25:31] is you know lowering your cost to serve [25:32] which allows you to reinvest and like [25:34] makes you a better business I think six [25:35] months ago our industry was really [25:37] focused on like rebuilding systems of [25:40] record internally and if your engineers [25:41] are focused on you know rebuilding a [25:43] system of record to save a couple [25:45] thousand a couple hundred thousand [25:46] dollars I expect there to be better uses [25:48] of like those resources yeah totally [25:51] agree [25:51] >> so Sarah referenced with with service [25:54] now uh seeing massive agenda authentic [25:57] usage. Um, agents are here. They are [26:00] performing tasks. Tasks require multiple [26:04] steps uh which require multiple model [26:06] calls. That helps explain the 14x growth [26:10] uh in agent token usage on open router. [26:13] Um those step steps though are becoming [26:16] more affordable. We're seeing caching uh [26:19] so the system can reuse background [26:21] information instead of processing it [26:23] from scratch each time. Uh I've seen [26:26] Hebia take advantage of this. They've [26:28] seen their financial chat workloads uh [26:30] become 10x cheaper to run. Um the same [26:34] budget that they had before can now [26:35] support more work. [26:37] >> Yeah, I think your your broader point is [26:38] just that lower costs make it practical [26:40] now for an agent to try check its work, [26:43] try again, and you can just open up a [26:46] ton of tasks where maybe reasoning or [26:49] tool use would have been just [26:50] prohibitively expensive. Um, and I think [26:53] this is especially important in cases [26:54] where reliability is maybe the de facto [26:58] reason you would or would not use an [26:59] agent. Um, and again, you know, we sort [27:01] of talked about type safe previously, [27:03] but imagine what happens when you go an [27:06] order of magnitude cheaper, two orders [27:08] of magnitude cheaper. Um, you know, I [27:10] think the amount of use cases that open [27:12] up are actually unimaginable. [27:13] >> Yeah, totally. And, you know, with much [27:16] uh greater improvement in latency. [27:17] >> Yeah, absolutely. [27:19] >> Yeah. So, so companies are definitely [27:21] thinking more and more today about how [27:23] to optimize latency or performance at [27:26] large with cost. Um, so take data bricks [27:30] as an example. They're leveraging [27:32] routing uh choosing the appropriate [27:35] model uh for each specific task to per [27:38] to improve performance and cost. And so [27:41] their smart router um performed uh [27:44] better. It solves more problems at 35% [27:47] lower cost than the strongest individual [27:50] model. Um, another approach we're seeing [27:52] a lot of right now is fine-tuning. We've [27:54] seen this with Harvey. We've seen this [27:56] with Elise AI. In the case of Elise, [27:58] they fine-tuned a smaller model. Um, so [28:01] it became much more affordable, 60% [28:03] cheaper. Um, but also had way lower [28:05] latency. So live use cases from an audio [28:08] perspective became tenable. And so uh [28:12] these engineering gains are making AI [28:14] more useful but also more affordable. [28:18] >> Yeah. And I think the if you you sort of [28:19] roll that up um and think about for [28:22] application business the relevant unit [28:24] for them is really the cost of getting [28:26] the customer's job done. And if you can [28:28] decouple that right you get the [28:29] customer's job done you charge for that [28:31] and you can use better routing to make [28:33] the product more economical. Um, that's [28:36] pretty magical because now you're not [28:38] using the most expensive model for each [28:40] step, but you are getting the job done [28:41] and that's gonna I think we're going to [28:43] see more margin improvements and and the [28:45] best app companies. [28:46] >> It's a nice segue. Yeah, we we've been [28:48] talking mostly about usage inside of the [28:50] enterprise, but important to talk about [28:52] consumer given obviously that's the [28:54] place where we've had some of the [28:55] largest outcomes and where we spend a [28:56] lot of our times. No, we think it's [28:58] still pretty early. There was a recent [28:59] survey that showed that only just over [29:02] 2% of US households actually have a [29:04] paying subscription for AI. You know, [29:07] maybe subscription is not going to be [29:09] the best place to monetize the majority [29:11] of households, but it it has been so [29:14] far. And the interesting thing about the [29:15] subscription revenue is it has some of [29:17] the best retention we've ever seen in [29:18] consumer. You know, uh, you know, Alex [29:20] likes to talk about smile curves and [29:23] it's really, really rare to see a [29:24] retention profile that not only flat [29:26] lines, but actually smiles as, you know, [29:28] the product improves and people come [29:29] back to it. But that's what we've seen [29:31] with AI, which just demonstrates the [29:33] value that the subscription is being is [29:34] is able to deliver to the to the person [29:36] who pays. [29:37] >> By the way, just on subscription, what [29:39] struck me about these numbers, [29:40] especially if you look at the bottom [29:41] right chart, is just how small they are. [29:44] Like think about Amazon Prime. That's [29:46] over 200 million households. Um, Netflix [29:50] is a great one. [29:51] >> 70 million households, right? These I [29:52] had to look it up. I'm like, is that K [29:54] correct? Um, and so totally agree that [29:57] it's just getting started. [29:58] >> Yeah. I mean, our partner Josh Elman [30:00] does have uh a definition I like that [30:03] would be counter to that a little bit, [30:04] which is consumer AI is what I use for [30:07] my daily life and not necessarily what I [30:10] expense. So in his view he would say [30:12] like it's maybe not super surprising [30:14] that 97% of households that are using AI [30:17] aren't yet paying for it. Um you are [30:20] right Santi I do like smiling retention [30:22] curves. Uh we do like to see users [30:25] coming uh back more and more. You know [30:27] another common characteristic of large [30:30] consumer platforms is time spent. Uh so [30:33] if you were to look at Facebook, [30:35] Instagram, Tik Tok, Snap, uh they all [30:37] get 60 or 30 to 60 minutes per day uh [30:41] from their active users. And while the [30:43] best AI assistants uh and there's many, [30:46] you know, up and coming right now, uh [30:48] all aspire to be my go-to application, [30:51] like I I hope that they are going to be [30:54] incredibly persistent and always on and [30:57] super proactive such that they may not [31:00] be the place where I spend the most time [31:03] in the future. Yeah, it's going to make [31:05] our job harder when we can't, you know, [31:07] look at engagement via outside in data [31:09] because the agents are working on the [31:10] background instead of us just looking at [31:12] like screen time on a social media [31:14] application. But it is important. I [31:16] mean, I remember like when the newest [31:17] models come out in December, even in our [31:19] workloads, you know, we would set off a [31:20] deep research task or an agent task and [31:23] then, you know, drive up to the city or [31:24] something and that's not driving Tesla. [31:27] >> That's not screen time that, you know, [31:29] being tracked as as would have been in [31:31] prior generations of consumer. [31:33] >> Yeah. Totally agreed. [31:34] >> It's early, but it is a really new and [31:38] exciting time in consumer. Uh Muse and [31:41] Instinct are no doubt the new kids on [31:43] the block, but they've grown really [31:45] quickly. Uh those along with chat GBT [31:48] are taking more and more of my queries [31:51] away from traditional search. Um, but [31:53] it's a really a dynamic time and so if [31:55] you are an existing consumer discovery [31:58] platform, an existing marketplace, you [32:01] really need to be thinking about your [32:02] chest moves right now. Um, there's a [32:05] couple questions I'd be asking. First, [32:08] how much incremental demand can I get [32:10] from an AI agent? How many more orders [32:12] can they bring me? And two, how much of [32:14] my business, how much of my profit pool [32:17] comes from owning the customer [32:18] relationship and discovery? Uh last week [32:21] there was a lot of news where Amazon [32:24] said you know no thank you to Muse uh [32:27] but Instacart said yes please. And if [32:30] you think about Amazon X AWS and [32:32] Instacart uh and look at their [32:34] advertising revenue it outpaces all of [32:37] their operating profit. And so [32:38] advertising revenue owning the customer [32:40] relationship is incredibly important. [32:43] But for Amazon, you know, how many [32:46] incremental new orders or certainly [32:48] customers am I going to get from [32:50] connecting to Muse? Not that many. [32:52] Whereas if you look at Instacart online [32:55] penetration of grocery, still relatively [32:57] early, a lot more orders to go get. And [33:00] so the optimistic possibility here is [33:03] that there's going to be a lot more [33:05] orders coming from [33:08] uh from from these applications. Uh [33:11] historically I've had to, you know, [33:13] click all these different ways through [33:15] to process an order. If I if I offload [33:18] that to an agent, all of a sudden [33:20] there's no clicks and hopefully more [33:22] GMV. The the trip I once wanted to book [33:25] but didn't gets booked. The order that [33:27] we all wanted to place for dinner [33:28] tonight happens. Um so I'm optimistic [33:31] there's going to be a lot more GMV and [33:33] that'll make up for some of this lost [33:35] advertising revenue. [33:36] >> Yeah, it's sort of this question though [33:37] that's open of how does it get [33:39] compensated for, right? So to your [33:41] point, you know, Amazon has an over 70 [33:43] billion advertising business that's [33:45] extremely high margin flow through that [33:48] uh you know is totally uh predicated on [33:52] the fact that consumers go to the [33:53] website and click on the ads and if they [33:56] don't do that anymore, you know what [33:57] happens? Um you know today, you know, [34:00] Meta and Meta and Google famously [34:02] advertise um you know probably best out [34:06] of any of the internet platforms. um you [34:08] know they each make call it 200 bucks [34:11] plus per per users in the US in [34:14] developed world um you know in the case [34:17] of Meta it's sort of a you know it's an [34:19] entertainment application we'll see you [34:20] know what happens with that it's [34:21] probably a little bit safer in the case [34:23] of Google you know um it's funny like [34:25] the whole talk track around Google two [34:28] years ago when all of this just happened [34:30] was oh my gosh what's going to happen to [34:32] Google's search business and it turns [34:34] out it's been really resilient part of [34:36] the reason it's been really resilient [34:38] is because their very high monetizing ad [34:42] terms were somewhat safe because they [34:45] were things like I need to buy insurance [34:47] >> or you know find me a hotel in this city [34:51] um you know or things like that that [34:52] that you can very directly monetize um [34:55] but that the AI was not yet capable of [34:58] going to take action for on your behalf. [35:00] If that changes that could be a very [35:02] different dynamic. [35:04] >> Yeah. The other thing just to add I [35:06] think the narrative last week was a [35:07] little bit like negative sum where it [35:09] was very negative you know like look at [35:10] these marketplaces are going to be [35:12] impacted. I do think like the positive [35:14] some view of this is number one maybe [35:16] that $70 billion that's spent on [35:17] advertising on Amazon just finds a [35:19] different channel and you know goes [35:20] elsewhere maybe spent directly on the [35:22] agents through a different form factor [35:24] or just finds better ways to target [35:26] people and then two I think you know [35:28] with the lower friction we might just [35:30] see more consumption like people might [35:32] buy more you know and that's just like a [35:34] positive flywheel that drives economic [35:35] growth and I think a little bit we're [35:37] locked into this oh this is bad for [35:39] profit pools but actually might be just [35:40] good [35:41] >> yeah totally [35:42] >> both could be true by the way. [35:43] >> Yeah. [35:44] >> Yeah. Yeah. [35:44] >> That it could be [35:45] >> there might be profit pools but it could [35:47] be additive overall to the economy. Like [35:49] I think we all would be disappointed if [35:51] we looked back three years from now and [35:52] there's not a meaningful productivity [35:54] improvement actually at the overall [35:56] macroeconomic level which would drive [35:58] you know consumer spending higher in the [36:00] case of advertising. But it might just [36:01] be that you know profit pools disappear [36:03] from certain places and get reallocated. [36:05] >> The the immediate reaction has been net [36:07] market cap positive across the [36:08] ecosystem. the meta gains have far [36:10] exceeded any of the losses from the [36:12] marketplaces. [36:13] >> Yeah, it's a great point. [36:14] >> Yeah, going going back to our discussion [36:16] on software and moving maybe more to the [36:18] the public side. Um I think the big [36:20] trend that jumps out at you from from [36:22] these charts is that the mix has really [36:24] shifted toward slower growing more [36:27] profitable companies. Roughly 75% of the [36:31] public software sample here is [36:32] profitable and only 30% is growing at [36:35] 20% or more [36:37] >> which is a staggering number that only [36:38] 30% of the public software companies are [36:41] growing at 20% plus [36:43] >> and not only 20% I mean that's 30%. If [36:45] you draw the line at 30% we've discussed [36:48] like the absolute number of companies is [36:50] less than five when you know in the [36:52] private markets basically every company [36:54] we see and spend time with is growing at [36:55] well in excess of 30%. [36:57] >> Yeah. And look, this is not um this is [36:59] somewhat obvious actually if you look at [37:01] just you talk to you know IT managers, [37:04] CIOS etc. Um if you look at the growth [37:07] that they're experiencing and what they [37:08] spend on AI% [37:10] >> like the easiest place for those dollars [37:12] to come from is not spending incremental [37:13] new dollars on new SAS software [37:16] projects. [37:16] >> Yeah, exactly. Um so you know I think [37:19] we're investors in in some of the you [37:21] know the SAS software companies [37:23] including the public markets and the [37:24] conversations that we're having with [37:26] them and the founders are very focused [37:27] on is okay how do I take my existing [37:29] distribution which in many cases is very [37:31] very strong you know locked in customers [37:34] um who you know to your point earlier [37:36] Santi wouldn't go anywhere and apply [37:38] really interesting new AI products to [37:40] them where I can drive my revenue growth [37:42] higher um I think you know in order to [37:44] sort of dispel the greatest fear fears [37:47] around the SAS apocalypse. Um, you know, [37:50] I think we just need to see a period of [37:52] time where the software companies [37:54] continue to post things like 98% gross [37:56] dollar retention, um, which was kind of [37:58] always the sticking point for why they [38:00] were so attractive as investments. Um, [38:02] continue to drive efficiency like we've [38:04] talked about, but most importantly [38:06] drive, you know, revenue growth [38:07] acceleration. um which shows that [38:09] they're sort of safe from a defensive [38:12] standpoint, but that the offensive [38:13] investments that they're making are [38:14] actually going to drive their business [38:16] to be better. [38:16] >> I thought your point in the blog post [38:18] that you wrote a couple months ago on [38:20] two paths was really interesting as a [38:22] rule of thumb. Maybe share more on the [38:24] finer point of the percentages. [38:26] >> Yeah, you know, the more I talked to [38:27] founders after it, um the more I felt [38:29] like uh you know, almost everyone that [38:32] we talked to is like, yeah, I'm trying [38:33] to drive revenue growth higher. Uh and [38:37] you know it's uh it's it's it's it's [38:39] clearly I think the predominant path for [38:41] the types of companies that we're [38:43] investors in and that we've backed over [38:44] the years. Um again I think you know [38:47] maybe relative to a year ago when the [38:50] fear was like okay everyone's going to [38:51] vibe code their software systems like [38:54] that is clearly not what's happening in [38:57] the market. Um but you know the onus is [39:00] high on delivering revenue growth [39:02] acceleration and so we had said you know [39:04] let's target 10% plus acceleration which [39:06] is a high number um but with a magical [39:08] product given the budgets that are [39:10] available to AI is seemingly doable and [39:13] so we'll see I think that there are a [39:14] few of them that we are close to where [39:16] we'll see that over the next 12 to 18 [39:18] months. Yeah, absolutely. And you know, [39:20] as with all things, um you can't lump [39:22] every uh every every company into the [39:25] software buckets sort of um you know, [39:27] talk about uh multiples coming down and [39:29] and sort of growth uh growth decline as [39:31] well. Um in fact, uh software has been [39:34] very differentiated and um if you look [39:36] at this chart here, um you can see that [39:39] cyber security observability have really [39:41] stood out um and uh vertical software as [39:43] well has generally held up um much [39:45] better than the horizontal applications. [39:47] Um, and I think the framework for how we [39:50] think about what's been, you know, [39:52] what's held up or even gone up nicely [39:54] versus, um, uh, in more secular decline, [39:57] if you will, um, is that, uh, really you [40:00] think about how AI changes the [40:01] customer's need for that product. Um, [40:04] right. And and I think the cyber [40:05] security risk um uh associated with AI [40:08] has been well publicized. But as you [40:10] think about more software and agents [40:12] creating new security and new monitoring [40:14] needs um this is uh creating greater [40:17] demand honestly for the incumbents in in [40:19] in these markets. And you can see um you [40:21] know the call out on crowd strike on the [40:23] on the right right hand. Um and then of [40:25] course on the flip side applications are [40:27] facing different degrees of of workflow [40:29] change right and it's sort of you know [40:31] if you think about um you know one of [40:33] our top CEOs Ali Goatsy likes to talk [40:36] about this chopping block of AI and like [40:38] what's first on the chopping block um [40:40] and you know you can kind of think about [40:42] uh on the flip side you know what is [40:44] actually needed for AI um and you you [40:47] pair those dynamics together and I think [40:48] you know it explains a lot of what we're [40:50] seeing in the public markets. Yeah, I [40:52] think a lot of this is pretty intuitive [40:53] with what we're seeing on on the private [40:55] market side. So, as you alluded to Sarah [40:58] on the on the security front, uh, agents [41:01] are accessing more and more systems, [41:03] taking more actions. The like open AI [41:05] hugging face incident was, you know, an [41:08] eye opening event for many. Uh, so [41:10] security is is paramount. And then as we [41:13] look at the vertical AI companies, these [41:15] are some of the fastest growing [41:16] companies we're seeing in the private [41:18] markets, right? Harvey, a bridge, a [41:20] lease, they are growing faster than any [41:22] of the precedents ever have in their [41:23] industry and and that's a reflection of [41:25] the customers recognizing it's not just [41:27] the model which we talked about earlier [41:29] but how you how you or orchestrate that [41:32] how you build the application around it [41:33] and vertical specific workflows command [41:36] uh command these needs [41:38] >> and I mean we spend a lot of time [41:39] looking at polar markets less so for [41:41] investments but we we underwrite exit [41:43] multiples right so we track that pretty [41:44] closely and I think if we were to recap [41:46] the year what's happened you know first [41:48] couple months months there was the SAS [41:49] apocalypse. We wrote a blog post about [41:51] like what kind of businesses would do [41:53] well. And if you fast forward to today, [41:55] the software index is actually back to, [41:58] you know, where it started the year, but [42:00] it's really bifurcated into like some [42:02] companies that are deemed AI losers and [42:04] some companies that are deemed winners. [42:06] And, you know, the public market [42:07] sometimes simplifies things a bit too [42:09] much. But to your point, the AI winners [42:11] are not just the ones that can improve [42:12] their cost structures. They're the ones [42:14] that are actually accelerating and [42:16] capturing some of this net new dollars [42:18] up for grabs that if you're not really [42:20] capturing, then you're probably not [42:21] riding the wave. [42:22] >> Yep. [42:23] >> Yeah. Um, you know, I think, um, on the [42:26] flip side, we wanted to show showcase, [42:28] um, some private company operating data [42:30] that gives us a view of what customers [42:32] are actually doing. Um, and there's a [42:33] little bit of a narrative violation in [42:35] Stripe's SAS customer data. um which [42:37] actually shows growth accelerating into [42:40] 2026 across both young and mature [42:43] businesses. [42:44] >> Stripe calls it the renaissance. Oh, the [42:46] renaissance. I actually quite like the [42:47] renaissance. [42:48] >> I like renaissance. Yeah, [42:49] >> it's it's very good. It's very good. [42:51] >> Um to if if you talk, you know, we've [42:54] been talking about public markets a bit. [42:55] You know, Sarah talked about private [42:56] market SAS acceleration. [42:59] We get a lot of questions around, you [43:00] know, companies staying private for [43:01] longer. So if you actually look at the [43:03] top six companies today, you know, [43:05] Anthropic, OpenAI, Data Brick, Stripe, [43:07] Wayland, Revolute by last round [43:08] valuation, they add up to about 2.4 [43:10] trillion. This is more than the combined [43:14] market cap of IPOs we've seen in the [43:16] last 10 years excluding SpaceX, which [43:18] adds up to 1.7 trillion. So just the [43:20] amount of activity that happens within [43:22] our market, again, six companies, 2.4 4 [43:24] trillion. It's almost as big as the [43:25] Russell 2000, which is, you know, a big [43:27] index in the public markets where [43:29] there's hundreds of public managers that [43:31] spend most of their time in. So, you [43:33] know, it's increasingly exciting to just [43:35] spend time in these late stage champions [43:37] that can keep investing in growth more [43:39] so than maybe um maximize short-term [43:43] profit. But David, you can talk a little [43:44] bit about kind of how we advise [43:46] companies on the IPO and when to stay [43:47] private versus public. Yeah, look, I [43:49] mean the IPO, you know, especially for [43:51] founder-led companies, one of the things [43:53] that we've talked about and and that [43:54] I've written about is that the founder [43:56] is the asset class at this point. And so [43:58] the bet that we make and part of the [44:00] reason that it's been a benefit for some [44:02] of these companies to remain private is [44:04] they can many times take bigger swings [44:07] um in the private markets that have [44:09] longer duration paybacks. You know, Zuck [44:11] and and Elon are kind of obvious [44:13] exceptions to the rule in the public [44:15] markets. Um but you know the way the the [44:17] way you see it manifest in the numbers [44:19] in companies like data bricks or stripe [44:21] um is you know massive massive new bets [44:25] um in new product areas um and you know [44:28] you can see revenue acceleration that [44:29] happens as a result uh and so you know [44:32] you can do that in the public markets um [44:34] but you know it's going to catch greater [44:35] scrutiny you know obviously Meta's stock [44:37] price um you know at the nater you know [44:40] got below 100 bucks a share um you know [44:42] when people were very very skeptical [44:43] about their investments in in ARVR. Um, [44:47] you know, for us, the way we talk about [44:48] it with our founders is just the IPO is [44:50] another financing event. And, you know, [44:52] what do you get out of being a public [44:53] company relative to what do you get out [44:55] of being a private company? Um, but [44:57] clearly you can reach big scale uh in [44:59] the private markets. And there's some [45:00] other companies that, you know, already [45:02] talked about running themselves like [45:03] they were a public company, you know, [45:05] with like high focus on efficiency and [45:08] just basically like every metro being [45:10] tracked. There's also some companies [45:11] that benefit from maybe like public [45:13] disclosing of financials and earning [45:15] customer trust both at the enterprise [45:17] level and on a consumer level. I mean [45:19] Navan went public and we've seen a [45:21] reaceleration just on the basis of some [45:23] of the larger enterprises actually [45:25] trusting them more because they're a [45:26] public business. [45:27] >> Yeah. And then of course there are [45:28] examples where you know capital needs [45:30] over time uh could be very very large [45:33] and so the p the the capital pools in [45:36] the private markets are very large uh [45:38] and can and can serve the needs of many [45:40] of these companies but at some point you [45:42] know the capital needs may even get too [45:43] big for the private markets. [45:45] >> That's the public markets. You know in [45:46] the private markets one of the things [45:47] that we we spend a lot of time in too is [45:49] is secondaries. So there's two dynamics [45:52] at play. You know, one is we see an [45:54] increasing amount of companies holding [45:56] tenders for their employees. So if they [45:58] choose to stay private, you know, [45:59] allowing employees to get liquidity [46:01] outside of the public markets. And an [46:02] interesting data point show here is, you [46:05] know, participation in tenders dropped [46:07] by Carta has only been 58%. So this is [46:10] employees choosing not to take liquidity [46:13] because they have so much conviction in [46:14] the performance of their business. So, [46:15] it's the opposite of I think sometimes [46:17] there's a narrative of like employees [46:18] cashing out and what we're seeing is [46:20] actually employees choosing not to cash [46:22] out because they believe in their [46:23] company so much. [46:24] >> Imagine SF home prices if that was 100%. [46:28] >> Exactly. [46:28] >> It'd be hard. Exactly. Yeah. The the [46:30] tender off the tender offers that [46:32] happen, you know, we're obviously like [46:34] participating in these and and leading [46:35] these uh pretty frequently. So, we think [46:37] they're a good thing. Um they do serve [46:39] two purposes. one, you know, for [46:40] employees and prospective employees, it [46:43] is sometimes hard to compete with the [46:45] liquidity appeal of public markets, RSUs [46:48] that hit your account on a net tax basis [46:50] every quarter. Um, so, you know, it is a [46:52] it is a um, you know, a weapon of [46:55] competition for the private markets that [46:56] want to compete for employees or [46:58] retaining their employees um, in the uh, [47:00] with with with the public market [47:01] companies. The other side of it is, you [47:04] know, we're advocates of sort of more [47:06] frequent resetting of your valuation. [47:08] Um, and so that keeps your sort of stock [47:10] price, you know, private stock price [47:13] fresh for a number of reasons. One, it's [47:15] easier to talk about that with employees [47:16] and prospective employees. Um, but two, [47:19] you know, in the event that you want to [47:20] do M&A like many of our companies have [47:21] done, um, you know, you have fresh [47:23] currency that you can use. [47:24] >> Yeah. No, maybe just the other dynamic. [47:26] I think people talk about the secondary [47:27] discount quite a bit because that's a [47:29] dynamic that we've seen over the last [47:30] five years where maybe coming out of [47:31] 2021 and 22, 23, 24, the median discount [47:35] to the last round price in secondary [47:37] markets was meaningful because there was [47:39] maybe like a round that was priced too [47:40] high and it was out of date. But [47:42] actually what you're seeing today is you [47:45] know when there are transactions [47:46] secondary markets the discount to the [47:48] last round price is basically zero. [47:49] Yeah. which just speaks to, you know, [47:51] these these valuations at which [47:53] companies have raised that are fresh and [47:55] there's always net new investors willing [47:57] to pay that same price, which wasn't the [47:59] case for last couple years. [48:00] >> Yeah, great point. So, um, this is a fun [48:05] slide that says everything is AI [48:07] computer. So, AI related companies [48:09] account for 86% of US VC deal activity [48:12] in this 2026 snapshot, up from 65% in [48:16] 2025. Um so you know obviously AI has [48:20] become the dominant destin destination [48:21] for venture dollars. [48:22] >> Yeah. But beneath that AI label the [48:25] opportunity has broadened a lot right. [48:26] So enterprise apps consumer apps uh but [48:29] also services semiconductors power [48:32] defense um a lot of the attention goes [48:35] to open AI space AI and anthropic but [48:38] beyond that um our opportunity set has [48:41] never been deeper and and broader. Last, [48:44] we thought it would be fun just to talk [48:45] about some of the areas that we're very [48:47] excited about right now. Um, you know, [48:50] obviously we touched on some of the [48:51] consumer work that will now be done by [48:54] agents. Um, but you know, you could call [48:56] it longunning agents, you can call it [48:58] autonomous agents, heavy use consumer. [49:00] Um, you know, if you if you take the [49:02] perspective of these uh that you know, [49:04] you could have consumer agents do all [49:06] the tasks that you wouldn't want to do [49:07] to give you things that you otherwise [49:09] wouldn't have spent time or money on. Um [49:11] it's very appealing and I think that the [49:13] uh the distribution of this could happen [49:15] pretty quickly. Robotics is an area [49:17] where we're spending a lot of time and [49:19] attention. We happen to think that it [49:20] could be even larger than LLMs. Um but [49:24] you know probably 3 to 5 years earlier [49:26] um and so we think over the next 5 years [49:29] this is going to be a massive area of [49:31] investment and excitement. Um autonomy [49:33] is here. Uh self-driving works. Uh it's [49:37] a very exciting time. uh if you just [49:39] take you know a step back and talk about [49:41] the auto industry and transportation [49:43] industry this is one of the biggest [49:44] industries in the world that we probably [49:46] don't talk about as much because we [49:48] spend so much time talking about AI [49:49] right now um but you know if you look at [49:52] miles traveled by Uber andyft or in [49:55] Ubers and lifts it represents about 1% [49:58] of miles traveled in the US I think with [50:01] um you know full networks of autonomous [50:03] driving cars that are 14 times safer [50:06] than human drivers we expect that to [50:08] expand by at least an order of magnitude [50:11] in the coming years. Plus, there's 17 [50:13] million new cars sold per year in the [50:15] US. And I think over the next 10 years, [50:18] those those will all be um be [50:20] autonomous. Um AI times bio, this is a [50:24] super exciting area. You know, obviously [50:25] the folks in the labs are talking about [50:27] this. Um but you know new drug discovery [50:30] solving some of the um you know most [50:32] debilitating illnesses uh or diseases in [50:35] the world I think is a promise that we [50:37] all are very hopeful for that we'll see [50:39] a lot of progress over the next 10 [50:40] years. Personal health. This is another [50:42] one that I'm very uh excited about for [50:44] myself uh you know as a as a health [50:47] maxer. Um, but you know, there's not [50:49] been a great place uh or avenue to take [50:52] all of the information about yourself, [50:54] put it into somewhere and get very [50:56] hyperpersonalized [50:58] uh advice. Um, and then, you know, [51:00] lastly, diffusion uh into the enterprise [51:03] beyond coding. This is one of the, you [51:04] know, the ones that Sarah you were [51:06] talking about just we're so early in [51:09] actual diffusion of the technology into [51:11] the enterprise um that I think it's [51:13] going to be, you know, super super [51:15] exciting. Um and then lastly um you know [51:18] there's another area that is not inside [51:20] the AI bucket which is you know what we [51:22] call American dynamism but um the sort [51:24] of retooling of the entire um sort of [51:28] American dynamism stack. We've you know [51:30] we've invested we've been large [51:31] investors in this area for a while. um [51:34] you know it's still just a small [51:35] fraction less than 5% of overall dollars [51:38] spent um you know in the military uh is [51:40] sort of you know newer vendors like [51:42] Andre uh or Seronic or Castellian um and [51:46] we expect that to grow dramatically uh [51:48] as needs change. So, so many areas of [51:51] excitement. Obviously, you know, we've [51:53] been very active in this AI space. Um, [51:55] and you know, we're optimistic about the [51:57] effect that it's going to have on the [51:58] overall economy. The buildout is [51:59] massive. Um, but we think it's going to [52:02] be massively productivity enhancing in [52:03] the US. So, uh, it's a blast to hang out [52:06] with you guys. Thank you.