How AI Is Rewriting the Power Law of Venture Capital

人工智能如何重塑风险投资的幂律法则 演讲嘉宾分析了前沿人工智能如何从根本上改变风险投资的经济模型，将资本直接转化为全球30万亿美元GDP板块中的算力与产品质量，而非导致运营规模臃肿。他们指出风投回报存在极其残酷的集中度——只有不到1%的风投机构能稳定实现3倍的净回报，并警告私募股权中负债累累的传统SaaS并购交易正面临严重的估值危机。为获取非对称超额收益（alpha），该小组主张LP应将资金集中配置于全生命周期平台型机构，同时预计下一个数万亿美元级别的增长前沿将扩展至能源基础设施、机器人、自主系统以及医疗健康领域。 在过去20年间的大约3,000家美国风投机构中，仅有20家稳定实现了3倍净回报，表明极端的幂律法则倾向于奖励对顶尖平台的集中配置。 (0:00) 生成式AI总收入仅用4年就达到了1,000亿美元，而企业级SaaS达到相同门槛则用了15年。 (2:37) 人工智能将软件潜在市场规模（TAM）拓展至规模达30万亿美元的全球劳动力经济中；仅在医疗健康领域，行政管理劳动力就代表着1万亿美元的市场，而传统IT支出仅为600亿至1,000亿美元。 (4:58) 随着上市SaaS公司的估值倍数回落至营收的2倍左右，软件私募股权正面临严峻的债务积压问题，这源于2021至2022年间以25至32倍EBITDA倍数进行的超2,000亿美元债务驱动型并购。 (37:45) 制约人工智能扩张的主要瓶颈并非终端用户需求，而是通电速度，其受制于输电许可和基础设施，而非总发电量。 (46:48)

Jen Kha, David George, and Aram Verdiyan argue that artificial intelligence is intensifying venture capital's power-law dynamics because capital converts directly into compute and product performance, while legacy software and leveraged buyouts face acute terminal-value risks. Capital-Compute Compounding and Expanded Addressable Markets (0:00) Historically, early-stage venture startups risked operational paralysis when flooded with excess capital due to coordination friction and management bloat. In frontier AI development, David George notes that capital converts directly into compute, which immediately improves model capabilities, reinforcing product dominance rather than creating organizational friction. Furthermore, generative AI reached $100 billion in aggregate revenue in four years, a milestone that SaaS required 15 years to achieve. This rapid compounding is driven by an expanded total addressable market. While traditional enterprise software sells seat licenses, AI targets completed labor tasks across a roughly $30 trillion slice of global GDP. Total U.S. labor expenditure exceeds enterprise software spending by roughly fortyfold; for example, healthcare IT spending is $60 billion to $100 billion annually, whereas administrative labor and billing exceed $1 trillion. Rather than mere software tools, frontier AI labs are capturing core economic pillars. Venture Dispersion and Asymmetric Institutional Incentives (10:00) Venture returns follow a punishing distribution: historical data cited by Aram Verdiyan shows that out of roughly 3,000 U.S. venture firms over the last 20 years, only 20 (fewer than 1%) have consistently generated 3x net returns. Average managers yield 1x to 2x net, trailing liquid public equities and buyout funds once illiquidity is considered. Top-decile fund economics accommodate high risk profiles—roughly 60% loss rates in early-stage funds and 10% to 20% in growth funds—demanding concentrated checks of 5% to 10% of a fund into category leaders. Institutional allocators, however, face misaligned career incentives. Jen Kha highlights that while venture general partners risk termination over errors of omission (missing a generational champion), limited partners are primarily penalized for errors of commission (visible losses). Consequently, allocators often over-diversify across 50 to 70 managers, blending their performance into the median benchmark rather than concentrating in the small cohort of platform firms and elite seed boutiques that control early access and pro-rata allocations. Terminal Risks for Leveraged Software Buyouts (30:00) Private equity software portfolios face structural vulnerabilities following the 2021–2022 market peak, when sponsors deployed $200 billion to $300 billion in enterprise software leveraged buyouts supported by more than $200 billion in private credit at 25x to 32x EBITDA. Today, only 15 to 20 public SaaS companies trade above 10x revenue, and public comps have contracted by roughly 50%. Aram Verdiyan emphasizes that public markets currently value a 1% increase in organic revenue growth as equivalent to 3% of EBITDA margin expansion, penalizing slow-growing, seat-based platforms. Superficial AI retrofits, such as bolting automated customer support onto legacy architectures, frequently fail by degrading customer experience and spiking churn, directly threatening debt covenants. The speakers note that unlike agile startups, leveraged private equity balance sheets rarely tolerate the corporate self-cannibalization required to rebuild core products from scratch. Diffusion Pace and Supply-Side Physical Bottlenecks (34:25) Although frontier adoption is intense, enterprise AI diffusion remains uneven. The median U.S. firm spends $12 per employee per month on AI tools, whereas the top 1% spend upwards of $7,000. David George points out that historic sector revenues have been realized across an estimated 10 million to 30 million active users out of 1.5 billion global knowledge workers, with future enterprise value expected to emerge from robotics, autonomous transport, and healthcare care delivery. Aram Verdiyan asserts that the critical constraint limiting this growth is not customer demand or chip architecture, but physical speed to power. In the United States, bureaucratic permitting and transmission line construction severely lag energy generation needs. Because next-generation AI data centers operate at more than ten times the power and compute density of legacy facilities, infrastructure execution dictates the operational ceiling of the machine age. While modern unicorns frequently remain private for ten or more years, creating liquidity overhangs for institutional funds, the speakers maintain that premature exits truncate power-law compounding. The ultimate expansion of AI value depends on overcoming domestic transmission permitting constraints and moving beyond early text interfaces into physical automation and workflow-native software.

Andreessen Horowitz 的 David George、Jen Kha 与 Accolade Partners 的 Aram Verdiyan 认为，前沿人工智能正重塑风险投资的底层幂次法则：资本不再带来管理冗余，而是直接转化为算力壁垒与产品优势。这一范式正在撬动高达 30 万亿美元的全球劳动力市场，促使机构投资者重构资产配置，并引发传统软件杠杆收购的深层危机。 算力复合循环与宏观劳动力市场扩容 (0:00) 历史经验表明，早期初创企业若注入过量资本，往往会因组织扩张导致协作摩擦与管理内耗；但在前沿 AI 实验室中，David George 指出，资本能够直接转化为基础算力，而算力又即时提升模型性能并巩固市场领导地位，形成持续复合的正向飞轮。这使得 OpenAI、Anthropic 和 SpaceX 等前沿梯队的潜在企业价值合计达 3.5 万亿至 5 万亿美元。Aram Verdiyan 补充指出，生成式 AI 仅用 4 年时间便实现了 1000 亿美元的行业总营收，而传统 SaaS 达到同等规模耗费了 15 年。 与此同时，AI 的潜在市场空间突破了传统企业软件的局限，正切入规模达 30 万亿美元的全球劳动力与专业服务经济。以医疗保健为例，全美年度 IT 软件支出仅为 600 亿至 1000 亿美元，但医疗账单与行政后勤等人工成本却超过 1 万亿美元。由于全美整体劳动力开支大约是软件开支的 40 倍，AI 正在从传统的按人头订阅许可模式，转向以完成具体工作任务为核心的价值交付，这使得机构投资者开始将 AI 视为核心资产配置而非边缘科技敞口。 风投极化分布与出资人的配置偏差 (10:00) 在过去 20 年的 3000 家美国风险投资机构中，仅有 20 家（不足 1%）能够跨多期基金稳定提供 3 倍净回报率。Aram Verdiyan 分析称，若机构出资人（LP）无法进入前 5 到 10 家顶级机构，其回报率往往会迅速滑落至行业平均的 1 至 2 倍，在扣除流动性折价后甚至落后于公开市场和常规并购基金。行业内部的分化促使市场出现“中间层消亡”：创始人日益青睐具备深厚运营资源的多阶段平台型机构，或是极度专注早期的垂类精品基金，导致领投 A、B 轮的中型基金生存空间被急剧压缩。 此外，投资生态中存在显著的职业激励不对称。Jen Kha 分析指出，风险投资合伙人（GP）最致命的职业错误是错失时代级企业的“遗漏错误”（Error of Omission）；而机构 LP 的首要关切则是避免重仓失败项目的“执行错误”（Error of Commission）。这种防御心理导致许多 LP 采取过度分散的策略，将资本平摊至 50 到 70 家基金经理。这种做法稀释了收益，即便偶然押中 10 倍回报的明星基金，在整体资产负债表上也仅能带来微弱的边际贡献，无法享受风险投资特有的幂次收益。 杠杆软件收购危机与真实拐点辨识 (30:00) 随着公开市场对软件企业的估值逻辑发生剧变，目前仅有 15 至 20 家上市 SaaS 公司的估值维持在 10 倍市销率以上。Aram Verdiyan 指出，当前市场对每 1% 有机收入增长的估值溢价已等同于 3% 的 EBITDA 利润率扩张，且仅有借助 AI 实现业务加速的资产才能获得溢价。这给 2021 至 2022 年间完成的软件杠杆收购（LBO）带来结构性危机：当时私募股权机构在 25 至 32 倍 EBITDA 的高位完成了 2000 亿至 3000 亿美元的交易，并动用了超 2000 亿美元的私人信贷；如今对标资产估值折半，股本垫严重受损，私人信贷市场开始显现赎回压力。 面对技术颠覆，许多传统企业试图将 AI 视为简单的降本插件，但简单替换人工客服往往导致客户流失与净推荐值（NPS）崩塌，进而在重债压力下陷入营收下滑螺旋。David George 与 Aram Verdiyan 强调，真正的转型需要如 Intercom 那样彻底重构产品底层架构。在甄别早期 AI 企业的真实牵引力时，投资人必须穿透以月化年（Run-rate ARR）堆砌的同行内销噪音，深入考察终端用户黏性。例如法律 AI 平台 Harvey 在引入深度推理模型后，企业态度才从担心幻觉转变为强制要求律所采用，完成了从概念试用到工作流落地的质变。 实体经济新前沿与供电基础设施约束 (44:32) 尽管独角兽企业的退出周期已被拉长至 10 年以上，且 IPO 后大额持股仍需 12 至 24 个月才能逐步变现，但 David George 强调，对于顶级赢家而言，长期复利的回报远胜于过早套现。展望未来的超级市值阶段，AI 的实际渗透率依然极低：全美企业在 AI 工具上的月度人均支出中位数仅为 12 美元，而前 1% 的深度用户达 7000 美元；华尔街金融机构仅将其用工成本的 1% 投入生成式 AI；前沿模型所覆盖的 1000 万至 3000 万活跃知识工作者，相对于全球 15 亿潜在受众仍处于起步阶段。万亿美元级的未来机遇将逐步向具身智能机器人、自动驾驶、医疗研发以及重工业等物理世界全面扩散。 然而，AI 扩张的核心瓶颈已从纯算力与算法需求，转移至物理供给侧的“通电速度”（Speed to power）。Aram Verdiyan 阐明，美国的核心挑战不在于发电总量，而在于电网传输容量不足、行政许可漫长以及监管繁琐。由于新一代 AI 数据中心的功率与计算密度达到传统设施的 10 倍以上，旧有资产无法通过简单改造实现复用。能否通过在新型半导体架构、先进内存、电力基础设施和高密数据中心领域的资本注入突破这一物理瓶颈，将决定整个生态系统能否真正步入实体层面的“机器时代”。 当前生成式 AI 的商业化正从概念验证与软件替换，演进为对宏观劳动力负债表与物理能源设施的系统性重构。然而，鉴于复杂知识任务的扩散节奏比代码编写更为渐进，且美国电网改造与高密度数据中心建设面临漫长的审批周期与供电瓶颈，技术资本的注入能否在私募信贷泡沫出清前转化为广泛的生产力增长，仍是横亘在技术与实体经济之间的未决问题。

## Background 本期播客由 a16z 的 Jen Kha 和 David George 携手 Accolade Partners 的 Aram Verdiyan 于 2026 年 9 月 10 日通过 RSS 平台发布。节目主要探讨了人工智能（AI）如何彻底重写科技投资领域的幂律法则、为什么前沿超级企业能以史无前例的方式持续复利放大自身优势，以及机构有限合伙人（LP）与资产配置者应如何相应重构其核心投资组合。 ## TL;DR 对话深入剖析了 AI 时代资本直接转化为算力壁垒的全新增长飞轮，指出 AI 仅用 4 年便完成了传统 SaaS 耗时 15 年才达成的千亿美元收入体量，其潜在市场空间正从单纯的企业软件预算扩展至高达 30 万亿美元的庞大实体劳动力与 GDP 支柱。投资人们直面全美 3,000 家风投中仅有 20 家能持续实现 3 倍净回报的残酷现实，揭示了“中游风投机构之死”与全栈平台型风投的必然性。此外，节目还尖锐对比了传统私募股权（PE）与私募信贷在面对未重构软件资产时的接盘与杠杆危机，并展望了未来在机器人、自动驾驶、医疗及能源基础设施等物理世界中孕育的数万亿美元级颠覆性机遇。 ## Atmosphere & style 整场对话充满了一线顶级投资机构内部研讨的高密度思辨感与行业内幕洞察，既有冷静理性的数学逻辑和严苛的数据拆解，又不乏行业老手相逢一笑时的默契幽默（例如调侃聊起“中游机构之死”就得碰杯）。Jen Kha 的敏锐穿透力、David George 的一线实操判断力与 Aram Verdiyan 宏观配置视角的严密逻辑相互碰撞，将创投行业在巨大代际变革面前的急迫感、对虚假繁荣的清醒警惕，以及押注“机器时代”物理重塑的宏大野心淋漓尽致地展现出来。 ## Bulletin - `04:47` 传统 SaaS 耗时 15 年才达到 1,000 亿美元年收入体量，而 AI 仅用 4 年便达成同等规模，正直接渗透并重塑规模达 30 万亿美元的全球 GDP 底层支柱。 - `00:00` 前沿 AI 实验室打破了以往资本增加导致组织内耗的定律，巨额资本可直接兑换为算力基础设施，形成自我强化的物理规模报酬递增飞轮。 - `10:00` 全美 3,000 家风投机构中，过去 20 年间仅有 20 家能稳定取得 3 倍以上的净回报，这要求资产配置者必须摒弃过度分散策略、集中重仓头部平台。 - `38:00` 2021 至 2022 年间私募股权以 25 至 32 倍 EBITDA 收购的传统软件资产，正因缺乏 AI 原生重构能力与估值暴跌面临失控的杠杆率及赎回暗礁。 ## Key takeaways - `00:00` AI 颠覆了传统初创企业资本效率递减的规律，美元资本能够直接转化为算力资源，并以物理级别的确定性飞速放大模型与产品的技术护城河。 - `04:47` 科技投资的幂律法则正在走向极致，SpaceX、OpenAI 与 Anthropic 三者潜在价值已达 3.5 万亿至 5 万亿美元，促使 AI 必须成为机构的核心配置而非卫星边缘仓位。 - `06:40` 重新定义潜在市场空间（TAM）：宏观经济中的劳动力支出约是纯软件采购的 40 倍，AI 解决具体任务和劳动力重构的商业潜力远超上一代企业软件。 - `08:20` 技术赛道内部由单一领军者通吃绝大部分份额与市值，但 AI 正在催生出品类数量（Number of categories）的指数级爆炸，切忌陷入零和博弈思维。 - `10:00` 全美 3,000 家风投机构中仅有不足 1% 能够持续跨周期取得 3 倍净回报，行业平均表现平庸，LP 盲目分散配置将失去流动性锁定的风险补偿溢价。 - `13:00` 风投行业呈现两极分化加剧的“中游机构之死”，只有极敏捷的早期专业垂直基金与具备全生命周期赋能的巨型平台才能保持竞争力。 - `17:00` 大型平台型风投通过早期投资触角感知一线微观纹理与技术动向，以此反哺成长期阶段下重注并锁定超级赢家，早期与后期形成紧密互锁闭环。 - `20:00` 早期初创公司在加速器内部互买或将单月数据乘以 12 制造出虚假 ARR 繁荣，投资人唯一的验证手段是深潜一线客户对话并检验真实产品渴望度。 - `26:17` GP 最大的职业风险是错失超级巨头的“遗漏错误”，而传统 LP 往往受限于防守考核心态；在极端幂律资产中，LP 唯有对头部基金敢于集中重仓才能拉动整体收益。 - `30:00` 公开市场目前仅剩 15 至 20 家 SaaS 公司能维持 10 倍市销率以上的估值，缺乏内生增长加速和 AI 防御韧性的传统软件资产正面临严重的倍数压缩。 - `40:00` 传统 PE 试图通过“空降 AI 客服”等浮于表面打补丁的方式实现降本，若缺少彻底自杀式重构底层工作流的决心，极易引发客户流失与债务违约反噬。 - `34:30` 当前 AI 公司的爆炸式增长仅仅由几千万知识工作者驱动，企业端平均 AI 支出依然处于极早期水平，向更广泛商业任务扩散的长远空间巨大。 - `42:47` 顶尖风投更倾向于长期持有核心赢家以赚取数十年以上的复利红利，同时会在技术周期演进中通过早期并购和分阶段退出主动制造流动性。 - `44:37` 下一代 10 万亿甚至 100 万亿美元市值的超级巨头极大概率诞生于机器人、自动驾驶、医疗研发以及制造业和国防等重度物理世界赛道。 - `47:18` 当前 AI 发展最大的阻碍不是需求侧而是供给侧，美国在芯片审批之外，面临电网扩容审批慢、高功率数据中心建设等严峻的供电速度瓶颈。 ## Detailed walkthrough ### `00:00` 极端幂律时代的重塑：资本直接转化为算力壁垒 在风险投资行业，幂律法则（Power Law）一直统治着回报分布，但如今这股力量正在被推向历史顶峰。全美 3,000 家风险投资机构中，过去 20 年间仅有 20 家能够稳定实现 3 倍以上的净回报。如今，以 SpaceX、OpenAI 和 Anthropic 为代表的前沿公司，潜在企业价值已经合计达到 3.5 万亿至 5 万亿美元之间。令人吃惊的是，在这些庞然大物崭露头角甚至筹备走向公众市场之前，许多机构有限合伙人（LP）与主流资产配置者在这一领域几乎没有任何实质性敞口。传统上，给早期初创企业盲目注入过多现金往往是毁灭性的——一旦公司盲目扩张招募千名员工，随之而来的便会是严重的组织协调内耗、官僚开销与优先级冲突。然而，生成式人工智能（AI）彻底打破了这一定律。 ```mermaid flowchart TD accTitle: Capital to Compute Compounding Flywheel accDescr: 资本转化为算力并强化模型与产品优势的良性循环飞轮 subgraph CapitalEngine["资本驱动引擎"] A["巨额资本注入"] --> B["大规模采购算力资源"] end subgraph ProductAdvantage["产品与市场壁垒"] B --> C["前沿模型能力与产品体验飞跃"] C --> D["市场垄断级吸引力与收入暴增"] end D -->|"滚雪球式强化优势"| A ``` 前沿 AI 实验室的逻辑截然不同：美元资本不再被消耗在低效的人力膨胀中，而是可以直接兑换为算力基础设施，而更充沛的算力能够立竿见影地提升模型智能与产品性能，从而不断自我强化前沿阵营的护城河。规模报酬递增（Increasing returns to scale）在以往更多体现在消费级互联网的网络效应或企业软件的品牌声誉与资源积累上，但如今在前沿实验室，规模效应变成了物理算力上的确定性飞轮。 > “在我的职业生涯中，这是第一次：你可以直接把资本砸向一家公司，而这笔资本会不断复合放大他们的优势。” —— David George - 全美 3,000 家 VC 机构中，仅有 20 家在过去 20 年中持续取得 3 倍以上的净回报（3x net returns）。 - 当前科技投资中的幂律分布比过去 10 到 20 年更为极端，甚至超过了早期消费互联网爆发的时代。 - SpaceX、OpenAI、Anthropic 三家公司所代表的潜在企业价值已达到 3.5 万亿至 5 万亿美元规模。 --- ### `04:47` 4年走完SaaS 15年之路：从边缘配置迈向超级核心配置 回顾整个软件行业的发展历程，传统 SaaS 行业耗费了整整 15 年才实现 1,000 亿美元的年收入规模，而 AI 浪潮仅用了区区 4 年就达成了这同一里程碑。更重要的是，市场目前仍处于极早期，对模型推理（Inference）的需求几乎深不见底。技术历史上的任何一次浪潮都无法与当下匹敌——AI 并非仅仅局限于科技圈内部的工具升级，而是在全面侵入、重构全美乃至全球高达 30 万亿美元的 GDP 基础支柱，横跨交通运输、劳动力、专业服务、资本运作以及组织协调等各个环节。面对如此惊人的渗透速度与广度，机构投资者若再将 AI 视作投资组合中的边缘卫星仓位（Satellite position），在逻辑上已经完全立不住脚。 作为资产配置者，必须清醒地认识到风险投资与成长型股权投资领域的底层资产结构已经永久改变。大约 18 年前，科技投资还只是另类资产中极小众的“手工业”（Cottage industry），如今整个资产类别的规模已膨胀至 5 到 6 万亿美元。企业在私募市场保持未上市状态的时间大幅拉长，这种趋势根本不可能发生逆转。过去，顶层十分之一分位数（Top decile）的超级大案往往在市值达到 100 亿美元左右时就已见顶，随后走向公开市场；而在当今的晚期私募市场，单个赢家的规模动辄达到 400 亿美元，并且随着前沿实验室的演进，这一门槛很快将被推高至 1,000 亿美元以上。 > “在以往，没有任何一个技术范式能够同时冲击 30 万亿美元的 GDP 蛋糕。” —— Aram Verdiyan > “你面对的是人类历史上增长最快的技术，它正在全面击穿 GDP 的各个板块，因此作为资产配置者，你很难不得出结论：它绝对不能只是一个边缘仓位，它必须在某种形式上成为投资组合的核心、乃至超级核心。” —— Aram Verdiyan - SaaS 达到 1,000 亿美元总收入规模用了 15 年，而 AI 仅用了 4 年。 - 上一轮科技周期创造了 25 万亿美元的新增市值，而由 AI 驱动的新周期预计创造的总体体量将远远超过这一数字。 - 私募市场顶级头部退出案例的估值体量已从过往的 100 亿美元跃升至约 400 亿美元，且即将迈向 1,000 亿美元时代。 --- ### `06:40` 重新计算 TAM：以重塑万亿劳动力为核心的扩张市场 许多市场参与者在评估 AI 的潜在市场空间（TAM）时往往犯下经验主义错误，将其狭隘地等同于上一代企业软件或医疗 IT 系统。以医疗行业为例，全美每年在传统医疗 IT 上的采购支出大约在 600 亿到 1,000 亿美元之间，但这仅仅是购买软件授权的预算；而 AI 真正瞄准的不是软件采购池，而是具体的劳动力和实际执行的任务价值。医疗流程中的保险索赔、计费结算、繁重的行政文档处理，是一个价值万亿美元的庞大劳动力市场。如果技术解决的是执行具体任务的经济价值，那么 AI 公司的目标市场空间将是传统企业软件的 10 倍以上。 全美经济体在劳动力上的总支出大约是软件支出的 40 倍。这并不意味着 AI 会瞬间让所有劳动力凭空消失，相反，人类的工作任务将被彻底重塑与重新想象。正如法律界当前的真实反馈：许多律师事务所接入 Harvey 等 AI 法律助手后，客户由于自己使用了 AI 检索，误以为自己也能看懂复杂的法律法条，反而跑来与合伙人就各类细节深入辩驳；结果原本客户会直接放手授权的普通事务，现在转变成了极高频的专业沟通，律师事务所的计费工时（Billable hours）非但没有减少，反而因为 AI 的普及而大幅上升。AI 带来的效应从根本上是极度扩张性的（Expansionary），远非简单的存量替代。 > “如果仅看美国经济在劳动力与软件上的支出对比，劳动力支出大约是后者的 40 倍。” —— David George > “因此，把 AI 简单等同于软件的下一代演进，这种看法实在是太有局限性了。” —— David George - 全美医疗 IT 年支出为 600 亿至 1,000 亿美元，但其背后的行政、计费、索赔等相关劳动力支出达数万亿美元。 - 整个宏观经济中，劳动力支出规模大约是纯软件采购支出的 40 倍左右。 - 法律 AI 工具 Harvey 的落地案例证明，AI 的普及不仅没有降低顶尖专业服务的需求，反而因为拉平了客户与专业人士的交互门槛而显著增加了可计费工作时长。 --- ### `08:20` 拒绝零和思维：单一品类的极端垄断与品类数量的指数级裂变 当前投资圈内充斥着针对 AI 架构各层级的焦虑：前沿模型实验室会不会向下吞噬垂直应用层？开源模型的快速进步是否会扼杀专有闭源实验室的商业化？然而，现实呈现出的却是整个生态系统各层级都在以惊人的速度同步狂奔。如果将技术投资视作零和博弈（Zero-sum game），就会完全错失真正的大机会。在这场巨变中，所谓“赢家通吃”（Winner-take-all）必须被放在正确的维度下去剖析：在某一个明确定义的技术品类内部，幂律法则确实极端残酷，头号赢家会掠走绝大部分市场份额与最终市值，第二名往往只能沦为残羹冷炙的争夺者；但技术革命的核心魅力，在于它会催生出品类数量（Number of categories）的爆炸式增长。 二十年前，客户关系管理（CRM）几乎算不上一个独立且庞大的主流赛道，市场内仅有 Siebel Systems 等少数参与者，规模极为有限；而如今，CRM 已经裂变并成长为一个庞大无比的独立产业生态。在 AI 时代，相同的品类裂变规律将以更剧烈的形式再次上演。对于风险投资而言，核心能力在于对失败和亏损保持极高的容忍度——在技术大爆发的周期里，如果不犯错、不承担投资失败的风险，反倒意味着投资人没有将筹码压在那些潜在拥有极端非线性回报的突破性赛道上。 > “在给定的特定细分品类中，我们绝对是幂律法则的坚定信徒：胜出者将攫取绝大部分的市场份额与市值，第二名只能去抢面包屑；但与此同时，我们将会看到品类总量出现极其巨大的指数级扩张。” —— David George > “从大体上看，认为‘如果开源做好了就必然对闭源实验室不利、反之亦然’的想法，实在太过局限。我们竭尽全力避免让自己陷入这种零和博弈的思维定势中。” —— David George ### `10:00` 风险容忍度、幂次法则与“3000家仅20家”的残酷真相 在风险投资的数学逻辑中，高失败率不仅是不可避免的副产品，更是判断一家机构是否承担了足够风险的试金石。David George 直言，如果在一定数量的投资中失败率不够高，说明机构承担的风险根本不够。回顾历史，在 a16z 表现最顶尖的几期早期风险投资基金中，损失率往往高达 60% 左右。进入成长期阶段（Growth Stage），损失率虽然会有所下降，但通常也维持在 10% 到 20% 的区间。这种适度的损失是合理的，因为只要承担了这种程度的风险，组合中就会跑出带来 10 倍甚至更高回报的超级赢家。David 总结道，做投资的关键就在于押注每一个具备可行性的赛道中的领军企业，如果赛道爆发，基金就会大获全胜；如果赛道落空，这也是投资人必须接受的正常风险。 围绕赛道热度与入场时机，Jen Kha 指出，投资界常陷入一种心态怪圈：人们总在某个当下感慨市场过热、估值太贵，但多年后回过头看，却发现当时的资产其实相当便宜。然而，市场上也确实存在真正的周期性异常泡沫。因此，机构出资人（LP）在风险投资这一资产类别中保持出资的一致性（Consistency）至关重要，因为没有人能够准确预判底层技术何时会实现跃迁。 ```mermaid flowchart TD accTitle: Top VC Persistence Funnel accDescr: 展现全美 3000 家风投机构中，仅有不足 1% 能够实现跨越 20 年的连续 3 倍净回报 subgraph AllVCs ["全美风投机构整体现状"] A["3,000 家美国风投机构"] B["行业 10 年平均净回报仅 1-2x"] A --> B end subgraph TopTier ["跨越 20 年的极端幂次法则"] C["仅 20 家机构 (<1%) 实现持续 3x 净回报"] D["核心特征：跨越周期持续锁定各年份定义赛道的超级赢家"] C --> D end B -.->|"极端离散度 (Dispersion)"| C ``` Aram Verdiyan 随即引入了一组极为残酷的行业调研数据： - Accolade Partners 调取并分析了全美 3,000 家风险投资机构的历史业绩。 - 在长达二十年的时间跨度里，能够持续实现 3 倍净回报（3x Net DPI/TVPI）的机构仅有 20 家，占比不到 1%。 - 这 20 家机构并不需要在 20 年间连续募集七八期基金都达到 3 倍，只要能交出 3 到 4 期 3 倍净回报的基金，就已经属于凤毛麟角。 正如 Aram 所强调的： > “Consistency in venture is really, really hard.”（在风险投资里，保持业绩的一致性简直难如登天。） Aram 指出，能进入这 20 家顶尖行列的机构，核心秘诀在于它们在每一个年份（Vintage）都能持续拿到那些定义赛道（Category-defining）的超级赢家份额。仅仅拿到 Logo 毫无意义：对于早期基金而言，基金规模再大也必须持有足够高的股权比例；对于后期成长期基金而言，头寸规模管理（Portfolio Sizing）则是生死线。顶尖成长期基金的最大单一标的必须占到基金总规模的 5% 到 10% 以上，只有这样，单笔投资才有可能还完整期基金（Fund Returning）。根据 Cambridge Associates 的数据，过去 10 年美国风险投资的行业平均净回报仅有 1 到 2 倍之间。如果 LP 拿不到最顶尖的那 20 家机构的份额，最终回报将远落后于私募股权（PE），甚至不如直接配置公开市场股票，白白承受长达 10 年的资金锁定流动性惩罚。 ### `13:00` 创始人偏好与“中游机构之死” 对话随后转向了风险投资基金规模为何会走向两极分化。Jen 援引了业内知名投资博主“Endowment Eddie”当天发布的一条推文：“大额风投基金的繁荣是由创始人推动的，而非 LP。创始人往往更渴望那些能够陪同企业一同扩展规模、具备全生命周期投资能力、并且能协助招募高管与拓展客户的大品牌基金。”这一论述打破了外界普遍认为“大基金是由追求配置体量的 LP 倒逼催生”的传统偏见。 由于科技巨头的终局规模和市场潜能比以往大得多，能够承载的基金体量也水涨船高。尽管在顶尖的 20 家机构中，依然存在极少数专注于利基垂直市场（Niche Verticals）或在极早期阶段深耕的精品基金，但这类小型基金的生存前提是必须严格克制自身规模与策略；一旦为了追求管理费而盲目扩大基金规模，就会迎面撞上全平台型大型风投，并在竞争中迅速丧失一致性优势。 三人笑称他们每一次聊到“中游机构之死”（Death of the Middle）这个词都恨不得举杯喝酒。David George 详细剖析了他眼中的“中游”困境与行业分层： - **微型与深度专业基金**：深耕垂直 AI 或特定技术赛道的领域专家（Deep Domain Experts），他们反应极其迅速，往往能在巨头尚未反应过来前率先咬下优质份额。 - **超大规模平台型风投（Large Scale Venture）**：以 a16z 及其少数同梯队机构为代表，拥有多条主力产品线，资金与能力覆盖从 Seed 轮一直到 IPO。 - **困境中的中游机构**：介于两极之间，既没有极早期的敏捷专业度，又缺乏全方位的平台赋能与跨周期开支票的能力，在争夺优质项目时处境极为艰难。 ### `15:00` 飞轮效应：全生命周期平台如何通吃优质标的 在深入探讨创始人的心理机制时，David George 用一句话给出了底层总结： > “The founder cares about taking capital from a partner that they think can de-risk the outcome for themselves.”（创始人真正关心的，是引入一个能够为他们自己的创业结果降低风险的合伙人。） 创始人固然在乎合伙人的个人品质，但这也是 a16z 不惜将收取的管理费全部反哺到运营资源、建立起拥有 700 名员工庞大体系的根源所在。这种平台资源具备双重功效：第一，切实帮助被投企业“改变命运曲线”（Bend the curve on the outcome）；第二，在面对白热化竞争的明星项目（Hot Deals）时，成为打动创始人的杀手锏。 ```mermaid flowchart LR accTitle: The a16z Platform Flywheel accDescr: 展现 a16z 依靠专业认知、资源赋能与品牌网络构建的正向飞轮 A["深厚专业认知<br/>与明确行业主张"] --> B["赢得创始人认可<br/>拿下核心 Deal"] B --> C["700人运营平台赋能<br/>降低风险/改变发展曲线"] C --> D["打造行业赢家<br/>产生标杆背书与飞轮效应"] D -->|"品牌声誉传导 / 新创始人集聚"| A ``` David 阐释了这一商业飞轮的运转机制： 1. **认知与赢单**：合伙人凭借深厚的行业认知与洞见，向创始人证明自己是最佳搭档，从而拿下份额。 2. **赋能与标杆**：通过平台支持助力企业脱颖而出，进而产生极强的品牌信号效应（Brand Signaling）。 3. **口碑与循环**：后续的顶尖创始人天生渴望与“赢家”站在一起，而老创始人的口碑背书（Killer References）会持续将新一代天才创始人源源不断地推向该平台，从而确保了超额业绩的持续性（Persistence of Returns）。LP 的出资只是该飞轮高效运转下的必然副产品。 ### `17:00` 从 Pre-Seed 到成长期：早期与后期投资的双向哺育 针对大机构与早期基金的生态共存问题，Aram 提出了一个命题：估值在 2000 万到 4000 万美元以下、基金体量低于 1 亿美元的 Pre-Seed 轮小型基金，是否能与大型机构共生？在项目的起跑线上，往往有六七家初创团队在做完全相同方向的 AI 探索。对于像 a16z 这样必须锁定最终赛道冠军（Category Winner）的巨型基金而言，在最早期往往很难判断谁能胜出，若盲目重仓极易沦落到第二或第三名；因此大机构可能更愿意等待一两轮融资，待行业格局相对明朗后，直接在 A 轮或 B 轮领投并加倍下注。这就为早期基金留出了独特的生态位。 对此，Jen Kha 表示认同，但强调这种界限正在变得模糊。a16z 不仅与整个创投生态中的 Seed 基金保持着健康的合作，本身也在涉足种子阶段投资。Jen 指出，a16z 设立的 Speedrun 创业加速营，正是把触角伸向极早期的典型体现： - 优秀的创始人对顶尖品牌有着本能的优先依附倾向（Preferential Attachment），他们希望尽早进入 a16z 的引力场轨道（Orbit）。 - 即使某些项目在极早期阶段 a16z 并不出手领投，早期团队也必须抢先站在最前沿（First to the Pole），唯有全面透彻地理解生态地貌与技术趋势，未来在后期出手时才能做出最精准的判断。 David George 强烈认同早期与成长期的双向共生关系，并指出这正是单纯的“中间阶段”投资机构无法立足的原因： > “I said that all the time. Like, our business starts and ends with early stage. And so that provides us a tremendous amount of advantages at the growth stage in terms of access, information, knowledge, relationships, etc.”（我一直这么说：我们的业务始于早期，也终于早期。这为我们在成长期阶段带来了关于准入渠道、信息感知、行业知识与人脉关系等维度的巨大优势。） 早期业务赋予了成长期基金无与伦比的信息与人脉网络；而成长期基金的资金体量，又让机构能够长久伴随优秀创始人不断下注，从而反向赋能早期合伙人在争夺 Seed 和 A 轮案子时的说服力。这种“两极互锁、向少数顶尖品牌收敛”的趋势，正在重塑整个科技创投的格局。 ### `20:00` 早期与后期投资的协同，以及风投资产配置的四种形态 讨论从后期基金的打法展开。David George 指出，风投领域最强大的后期品牌背后，几乎无一例外都有一个庞大的早期业务在支撑。当一家基金有能力在自己规模庞大的后期基金中，将 5% 到 10% 的资金重注于一家定义赛道的领军企业时，之所以能做到这一点，往往是因为早期的团队很早就与创业者及其核心高管建立了极其深厚的互信。如果一个团队只是凭空建立一个纯后期基金（de novo late stage firm），想要平地起惊雷、在激烈的竞争中直接开出一张 5 亿美元的支票，几乎是不可能的。 从有限合伙人（LP）的资产配置视角来看，Aram Verdiyan 将风投行业清晰地拆解为四种玩法：第一类是种子前与种子轮（Pre-seed / Seed），通常是规模在 1.5 亿美元以下的小基金，仅在全美就有接近 2000 家；第二类是所谓的“混乱中间层”（Messy middle），包含数百家规模适中的机构；第三类是头部的全栈式大机构（Big firms）；第四类则是纯后期基金（Dedicated late stage）。Accolade Partners 在配置上历来重仓大机构和精选种子基金，对“混乱中间层”极为挑剔，而完全避开了纯后期基金。 然而，AI 的爆发正在让投资人的评估工作变得前所未有地棘手。融资规模越来越大，节奏越来越快，行业中充斥着令人困惑的业务进展（traction）。一个典型的场景是：某家创业公司刚刚从孵化器走出来，便号称在短短一个月内实现了从零到 500 万美元的“ARR”（年化经常性收入）。然而，这类公司既没有经历过完整的客户续约周期，很多时候收入甚至只是在同一个加速器生态内互相买单，或者是把单月偶然的暴增直接乘以 12 伪装成 ARR。 在九家制造虚假繁荣的公司背后，可能只有一家真正拥有数百万美元扎实 ARR、估值高昂但未来能长成下一个 Cursor 的划时代企业。对于普通机构而言，穿透这些噪音几乎是不可能的；这也是为什么大机构往往表现更好，因为它们有足够的资本耐受力和时间去验证确定性。Aram Verdiyan 特别以 a16z 领投的代码生成工具 Cursor 为例，指出即便在交易宣布的当天早晨，外界依然有人唱衰 Cursor 已经死了，质疑为什么 a16z 愿意在约 4 亿美元的高估值下押注。 > “Like the ability to win is multiplied when you have an early stage franchise.” —— David George ```mermaid flowchart TD accTitle: Four Segments of Venture Capital Allocation accDescr: Diagram showing the four ways to play venture capital from an LP allocation perspective. subgraph EarlyStage["早期投资生态"] PS["种子前/种子轮基金<br>(规模 < 1.5亿美元，全美约 1000-2000 家)"] end subgraph MiddleAndLate["中后期与全栈"] MM["混乱中间层 Messy Middle<br>(数百家基金，定位易受挤压)"] BF["全栈式头部大机构 Big Firms<br>(覆盖全生命周期，资本与认知协同)"] DLS["纯后期基金 Dedicated Late Stage<br>(缺乏早期源头，开大额支票极难)"] end EarlyStage --> BF MM -.-> BF ``` --- ### `23:44` 识别虚假数据与真实需求：从 Cursor 到 Harvey 的信号演进 面对市场上让人眼花缭乱的指标膨胀，David George 坦言，识别人性与创始人判断力（Founder judgment）才是最核心的能力。对于早期投资合伙人来说，长期深入陪伴创始人、洞察其思考方式是不可或缺的基本功。行业里每隔一段时间就会跳出来重新定义什么是 ARR，但这往往是市场脱离实际的遮羞布。David George 强调，他的电脑屏幕上始终贴着一张便签纸，上面写着最根本的问题：“市场是否真的在渴望更多你的产品？” 当一家初创企业刚刚成立或开始销售仅仅几个月时，传统的财务报表和数据分析几乎是完全失效的。投资人无法仅靠财务模型得出结论，唯一的办法是跳进泥潭与一线客户深入对话，感知真实的使用情况。David George 以 legal tech（法律科技）领域的明星企业 Harvey 为例： - **早期阶段：** 创始人是技术研究者与资深律师的组合，商业拓展能力极强，成立之初便迅速签下了多家声名显赫的顶级律师事务所。然而如果穿透账面销售数据、深入看实际部署与高频使用情况，其表现其实相当平庸。客户虽然买了，但使用率不高。 - **转折点（推理模型出现）：** 随着具备高级推理能力的底层大模型问世，Harvey 迎来了爆发性的拐点。产品能够为律师解决极其繁复的专业工作，实际使用量和用户参与度直接起飞。 - **市场心态反转：** 市场态度从最初的“担心 AI 幻觉不敢用”，彻底转变为“所有委托客户强制要求律所必须使用 Harvey 提高效率”。 David George 总结道，任何普通分析师都可以做留存分析（cohort analysis）和续约率测算，但真正的顶级投资人必须深潜一层，理解市场的微观纹理、客户真实的痛点与备选方案。这也是必须拥有早期投资业务的原因，因为只有站在技术与产品最前沿的一线团队，才能早在趋势完全明朗之前捕捉到 Cursor 和 Harvey 这类非共识机会。 > “That is always the question that's a post-it note on my computer screen: is the market demanding more of your product?” —— David George --- ### `26:17` LP 与 GP 的激励错位，以及风投组合的极致集中原则 对话随后转向资金募集的现实博弈。Aram Verdiyan 向 a16z 负责募资的合伙人 Jen Kha 发问：面对当前的 AI 狂热与风投市场，LP 最大的顾虑和反弹究竟是什么？Jen Kha 指出，LP 最普遍的担忧是“自己是否在接一把正在下坠的飞刀”（catching a falling knife），担心市场整体估值过热、进入的时机不对。 更深层次的问题在于，LP 机构内部的激励机制与 GP 之间存在着巨大的结构性错位： 1. **错失风险 vs 犯错风险：** GP 的饭碗取决于能不能投中下一个 Facebook 或 Uber，错过定义时代的超级巨头是不可原谅的“遗漏错误”（error of omission），是足以导致合伙人被解雇的致命伤；然而对大部分机构 LP 而言，由于遵循经典的基准线防御心理，投资稳妥平庸的资产哪怕收益平平也不会丢掉饭碗（正如“没有人会因为采购 IBM 而被解雇”），只有主动投资了暴雷的项目才会面临丢工作的风险。 2. **上行收益不对称：** 对很多大型主权基金、养老金这类 LP 来说，单笔风投带来的超额上行倍数在庞大的总盘子里感知微弱，因此“你会错过下一代造富神话”的说辞很难真正打动他们。GP 必须帮助 LP 认识到 AI 对其全盘资产组合（不仅限于风险投资，还包括二级市场、私募股权等）的颠覆性重构。 Aram Verdiyan 进一步拆解了专业 LP 的三大核心职能：准入能力（Access）、标的选择（Selection）和投资组合构建与仓位把控（Portfolio construction sizing）。风投行业是一个极端符合幂次法则的领域——在全市场 3000 多家风投机构中，真正能持续创造超额回报的不过 15 到 20 家。如果一个 LP 的风投组合分散投资了 50、60 甚至 70 家风投机构，其整体回报必然被拉平至行业平均水平。而风投行业的平均回报在流动性锁定、高亏损率的风险对比下，根本无法与成熟的公开市场或 PE（私募股权）相比。 Aram Verdiyan 痛斥了常见的组合构建硬伤：许多 LP 好不容易挖掘出了一只极具潜力的黑马基金，认知完全正确，却谨小慎微地只投了总盘子 1% 的仓位。结果就算这只基金未来实现了 10 倍的神话回报，折算到整个基金上也仅仅贡献了区区 10% 的总收益，“根本无法真正撼动回报的指针”（not moved the needle at all）。在风投这种极度非线性的资产类别中，看准顶级基金并敢于集中重仓，才是 LP 跑赢大盘的唯一解。 > “A GP can get fired for missing out the next Facebook, the next Uber, right? Like that is the error of omission and that's fireable. But LPs on the flip side only get fired if you invest into a mentor.” —— Jen Kha > “I've seen too many times an LP or an allocator find an interesting fund, actually get it right, and put 1% of their fund into it. Great, you 10x'd it, it returns 10% of your fund. It has not moved the needle at all.” —— Aram Verdiyan | 评估维度 | 全球 GP（风投基金合伙人） | 传统机构 LP（出资人/配置方） | |:--- |:--- |:--- | | **主要职业风险** | **遗漏错误**：因保守而错失下一个定义时代的超级巨头（如 OpenAI、Facebook） | **实施错误**：主动押注了高风险创新项目最终暴雷蒙受损失 | | **考核基准心态** | 必须追逐极端幂次法则下的极少数颠覆性赢家 | 只要跟住基准线或略低于基准线，就能保住工作 | | **上行收益感知** | 极高（Carry 与基金品牌直接挂钩） | 相对平淡（超额回报难以显著拉动整个庞大资管总盘） | | **最佳组合策略** | 深入早期技术一线，通过主权洞察重注核心项目 | 放弃过度分散（50+家），集中重仓前 15-20 家顶尖 GP | ### `30:00` 资产配置的倾斜：公开市场、PE 与风投中的 AI 溢价 当讨论到机构投资人整体资本在风投/成长型投资（Venture & Growth）与私募股权（PE）、公开市场或实物资产之间的分配比例时，专注风投的投资人们旗帜鲜明地主张加大对风投的配置（“supersize”）。在公开市场中，通过评估上市公司的“AI 韧性（AI resiliency）”可以清晰发现：SaaS 板块目前仅有最多 15 到 20 家公司的交易估值倍数能维持在 10 倍市销率（10x revenue）以上。而在数年前，这样的公司比比皆是。如今能维持高估值的少数标的，几乎无一例外都展现出了 AI 驱动的业务加速（如监控、安全、AI Agent 部署等）。 这种“AI 联动性（connective tissue）”已渗透进所有资产类别。即便在传统私募股权领域，投资人也不再有兴趣收购仅增长 10% 的老派按席位付费（seat-based）的工作流软件；他们渴望的是具备 AI 原生管理团队、能展现出增长加速的核心记录系统（system of record）。然而，风投仍然是捕捉这一趋势最强劲的工具，但这高度考验投资机构的筛选能力、投资额度获取权（access）以及投资组合构建策略。 - **公开市场 SaaS 估值挤压**：仅剩 15–20 家头部公司交易倍数在 10 倍营收以上，其余传统标的大幅回落。 - **PE 投资逻辑转变**：抛弃低增速的席位制工作流工具，转向可由 AI 赋能并重拾增长的底层平台。 > Aram Verdiyan 指出：“So it goes back to the same principle where if you are in some shape or form tied to AI... which is the fastest growing facet of all the elements of the GDP, then you should supersize that in your portfolio.”（归根结底原则都是一样的：如果你以某种形式与 AI 深度绑定——而 AI 是目前整个 GDP 中增长最迅猛的板块——你就应该在投资组合中超配它。） --- ### `32:00` PE 的退出困境与软件资产的“自杀式重构” 风投领域的退出规模正在颠覆传统认知，甚至远超私募股权。回顾当年的大体量交易，PE 领域的大型退出包括对 EA（约 500 亿美元）和 Medline（约 500 亿美元）的杠杆收购，但部分风投支持的项目（例如 Cursor 并入 SpaceX 等并购）规模已超越传统的 PE 交易。 反观 ChatGPT 问世前（2016 至 2021 年 vintage）由 PE 收购的大量软件资产，当时买方支付了 15 到 20 倍 EBITDA，而其增长率仅有 10% 到 20%。如今，这些资产在公开市场的估值仅剩 2 倍市销率（2x revenue）。更致命的问题不仅是估值缩水，而是**失去了接盘者（buyer）**。当前评估软件公司的第一要素就是其面对 AI 颠覆时的“终值（terminal value）”和韧性，而最好的证明就是内生性收入加速。公开市场数据显示，**1 个百分点的营收增长对市值的拉动相当于 3 个百分点的 EBITDA 改善**。与 2021 年狂热期或疫情期间对“必须盈利”的反转要求相比，现在的市场定价完全与风险承受能力和高增长率绑定。 绝大多数传统 PE 收购的软件企业既缺乏高增速，也缺乏重构业务的管理团队。像 Intercom 这样的例子是凤毛麟角：创始人重掌帅印，全面翻新业务底座，打造出 AI 原生产品并重新规模化。对于追求稳健现金流的 PE 模式而言，这种为了适应 AI 而必须主动“杀死现有业务（suiciding your existing business）”的激进变革，在操作上极度艰难。 ```mermaid flowchart TD accTitle: 软件资产在 AI 冲击下的两种命运演变 accDescr: 对比传统 PE/SaaS 模式的估值衰退与 AI 原生转型策略 subgraph Legacy["传统/前 ChatGPT 软件资产"] A["高倍数收购 (15-20x EBITDA)"] --> B["增长停滞 (10-20% 甚至更低)"] B --> C["估值倍数严重压缩 (降至 2x Revenue)"] C --> D["缺乏接盘方 / 退出陷入停滞"] end subgraph Pivot["激进转型路径 (如 Intercom)"] E["创始人回归 / 换血核心团队"] --> F["主动淘汰/自杀式改造旧业务"] F --> G["打造 AI 原生产品架构"] G --> H["内生增长加速 / 估值倍数重获溢价"] end ``` > David George 谈及那些敢于彻底颠覆自我的创始人时感叹道：“It's an end of one right now. But there's a bunch of really good founders who are capable of those businesses, public markets, private markets, who I think are going to take a crack at it.”（目前这确实是极其罕见的孤例。但无论在公开市场还是一级市场，都有一批非常有实力的创始人有能力驾驭这类企业，我相信他们会放手一搏。） --- ### `34:30` 遗留 SaaS 的真实处境与 AI 扩散的初级阶段 面对投资人手中大量在 2016–2021 年投资、年增速 30% 左右、风投账面估值仍挂在 10–20 倍营收却无法上市且无 PE 愿意接盘的遗留 SaaS 资产，真实的答案是“前景仍未定（very TBD）”。 一方面，AI 是代际级的工业重塑；但另一方面，那些能够积极求变的软件公司依然具有沉淀价值。当前 AI 在代码生成（coding）领域的爆发虽然惊人，但代码任务具有三项特殊属性：**完美的数据、可验证性（verifiable）以及可模拟性（simulatable）**。然而，商业社会中绝大多数白领工作并不具备这些特征，这意味着 AI 向更广泛知识工作的扩散速度可能会比编程慢得多，这也给了一部分传统软件公司调整商业模式、融合 AI 的自救窗口。 但从长期来看，AI 的渗透程度依然处于极端早期： - **渗透深度极度不均**：美国普通企业员工在 AI 上的支出中位数仅为每人每月 12 美元；而前 1% 的顶尖先进组织每人每月的 AI 支出高达 7,000 美元。 - **行业渗透刚起步**：即便是走在最前沿的投资银行，目前在 AI 工具上的开支也仅占其人力成本的 1% 左右。 - **庞大的人口基数红利**：当前驱动这些历史级增速 AI 公司的，仅仅是 1,000 万到最多 3,000 万活跃用户；而全美拥有 15 亿人次的知识工作任务体量，长远空间极其惊人。 大型机构也在激进调仓，例如加州公务员退休基金（CalPERS）曾因未充分配置风投而错失数十亿美元收益，如今正大幅扭转资产组合比例（将其传统资产比例从 91% 大幅削减至 58%，并将风投和成长型投资的配置从 9% 激增至 43%）。投资人切忌陷入非黑即白的“零和博弈（zero-sum thinking）”思维，传统平台依然拥有可被剥离或转化的资产价值。 - **AI 支出断层**：中位数企业仅 $12/人/月，前 1% 领先企业高达 $7,000/人/月。 - **CalPERS 资产大腾挪**：风投及成长型资产配置比例从 9% 强势提升至 43%。 > David George 在剖析知识工作的复杂性时指出：“Coding is perfectly documented, right? So it has like, perfect data. It's verifiable and it's simulatable, right? And so like most tasks in business do not share those three attributes.”（编程有着极度详尽的文档沉淀，具备完美的数据。它是可被验证的，也是可被模拟的。而商业环境中的绝大多数任务，并不具备这三个属性。） --- ### `38:00` 私募信贷的暗礁：高杠杆遇上软件“CES 浩劫” 在过去十年中，除风投、PE 和公开市场外，私募信贷（Private Credit）作为一种全新的重要资产类别迅速崛起，其中相当大的规模沉淀在 PE 收购的软件资产组合中。 在 2021 至 2022 年间，全市场发生了约 2,000 亿至 3,000 亿美元的软件 LBO（杠杆收购）交易，伴随了超过 2,000 亿美元的高额负债融资。当时这批软件公司的收购估值倍数普遍高达 25 到 32 倍 EBITDA。然而，经历了软件板块的估值崩塌（被称为“CESpocalypse”般的剧烈回调），这批资产当下的实际价值可能已经腰斩。软件估值的剧烈收缩直接导致实际杠杆率（leverage ratios）失控飙升，这也正是目前私募信贷市场开始出现资金赎回潮的核心导火索。 无论是在股权层还是债权层，如果一家软件企业无法证明自身对 AI 的防御韧性，其面临的偿付与估值压力都是致命的。更为关键的是，即便 PE 试图给传统业务贴上 AI 标签，也无法实现真正救赎——正如当年“把传统百货 Sears 搬上网，并不能让它变成亚马逊”一样，缺乏原生重构的修修补补并不能免疫被颠覆的风险。 - **LBO 债务隐患**：2021–2022 年软件收购伴随超 2,000 亿美元负债，当时估值倍数达 25–32x EBITDA。 - **杠杆率被动走高**：底层软件资产价值折半，导致私募信贷机构面临资产穿透后的严重偿债风险与赎回压力。 > Aram Verdiyan 针对当前软件债权市场的脆弱性警示道：“If you are a software company that is in somewhat not resilient to AI, I think you're challenged both in terms of your equity position, also credit as well.”（如果你是一家在面对 AI 时缺乏韧性的软件公司，我认为无论在股权端还是债权端，你都将面临极其严峻的挑战。） ### `40:00` PE 模式的 AI 困境与风投流动性周期之辩 将 AI 技术生硬地嫁接到现有企业中并非万能药，尤其是对于私募股权（PE）杠杆收购的企业而言。Jen Kha 指出，亚马逊之所以强大，是因为它在物流等底层基础设施上从零做起，而不仅仅是做了一个前端网站。许多被 PE 收购的企业试图摘取“低垂的果实”，例如第一步就是用 AI 客服替代人类团队。然而事实证明，如果底层业务工作流没有真正重构，习惯与真人沟通的客户会迅速流失；净推荐值（NPS）的每一下滑，都直接对应着营收的锐减，对于背负债务杠杆的企业来说，这极易引发致命的恶性循环。David George 补充道：“你不能只是空降一个运营合伙人到公司，然后说‘让我们把 AI 装上去’，这根本行不通。”（“Yeah, you can't just throw an operating partner at the company and say, let's put AI on it. It just doesn't work.”）只有当管理层具备真正的创始人思维、董事会和所有投资人目标完全一致，并且愿意像 Intercom 那样做出艰难的彻底重构时，变革才有可能发生。 面对风投与成长型投资面临的合理质疑，Aram Verdiyan 坦言最常见的阻力在于“流动性周期”（timelines to liquidity）。普通独角兽企业在私募阶段停留的时间通常长达 10 年以上，后续多轮融资接踵而至，同一个明星项目的 Logo 会出现在五六家不同机构的投组中，LP 不免质疑资金何时才能退出。即便企业完成 IPO，也绝不意味着能立刻套现，往往需要再等 12 到 24 个月以上；若风投在上市时仍持有 10% 到 15% 的股权，退出过程会更加漫长。 对这一质疑的反驳则直接回到了风险投资的“幂次法则”本质：全行业约 3,000 家风投机构中，能持续表现优异的仅有 20 家左右。如果你身处这前 1% 的顶尖机构，手握某一赛道的终极赢家，你首要考虑的应当是让它的复利优势持续放大。David George 举例反问，三四年前会有谁愿意卖掉手中持有的 Stripe 或 Databricks 吗？答案显然是一致的“绝不”。 ```mermaid flowchart TD accTitle: AI 整合路径与回报分化 accDescr: 对比传统 PE 式打补丁与 VC 式底层重构带来的不同商业后果 subgraph PE_Approach["传统 PE 式打补丁"] A1["仅替换前端界面<br/>（如直接裁员上 AI 客服）"] --> A2["缺乏底层工作流重构"] A2 --> A3["客户体验崩盘 / NPS 下滑"] A3 --> A4["营收缩水 + 债务杠杆压垮"] end subgraph Native_Approach["原生 / 创始人驱动重构"] B1["从底层架构重写系统<br/>（如 Intercom 式坚决转型）"] --> B2["创始人思维 + 董事会高度协同"] B2 --> B3["构建长期防御壁垒"] B3 --> B4["10 年以上持续复利增长"] end ``` ### `42:47` 顶尖 GP 的退出策略与 LP 的诉求分歧 风投生态中，优秀项目的退出速度并非千篇一律。Anthropic 在 2021 年完成早期融资，大约 5 年左右就走向上市流程；而 Cursor 从最初融资到被收购的时间周期同样极其短促。顶尖风投机构完全有能力像 PE 一样甚至更快地向 LP 返还整个基金体量的流动性，但这仅限于极少数顶尖机构。 关于是“继续复利”还是“落袋为安”，不同性质的 LP 态度截然不同： - **大学捐赠基金（Endowments）**：通常倾向于长期持有，让标的继续奔跑复利。 - **家族办公室（Family Offices）**：出于避免触发高额资本利得税的现实考虑，往往不希望过早收回本金，宁愿资产在账面上继续复利。 - **a16z 的真实案例**：大约一年半前，a16z 针对其成立已有 16 年的 1 期基金（Fund 1）向 LP 发起征询，询问是否要清算在种子期就重仓持有的 Stripe 股份。“当时每一位 LP 的回答都是‘不，我们宁愿让它继续复利’。”（“And every single one of those LPs said, no, we'd rather let this continue to compound.”）直到基金存续期进入第 17 年，出于最终关闭老基金的合规需要，才必须执行清算退出。 顶尖的普通合伙人（GP）会在长周期中主动“制造”流动性。2021 年市场狂热期未能在高点套现是一个重要警示；而在当下的新周期中，优秀的 GP 正在通过早期的战略并购（M&A）实现部分本金落袋，同时保留核心赢家长期持有并走向 IPO，以此在获取阶段流动性与赚取终极复利之间取得平衡。 ### `44:37` 下一个十万亿乃至百万亿美元赛道 面对“谁会成为下一个市值达到 10 万亿甚至 100 万亿美元的公司”这一终极假设，David George 认为这种量级的跃迁很可能需要跨越两整个技术周期，其诞生的土壤必然是目前 AI 技术尚未真正触及的全新物理与实体领域。过去一年市场热炒消费级应用，如今消费级 AI 似乎无人谈及，但当前的聊天机器人界面只是早期的“拟物化”（skeuomorphic）过渡产品；未来原生的消费级 AI 将主动替人类执行任务。 David George 梳理了当前被严重低估、甚至“几乎还未破局”的深水区： 1. **机器人技术（Robotics）**：目前几乎处于起步阶段，但其商业规模在未来 10 年内将超越纯语言模型。 2. **自动驾驶（Autonomy）**：全美目前投入运营的 Waymo 尚不足 10,000 辆，无人驾驶出租车规模依然很小，存在巨大的空白市场空间。 3. **医疗健康（Healthcare）**：医疗支出占据美国 GDP 的 18%，但无论是在临床护理交付还是药物研发（drug discovery）领域，AI 创造的价值才刚掀开冰山一角。 4. **实体世界的全面重塑**：从国防军工、制造业到下一代超高密度数据中心，未来 10 年回顾当下，会发现真正的价值诞生于对物理世界的深度融合之中。下一个像 SpaceX 或 OpenAI 这样级别的超级巨头，极大概率会诞生于这些领域。 ### `47:18` AI 的真实瓶颈与“机器时代”的基础设施重构 Aram Verdiyan 针对当前 AI 生态的底层制约给出了关键诊断：当前 AI 的真正瓶颈根本不在于需求端，而是在供给端。从整个产业链路来看，左侧是能源、电网与数据中心，中间是芯片，右侧则是前沿模型与应用层。美国在右侧（芯片及以上应用）拥有全球领先的风投体系支持，但在左侧的能源和物理基础设施上面临严峻挑战。 美国并不缺发电能力，真正的痛点在于“供电速度”（speed to power）——这涉及繁琐的审批许可、输电网络扩建以及监管障碍；相比之下，其他国家每年的可再生能源新增装机容量可达美国的 10 倍。此外，新一代 AI 数据中心的功率密度提升了 10 倍以上，绝不可能简单地通过改造旧数据中心来直接适配。面对部分 LP 认为“这轮 AI 繁荣与互联网泡沫或疫情期间的虚假繁荣如出一辙”的质疑，Aram Verdiyan 明确反驳道：这绝不是疫情期间转瞬即逝的虚假收入，终端需求是完全真实且被验证的，只要能攻克供给端的能源基础设施、下一代芯片及新型存储等瓶颈，就会释放出不可限量的商业机遇，“机器时代（Machine Age）已经到来”。 ## Notable quotes > `00:00` 在我的职业生涯中，这是第一次：你可以直接把资本砸向一家公司，而这笔资本会不断复合放大他们的优势。 — David George > `04:47` 你面对的是人类历史上增长最快的技术，它正在全面击穿 GDP 的各个板块，因此作为资产配置者，你很难不得出结论：它绝对不能只是一个边缘仓位，它必须在某种形式上成为投资组合的核心、乃至超级核心。 — Aram Verdiyan > `08:20` 从大体上看，认为‘如果开源做好了就必然对闭源实验室不利、反之亦然’的想法，实在太过局限。我们竭尽全力避免让自己陷入这种零和博弈的思维定势中。 — David George > `04:47` 在以往，没有任何一个技术范式能够同时冲击 30 万亿美元的 GDP 蛋糕。 — Aram Verdiyan ## Mentioned David George, Jen Kha, Aram Verdiyan, a16z, Accolade Partners, SpaceX, OpenAI, Anthropic, Cursor, Harvey, Intercom, Stripe, Databricks, Waymo, Speedrun, Cambridge Associates, CalPERS, Electronic Arts, Medline, Siebel Systems, Endowment Eddie ## Video description a16z’s Jen Kha and David George sit down with Accolade Partners’ Aram Verdiyan to discuss how AI is changing the power law of technology investing, why the largest companies can compound advantages in ways that weren’t possible before, and what that means for how investors construct portfolios. They explore why AI may be much bigger than traditional software, with applications reaching into labor, healthcare, transportation, services, and other major parts of the economy. David explains why capital itself can now reinforce an AI company’s advantage by buying more compute, while Aram makes the case that AI should increasingly be treated as a core allocation rather than a satellite position. The conversation also gets into the changing economics of venture and growth investing, how to distinguish real AI traction from early hype, what AI means for legacy software and private equity, and why some of the largest opportunities may still be ahead in robotics, autonomy, healthcare, energy, and physical infrastructure. [Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/d0bf22e9-4ae1-41c6-949b-3d6333d7aa0b/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=d0bf22e9-4ae1-41c6-949b-3d6333d7aa0b&feed=JGE3yC0V)

[0:00] David George: 在我的职业生涯中，这是第一次：你可以直接把资本砸向一家公司，而这笔资本会不断复合放大他们的优势。 [4:47] Aram Verdiyan: 你面对的是人类历史上增长最快的技术，它正在全面击穿 GDP 的各个板块，因此作为资产配置者，你很难不得出结论：它绝对不能只是一个边缘仓位，它必须在某种形式上成为投资组合的核心、乃至超级核心。 [8:20] David George: 从大体上看，认为‘如果开源做好了就必然对闭源实验室不利、反之亦然’的想法，实在太过局限。我们竭尽全力避免让自己陷入这种零和博弈的思维定势中。 [4:47] Aram Verdiyan: 在以往，没有任何一个技术范式能够同时冲击 30 万亿美元的 GDP 蛋糕。

[0:00] David George: For the first time, you can take capital and throw it at a company, and it compounds their advantage. [44:32] David George: The end use case for consumers is not going to be a chatbot interface. That's the skeuomorphic version. We're gonna have a native version. It's gonna be proactive. It's gonna do work on our behalf. [46:48] Aram Verdiyan: Because the U.S. doesn't have a problem with energy generation. It has a problem with speed to power. That's permissioning, transmission, that's regulatory. [20:00] Aram Verdiyan: We’re big believers that the strongest late stage franchises have a huge early stage franchise attached to them. [27:30] Aram Verdiyan: Because if you have an asset class where 20 firms out of 3000 do well, you should concentrate in those 15, 20 firms pretty consistently.

[0:00] We've looked at the data of 3,000 venture capital firms in the U.S. Only 20 have achieved consistent 3x net returns over the last two decades. Right now, clearly, the power law is more extreme than it has been in the last 10 to 20 years of technology investing. For the first time, you can take capital and throw it at a company, and it compounds their advantage. AI is attacking every facet of the GDP. Transportation, labor, services, capital, coordination. There hasn't been a technology paradigm that hits on 30 trillion in GDP at the same time. Elon has talked publicly about BrockBot on Sam's side. He's talked about Astra and some of the long-running capabilities that are going to come out soon. What do you think is going to be the next $100 trillion market cap company? It is possible that... AI isn't just creating faster-growing companies. It may be changing the power law itself. In this episode, A16Z's Jim Kaa and David George sit down with accolade partners Adam Verdian to unpack what that means for technology investing and portfolio construction. For most startups, too much capital can become a liability. But David argues that Frontier AI is different. Dollars can be converted directly into compute and compute can make the product better reinforcing advantages of companies already at the frontier They discuss why AI addressable market could extend far beyond software why Aram believes AI is becoming a core rather than satellite allocation and why access, selection, and position sizing matter even more as the power law becomes more extreme. They also look at what comes next, from robotics and healthcare to energy, chips, and data centers, and why some of the biggest outcomes of the AI era may still be in categories that barely exist today. Welcome back to the A16Z podcast. Something fundamental has changed in how value gets created. PowerLaw used to be just a feature of a cottage industry and venture capital, and now it's systemic throughout, and particularly the three frontier model companies, SpaceX, OpenAI, Anthropic, represent somewhere between $3.5 to $5 trillion of potential enterprise value. And shockingly, before SpaceX went public, a lot of our LPs and also the broader institutional allocator community didn't have a lot of exposure to it. And so we talk about today why potentially portfolio construction and asset allocation may have changed why power law is not just only in the venture capital industry and then particularly where and how value actually compounds today David George Aram Viridian thank you for joining me Great to be here Thanks for hanging out Thank you for having us here Awesome Awesome Awesome Okay So DG so if you add up every venture IPO for the last six years, all of them together, where does it go from here? Right now, clearly, the power law is more extreme than it has been in the last 10 to 20 years of technology investing, probably going back to the emergence of the network effect-driven consumer companies. There are many reasons why that's the case. Increasing returns to scale have always been a dynamic in our business. Obviously, it's well covered how a network effect business can have increasing returns to scale, but so can software businesses, right? And they can take different forms, but brand, reputation in the market, the accumulation of resources, all provide competitive advantages. That all still is the case. But right now, especially with the labs, for the first time in my career, you can take capital and throw it at a company and it compounds their advantage. And this is a thing like, how do you screw up a startup? Well, throw too much money at it and have them hire a thousand people. And then you create all these coordination issues and overhead issues and dueling priorities. And it sort of gets messed up because you can't hire enough people to do enough things fast enough. Now that not the case You can throw dollars at compute and compute can make products and the businesses better And so to me it not terribly surprising that the power law is more extreme Right now economies of scale are a very real thing in the AI market I think it will continue to be the case. So, Ram, so first of all, you're not just one of our longtime LPs at Accolade, but incidentally, it's been exactly 10 years since you were actually an employee of A16Z. And so for your 10-year anniversary since you were last year, I've brought this gem back. Oh my gosh, this is amazing. How do you still have this? So we dug in the catacombs and we made this extra large version just for posterity here. We got this from the catacombs. But incidentally, during the last 10 years, a lot has changed in the world. And if you remember, at that point in time, people were belly aching about fund sizes being too large back then. And you had one out of a billion. Yeah, the first one out of a billion. Venture fund three? Exactly, exactly. And so a lot has happened since then. How do you think about your venture portfolio juxtaposed against your private equity one? And then just also generally asset allocation. We've talked about this on the way. And if you were to start from a blank sheet of paper again, knowing what you know now, how would you have constructed differently? Let's take venture today. We've reached 100 billion in revenue in AI. It took SAS 15 years to get to the same point. AI did that in four years. [5:00] and we're not even close to anywhere in terms of the penetration of demand. The reason DG is saying you can throw capital at it, that's a function of unlimited demand for inference. And we are at a point where AI is attacking every facet of the GDP, transportation, labor, services, capital, coordination. There hasn't been a technology paradigm that hits on 30 trillion in GDP at the You have the fastest growing technology. It's hitting on all parts of the GDP. So as an allocator, It's hard not to make the case. It's not a satellite position. You should be core or a super core in some shape or form. I'm very biased, but if you just think about the shape of the markets and how they've changed. Since I started my career in private equity and growth equity, that was, I guess, 18 years ago, this was a cottage industry. And now it's not. And I talk about this all the time, but our asset class is $5 to $6 trillion of value. And the dynamics around companies staying private longer... they're not going to reverse. Yeah, I mean, you had that insight in 2019 when you left GA. It's venture-like outcomes in late stage, which is now happening. So it's no longer just early stage and you IPO when you have 100 million in revenues. Yeah, the top decile outcomes, I think, used to be 10 billion, and now they're like 40 billion or so. Yeah, it's soon to be probably 100 billion by the time Anthropic and then OpenAI come out. Yeah, yeah, yeah. And look, this makes sense, right? The last cycle created 25 trillion in market cap, new market cap, and a bunch of that went to the incumbents. Yeah. But a lot of it went to startups and the new startups. And each one of these subsequent waves gets bigger than the prior one. And so our expectation is take the 25 trillion and it's going to be a larger number. Yeah. Yeah. And I'm... constantly confused about the theme of AI. Love your thoughts on this. Take healthcare. Healthcare spends 60 to 100 billion on healthcare IT per year. But AI is hitting on actual labor, and the value of tasks that are being performed in healthcare. That's claims, billing, administration. That's a trillion dollar industry. So the TAM of AI can be 10x plus bigger than traditional SaaS or healthcare IT. And what is that value? Well, what's the economic value of a task that's being performed? That's the time you're looking at. And then there's some capture rate that the AI company will take. But we have no idea... how big the 10 can get, to your point. You look at every wave, the incumbents, are 10x smaller over time. So it's hard to estimated, but I can tell this for myself, I've been chronically wrong about how big these outcomes can get. Yeah, same here. Yeah, labor, I mean, look, if you just look at how much... of dollars are spent in the U.S. economy on labor versus software. It's something like 40 times more. Now, that doesn't mean, importantly, that labor is going to go away. I think labor is just going to get reinvented. And so we'll end up with sort of reimagination of the tasks that humans do. But I think that's actually the whole point of AI is you're going after this different thing. And so to equate it to software and say, oh, it's the next evolution of software is far too limiting. Our legal counsel says to us often, it's like, I love Harvey. All my clients think they're lawyers with me on topics where they would have probably been like, I don't really understand this. I'm just going to defer to you. So my billable hours have only gone up with the advent of the usage of AI. And so all those use cases are massively, massively probably underappreciated. We don't still even know. Yeah, they're expansionary. Yeah, that's the whole point. Yeah. There was this whole thesis where, well, Frontier Labs are going to cannabis the apps, which layer is going to win. It turns out everyone is sort of growing. Yeah, yeah, yeah. We just... A friend of mine did a podcast where he described everything is going to work. kind of thing. And I describe it slightly differently, but we get questions all the time from LPs when they ask us like, you know, which layer in the stack is going to work. And I'm kind of like, I don't know, the market is going to be so big. Like I think everything might work. Now there's going to be a lot of companies that don't work and there may be a decent credit categories that don't work. But I think by and large, it's far too limiting to think, oh, if open source does a good job, it's bad for the labs and vice versa. So we try and remove ourselves from thinking in a zero-sum way like that. Extract it out, because why do people think it's going to be a winner-take-all in We can hypothesize, but the last year of technology was probably win or take all in a lot of categories. But this feels categorically different because we're re-underwriting a lot of the fundamentals. So maybe extract it out. Yeah, look, win or take all is an interesting way to describe it. Because if you just look at the market cap growth of all the leading technology companies, There are many, many, many that were successful. It wasn't winner take all, right? Now, there's an important distinction. We very much are believers in the power law within a given category. So the winners will capture the vast majority of the market share and market cap. And second place is plan for scraps. But I think there will be a massive expansion of the amount of categories that we have, right? And so if you go back 20 years, CRM was not really a category. I mean, it was small. It was like Siebel Systems and things like that. But now it's a massive category. And so I think the same thing will happen. We've seen it in every technology market that we invest in. Again, our approach... is in our business we can tolerate loss, right? And if we're not losing money in a given [10:00] on a given amount of investments. we're not taking enough risk, right? And so if you look at our best performing venture funds over time, I think the loss rate is 60% or so. On early stage, yeah. On early stage. Now, at the growth stage, the loss rate will be lower, but it's probably going to be in the 10 to 20% range, and that's appropriate because with that, you will get investments that we make that 10x or more in returns. And so if we're doing a good job, we're backing the leading company in every category. that is a credible category. And if the category works out, Well, I don't know. then we do a great job. And if the category doesn't work out well, That's okay. That's kind of the risk that we live with. Yeah. And embedded in that is also kind of timing, because I know, Rami, you and I lament on this, in that. I think a lot of folks... oftentimes think that things are overheated in that moment in time. And then you look back in retrospect and it turns out everything was actually quite cheap. But then there's these aberrations in the market where it's probably actually true. And so I know you advise a lot of your LPs on the importance of consistency in venture capital, probably even more than any other asset class because you just never know when these technologies can come out. Maybe walk through that because there's a lot of institutional allocators out there who actually don't have access to a lot And in some instances, probably maybe introducing some adverse behavior to that. It's a little bit too reflective of things being a little bit too frothy. So maybe unpack that for us. Yeah, I mean, the extremeness of the power law that DG talked about. If you, as an allocator, have not had access to the top five to 10 companies over the last five to 10 years, you're significantly behind in terms of returns. And let's take a step back. We've looked at the data of 3,000 venture capital firms in the U.S. Only 20 have achieved consistent 3x net returns over the last two decades. Sorry, say that one more. 20% have achieved... No, 20. Less than 1%. Wow. Consistent 3x net returns. That's incredible. And it's actually, you don't need seven, eight funds in those 20 years. Do you have three to four 3x net TPPI funds over a 20-year period? We found only 20 firms that have done that. Wow. Consistency in venture is really, really hard. Yeah. But what's interesting is the consistent ones... consistently had access to the category defining companies every vintage yeah um Now, there are exceptions. And by the way, just having the logo is not sufficient enough. If you're early stage and you have a large fund, you need to own enough. If you're late stage, DJ, I'm curious if you'll agree, sizing is really critical. Venture-like returns are possible in late stage, but your best company should be 5%, 10% plus of your fund. That way you can actually return the fund on a single company. Fund returning math in late stage didn't exist before. It now does. So we have found the right portfolio sizing and the ones that consistently have gotten access. are in the top 20. out of 3,000. So if you don't have them, there's a huge dispersion of returns. And if you don't have those, you're getting the average venture return. If you look at Cambridge data, the average venture return over the last 10 years is 1 to 2x net. You will do better in private equity. You'll definitely do better in the public markets. You don't need to lock up your money for 10 years. Yeah, for sure. Yeah, I just pulled up a tweet from our friend Endowment Eddie who posted actually this morning. He said, interest in big VC funds has been driven by founders, not LPs. Founders, more often than not, want the brand that can scale, be a lifecycle investor, and help land customers slash hires. because it's actually, it's counter to conventional wisdom. Yeah. And the outcomes are larger. So funds can be larger. There are some exceptions. In those 20, there are some small firms that are focused on niche vertical markets or they're playing at a stage that's so much earlier than the bigger firms where there is not a lot of competition with the bigger firms. Now, the problem with that strategy is you have to stay consistent in terms of fund size and their strategy. If you start getting bigger over time, then you bump into the big firms. And I think it becomes really, really hard to stay consistent. middle. Death of the middle. We said we were going to drink every time we said death of the middle. I'm very complimentary of many of our peers in the venture ecosystem. You know, like, but this death of the middle thing. How do you define it? Like, what's the middle? So I think... like sort of what you described, like the highly specialized funds, some of the ones that were very early-level, to AI with like deep, deep, deep domain experts, um, have done a pretty good job, right? Like they've done a good job and, and, you know, sometimes they can move fast, get into things or take shares of deals that we want to do. And you know, that, that is, that is a reality. Um, yeah, Then I think there's like the, you know, whatever we're, you know, large, you know, large scale venture, right. In the sense that we have main product lines and we can scale all the way from, you know, a seed all the way through to when you go public. And I think we have some peers who employ a similar strategy. Right. And, you know, there's I'd like to think that we're the best, but there's there's a few other folks who do that. Everything else is... In between that, I think, [15:00] struggles to compete a little bit for the reasons that Eddie said, right? The, what does the founder care about? The founder cares about, um, you know, sort of, taking capital from a partner that they think can... de-risk the outcome for themselves. Like that, if you were just to simplify it, That is the simplest way to describe what the founder really cares about. Now they care about the partner, right? They care about the person, so you have to be a good actor and all those things. But there's a reason why we build up a tremendous amount of resources, right? Like, it's why we have 700 employees. That's why we take, you know, the management fees that we make on our funds and we invest them in operating resources because we think that it will, one, bend the curve on the outcome, and two, help us to win deals. Um, and so, When founders select their partners, often if it's a hot deal, they'll have many alternatives. That's the revealed preference. And as Eddie said, we're doing an okay job with that. What I agree with him about then is, our LPs coming to invest in us is a byproduct of that, right? And so, you know, our business... our business is a flywheel. The flywheel starts with, you know, are we deep domain experts? Are we going to have a point of view that is the right point of view? Can we demonstrate to the founder that we... you know, are the right partner for her or him. If so, we win the deal. If we can help to make the outcome better, That's great. And then if we do make the outcome better, there's two things that happen. One, um, our business has persistence of returns partially because, um, the new founder wants to be around the winners. They care about that because there's important brand signaling and that has knock-on effects for them. And then secondly, by being a part of the winners and helping them in small ways, we create killer references. And the founders then tell the other founders, hey, you should work with these folks. And so that's the way the flywheel works in our business. One theory, curious to get both of your takes is, how take pre-seed seats, so sub-20, 30, $40 million valuation, sub-100 million funds, they can coexist with the big firms. Because at inception stage, say there's seven AI companies kind of doing the same thing. I would think a large firm like Andreessen would want to wait for a round or two until there's more relative certainty. Because one thing you don't want to do is be in the number two or number three. Like you said, you have to be in the category winner. So you'd rather wait for that round and actually double down and lead the A or the B. So the small firms, they can... carve out a niche for themselves. a clip or two earlier than the big firms and actually have a right to win, they can do really well and be complimentary to the big firms. Do you agree with that? Yeah. And look, we have very healthy relationships with seed funds across the ecosystem. We also do seed ourselves, right? But pre-seed, for sure, sort of earlier than we typically do. The seed you would do, I think, is like chunkier, bigger seed. They used to be A's. That said, we have our speedrun program, which we just came from, actually, earlier this this morning and I don't know, I think, I think the market is evolving where, where, um, Because founders have such a preferential attachment, as DG was saying, to the brands, we get the look at everything. And sometimes it does make sense for us to do the pre-seed and seed. And so, you know, you kind of want that flex of capability. And for the founder, to your point, they kind of don't really care where your focus is. They just want to be in that orbit and like you'll find the funds to kind of match to it. And so I think we can coexist in this world. But I think it's also very important. early stage business, we have to be first to the pole. And we may not actually do the investment, but we have to at least understand the landscape and market to be able to actually make informed decisions later on. Yeah, I spoke to a few founders at Speedrun today, and they very much are hoping to stay in the orbit. Yeah, right, right. And so this is, you know, a new phenomenon that worked 10 years ago. Again, this is starting to really take shape as more of the early stage folks started doing later stage and then extending across the stack, but not quite in the way that it is today. agree with that. Like it's just virtually impossible to have this sort of mid-stage business effectively without having the early stage and then also the late stage to come behind it as well. Yeah, I mean, look, I'm very biased. But I think the reason that we've been successful as a growth fund is because of our early stage business. I said that all the time. Like, our business starts and ends with early stage. And so... That provides us a tremendous amount of advantages at the growth stage in terms of access, information, knowledge, relationships, etc. And I would think that our early stage partners would probably say that the growth business provides them benefits too. Because it allows us to scale up and deepen partnerships with founders over time. And that helps to win deals at the early stage. Yeah, totally. And you can't, there's both sides. You can't only do the early. You can't also only wait to the late too. increasingly, that firms are converging to a handful of names. That's probably true because, again, [20:00] power law dynamic, but also at the same time, Almost majority of those logos, so to speak, we have to get at the early because that's the only way we maintain the ball control and also participate in the pro rata and then some. And so that's obviously the business. Yeah, we're big believers that. the strongest late stage franchises have a huge early stage franchise attached to them. Like the ability to win is multiplied when you have an early stage franchise. Yeah. And we talk about sizing in late stage where you can... Whatever the size of your late stage fund is, if you can at scale put five to 10% of your fund in one of the category defining companies, the way you could do that is because you had an early stage franchise that developed that relationship with the entrepreneur and the management team early on. It's really hard to come in as a de novo late stage firm and write a $500 million check. Yeah. Yeah. No, it's very hard. Yeah. I've lived that world. How do you think about from the LPC how venture is fundamentally perhaps a structurally different job than maybe when you started your career also as well? And, you know, how do you think about also asset allocation within venture? Because there's actually subclasses within venture as you think about portfolio construction as well. Yeah. I mean, in a very simplistic way, in our mind, there's four ways to do venture. Pre-seed seed. So think sub-150 funds. There's 1,000, close to 2,000 today in the US alone. Messy middle we talked about. And then there's a lot of firms that are better. Thousands or not thousands, hundreds of firms. And then the big firms. And then dedicated late stage. So there's four ways to play it. We have... done the larger firms for decades now. We've done the seed firms. We've selectively done a few in the messy middle and we haven't done dedicated late stage for the reasons we talked about. Um, how is it changing? AI is actually making our jobs harder than ever before. Um, It's making it harder because rounds are larger. Mm-hmm. in general, they're faster. the traction that's happening in the industry, is confusing. And here's why it's confusing. You can have a company. I'm actually really curious to hear this from you because we hear this a lot. Company comes out of pick your accelerator. I went from zero to five minute ARR in a month. Yeah. There's no renewal cycle yet on that company. Yeah. And they're raising off of that traction at huge multiples. And a lot of times they're selling to each other in a cohort potentially. And it's not even ARR, but they're multiplying by 12. So... But for nine companies like that, there's one really special one. that's doing a couple of million in ARR, actual ARR, that is a huge valuation that will go on to be the next cursor. So it is really, really tough actually today to parse out like what's real traction, what's not. Valuations are really high. this is why the big firms do well. I actually think they can wait or they have enough time. relative certainty in the next round, then lead that round. But even then, there's a lot of certainty. When you guys did cursor, I don't think there was a lot of certainty. And how many, for how many months were people saying Cursor is dead? Oh, even the morning of the acquisition announcement, people were still saying that Cursor is dead. I'm like, they just announced that they were going to be acquired by SpaceX for 60 months. Tell me if I'm wrong. $400 million round or somewhere maybe around there. Yeah, yeah, something like that. Yeah. A lot of people would say, like, that's crazy. Why did they do that deal? Yeah, yeah, yeah. Well, look, okay, so, like... Founder judgment is a very important thing, right? And so getting to know founders over time, spending a lot of time with them, seeing how they think, I'd like to think that, especially my early stage partners are pretty good at that. Secondly, like, It is hard to parse out real versus misleading interaction. And misleading in the sense that you can't take market signal from it, not that anybody's doing any misleading. Probably some of that, too. Yeah, there's probably some of that, too. But, like, I go back to... I love every now and then people are, like, helping to redefine what ARR actually means. I'm like, this is a very helpful PSA for the industry. This is always helpful, yes. Yeah. But, like... I don't know, come back to, you know, is the market demanding more of your product? That is always the question that's a post-it note on my computer screen. How do you know that when the company's been only operating or selling for a couple of months? You're not going to be able to do it with financial analysis. You'll have to do it by really understanding the customers, talking to the customers. And then, you know, it's like one of the things that I said about Harvey Overtime as an example, right? They did a really good job commercially early days, like because they were... They were smart. There was like a research plus lawyer combo. They got some momentum and some high profile law firms to sign up early days. But, you know, the usage was not very good. And so if you looked at the actual deployment of it, it looked mediocre compared to some other software firms. From a retention standpoint? Not retention. No, it's actual usage of it. Now, fast forward, post reasoning models, like that totally flipped. And you could see absolute takeoff of adoption, right? And so a bunch of different things happen at the same time. Lawyers got way more value out of the product. You could see it in usage and in [25:00] and then it became, because it was like high utility usage and engagement, it almost became a flip from what was previously like, oh, we're scared of things like hallucinations to no, no, no. Every client is actually demanding the law firms use the product. Um, and so, you know, I think we look for markets like that. We try to catch them early. Like we try to catch them earlier than when, than we did at Harvey. Um, but you know, that's the kind of signal that we look for. You gotta go, you gotta go layer down. Like everyone can do cohort analysis. Everyone can look at renewal data. Um, but like understanding the texture of the market and what the customers actually want and need and their alternatives. I think, you know, that's how you make the decision. That's why it's so important to have the early state business. Um, because they're the deepest in the, in the technology and the products. Obviously saw it in cursor and they've seen it in many other things. Jen, what's the biggest pushback? I mean, you're the most prolific fundraiser I know. So what's the pushback you get from LPs? They keep you very gainfully employed at this firm. I agree that you are a prolific fundraiser. What's the biggest pushback you're getting from LPs on the state of AI, the state of venture? So a lot of it is worries around... you know, are we catching a falling knife here? Just like the timing of the market where we are, like are things overheated, et cetera. And so we talked a lot about this at the outset, you know, around valuations and what the potential of the market is. But I do get... a lot of sentiment from the LPs that their job is also about to fundamentally change as well. And you mentioned earlier one aspect of it around AI making it more challenging to evaluate opportunities and funds. But the other aspect of it, as well as The LP historically has not been incentivized to actually embrace change in some respects, right? You know, this is very much a job where... the end goal is actually somewhat diametrically posed with the risk tolerance of the GP. And this is just the mechanics of the industry, but oftentimes say, you know, a GP can get fired for missing out, you know, the next, you know, Facebook, the next Uber, right? Like that is the error of omission and like that's fireable. But LPs on the flip side only get fired if you invest into a mentor. So in some respects, like the incentive outcomes are actually completely opposite of the GP. You don't get fired for investing in IBM if you're an LP. Exactly. Yeah. And in fact, like, you know, potentially even get fired for not investing at all. Yeah. If you miss the frontier models, back to your first question as an LP, but you kind of are along the benchmark, maybe slightly below the benchmark, you're keeping your job. Right, right. That's fascinating. And also, for most folks, and I'll leave fund of funds out of the equation because it's a different piece, but for a lot of folks… the upside actually is not that interesting for them. So the pitch of like, hey, you're going to miss out on the next potential of generation. It's incentive misalignment. It's actually quite, So where do we go from there in terms of the LP kind of role and see if like, I think we also have an important role and function. We oftentimes talk about in the context of our job as a leader of the venture capital industry is we have to help folks understand where the future is going. And part of that is understanding how to infiltrate not just within their venture capital allocation, but across their entire portfolio. And that, I think, is way more interesting than just saying, hey, like you might miss out on this next generation returns or the optimization of like the next frontier model or, you know, one or two power law companies, etc. what LPs do. So access, you could argue you have the data to figure out who has done well historically. Mm-hmm. Out of those 20 firms, out of 3,000, like, you are not going to see consistency, right? Maybe half of them are consistent. But... the LP's job is also to find the next gen firms as well as continue accessing that. So one is you access, two is selection, three is portfolio construction sizing. And it's critical from an LP standpoint. Because if you have an asset class where 20 firms out of 3000 do well, you should concentrate in those 15, 20 firms pretty consistently. So when I see a portfolio with 50, 60, 70 venture capital firms, It's very hard for me to imagine that the overall portfolio can generate better than the average. And again, going back to the average adventure, that's just, It's not compelling enough for the liquidity. Relative to any of the other asset classes, public markets, private equity. I mean, private equity can probably get you like one and a half to two X net without the lockup, without the risk you're taking in venture. You talked about 6% loss ratio. P doesn't have that. P has other problems we can talk about today when it comes to AI software, but... So portfolio construction sizing for an LP is critical. I've seen too many times, an LP or an allocator, find an interesting fund, actually get it right. and put 1% of their fund into it. Great, you 10x'd it, it returns 10% of your fund. It has not moved the needle at all. Yeah, yeah, yeah. Where do you all think we are today in that evolution that Jen was talking about of, how you would, I know you're biased, [30:00] Free it. percentage of overall capital. allocated to venture and growth compared to private equity or public markets or real assets. credit, whatever. Yeah, this is hard for me to answer because All I do is venture growth. So I would be biased to say it should be supersized. We like that answer, though. Look at the public markets today. We vibe coded like something that created a great way for us to assess AI resiliency of public companies. And now we're doing that on the private side and helping us hugely in our growth equity portfolio. But. In the SaaS public markets, there are only like 15 to 20 companies max trading above 10 times revenues, which is quite an insane number because it used to be dozens and dozens a few years ago. And every one of those companies, for the most part, is showing acceleration of growth from AI. You read them monitoring security, deployment of agents, etc. So it goes back to the same principle where if you are in some shape or form tied to AI... which is the fastest growing facet of all the elements of the GDP, then you should supersize that in your portfolio. That will have impacts in the public markets. Even private equity today, when they're doing a new investment, they're looking for something that's AI native. They're not looking to buy a workflow software company growing 10% that's seed based. That's just not happening. They're looking for the system of record that can show acceleration with an AI native management team. So... That connective tissue of AI is actually across every asset class today. And one of the strongest way to play it is, I mean, it's probably through venture. But that's where selection access and portfolio construction is really critical. Yeah. Well, even the exits that we were talking about, you know, Silver Lake potentially, you know, buying Workday, for example, just like doing more provocative things, for example, in private equity land when you have the capabilities to potentially infuse and bring them into the future as a part of that. The other version of it is also... Exits in venture now way exceed private equity. Like I was looking this up last night. Private equity this year, the biggest exits are the buyout of EA, which was like around $50 billion, and Medline, which is around $50 billion. Cursor, let's exclude the IPOs. Cursor was, the M&A sale to SpaceX was way bigger than that. The problem with that is... you look at the pre-chat GPT vintages in private equity, you would have paid, I don't know, 15 to 20 times EBITDA for a software asset that's growing exponentially. 10, 20% max. Yeah. If you look at the public markets today, that asset is trading at two times revenue. Yep. And the problem is, not just the valuation, there might not be a buyer for that company. Because if you're looking at a software company today, the first thing you think about, what is the terminal value of, Is it resilient from AI? The best way to show that is organic growth acceleration. Our data shows one percentage of growth in the public markets is equivalent to three percentages of EBITDA. So, by the way, it's funny because in COVID, everyone was like, we need to be profitable. It was the inverse. And now it's the opposite. No, no, 21, it was the inverse. Inverse, exactly. And in post-COVID, it's basically like fully aligned with risk, right? Exactly. It's correlated with like risk in the public. Exactly. So, unfortunately, a lot of those private equity deals, they're not growing fast enough. They're not showing that acceleration. And they may not have the management teams to revamp the business, like what Intercom did. is a great example. Bring the founder back, revamp the whole business, Create an AI native product, scale it, and then sell. It's almost like you're suiciding your existing business, which in private equity is really hard to do. It's really hard to do. I just spent a little time with the founder, and I went up to him at an event, and I was like, I just gave him a big high five. And I'm like, you did it, man. He did. This is the thing that is really, really hard to do. It's an end of one right now. But there's a bunch of really good founders who are capable of those businesses, public markets, private markets, who I think are going to take a crack at it. So we'll see. Yeah. Maybe on that thread, though, DG, because we also sometimes get the pushback as well. Like for folks who have been in venture and allocated to venture, they might also have a similar problem where they do have the legacy SaaS businesses also as well. Like what's the balance between how you think about the historical stuff? Let me ask you that question. So I'm going to piggyback off of John's question to you. You got a... pick a 2016 through 2021 pre-chat GPT vintage software company that, was fine, but doesn't have that AI native features anymore. It's not accelerating. It's growing like 30%. On the venture books, it's like 10, 20 times revenue. Issa can't go public anymore. No one cares to take that public. Yeah. Silver Lake has no interest in that company anymore. They would have a year ago. They don't. What happens to that company? Thank you. Um, We have a lot of exposure to those companies too. I would say it's very TBD, right? I was with one of our CEO founders this weekend, and he was like, give me the straight scoop. What do you actually think is happening? He actually said to me, don't give me a pie. [35:00] answer, which is ironic. And, you know, I think both things can be true, that AI is like the biggest generational change that we've ever seen, and it's going to transform industries. And also there will be some enduring value of software companies that are able to adapt, right? Part of the thing that we're monitoring, which makes us extremely bullish about AI, is just actual diffusion into the real economy, right? So coding, I think, if you were to paint the bullish scenario for regular software and for slower pace of change, you would say, you know, coding hit, but that's kind of a head fake, right? Like coding is perfectly documented, right? right? So it has like, Perfect data. It's verifiable and it's simulatable, right? And so like most tasks in business do not share those three attributes. And so, you know, maybe the diffusion into other knowledge work beyond, you know, coding will take a lot longer. That would be the case to make for the software companies. And then some of them will evolve and have AI, you know, solutions and they'll change their business models. And I think that's a must. But that would be the case for why maybe, you know, it's a little bit overblown. I think if you look at the way that a lot of the public SaaS companies have reacted over the last months, like that's, I think there's a little bit of a growing realization in that. makes me super, super, super bullish on AI though, right? So like if you look at our portfolio, you know, we have some of those companies put about 95% of our nav is not in those companies, right? It's in the companies that are growing very fast, accelerating, et cetera. The average, I think it was the median average, company in the US is spending $12 per employee on AI. per month. The top 1% of the data set that we've seen is spending $7,000. per employee on AI per month. So, um, Not only have we had limited diffusion beyond coding, but if you just look at diffusion of the shape of who is consuming tokens and actually getting real value out of AI today, we're super early. Banks, the most cutting edge banks are probably doing 1% of headcount cost on AI tools. And so the reason this makes me very bullish is... these are the fastest growing companies we've ever seen, like of all time. Again, they're adding more revenue per month than the mega cap tech companies. And yet it's probably on the back of adoption of like, 10 million users, maybe 20, maybe 30 max. And there's one and a half billion knowledge workers in the US. And I think it's going to transform the way we do a lot of work. Yeah, Calper's famously... lost out on billions of gains by not investing in their back. And they're making up for lost time now. They've converted their portfolio from 91% to 58% and venture in growth from 9% to 43%. There's probably some balance in between those things, but they're leaning hard into. But there's going to be a lot of value still that's going to be accreted in some of these historical companies. And I started joking around about Bendix, but that was probably a great outcome for Airtable, outside of the fact that they're going to actually spin off the hyper-agent piece of the business interesting things in that, but in the scheme of things, I think there's going to be a lot of homes for a lot of things, and this zero-sum thinking, I think, is probably the pitfall of which we would advise against. So we talked about... We've talked about venture growth, we've talked about private equity, we've talked about public markets. And by the way, the composition of all of those have radically changed over the last 10 years. We also have had the emergence of entirely new categories that are available in the private markets, like private credit. Do you have a view on sort of outlook of those on a relative basis? in software in particular? Yeah, I'd say in technology. The vantage point we have is looking at private credit which resides in a lot of private equity software portfolios, which is hundreds of billions. I'll give you one statistic. So you look at 21-22, about... two to 300 billion in LBO software transactions happened, with over 200 billion in debt taken out. The average valuation for the software deals were 25 to 32 times EBITDA. those companies today are worth probably half that. The reason you're seeing redemptions in the credit markets and private credit is exactly that. They're looking at the public markets. You've had CESpocalypse. It's been a massive correction in software. And you can see a contraction in valuations, which means the leverage ratios have gone up dramatically. So... If you are a software company that is in somewhat not resilient to AI, I think you're challenged both in terms of your equity position, also credit as well. By the way, even the... AI version of private equity is not completely insulated. We oftentimes talk about just because you put Sears on a website, didn't make it Amazon, right? [40:00] to have the benefit of building Amazon from the studs logistically to make it Amazon. It's not just the website. And in a lot of instances with the private equity backed companies that are now just infusing AI. We've seen it actually in some of our companies as well, where the peer competitor is like, oh, the first thing I'll do is, of course, hire AI customer service agents because that's like an easy, low-hanging fruit. Like, it turns out, if you don't actually build on the workflow, you start to turn customers very quickly if they're used to talking to a human. And for every dollar, every drop in NPS is like a direct correlation with drop in revenue. And then you start to spiral, especially if you have debt laid on top of it. from folks going, well, I'll just do the AI, you know, kind of version of private equity. It's not a panacea for people. generating returns, especially when it's just so categorically different from a technological perspective to actually infuse that throughout the company as well. Yeah, you can't just throw an operating partner at the company and say, let's put AI on it. It just doesn't work. You need to completely, and if you have, by the way, if you do have a founder mentality at the management team, like it is possible, but the board has to be aligned, all the investors have to be aligned, and you do have to make some really hard decisions the way Intercom did. Yeah, yeah. Yeah, so obviously this is a group that is very pro, you know, venture and growth story, let's talk about the legitimate opposition to it. And what is the case, you know, for why maybe the, you know, the risk that you're taking or whatever it may be with venture and growth, you know, doesn't justify it. Thank you. I mean, the pushback we get a lot is timelines of liquidity. So... It takes the average unicorn is private for 10 plus years, typically, and then you got all these follow on rounds that are happening pretty quickly one after the other, you see maybe the same logo in five, six different firms. And the question is, how do you get out of it? Um, And an IPO isn't actually a distribution. It could take 12, 24 plus months before you actually get liquidity out of an IPO. Especially if you own 10, 15% at IPO, it's going to take a long time if you're in the generational company. So we get that pushback a lot in terms of timeline to liquidity. And then what is the sort of counter to that pushback? Well, counter to that pushback is, Going back to 3,000 firms, 20 do well consistently. If you're in the top 1% of those firms and you have a category winner, You want to make sure that compounds, actually. Yeah. Would you have wanted to sell Stripe, Databricks or any of those other companies three, four years ago? Like the answer is unanimously no. Now, could some of these companies go public earlier than not? Sure. Anthropic was really first funded in 2021. It's about to go public five years later. Yeah. Cursor. from first acquisition, from first financing to acquisition is short. Yeah. Like, so the best venture firms actually have fund returning liquidity pretty quickly, maybe even quicker than private equity. But that subset of firms is tiny. Yeah. Actually, very famously, a year and a half ago or so, we went to our fund one LPs. At that point in time, the fund one was 16 years old, and we had this position in Stripe that we invested at the seed stage. And we asked all of our LPs, like, hey, do you want liquidity out on this? We came to do, is now done, 16 years in. Like, do you want liquidity back on this? And every single one of those LPs said, no, we'd rather let this continue to compound. And then ultimately, a year later, you know, we decided to make that answer because it's 17 years in. We've got to get this liquidity out. We've got to wrap up the fund, et cetera. But, you know, so many... But LPs, I think, is very specific to certain categories, right? Endowments would prefer to let it run. Family offices, quite frankly, don't want the money back because they don't want to pay taxes on it. They'd rather have it continue to compound. And so there is specific nuance with each LP group where it's very hard to paint a broad brush stroke on like across the board on everyone wanting the same thing. But I also think, to your point, The very best LPs, excuse me, the very best GPs have manufactured along the way liquidity. And particularly in 2021, when a lot of folks didn't take money off the table, you know, I think that was a good sign of the first indicator. And now in this next cycle, it's can you actually get some early liquidity out through M&A and then let, you know, potentially the winner's IPO over time with the fullness of compounding as well. Yeah, exactly. I'm going to end on this note because I thought this was an interesting question that you and Gavin were tossing back and forth, DG. He didn't want to answer the question on what will be the next $10 trillion company, but he had a certainty around what will be the next $20 trillion market cap company. So I'm going to ask both of you, what do you think is going to be the next $100 trillion market cap? Oh, my God. Here we go. I can't even think of those things. Yeah, that's probably two tech cycles away, not just one. uh, It is possible that We... have entirely new companies that get created. And I think a lot of the market cap creation that you would talk about that would drive a $10 trillion outcome or more is [45:00] product areas that haven't yet been touched. Right? So like what, I talked about the sort of diffusion of the technology into the enterprise, like we're nowhere, right? A year ago, everyone talked about consumer all the time. No one even talks about consumer AI anymore. But that is going to... The end use case for consumers... is not going to be a chatbot interface. That's the skeuomorphic version. We're gonna have a native version. It's gonna be proactive. It's gonna do work on our behalf. It's gonna create a ton of value for consumers. And we're kind of nowhere on that. I mean, yeah, there's a billion ChatGPT users, but that's like scratching the surface. We are... Nowhere on robotics. But I think robotics is going to be bigger than the language stuff. And I think it's going to happen in the next 10 years. We are almost nowhere on autonomy, right? Like there's fewer than 10,000 Waymos. live in the US, way fewer robo taxis. And then, you know, a lot of open space for others to build in that area too. We, you know, healthcare is 18% of GDP. - Yep. - Like, we've done nothing to scratch the surface, either on care delivery, or on drug discovery yet. I mean, there's some companies that are working on it, showing some early signs of progress. But I think the progress that we make there in the next 10 years is going to be massive. And then, you know, we're in this interesting era of reimagining all things physical world, from defense to manufacturing to data centers. And so I look at the confluence of all these trends today, And I'm like, yeah, It may feel like we've done a lot with AI already, but 10 years from now, we're going to look back and say, oh my gosh, those other major areas created a ton of value. And so I'm excited that I think the next SpaceX AI or OpenAI are probably going to get created. And they'll probably be in those kinds of domains. Yeah, I'll add one category that... to me is both a concern but a huge opportunity. It's a plug for your new fund on the opportunities fund that you did. I think very simplistically about like the bottleneck in AI today is not demand. It's on the supply side. So you got energy, the grid data center, then you got chips, then you got frontier models and apps. The U.S. is amazing at the right side of that. So like chips and onward. The VC ecosystem supports that well. I think the new fund you have is really going to help on the left side as well. Because the U.S. doesn't have a problem with energy generation. It has a problem with speed to power. That's permissioning, transmission, that's regulatory. Other countries are putting out 10x more renewable capacity a year. So that is a real bottleneck. And that means reimagining the data center. You talked about the density being 10x plus. You can't just repurpose an old data center for a new AI facility. it is as well. That can really solve the bottleneck. And I think that is a real concern because demand is not a concern. I've heard LPs say, this is like the dot-com or this is COVID. It's not because the traction is real and it's not ephemeral revenue like COVID. The bottleneck could be supply, but if you have the right inputs, like the fund that you're now backing those companies, next generation chip companies, memory, et cetera, that's a huge opportunity. It's time for machine age. Let's bring the machines. I love it. All right. Let's Thank you for having us. Awesome. See ya. Thanks for listening to this episode of the A16Z podcast. If you liked this episode, be sure to like, comment, subscribe, leave us a rating or review, and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts, and Spotify. Follow us on X at A16Z and subscribe to our Substack at a16z.substack.com. Thanks again for listening, and I'll see you in the next episode. Reminder, the content here is for informational purposes only. It should not be taken as legal business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see A16Z.com forward slash disclosures. Bye.