Big Tech Taps Debt to Fund AI Capex as Investors Weigh Bubble Risk

大型科技公司发债资助AI资本支出，投资者权衡泡沫风险 超大规模云服务商正通过债务市场为庞大的AI基础设施建设融资，而前沿实验室的收入仍集中在狭隘的编程任务上，若企业端采纳停滞，更广泛的股票和信贷市场将极易受到严重波及。 Ezra Klein主持了一场严谨的经济学辩论，探讨当前的AI繁荣究竟是一场历史性的投机泡沫，还是一轮受产能受限但具备生产力的基础设施周期。Paul Kedrosky认为，超大规模云服务商的过度支出、举债建设数据中心、狭窄的客户集中度以及对实体资源的挤占，与19世纪铁路和电信等过去的产业崩盘高度相似。Azeem Azhar则反驳称，算力与能源的结构性供不应求、实打实的企业端采纳，以及前沿模型实验室的成败与持久的实体算力资产之间的区隔，都降低了灾难性泡沫的风险。讨论还延伸至劳动力市场冲击、监管风险，以及能够证伪两位嘉宾各自论点的宏观经济指标。 预计AI基础设施支出在数万亿至10万亿美元之间，科技超大规模云服务商正从依赖经营现金流转向成为美国最大的投资级债券发行主体。 (1:58) 市场集中度带来了严重的系统性风险：标普500指数中排名前10的股票约占其总市值的40%，而美国家庭净资产的50%与股票挂钩。 (1:58) 数据中心建设正在吸纳熟练劳动力和原材料，引发了荷兰病效应，显现出对国内住宅建设和商业工厂建设的挤出效应。 (40:14) 前沿AI实验室的收入依然狭窄且脆弱，软件开发约占总Token消耗量的80%，极少数企业客户主导了整体营收。 (12:45)

Venture capitalist Paul Kedrosky and tech researcher Azeem Azhar debate whether the multitrillion-dollar artificial intelligence buildout represents an unstable, debt-heavy bubble or a productive, supply-constrained infrastructure revolution anchored by real enterprise demand. Debt Financing and Systemic Equity Concentration (1:58) Paul Kedrosky argues that historical capital expenditure waves around general-purpose technologies—such as canals and 19th-century railroads, where roughly half of all track laid proved useless—consistently end in speculative oversupply and severe write-downs. With projected AI buildout commitments heading toward $10 trillion, Kedrosky warns that hyperscalers are transitioning from organic cash flow to aggressive debt issuance, private credit, and off-balance-sheet vehicles. This debt directly competes with sovereign Treasuries while tech becomes the largest issuer of investment-grade corporate bonds. Macro vulnerability is heightened by public exposure, as the top ten stocks account for roughly 40% of the S&P 500 and 50% of American household net worth is held in equities. Azeem Azhar counters that the financing environment does not yet resemble historical credit crises, noting that external debt financing is only approaching a 30% threshold by late 2026. For Azhar, if hyperscalers and credit funds suffer poor returns from overpaying for hardware at the top of a cycle, the outcome represents bad corporate investing and valuation adjustments rather than a true systemic bubble burst. Compute Bottlenecks Versus Collapsing Token Prices (9:20) Azhar maintains that the AI sector currently faces an acute compute undersupply rather than an asset glut. Data center operators are experiencing rapid capital recovery cycles of 18 to 24 months due to intense capacity shortages. Because specialized GPUs represent approximately 65% of total data center costs and non-frontier enterprises deploying open-weight models now account for half of leased compute, Azhar argues that physical computing power remains an enduring, general-purpose asset even if high-profile model labs face margin pressure. Kedrosky dismisses token volume growth as a vanity metric that confuses technical scale with economic returns. He points out that performance-adjusted token prices have declined roughly 13-fold on an annualized basis. Furthermore, Kedrosky highlights an early-adopter mismatch: software developers have driven an estimated 80% of frontier lab revenues through "expansive" workflows where a brief prompt generates massive code outputs. By contrast, mass-market enterprise knowledge work is fundamentally "compressive"—summarizing documents into concise outputs—requiring far fewer tokens and providing weaker revenue foundations. Resource Siphon, Cognitive Repricing, and Friction (40:14) Beyond balance sheets, the panel analyzes how data centers physically crowd out the broader economy, absorbing construction labor and materials while domestic housing and manufacturing investment slow. Azhar likens this imbalance to 1970s "Dutch disease," though he notes the surge in demand for electricians and HVAC technicians could leave the economy with expanded vocational capacity over time. Conversely, Kedrosky points to severe local distortions, citing New Mexico projects where public subsidies and abatements reached $4 million per job created for facilities employing only a handful of permanent staff. The participants also divide over labor and societal fallout. Kedrosky asserts that generative tools act primarily to reprice white-collar cognitive labor, destroying jobs like call-center support far faster than the economy can invent replacement roles. Azhar views adoption as a slower, 5-to-10-year organizational transition where human verification choke points will command higher wages. However, both guests and Ezra Klein agree that removing social friction with automated agents risks a public backlash, creating what Kedrosky terms a "denial of service attack on modernity" that floods legal systems, merchants, and institutions with synthetic noise. The dispute rests on distinct empirical hurdles: Azhar states his confidence would weaken if American enterprise AI spending growth halved for two consecutive quarters, whereas Kedrosky maintains he will only concede the boom is non-speculative if macroeconomic data demonstrates decisive, above-trend gains in aggregate U.S. labor productivity, labor force participation, and real GDP growth.

围绕当前的人工智能热潮，风险投资家Paul Kedrosky指出该产业正陷入由过度资本开支与杠杆债务驱动的投机泡沫，而科技分析师Azeem Azhar则认为算力短缺和真实企业需求的扩张表明技术繁荣仍具备基本面支撑。 资本过度扩张与宏观金融脆弱性 (1:58) Kedrosky认为当前AI资本开支与经济现实脱节，全球相关投资预计将达到10万亿美元，其演变路径重蹈了19世纪铁路等通用技术的覆辙——当时约50%的铁轨最终被证明毫无用处。如今科技巨头深陷囚徒困境，在赢者通吃的寡头预期与中美竞争压力下过度投资，Mark Zuckerberg等高管宁愿承受数千亿美元的过度支出也不愿承担落后风险。更严重的是，标普500指数前十大股票市值占比已达约40%，美国家庭约50%的净资产配置于股市，这使普通散户与退休金深度暴露在单一主题的系统性风险中。 融资结构的变化进一步加剧了金融系统的脆弱性。Kedrosky指出，科技行业已从依靠自有现金流转向大举借债，成为美国最大的投资级企业债发行群体，微软和甲骨文等龙头甚至面临评级下调压力，Meta还借助表外特殊目的载体隐匿债务；科技债的大规模发行甚至与10年期美国国债争夺流动性并推高主权借贷收益率，当前的席勒CAPE比率已接近1929年大崩盘前的峰值水平。Azhar虽承认科技巨头外债融资比例在2026年下半年接近30%的警戒线，但他认为只要信贷基金仅面临低回报而非全面违约，就属于普通的不良投资而非真正的资产泡沫破裂。 算力供需矛盾与代币经济的真实价值 (9:20) 在底层资产与需求层面，两位分析师对算力市场的成熟度产生分歧。Azhar指出，数据中心约65%的资本开支流向GPU等硬件，受算力紧缺推动，算力提供商的资本回收周期仅为18至24个月。不仅OpenAI和Anthropic仅占早期云算力租赁收入的约一半，另一半及未来增量正在流向运行开源权重的普通商业机构。他主张将前沿模型开发商与算力基础设施拆开评估，即便私有前沿模型遭遇商品化冲击，底层数据中心与电网依然是长期的通用公用事业。 Kedrosky则强烈反驳代币经济的繁荣假象，指出按模型性能调整后的代币价格每年暴跌约13倍。前沿实验室高达80%的代币消耗与收入来自软件开发者这类特殊早期用户，其工作流具有“扩展性”（数个词提示可生成百万行代码），而广大普通企业的日常办公具有“压缩性”（将多篇资料浓缩为简报），对昂贵的高级推理模型需求微弱。Klein亦质疑基础设施与模型公司无法完全切断联系，一旦企业流向DeepSeek等低成本开源架构并引发前沿实验室估值崩塌，巨额资本开支的重定价必将波及整个科技板块与信贷市场。 实体经济挤出与劳动力重定价争议 (40:14) AI资本繁荣对实体经济的扭曲正在显现，数据中心建设大量消耗劳动力和建筑材料，导致住宅制造等实体投资受到挤压。Azhar将这种现象类比为1970年代荷兰由于北海天然气开发而削弱其他工业部门的“荷兰病”，水暖工、电工等蓝领工种遭遇严重用工短缺，唯有政策干预与行业技能培训才能缓解失衡。此外，数以十亿美元计的数据中心通常仅创造2至6个永久就业岗位，在新墨西哥州甚至出现每个就业岗位摊派高达400万美元公共补贴与税收减免的极端失衡。 关于就业替代，Azhar坚持通用技术需要5至10年逐步扩散，虽然企业通过提高人均创收放缓了补充性招聘，但人类在关键决策环节的审核把关将推高专业节点薪酬，并最终催生类似当年“网站管理员”的全新职业形态。Kedrosky则直斥这种技术乐观主义是盲目教条，强调生成式AI是有史以来渗透最快且唯一以重定价白领认知劳动力为核心的技术，客服等工种遭受的冲击速度远超新岗位创造速度，且宏观经济至今未出现全要素生产率跃升或劳动参与率改善的实质证据。 社会摩擦消失与不连续性监管反弹 (50:16) 无节制的自动化正在侵蚀现实社会的微观运行秩序。Klein指出，自主软件智能体大量制造无意义文本与虚假预订，已严重干扰中小商户的正常运营；Azhar观察到仅有约2.5%的美国人直接为AI付费（人数甚至低于GLP-1减肥药的自费群体），绝大多数人只是被迫接触免费或内置功能，反过来被迫购买反骚扰筛选服务。Kedrosky警告盲目消除社会摩擦相当于对现代社会发起“拒绝服务攻击”（DoS），促使自动化诉讼塞满轻罪法庭，导致防守型支出反常推高名义GDP。 伴随社会不适感上升，监管与抵制风险正急剧聚集。Klein认为民众对被算法替代的天然抗拒可能催生类似核电行业面临的严苛红线，医生和教师等成熟利益群体也将针对医疗失误或教育应用发起封杀。Kedrosky将前沿大模型的无序发布比作“未经许可的人体药物试验”，并计算指出若面临网络安全崩溃、关键基础设施瘫痪等12类独立风险（哪怕每项发生概率仅为5%），系统爆发灾难的复合概率已突破50%，未来必将招致类似FDA或原子能委员会式的惩罚性管制；Azhar则认为这种抗议是健康的民主反馈，正推动企业通过设立地方年金等方式换取“社会经营许可”。 本场辩论最终指向检验AI基本面的量化标准：Azhar认为若美国企业在AI软件上的支出增速连续两个季度减半，其增长论断便告破产；Kedrosky则主张唯有看到全美劳动生产率显著打破历史均线、劳动参与率大幅提振及实际GDP加速上行，才能证明这并非一场终将引发宏观债务阵痛的超级泡沫。

## Background 本内容由《The Ezra Klein Show》于 2026 年 10 月 16 日发布在 YouTube 平台。页面简介指出，对 AI 前景的庞大豪赌正驱动着美股与宏观经济增长，其数据中心建设甚至挤压了民生住房等资源，本期节目邀请投资人 Paul Kedrosky 与科技分析师 Azeem Azhar 就“AI 究竟是能兑现回报的生产力飞跃，还是一场注定破裂的泡沫”展开深入辩论。 ## TL;DR 节目围绕 AI 产业是否存在系统性金融泡沫展开对抗性研判。Paul Kedrosky 结合历次通用技术资本开支周期指出，过度依赖信贷发债、标普 500 极端集中、家庭资产高配美股以及代币单价暴跌等特征，预示着严重的产能过剩与类似 1929 年前夕的脆弱性；Azeem Azhar 则持谨慎乐观态度，认为终端企业对算力的真实需求依旧强劲，GPU 投资回本周期良好，底层算力基础设施建设属于健康的“生产性繁荣”，即便前沿模型实验室受挫也不等同于基础设施破裂。双方进一步剖析了科技巨头表外负债与“荷兰病”挤出效应、实体企业落地与认知劳动力重塑的阵痛，以及断裂式监管和消除社会摩擦可能引发的全社会强烈反弹。 ## Atmosphere & style 整场对谈兼具华尔街金融分析的严密推演与科技社会学的宏大思辨，节奏紧凑且交锋激烈。主持人 Ezra Klein 穿针引线、不断用尖锐的现实困境倒逼双方，Kedrosky 语言犀利辛辣、充满历史宿命论的警世讽喻（如“马粪堆里找小马”），而 Azhar 则严谨克制、试图以数据看板和长期制度演化进行抗辩。听众能清晰感受到狂欢市场表面下暗流涌动的债务重压、巨头博弈的囚徒困境，以及技术浪潮对普通人真实生活的不可逆挤压与时代焦虑。 ## Bulletin - [01:28](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=88s) 本轮针对 AI 的资本支出总额预计将攀升至 10 万亿美元，但美国家庭已有高达 50% 的净资产绑定在股市，且标普 500 指数前十大股票市值占比逼近 40%，系统性单一因子风险极高。 - [36:50](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=2210s) 超大规模云服务商的融资模式正从消耗自由现金流转向发行投资级债券与表外 SPV 融资，导致 Oracle 等企业信用评级承压，并与 10 年期美债产生流动性竞争。 - [12:49](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=769s) 以代币单价每年骤降约 13 倍为背景，当前前沿模型约 80% 的消耗集中在程序员的代码生成上，这种膨胀型需求难以代表普通大众的压缩型日常任务。 - [40:14](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=2414s) AI 算力基建正在引发实体经济层面的“荷兰病”，大量建筑材料与劳动力被抽调至数据中心，直接挤压了民生住宅建设与传统制造业厂房的投资占比。 ## Key takeaways - [01:28](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=88s) 西方经济史上的 8 至 9 次通用技术（GPT）基建狂潮均以产能过剩和资产暴跌告终，历史上的铁路建设最终约有 50% 的铁轨被证明是多余无效的。 - [05:08](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=308s) 各大科技巨头深陷寡头通吃博弈的囚徒困境，宁可浪费数千亿美元过度扩建，也不愿承担落后挨打甚至被挤出市场的潜在风险。 - [07:20](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=440s) 多头认为眼下仍是真实的算力饥渴与“生产性繁荣”，硬件投资依靠动态加价机制能在 18 至 24 个月内回本，短板主要受制于先进制程芯片与电力供给。 - [12:49](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=769s) 前沿模型实验室过度依赖程序员的膨胀型代币消耗，将早期高溢价客户的特性误判为大众市场需求，存在“马粪堆里翻找小马”的价值错觉。 - [20:15](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=1215s) Azeem 认为历史上的泡沫破裂要求投资规模腰斩且股市跌幅达 50% 以上，而买高资产或低回报率仅属普通投资失误，不能算作泡沫爆裂。 - [21:43](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=1303s) 当前美股席勒市盈率（CAPE Ratio）已触及 1929 年大崩盘前夕的极端水平，若 OpenAI 与 Anthropic 等超级巨头上市，其总估值体量将压倒当年互联网泡沫峰值。 - [23:43](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=1423s) OpenAI 与 Anthropic 的潜在 IPO 本质上已非扩张性融资，而是早期股东与内部人士在高位落袋为安的套现离场通道。 - [31:49](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=1909s) 苹果每年仅维持约 150 亿美元资本支出并拒绝加入前沿基础模型竞赛，却凭借对硬件终端的垄断成为坐享外部基建红利的终极“搭便车者”。 - [34:41](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=2081s) 前沿实验室的资本扩张外溢到了网络安全等配套服务商，如 Cloudflare 因全网抵御 Bot 恶意流量的需求激增而在过去 18 个月股价翻了三倍。 - [44:20](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=2660s) AI 对传统骨干企业的重塑极其缓慢，企业获得生产力提升后，在总营收未爆发的情况下往往直接体现为对人工岗位的压缩与替换。 - [49:02](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=2942s) 硅谷长期信奉“技术创造的岗位永远多于所摧毁的”，但在认知劳动力被快速重新定价的冲击下，社会适应的周期远远落后于职业被替代的速度。 - [55:37](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=3337s) 巨额补贴的数据中心往往仅创造极少数长期就业岗位，若发生系统性失控事件，公众反弹可能促使监管从医药审批模式骤变为全面叫停的核电管制模式。 - [57:59](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=3479s) AI 代理盲目消除了社会生活与商业交互中必要的“摩擦力”，零边际成本的自动化外呼与滥诉正在演变成一场针对现代性社会的拒绝服务攻击（DoS）。 - [01:01:57](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=3717s) 要从根本上反驳泡沫论并证明当前天量支出的合理性，必须在宏观数据上观察到全美劳动生产率、劳动参与率以及 GDP 增速同时显著突破历史趋势线。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=0s) AI 浪潮下的经济赌注与两派分歧 在当前的美国经济中，人工智能（AI）已成为推动美股上涨与宏观经济增长的核心引擎。然而，数据中心建设对资源的吞噬已达到极为夸张的程度，甚至开始挤压住房、制造业设施等传统基础设施的建造资源。这就带来了一个关乎数万亿美元的金融核心命题：这场空前的豪赌究竟能否带来回报，还是说 AI 本身就是一个注定破裂的泡沫？一旦某家核心巨头业绩不及预期或评级被下调，泡沫破裂便可能将普通大众的养老金组合拖入深渊。 为了厘清这一争论，Ezra Klein 邀请了两位持相反立场的观察者展开辩论： - **Paul Kedrosky**：风险投资机构 SK Ventures 的合伙人、麻省理工学院（MIT）数字经济倡议研究员，同时主理知名个人分析通讯。他坚定地认为，当前行业正处于极其明显的 AI 泡沫之中。 - **Azeem Azhar**：英国科技创业家、分析师，著名科技通讯与研究平台 *Exponential View* 的创办人。其团队长期通过构建数据监控看板追踪 AI 产业的财务压力，整体观点相对更为乐观。 ```mermaid flowchart TD accTitle: AI投资驱动与宏观风险传导链路 accDescr: 展示从AI资本开支到宏观资源挤压及资产负债端风险敞口的传导路径。 subgraph CapitalDrive["资本驱动与实体挤压"] A["巨额资本开支 (CapEx)"] --> B["数据中心疯狂扩建"] B --> C["挤压住房与制造业建设资源"] end subgraph MarketRisk["金融敞口与脆弱性"] A --> D["美股及宏观经济增长引擎"] D --> E["标普500权重高度集中 (前十大占40%)"] E --> F["全民401(k)与净资产高杠杆暴露"] end ``` ### [01:28](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=88s) Paul Kedrosky 的空头论据：通用技术过度建设的宿命 针对 Ezra Klein 提出的“为何高盛、贝莱德等顶尖机构以及谷歌、Meta 等超大规模云服务商（Hyperscalers）全错了，唯独你看清了”的质疑，Paul Kedrosky 首先笑称“我倒不会直接说他们全是一群笨蛋”，但他坚决指出，当前的资产价格已经完全脱离了经济现实。从西方经济史来看，人类一共经历过大约八到九次针对通用技术（General-Purpose Technology, GPT）的巨额竞争性资本开支周期——从运河、铁路、电气化到州际公路及光纤网络。 Paul Kedrosky 强调，历史上没有任何一次大规模通用技术资本开支周期不是以连环危机、严重膨胀的资产价格以及产能大过剩而收场的。他以铁路为例解释道，这绝非意味着 AI 技术本身无用或轻浮，“铁路极其有用，但最终修建的所有铁轨里程中，依然有大约 50% 证明是完全无用的。”当前的泡沫本质在于：资本的极度充裕促使资本供给方基于对未来需求的狂热预期无节制扩产，而如今这种过度供应正越来越多地依靠信贷支撑。 债务与信贷已经渗透进整个经济体系，隐藏在私人信贷、保险公司和主权财富基金内部，目前的资本规模正朝着 10 万亿美元迈进。更致命的是散户层面的脆弱性：当前美国家庭净资产中高达 50% 绑定在股票上，创历史新高；而标普 500 指数中排名前十的股票市值占比接近 40%。这意味着绝大多数持有指数基金或 401(k) 养老金的普通人，在本质上都处于对 AI 的极致单向做空对冲缺失（Extreme Long）状态。关于整个宏观层面的系统性风险，Paul Kedrosky 警示道：“The bubble is this vast oversupply of something largely fueled now by credit which in turn will lead to massive business failures and write down of assets and all of that credit has spread through the economy.” - 西方经济史上大约发生过 8 至 9 次大规模通用技术资本开支浪潮，全部以产能过剩和危机告终。 - 历史上的铁路大扩张中，最终约有 50% 的铁轨铺设被证明是无用的资产。 - 预计本轮 AI 资本支出总额在尘埃落定前可能攀升至 10 万亿美元规模。 - 标普 500 指数前十大重仓股占据了整个指数市值的约 40%，呈现极端的单一因子（One-factor）风险。 - 美国居民净资产中已有 50% 配置于股票市场，个人财富与 AI 交易链条深度绑定，脆弱性极高。 ### [05:08](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=308s) 巨头寡头博弈：囚徒困境与不可逆的竞速动态 Ezra Klein 随后深入分析了核心玩家的决策逻辑。无论是 Meta、Alphabet 还是 OpenAI、Anthropic，管理层普遍认定未来只会产生极少数赢家，甚至是一个或两三个基础大模型主宰全球的寡头市场。马克·扎克伯格（Mark Zuckerberg）此前公开表达过这种心态，直言如果花冤枉了几千亿美元固然不幸，但落后挨打的风险要严峻得多。这种对极端集中市场的预期，导致每家公司都在按照超越整体市场承载力的规模进行过度投资。 Paul Kedrosky 对此完全赞同，称这正是历次基建狂潮中经典的囚徒困境。当年铁路公司为了不被挤出市场，也在不同城市之间拼命修建平行的重复线路。即便 Anthropic 的首席执行官 Dario Amodei 曾提出所谓“为技术前沿定速（Pace the frontier）”以减缓军备竞赛，也立刻遭到了同行的无视；其他竞争者不仅会借机超车，更会拿出“即使美国公司减速、中国也不会停下”的论调来合理化扩张。关于这种身不由己的军备竞赛，Paul Kedrosky 直言：“As long as you have a supply of capital and a race dynamic you're guaranteed over capacity much like whether it's canals, railroads, electrification, take your pick.” ```mermaid flowchart LR accTitle: AI巨头扩张的囚徒困境模型 accDescr: 展示竞争恐惧与资本供给如何迫使企业无法减速，最终导致全面产能过剩。 A["寡头通吃预期<br>(仅剩2-3个顶级模型)"] --> B["落后即出局的恐惧<br>(扎克伯格逻辑)"] B --> C["个体理性投资决策<br>(宁可浪费千亿也要超额投入)"] C --> D["外部竞争推波助澜<br>(中美竞速、对手绝不降速)"] D --> E["宏观必然结果:<br>产能严重过剩与资产减记"] ``` ### [07:20](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=440s) Azeem Azhar 的多头视角：真实的算力饥渴与“生产性繁荣” 面对 Paul Kedrosky 描绘的崩溃图景，Azeem Azhar 提出了截然不同的评估框架。他认为研判一场技术周期的健康程度必须抓住两个基本维度：其一是终端客户是否真正觉得该技术有用、是否正在为此掏出真金白银；其二则是融资机制是否存在短期债务错配和再融资断裂。历史上的泡沫崩溃，几乎都直接导源于短期信贷融资链条的断裂，而非单纯的技术演进本身。 根据 Azeem Azhar 团队建立的指标看板，眼下的现实情况与所谓的“产能过剩”大相径庭。美国及全球企业端对 AI 的需求依然极其强劲，算力供给并非过剩，而是处于严重的绝对短缺状态。全行业的真正瓶颈在于未来两到三年能否获取足够多的先进芯片，以及是否有足够的电力来驱动这些芯片。此外，当前的资金大头依然来自超大规模云服务商在过去二十年积累的庞大自有现金储备，尽管债务融资开始入场，但距离引发系统性金融危机的杠杆深度仍有相当长的距离。 关于当前周期的真实定位，Azeem Azhar 总结道：“So for me today it's you know it's in that productive boom with the hints of maybe getting for forgive my French a little bit stinky.” 他认为当前仍处于健康、具有实际生产力的繁荣阶段，即便边际上开始出现某些过热的不良苗头，也远未到达大破灭的临界点。 ### [09:20](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=560s) 算力层与代币需求：基础设施投资的闭环争议 讨论转向了市场需求的核心定义。Ezra Klein 提出，当人们谈论人工智能的需求时，最直接的表现是对“代币”（tokens）的消耗——用户提问或让模型编写代码时，本质上都在购买和消耗代币，例如像 OpenAI 或 Anthropic 这样的公司出售的正是代币使用量。然而，当前代币单价正在大幅跳水，这带来了一个严重疑问：即便总需求很高，人们是否真的愿意为昂贵的前沿模型买单？ Azeem Azhar 回应称，必须穿透代币看本质，回到更底层的资产结构。目前巨额资本所投向的是数据中心，其中大约 65% 的成本都花在 GPU 等 AI 加速芯片上。像亚马逊、微软、谷歌等超大规模云服务商（Hyperscalers）以及新兴的 Neoclouds 斥资数百亿美元购买芯片并出租算力。当前的算力投资之所以能维持极高的回报率，是因为 GPU 供不应求，出租方能够收取类似“Uber 动态加价”（Uber Surge pricing）的高额租金，在 18 到 24 个月内就能收回资产成本。 针对资金来源的健康程度，Azhar 解释了算力租金的客户构成演变：过去约 50% 的收入依赖于 Anthropic 和 OpenAI 两家前沿模型公司的巨额开销，另外 50% 来自其他 AI 企业及直接在云上运行开源/开放权重模型的机构。Azhar 坦承，如果未来三到四年依然是那两家公司支撑半壁江山，整个生态确实更像是“左手倒右手”的资金游戏；但现实情况是，该比例正在不断下降，越来越多真实的企业客户在直接购买云端算力，这证明市场每个季度都在走向成熟。 ```mermaid flowchart TD accTitle: AI 资本支出与算力变现链条 accDescr: 展示从底层数据中心资本支出到最终用户的价值流动与客户结构 subgraph CapEx["基础设施与资本支出"] DC["数据中心建设投入"] --> Chip["芯片/GPU 采购 (占成本 65%)"] Chip --> Cloud["超大规模云厂商 / Neoclouds 算力池"] end subgraph Revenue["算力出租变现 (18-24 个月回本)"] Cloud -->|"约 50% 收入 (比例正下降)"| Frontier["前沿实验室 (OpenAI / Anthropic)"] Cloud -->|"约 50% 收入 (比例正上升)"| RealBiz["企业客户与开源模型运行商"] end Frontier --> TokenEnd["最终用户：代码生成与复杂推理"] RealBiz --> NormalEnd["最终企业应用：日常自动化与压缩任务"] ``` --- ### [12:49](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=769s) 马粪里找小马：代币降价与早期用户误区 Paul Kedrosky 对这种乐观情绪提出了极具讽刺意味的质疑。他讲了一个经典笑话：一个小女孩在圣诞树下看到一大堆马粪，便兴奋地在里面拼命翻找，母亲问她在干什么，她回答“既然有这么大一堆马粪，里面肯定藏着一匹小马”。Kedrosky 认为，鼓吹代币暴增就等于创造价值，完全是在“马粪堆里翻找小马”的幻觉，混淆了规模与价值的区别。 Kedrosky 指出，算力收入的核心公式是“代币数量 × 单价”。在过去几年中，经性能调整后的代币价格每年暴跌大约 13 倍（13-fold）。虽然前沿模型公司坚称能够“入侵银行”的高难度旗舰模型依然享有高溢价，但大部分普通用户和企业根本不需要如此昂贵的算力。他以 Meta 的 Muse 等消费级爆款工具为例，这些应用仅使用相对原始轻量的模型就能风靡各大应用榜单，普通用户使用起来完全足够，根本不在乎是否是前沿模型。 更致命的结构性问题在于“早期客户误导”（early customer confusion）。Kedrosky 强调风险投资领域的一条铁律：“千万不要把你的早期客户误认为是真正的大众市场。”目前前沿模型公司大约 80% 的代币消耗与收入都来自于程序员： * **程序员极度偏好前沿模型：** 他们面临高度复杂的逻辑问题，因而愿意付出极高溢价。 * **代码生成属于“膨胀型”消耗（expansive）：** 用户只需输入 7 个单词的提示词，模型就会生成数百万行代码，瞬间消耗海量代币。 * **大众任务属于“压缩型”消耗（compressive）：** 正常人与普通企业的日常工作完全不同，更多是一对一交互或信息浓缩（例如把高盛发来的 25 封杂乱邮件提炼成简报），这使得单位任务的代币消耗极小。 Kedrosky 补充道，这种“代币狂欢”（token maxing）与内部比拼代币消耗的阶段已经破灭，普通企业在几个月内就会烧光预算，Meta 和微软已经大幅削减或限制内部工程师调用外部昂贵前沿工具的额度，市场正在迅速从浪费型代币消耗中清醒过来。 > “很多关于代币巨大规模的讨论，在我看来就很像在马粪堆里挖东西，觉得里面肯定有匹小马——因为人们把代币的庞大体量与实际价值混为一谈了。” —— Paul Kedrosky > “永远不要把你的早期客户误认为是未来真实市场的样子。” —— Paul Kedrosky --- ### [16:37](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=997s) 基础设施与模型实验室的脱钩？泡沫破裂的传染风险 面对 Kedrosky 的抨击，Azeem Azhar 反驳其在辩论中玩了“障眼法”（sleight of hand）。Azhar 强调必须把两件事严格区分开：一边是前沿模型实验室（如 OpenAI、Anthropic）自身的商业模式风险——即构建性能越来越强、但受众越来越狭窄的高端模型；另一边则是投入逾万亿美元构建的实体算力基础设施（GPU 与数据中心），这些基础设施承载着未来全美乃至全球企业的数字化转型。即便前沿实验室面临商业模式调整，底层数字基础设施依然会找到广泛的企业级和消费级应用场景。 Ezra Klein 对 Azhar 将这两者割裂看待的逻辑表示了强烈质疑。Klein 指出，一旦前沿实验室的商业闭环崩溃，其后果绝不会仅停留在软件层：如果事实证明企业更倾向于使用来自中国的 DeepSeek 这种成本极低的模型，导致 Anthropic 或 OpenAI 轰然倒塌，这种巨大的估值泡沫破裂将直接演变成整个股票市场的传染性灾难（contagion），甚至重创未来的信贷环境与整体宏观经济。 > “在那个巨大的马粪堆里，Paul 其实也玩了一套障眼法。Paul 讨论的是前沿模型和代币，而我谈论的是我们一直在聊的资本支出与产能——也就是 GPU。我认为这是两码事，理应分开讨论。” —— Azeem Azhar - **基础设施投资规模**：全球预计将有超过 1 万亿美元（trillion plus dollars）的资本流入数据中心与 GPU 基础设施建设。 - **市场溢价与投资回报**：云厂商目前依靠类似动态加价的算力租金，能够在 18 至 24 个月内实现硬件资产的回本。 - **大客户依赖度降温**：Anthropic 和 OpenAI 占云巨头算力支出的比例虽然目前在 50% 上下，但正呈现逐步下降趋势，外部真实客户份额在稳步提升。 - **代币通缩与预算缩减**：代币价格年均下降约 13 倍，企业正逐步终止内部无节制的“代币竞赛”，Meta 等巨头已将获准使用外部前沿工具的内部开发人员数量砍半。 ### [18:51](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=1131s) 基础设施投资风险与模型厂商的护城河 Azeem Azhar 建议将视野拉远，重新审视任何大型基础设施建设中涌现的根本风险：核心风险并非竞争对手变多，而是没有足够的终端客户去索求产品并为底层资产支付充足的溢价。如果企业客户选择转向由云服务巨头（Hyperscalers）直接托管和运行的开放权重模型（Open-weight models），这种需求转移虽然可能压缩像 Anthropic 这样的模型实验室的利润率，但底层的资本支出（即芯片和数据中心算力）依然得到了变现。 面对 Ezra Klein 和 Paul Kedrosky 提出的“Anthropic 可能会倒下，但英伟达（NVIDIA）依然安然无恙”的质疑，Azeem 强调模型厂商本身拥有极大的灵活度与自由度。他指出，就在节目录制前的 24 到 48 小时内，Anthropic 已经再次降价，并推出了效率更高、更轻量级的模型，而这些模型依旧保持着相当可观的利润空间。更重要的是，模型厂商正在复制传统软件行业数十年来的打法——围绕核心模型构建垂直应用场景与功能生态，例如推出面向办公协同的 Claude CoWork 或面向法律行业的专有工具。Azeem 认为，只要终端客户的真实需求依然在澎湃增长，整场技术投入就绝不会以惨剧收场。 ### [20:15](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=1215s) 什么是真正的“泡沫破裂”？ Ezra Klein 随后抛出了一个核心定义问题：究竟什么才算作“泡沫破裂（Pop）”？如果股市遭遇剧烈回调，但新建的数据中心依然被企业满负荷使用，这到底算泡沫爆裂，还仅仅是一场普通的市场修正？ Azeem Azhar 提出了一个严格基于金融历史的经典判断标准。在他看来，历史上的泡沫破裂必须同时满足两个特征： - 针对该特定底层资产的投资规模出现断崖式萎缩（通常要求整体投资水平暴跌 50% 以上）。 - 相关资产市场（如股票市场）发生全面崩盘，跌幅普遍达到 50% 或更高。 Azeem 认为，未来市场完全可能出现这样一种局面：信贷基金和超大规模云服务商因为之前在高位过度激进地投资，不得不忍受极度低迷、甚至令人尴尬的惨淡回报率。但 Azeem 直言：“这种程度的阵痛并不是泡沫破裂。那只是某些人做了一些糟糕的投资，在应该以更低价格买入时买高了而已。”（“That degree of pain is not a bubble bursting. That is just somebody made some bad investments and they bought high when they should have bought at a lower price.”） ```mermaid flowchart TD accTitle: 泡沫破裂与普通市场修正的判定分歧 accDescr: 对比Azeem Azhar基于历史指标的客观定义与Paul Kedrosky基于全市场脆弱性的系统性定义 subgraph Azeem["Azeem Azhar：历史经典定义"] A1["投资规模腰斩<br/>（CapEx 下跌 ≥ 50%）"] A2["相关资本市场崩盘<br/>（估值跌幅 ≥ 50%）"] A3["低回报与买高资产<br/>＝ 普通错误，非泡沫破裂"] end subgraph Paul["Paul Kedrosky：系统脆弱性定义"] P1["美国家庭净资产 50% 押注股市"] P2["美股重仓科技，科技全力押注 AI"] P3["巨型 IPO 与 CAPE 估值达到 1929 年前夕极值"] P4["任何链条微调均可触发<br/>系统性灾难崩溃"] end ``` ### [21:43](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=1303s) 超级 IPO 悬顶与 1929 年前夕的估值极值 Paul Kedrosky 对 Azeem 的定义表示全盘反对，认为设置所谓“投资跌 50%”的标准完全是武断的说辞。Paul 警告称，当前整个金融系统的脆弱性已经达到了历史罕见的程度：美国人目前约 50% 的净资产直接与股票挂钩，而股市的大头是科技股，科技股的重心全部压在 AI 上。这种极度集中的系统性风险，将随着 OpenAI 和 Anthropic 等超级巨头的上市而进一步恶化。 Paul 指出，如果将 OpenAI、Anthropic 以及 SpaceX 算在一起，这三家公司待上市的体量已经超过了“二战以来所有 IPO 融资规模的总和”，甚至其合计估值超过了当年纳斯达克在互联网泡沫巅峰时期的总市值。针对市场对此视若无睹的态度，Paul 情绪激动地抨击道：“伙计，仅仅因为泡沫长得不像你预期的样子，并不意味着它的脆弱性就会减少分毫。”（“Dude, just because it doesn't look the way you want it to look doesn't make it any less fragile.”） 此外，Paul 引用了衡量美股长期估值水平的席勒周期调整市盈率（CAPE Ratio）。他强调，当前的 CAPE 指数已经达到了与 1929 年大崩盘前夕完全相同的水平，处于整个人类金融史前所未有的极端区间。Paul 严厉批评那些认为“这次不一样”的乐观派缺乏想象力，并称本轮 AI 繁荣在逻辑上是“过度决定（Overdetermined）”的失败——存在太多诱发系统性断裂的导火索。即便仅仅是 Anthropic 的 IPO 从 11 月推迟到次年，都会立刻重挫一级市场的募资定价以及整个上游供应链的估值链条。 - **家庭资产集中度**：美国居民约 50% 的净财富捆绑在股市之中。 - **历史级 IPO 规模**：OpenAI、Anthropic 与 SpaceX 的预期合并体量，超过二战后所有 IPO 的总和，其估值体量已压倒当年的纳斯达克峰值。 - **估值指标见顶**：席勒市盈率（CAPE Ratio）当前位移已触及 1929 年股市崩盘前夕的历史极值。 ### [23:43](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=1423s) 资本市场的博弈：万亿预期与“博傻”套现 针对 CAPE 指数的高企，Azeem 反驳称该指标采用的是 10 年回溯期，完全无法作为预测崩盘时点的择时工具，且往往滞后于底层经济制度的范式转变。Azeem 进一步举例称，市场其实正在进行自我调节与理性定价，比如最近几周澳大利亚一家数据中心公司的 IPO 就因为投资者不买账而直接流产。但 Paul 随即打断并反驳指出，该数据中心 IPO 流产的真实原因并非投资者理性甄别，而是因为天空中盘旋着两个“万亿美元级别”的巨型 IPO 幽灵，它们吸干了市场上所有的风险资本，导致其他次要项目根本无法获得资金承接。 Ezra Klein 借此向 Paul 发问：如果这一切真是一场赤裸裸的泡沫，为什么市场上无数拥有真金白银、极具专业素养且利益攸关（Skin in the game）的成熟机构投资者，依然前赴后继地扎堆涌入这些项目？这难道仅仅是数学计算的错误吗？ Paul 解释了现代超级 IPO 的底层博弈逻辑。他指出，过去企业上市是为了给扩张融资，而如今的 OpenAI 和 Anthropic 在上市时可能已经拥有 700 亿甚至 1000 亿美元的年化营收跑道（Run-rate revenues），它们根本不再需要通过 IPO 融钱来搞业务发展。因此，这根本不是什么发展型融资，而是彻头彻尾的“退出事件（Exit event）”： - 早期内部人士、创始员工与机构股东在高位套现离场，他们内心清楚，在此等天价估值下，持有现金远比持有 Anthropic 或 OpenAI 的股权要安全得多。 - 外部接盘资金之所以蜂拥而入，也不是单纯出于对公司长期发展前景的信仰，而是寄希望于在解禁期（Lock-up）结束后，能以更高价格转手卖给下一个狂热者。 Paul 借用此前 SpaceX 早期股权交易后的波动趋势警示道，一旦禁售期结束，市场上将出现由原始股东抛售汇聚而成的“抛压之墙”。他用扑克桌上的经典谚语作结：如果所有人在入场时都只想着稍后转手卖给别人，“在扑克牌局里，如果你看了一圈都不知道桌上的‘更大的笨蛋（Greater fool）’是谁，那这个笨蛋就是你自己。” ### [27:46](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=1666s) 私有巨头的黑盒与 Anthropic 的客户集中度风险 Azeem Azhar 指出，当前关于 Anthropic 等企业能否维持健康商业模式的争论，核心在于它们仍是私有公司，外界只能依赖零散报道、传闻乃至偶发的内部泄密来拼凑真相。他认为这种信息不对称正是过去两三年市场的痛点所在：尽管这些前沿实验室撑起了科技板块占大盘近 38% 的估值狂潮，公众却无法像审视上市公司那样检验其基本面。只有等提交招股书（S-1 文件）时，外界才能真正看清其营收走势、利润率轨迹以及客户留存率。 ```mermaid flowchart TD accTitle: AI 前沿实验室 IPO 信息透明化对比 accDescr: 对比私有阶段与 IPO 提交 S-1 阶段的信息透明度及对市场的影响 subgraph Private["私有公司阶段（当前现状）"] P1["信息严重匮乏 / 依靠传闻与泄密"] P2["缺乏透明财报披露机制"] P3["高估值支撑科技板块（占比近38%）"] P1 --> P3 P2 --> P3 end subgraph Public["IPO 阶段（提交 S-1 招股书）"] S1["披露真实营收轨迹与利润率"] S2["揭示客户集中度与留存率"] S3["检验商业模式真实脆弱性 / 风险验证"] S1 --> S3 S2 --> S3 end Private -->|"迈向公开市场检验"| Public ``` Paul Kedrosky 随即抛出一个极具戏剧性的例证，形容这就像惊悚片里“凶手就在屋里”的情节：Anthropic 此前在招聘“巨型大客户总监”（mega account executive）时，意外泄露了其最大客户的身份，外界顺藤摸瓜推算出该客户正是 Meta，且 Meta 一家就贡献了其约 10% 的营收。虽然 Azeem 辩称这是 2025 年的旧账，且 Anthropic 仅今年第一季度的营收就已经超越了去年全年的总和，但 Paul 坚持认为，这种高度的客户集中度在同等体量的拟上市公司中是前所未闻的。他推测，Anthropic 可能仅靠 8 家公司就支撑了 80% 的营收，而且这些客户几乎全都在做同一件事——写代码。 - 传统美股监管规定：若单一客户贡献的营收达到或超过 10%，公司必须在招股书（S-1）中予以明确披露。 - Paul 提出极具可证伪性的预测：如果 Anthropic 递交 S-1 招股书，外界将能核实是否存在“8 家客户贡献 80% 营收”的异常集中现象。 - Paul 从前股票分析师的视角犀利定性：“The dependency on such a small number of customers who are mostly using it for the same thing is historically unprecedented to have such customer concentration.”（依赖如此少量的客户，而且他们大多将其用于同一用途，这种客户集中度在历史上是前所未有的。） --- ### [31:49](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=1909s) 缺席前沿大模型竞赛的苹果：失误还是搭便车？ 主持人 Ezra Klein 转移话题，提出了一个耐人寻味的对比：当 Alphabet、Meta、微软以及负债累累、信用评级濒临垃圾级的 Oracle 每年豪掷 1000 亿至 1200 亿美元进行资本支出（capex）狂揽算力时，苹果却反其道而行之，不仅没有研发自己的前沿基础大模型，年均资本支出也仅在 150 亿美元左右，但其股价年内表现却跑赢了所有竞争对手。这是否意味着不参与前沿大模型“奥运会”反而是更优策略？ Azeem 分析称，苹果的商业模式建立在硬件销售（iPhone、Mac）与订阅服务之上，无论 AI 浪潮如何演变都能稳定变现。此外，苹果的企业文化对细节和掌控力有着近乎严苛的执念。他回忆起乔布斯在初代 iPhone 发布时，曾因谷歌 Logo 里的黄色色号不对而痛骂谷歌高级副总裁的轶事；相比之下，大语言模型本质上是不可控的、容易“陷入窘境”（hapless）。面对庞大的十亿级设备生态，苹果绝不愿冒着破坏品牌体验的风险去交付一个混乱的 LLM。 Paul 则持有截然不同的刻薄观点，认为外界高估了苹果的远见。在他看来，苹果并非谋定而后动，纯粹是内部多次尝试受挫、项目解散、Siri 整合反复无常后的“因祸得福”。通过不作为，苹果成了终极的“搭便车者”（free rider），坐收其他人用大火烧钱砸出来的基础设施成果，自己则全身而退。 - 对比鲜明的资本支出：其他科技巨头年均 capex 达 1000 亿至 1200 亿美元，而苹果年均 capex 仅约 150 亿美元。 - Paul 直言不讳地揭露真相：“I don't think Apple was very smart about this. I think they mostly screwed it up and got lucky.”（我不认为苹果在这件事上有多聪明。我认为他们主要是搞砸了，然后运气好罢了。） --- ### [34:41](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=2081s) 通用技术的配套经济学与巨头的资产负债表底气 Paul 进一步指出，前沿实验室砸下的巨额资金必然会外溢给外部生态。除了苹果这种高居生态位顶端的掠食者外，像 Cloudflare 这样原本被市场忽视的基础网络安全公司，在过去 18 个月内股价暴涨三倍。原因在于全网激增的自动化 Bot 和 AI Agent 构成了全新的流量冲击，迫使企业纷纷采购 Cloudflare 的防护服务。这与 1920 年代电气化浪潮如出一辙：1929 年公用事业公司纷纷崩盘后，真正的赢家变成了制造空调和电器的下游企业。 Azeem 对此表示认同，但他更倾向于用经济学中的“互补性服务与互补性商业”（complementary services and complementary businesses）来定义这一现象。他解释称，任何通用技术（General Purpose Technology, GPT）的演进轨迹都是如此——技术底座不断迭代，随后衍生出维护、防御与全新应用场景的配套生态。 话题最后，Ezra 抛出了泡沫辩论中的另一个核心分歧：历史上的铁路泡沫等危机中，推动基建的企业往往背负着沉重的致命杠杆；然而眼下的 Alphabet 和 Meta 却拥有极其健康的资产负债表与充沛的自由现金流。当马克·扎克伯格（Mark Zuckerberg）表态称“即便错掷数千亿美元也比错失浪潮要好”时，坐拥庞大现金储备的科技巨头是否真能对这场狂赌免疫？ - 科技外溢受惠典型：Cloudflare 凭借抵御 Bot 与 Agent 恶意流量的需求，股价在过去 18 个月翻了三倍。 - 历史对照：1920 年代美国电气化建设中公用事业公司破产，利润却流向了空调等电气设备制造商。 - 经济学范式：Azeem 强调通用技术（GPT）必然催生跨行业的互补性需求，而非单纯投机性的“搭便车”。 ### [36:50](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=2210s) 债务驱动的狂欢：科技巨头资产负债表的恶化 主持人 Ezra Klein 抛出了市场上的一个主流论调：以往的技术狂潮往往通过高杠杆的金融衍生品（如 2008 年金融危机）席卷整个经济，而眼前的这场 AI 热潮，似乎是由拥有巨额现金储备的巨头主导；哪怕这些巨头挥霍数千亿美元，它们自己的资产负债表也足以承受损失，不至于对外部经济产生传染风险。然而，Paul Kedrosky 严厉反驳了这种观点，并指出了危险的转变：一年前人们还在嘲笑他的担忧，声称大厂现金流无敌；但如今，巨头的资本支出正迅速从“消耗自由现金流”转变为“大举借债”。 华尔街擅长的事情正是把干净的资产负债表弄脏，因为承销发债可以赚取天文数字的佣金。随着巨头们不断发债，其信用评级开始面临下调风险——甲骨文（Oracle）已经显露颓势，微软（Microsoft）等其他超大规模云厂商（Hyperscalers）也已逼近降级边缘。为了掩人耳目，Meta 等公司开始玩起类似金融危机的“表外融资”游戏，利用特殊目的载体（SPV）把庞大的债务隐藏起来。科技行业过去赖以成名的“高增长、无负债”标签已经荡然无存，如今科技巨头反倒成了美国投资级债券的最大发行主体。其发债规模之庞大，甚至与美国 10 年期国债形成了边缘竞争，迫使国债收益率走高并引发债市震荡。Paul Kedrosky 发出警告，这场资本游戏最终会演变成难以收场的债务积压： > “This kind of all these corporations with giant balance sheets loaded with debt that are trying to work it off.” —— Paul Kedrosky ```mermaid flowchart TD accTitle: 科技巨头 AI 融资与债务外溢机制 accDescr: 展示科技巨头如何从消耗自有现金转变为发行巨额债券与表外负债，进而对宏观债券市场造成冲击。 subgraph CapitalShift["资本支出模式转变"] Cash["自有巨额现金流支出"] -->|"资金需求超出负荷"| Debt["大举借债融资"] end subgraph WallStreet["华尔街运作与结构化手段"] Debt -->|"赚取承销费用"| Bonds["大规模发行投资级债券"] Debt -->|"规避资产负债表恶化"| SPV["设立 SPV 等表外融资架构"] end subgraph MacroImpact["宏观经济与信用外溢"] Bonds --> Downgrade["企业信用评级承压 (如 Oracle)"] Bonds --> BondShock["挤压 10 年期美债流动性，推高国债收益率"] SPV --> Overhang["形成类似 1990 年代日本的沉重债务包袱"] end ``` ### [40:14](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=2414s) 挤出效应：AI 算力基建对实体经济的“荷兰病” Ezra Klein 进一步指出，除了资本层面的债务竞争外，AI 的大跃进还引发了严重的“实体资源挤出”。在住房严重短缺的美国，大量建筑材料、工人和资本被源源不断地抽调至数据中心建设，导致住宅建设支出下降，普通制造业厂房的投资占比也在萎缩。把本可以投向能实质改善民众生活、回报更为稳健的民生住房与制造业的资源，全部押注在可能落空的 AI 基础设施上，存在巨大的社会成本。 Azeem Azhar 承认这种挤出效应是一个极度棘手的复杂难题。他提到，他在 2025 年夏天曾提出一个衡量“繁荣还是泡沫”的门槛：一旦超大规模云厂商的外债融资比例突破 30%，系统性风险就会急剧上升，而 2026 年底的市场恰恰逼近了这条红线。他将眼下的美国经济与 1970 年代北海发现天然气后的荷兰相比——当时整个荷兰的劳动力与资源全被天然气开采吸干，从而形成了削弱其他工业部门的“荷兰病”（Dutch disease）。针对完全依赖市场调节无法纠偏的失衡局面，Azeem Azhar 强调了政策干预的必要性： > “And I'm afraid my only answer is that it requires something of a policy stimulus to make sure that those sectors get the activity that they need because I'm not persuaded that market forces on their own would do that.” —— Azeem Azhar 对此，Ezra Klein 带着苦笑吐槽道，在当前政治领导飘忽不定、国家负债高企且借贷成本飙升的背景下，指望推出精准的政策刺激无异于痴人说梦。Azeem Azhar 则给出了一个带有长远希望的转折点：虽然眼下数据中心建设吸空了电工、暖通空调（HVAC）技工与建筑劳动力，但这也倒逼了全社会的职业技能培训与产业升级；两三年后，这些大量培养出来的熟练产业大军和扩产后的建筑工程公司，将转化为整个经济体共享的基础能力。 ### [44:20](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=2660s) 现实落地困境：从通用电气到可口可乐的 AI 转型 抛开资本市场的炒作与金融账目，Ezra Klein 将视角拉回到实体产业：如果这场耗资数万亿美元的豪赌最终要产生回报，那么像通用电气（GE）、可口可乐（Coca-Cola）这样的传统主力企业，到底需要展现出怎样的图景？企业如果不把预算花在人力上，而是全部拿去采购 AI Token，整个商业社会的运作逻辑会发生何种根本改变？ Azeem Azhar 分享了他每年与欧美数十位企业高管沟通的实操反馈。企业在推进 AI 时的真实心态非常一致：“进展比想象中困难，但我们必须坚持，而且已经开始初见成效。”然而，这些成效在财务层面的兑现机制却是残酷的：除非业务本身能够实现爆炸式增长，否则企业从 AI 获得的效率提升，绝大部分将直接体现为对劳动力的缓慢裁撤与替代。人均创收提高了，但如果公司总业务盘子没有扩张，员工总数就会不可避免地下滑。不仅如此，企业真正的“AI 原生化”改造极其缓慢，就像当年福特汽车（Ford）建立端到端流水线实现了“电力原生”，但美国绝大部分企业直到 15 到 20 年后才真正完成电气化普及一样。在当下主流实体经济中，人们甚至几乎找不到一家真正脱胎换骨的非科技类“AI 原生”企业。 - **AI 回报机制的核心**：企业通过应用 AI 提高单位员工产出，在营收未发生同等倍数暴增的前提下，主要通过自然缩招或减少员工编制来兑现利润。 - **历史技术的扩散周期**：电力的普及历经数十年，福特汽车的流水线只是起步范式，而当今的传统骨干企业（Main Street firms）在 AI 基础设施和业务流程的重塑上，依然处于极度早期的摸索阶段。 ### [46:20](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=2780s) AI 原生企业与劳动力替代的争论 主持人提问：如果所谓的“AI 原生公司”核心就是用 AI 全面替代人类员工，那整个经济体系是否会迅速滑向所有人都在担忧的“净失业”局面？毕竟新成立的初创公司在理论上会更加 AI 原生，因而对人力劳动的依赖更少。如果在这种工作被大量剥夺的经济环境下，究竟还有谁有钱去消费和支撑市场？ Azeem Azhar 对此持有较为长期的演进视角，他认为现实中存在多重机制： 1. **人工校验与卡位溢价**：AI 依然有无法胜任的领域，需要人类从事最后一环的验证和审批。“处在这些关键卡位节点上的人，最终收入反而会越来越高。” 2. **全新职业工种的诞生**：未来会出现大量当下无法想象的角色，就像 1993 年谁也想不到会出现“网站管理员”（Webmaster）一样。 3. **通用目的技术（GPT）的文化与衍生效应**：正如亨利·福特（Henry Ford）普及汽车后，带来了郊区生活方式、度假需求以及庞大的周边服务业链条。 Azeem Azhar 坦言，新岗位的创造周期确实可能比技术淘汰旧岗位的速度更慢，但他强调自己的观察时间跨度是 5 到 10 年，而不是短短的一两年。管理层消化和重塑业务流程需要时间，不是从 OpenAI 拿来一个顶级模型下令“修复我的公司”，企业就能瞬间脱胎换骨的。 ```mermaid flowchart TD accTitle: Azeem Azhar 的劳动力吸收模型 accDescr: 展示技术普及到产生新岗位与卡位高薪的演化链条 subgraph ChokePoint["短期承压与关键节点"] A["引入顶级模型"] --> B["企业无法瞬间改变"] B --> C["关键人类验证层 / 签字审批"] C --> D["卡位节点获得更高薪酬"] end subgraph LongTerm["长期结构性创造 (5-10年)"] D --> E["催生全新工种 (类似 1993年后的网站管理员)"] E --> F["衍生全新生活方式与周边服务业"] end ``` ### [49:02](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=2942s) “重塑劳动力定价”与西海岸的信条裂痕 Paul Kedrosky 强烈反对 Azeem Azhar 的乐观态度，直言当下的本质就是对人类劳动力进行暴力的“重新定价”（Repricing Labor）。他指出： “我们基本上是在给劳动力重新定价，对吧？这是历史上普及速度最快的技术，而它的核心功能就是给认知劳动力重新定价。” Paul Kedrosky 尖锐地批评了硅谷的意识形态： - **西海岸信条的失效**：“西海岸在这方面的教条是‘技术创造的工作岗位永远多于它所摧毁的’，这成了一句人们反复念诵的真言，寄希望于它会成真。” - **无前例的技术冲击**：人类历史上从未见过一种通用目的技术，既具备如此广泛的适用性，普及速度又快如闪电，同时还在直接拉低认知劳动的边际成本（例如让呼叫中心的成本呈断崖式下跌）。 - **阵痛远快于适应**：在这种急剧冲击下，社会根本来不及摸索出下一代“网站管理员”究竟是什么，劳工阶层就已经面临颠覆。轻描淡写地宣称“未来岗位会更多”，完全背离了受冲击群体的真实生活体验，无异于轻蔑地安慰大众“别担心，一切都会好起来的”。 ### [50:27](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=3027s) 抵制浪潮、职业壁垒与断裂式监管风险 讨论进一步延伸到由社会反弹引发的“AI 泡沫破裂”风险：如果企业在高利率或衰退压力下，急于用 AI 替换需要医保、会组织工会、有自主意见的人类员工，必然遭遇猛烈的社会反噬。 - **行业抵制已有先例**：教师群体完全可以联合起来宣布“禁止 AI 进入教室”；医生群体也能在几起 AI 误诊事故后，推动将算法彻底锁在医院大门之外；正如人们因环保和生活受扰而抵制在居住地周边兴建数据中心。 - **灾难事件触发系统性干预**：Anthropic 在其招股书的风险提示中直白列出“对人类造成生存危机”；若某天 OpenAI 或相关模型发生失控，导致类似富国银行（Wells Fargo）等金融命脉或海外关键基础设施被黑客攻破，监管将会“不是渐进式、而是一次性全面砸下”。 Paul Kedrosky 对这种“断裂式监管”（Discontinuous Regulation）深信不疑，认为从一开始科技公司就是在公众身上做“未经许可的药物临床试验”： “科技公司能够不受惩罚地这么干，仅仅是因为我们刚好生活在一个它们目前还没害死我们的时间线上，这种逻辑荒谬至极、站不住脚。” Paul Kedrosky 用概率论模型说明监管的必然性： - 假设存在 6 种导致严重后果的情境，每种发生概率仅 5%，其中至少发生一起的概率就是约 20%。 - 如果列出 12 种恶性情境，每种概率同样只有 5%，那么至少有一件暴雷的复合概率将超过 50%。 - 现实中潜在的严重事故场景远超一打，因此发生恶性事件并倒逼激进监管的概率高达 50% 到 60%。未来的监管形态很可能类似 FDA（美国食品药品监督管理局）或原子能委员会（Atomic Energy Commission），带来极其刚性的行政硬约束。 ### [54:07](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=3247s) 政治风险、民主表达与社会许可 Azeem Azhar 从另一个侧面看待政治与合规风险。他认为行业在一定程度的监管下其实能发展得更好，正如制药业的发展轨迹一样。此外，从底层大模型研发团队到超大规模云服务商（Hyperscalers），整个 AI 产业链聚集了巨额利润，产业界完全具备承担合规与社会成本的资金实力。 针对美国各地针对数据中心建设的抗议潮，Azeem Azhar 反倒持积极看法： - **民主诉求的自然表达**：民众抗议数据中心落地，是对技术巨兽及其显性负面影响的不安与制衡，是一种健康的民主意志表达。 - **重构“社会许可”（Social License）**：这种博弈能倒逼出更受公众认可的社会许可机制。 - **地域治理的差异化探索**：美国可能依靠各州层面的独立治理逐步化解冲突；而在英国，政府已在探索要求数据中心开发商将一部分年金或收益直接返还并投资于当地社区，以弥补巨型数据中心占用了大面积土地和资源、却几乎不创造本地就业岗位的结构性缺陷。 ### [55:37](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=3337s) 数据中心的高昂补贴与反噬：监管风险的两个极端 讨论从数据中心极其夸张的就业与补贴成本切入。建设耗资数十亿美元的数据中心，最终往往只能提供两到六个长期岗位。Paul Kedrosky 举例提到新墨西哥州的一个项目：如果把税收减免以及为配合数据中心而对当地公用设施进行的强制改造全部算上，平均创造一个就业岗位的补贴成本就高达 400 万美元。对此，Paul Kedrosky 毫不客气地评价道：“The numbers are insane.”（这些数字简直荒谬透顶。） Azeem Azhar 承认这种模式在长期来看根本无法维持，但他指出，眼下产业链上下游都在赚取真金白银的利润，这笔账单最终必须由相关利益方承担。自我监管绝不可能奏效，必须在地方、国家乃至国际层面建立硬性规则，规范其运行方式及可能造成的外部损害。 随后，Ezra Klein 进一步指出了监管走向的极端不确定性，认为它既可能像医药监管，也可能滑向核电监管的死局： ```mermaid flowchart TD accTitle: AI 监管可能走向的两种路径 accDescr: 对比医药行业监管模式与核电行业监管模式下对待新技术的不同态度与结局。 subgraph Med ["医药监管路径 (Pharmaceutical)"] M1["目标：确保药品安全与功效"] --> M2["容忍研发与应用，虽审批缓慢但允许通行"] end subgraph Nuc ["核电监管路径 (Nuclear Power)"] N1["公众恐慌与极端严格约束"] --> N2["最终诉求演变为：彻底不再建任何新核电站"] end Tech["AI 技术与社会反弹"] --> Med Tech --> Nuc ``` Ezra Klein 警告，如果公众对 AI 产生强烈的抵触情绪，监管诉求可能不是让它变得更安全，而是演变成民意层面的彻底拒绝。正如 Ezra Klein 所言：“We don't want to be replaced by robots, which actually when you say it aloud like that seems like a fairly straightforward thing and probably the kind of thing that people are going to express in a democracy.”（“我们不想被机器人取代”，当把这句话大声说出来时，这其实是非常直白的心声，而这很可能正是人们在民主社会中会表达出的态度。） - **数据中心人均就业成本**：以新墨西哥州为例，算入税收减免与公共基础设施投入，每个新增就业岗位的政府补贴高达 400 万美元，而一座数十亿美元的设施通常仅雇佣 2 到 6 名常驻员工。 - **监管光谱的双重端点**：一端是旨在确保产品安全、虽然放缓但仍在向前推进的医药模式；另一端则是由于社会反弹和严格限制，导致新项目全面停摆的核电模式。 --- ### [57:59](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=3479s) 消失的社会摩擦力：“针对现代性的拒绝服务攻击” 除了担心被替代，社会公众对 AI 代理（Agent）侵入日常生活也产生了巨大的疲惫与反感。Ezra Klein 列举了现实中的典型乱象：餐厅前台被各种私人 AI 代理的订位电话狂轰滥炸，导致人类员工根本无法再接听电话；企业经理被泛滥的 AI 自动生成报告淹没；员工的收件箱里全是由大模型生成的寒暄询价，导致对真正邮件的回复效率全线崩溃。正如 John Herrman 在《纽约杂志》（*New York Magazine*）中所描写的，商业机构正在承受来自无数 AI 代理涌入的第一波痛苦，而整个社会根本没有准备好迎接每个人都拥有一个机器人代理四处打电话压价、订位或抢票的世界。 尽管 Azeem Azhar 提到付费 AI 订阅的渗透率其实很低——全美仅约 2.5% 的人购买了 AI 付费服务，甚至还不如购买 GLP-1 减肥药的人多（因为减肥药几周内就能让人感受到实打实的生理变化）——但数以亿计的人正在通过免费工具接触 AI。随着 Meta 等巨头强推，即使人们在社交媒体上抱怨，这种无孔不入的工具依然会强行渗透到每部手机中。骨科医生收到患者用 ChatGPT 生成的成千上万字的病例描述，或是普通人为拦截 AI 骚扰电话每月向电信运营商支付一美元，这些防范成本在账面上甚至还会被计入 GDP。 Paul Kedrosky 提出了一个深刻的洞见：西方资本主义总是一味推崇效率，但人类社会在本质上是依赖“摩擦力”（frictions）来维持平衡的。 ```mermaid flowchart LR accTitle: 盲目消除摩擦力引发的现代性 DoS 攻击 accDescr: 展示 AI 消除社会日常摩擦后，如何演变成对社会体系的拒绝服务攻击。 A["消除社会摩擦力 (无成本提交诉讼/自动外呼/AI公文)"] --> B["边际成本降为零的无底线调用"] B --> C["系统过载 (对现代性的拒绝服务攻击 DoS)"] C --> D["双向AI博弈 (AI对接AI，体验恶化)"] D --> E["社会强烈反弹：全面抵制某些应用场景"] ``` Paul Kedrosky 强调，一旦技术将那些人们习以为常的社会阻力盲目抽干，后果将是灾难性的。例如借助大模型向小额索赔法院疯狂递交诉状，或是向餐厅发起无休止的订位轰炸，这种泛滥无异于一场技术层面的系统瘫痪。Paul Kedrosky 概括道：“Blindly removing all of them is kind of creating in tech parlance almost a denial-of-service attack on modernity.”（盲目消除所有这些摩擦，用技术术语来说，几乎就是在对现代性发动一场拒绝服务攻击。）最终的结局可能就是两个 AI 在前台与后台互相扯皮踢皮球，构建出一个越来越令人不快的经济生态，并最终招致公众的全面抵制。 - **AI 普及率对比**：全美仅有约 2.5% 的人口为 AI 服务付费，该比例甚至低于全美自费使用 GLP-1 类减肥药物的人数。 - **无感成本增加**：用户不得不向电信运营商每月支付约 1 美元购买防伪造防骚扰服务，这类由 AI 衍生出的防范费用最终在经济统计中反被算作 GDP 增长。 --- ### [01:01:57](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=3717s) 证伪时刻：怎样的宏观指标能改变他们的观点？ 行将结束时，Ezra Klein 向两位嘉宾抛出了一个核心问题：究竟看到什么样的数据或迹象，才会让他们推翻自己目前的立场？ Azeem Azhar 从其乐观的立场出发，表示支撑他信心的核心在于真实的商业客户是否在投入真金白银并获得实际成果。如果他看到美国企业在 AI 上的支出增长速度出现断崖式下跌，例如连续两个季度的增速腰斩（halving），那就意味着客户的胃口已经达到饱和，后续的营收将不足以支撑健康的商业运转，这便会让他感到深切的不安。 而长期持怀疑态度的 Paul Kedrosky 则明确表示，自己一切以宏观经济硬指标为准绳。他开玩笑称自己坚信自己在大部分事情上都是错的，因此很乐意看到更多能打脸自己的证据。他指出，如果要证明 AI 的支出是合理的，必须看到真正的宏观生产力震荡： > “So if we began to see I don't know productivity growth go well beyond trend that would be striking, right? I mean so far we're sitting at around trend data with respect to U.S. productivity growth. There's no sign of, you know, A.I. improving productivity.” —— Paul Kedrosky Paul Kedrosky 列举了能够彻底逆转其悲观态度的三大宏观指标： 1. **劳动生产率大幅突破历史趋势线**：目前美国生产率增长仍然仅仅徘徊在长期趋势线附近，毫无 AI 赋能的实质性抬升迹象。 2. **劳动参与率剧烈上升**：证明 AI 是一种具包容性的技术，真正能把更多劳动力重新带回实体经济之中。 3. **GDP 增速显著超出历史趋势水平**：当前经济增速仍处于二战后长期趋势区间的低端。 Paul Kedrosky 强调，如果这三项指标（甚至需要三者同时出现）发生逆转，那就说明 AI 的巨额资本开支得到了经济现实的支撑；但如果这些数字始终没有起色，当下的狂热与透支结局将更像 1930 年代的大萧条，而不是 2000 年代尚有缓冲的互联网泡沫破裂。 --- ### [01:04:01](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=3841s) 最后的推荐书单：Paul Kedrosky 谈《米德尔马契》 在对谈的结尾环节，Ezra Klein 按照惯例请两位嘉宾推荐三本著作。Paul Kedrosky 首先提到了乔治·艾略特（George Eliot）的经典小说《米德尔马契》（*Middlemarch*）。 Paul Kedrosky 解释道，这部作品是他观察普通人在被历史滚滚车轮卷入时的首选读物。全书的时代背景正值第一次工业革命全面确立并重塑社会的关键历史节点，生动刻画了身处重大技术与制度变革狂潮中的芸芸众生，记录了普通人如何在时代剧变中被不可抗拒的力量推着向前。 - **推荐书目**：《米德尔马契》（*Middlemarch*），乔治·艾略特（George Eliot）著。 - **推荐理由**：描摹工业革命全面铺开时普通人在巨大技术洪流中被席卷与重塑的命运轨迹。 ### [01:04:26](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=3866s) 历史转折与商业崩塌：Paul Kedrosky 的书单推荐 在节目的常规荐书环节中，SK Ventures 合伙人 Paul Kedrosky 结合整场关于技术颠覆、资本狂热与社会变革的讨论，分享了三部历史视角浓厚的作品。他首先推荐了历史学家 Emma Griffin 的著作《自由的黎明》（*Liberty's Dawn*）。这本书通过收集英国工业革命时期大量普通工人的自传、日记和一手记录，展现了微观个体如何在巨大的技术和经济变革巨浪中挣扎与生存。Kedrosky 感慨道，当今天的学者和分析师们滔滔不绝地谈论宏观经济数据时，这本书却将目光投向了那些亲历剧变的普通人：“what it was like in the middle of that wrenching transition and as you know people are talk about economic data and everything else what it was like to live through it what did it feel like what was the pain I felt and you know what were my hopes for the future it's just like a wonderful book about that.” 接着，Kedrosky 推荐了 Forrest McDonald 所著的传记作品《因萨尔：一位公用事业亿万大亨的兴衰》（*Insull: The Rise and Fall of a Billionaire Utility Tycoon*）。传记主角 Samuel Insull 是 1920 年代美国商业史上举足轻重却在今日几乎被遗忘的企业家，他曾经利用刚刚普及的通用目的技术（GPT）——电力，打造了庞大无比的商业与公用事业帝国。Kedrosky 指出，这段历史与当下科技巨头围绕 AI 构建帝国的路径如出一辙，但即便初期大获成功，盲目扩张和金融杠杆依然可能引发惨痛的结局：“He was the probably one of the most important entrepreneurs of the 1920s whose empire and electricity shows what happens that you can build building an empire on top of a GPT and even doing that successfully can still end up in collapse and calamity but it's a wonderful story.” - **《自由的黎明》（*Liberty's Dawn*）**：作者 Emma Griffin，基于工业革命时期蓝领工人的第一手自传记录，重构普通人在剧烈转型期的切身体会、痛苦与希望。 - **《因萨尔：一位公用事业亿万大亨的兴衰》（*Insull: The Rise and Fall of a Billionaire Utility Tycoon*）**：作者 Forrest McDonald，深入记录 Samuel Insull 凭借电力基础设施扩张崛起并在金融风暴中坍塌的商业史。 --- ### [01:05:28](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=3928s) 知识革命、代际冲突与批判启蒙：Azeem Azhar 的书单推荐 《Exponential View》创办人 Azeem Azhar 随后带来了他的三本推荐读物，视野横跨科学史、经典文学与哲学讽刺。第一本是英国 BBC 知名科学记者 James Burke 于 1980 年代初出版的经典之作《宇宙改变的那一天》（*The Day the Universe Changed*）。Azhar 透露自己少年时期就读过这本书，最近重新温读，依旧深感震撼。该书贯穿西方与亚洲上千年的历史，探讨了重大知识范式转移如何彻底重塑人类对自我身份与存在意义的认知：“it is all about how as we develop new types of knowledge we start to rethink our role in the world, what matters to us and how unsettling that is.” Azhar 推荐的第二本书是 19 世纪俄国文豪伊万·屠格涅夫（Ivan Turgenev）的长篇小说名作《父与子》（*Fathers and Sons*）。故事发生在新兴科学与现代性思潮全面涌入传统俄国社会的转型节点，展现了追求理性实证的年轻一代与笃守传统价值的父辈之间剧烈的代际冲突。在 Azhar 看来，即便面对最前沿的科学突破，有些关于人性的核心本质依然无法被科学直接解答：“What even the the most recent breakthroughs in science can't tell you about what it is to be human. I think very relevant.” 节目的最后，Azhar 简短提到了他正在重读的第三部作品——法国启蒙思想家伏尔泰（Voltaire）的哲理讽刺小说名作《老实人》（*Candide*），主持人 Ezra Klein 随即向两位对谈嘉宾表达了真挚的感谢，为这场兼具经济深度与历史反思的 AI 泡沫辩论画上了句号。 - **《宇宙改变的那一天》（*The Day the Universe Changed*）**：作者 James Burke，追溯人类在重大科学与知识发现爆发时，自身世界观与文明秩序所经历的动荡与重构。 - **《父与子》（*Fathers and Sons*）**：作者伊万·屠格涅夫，描摹科学理性浪潮冲击传统道德秩序时的代际摩擦，以及科技无法取代的人性羁绊。 - **《老实人》（*Candide*）**：作者伏尔泰，经典启蒙哲理小说，深刻讽刺了盲目乐观主义，呼吁脚踏实地面对现实困境。 ## Notable quotes > [49:02](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=2942s) 这是历史上普及速度最快的技术，而它的核心功能就是给认知劳动力重新定价。 — Paul Kedrosky > [05:08](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=308s) As long as you have a supply of capital and a race dynamic you're guaranteed over capacity much like whether it's canals, railroads, electrification, take your pick. — Paul Kedrosky > [12:49](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=769s) 永远不要把你的早期客户误认为是未来真实市场的样子。 — Paul Kedrosky > [01:28](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=88s) The bubble is this vast oversupply of something largely fueled now by credit which in turn will lead to massive business failures and write down of assets and all of that credit has spread through the economy. — Paul Kedrosky > [55:37](https://www.youtube.com/watch?v=9GLdGc9OX-A&t=3337s) We don't want to be replaced by robots, which actually when you say it aloud like that seems like a fairly straightforward thing and probably the kind of thing that people are going to express in a democracy. — Ezra Klein ## Mentioned Ezra Klein, Paul Kedrosky, Azeem Azhar, SK Ventures, Exponential View, OpenAI, Anthropic, Meta, Alphabet, Mark Zuckerberg, Dario Amodei, DeepSeek, NVIDIA, SpaceX, Oracle, Microsoft, Apple, Cloudflare, Steve Jobs, General Electric, The Coca-Cola Company, Henry Ford, Wells Fargo, FDA, Atomic Energy Commission, George Eliot, Emma Griffin, Samuel Insull, Forrest McDonald, James Burke, Ivan Turgenev, Voltaire, Middlemarch, Liberty's Dawn, Insull: The Rise and Fall of a Billionaire Utility Tycoon, The Day the Universe Changed, Fathers and Sons, Candide ## Video description The gargantuan bet on the future of A.I. is driving the U.S. stock market and economic growth. And the building of data centers is sucking up so many construction resources, it’s crowding out the construction of homes. The multitrillion-dollar question is: Will this bet pay off? Or is it a bubble, destined to burst and pull the economy down with it? Paul Kedrosky is a partner at SK Ventures, a fellow at M.I.T., and the author of a newsletter about A.I.. He thinks we’re very obviously in a bubble. Azeem Azhar is a technology entrepreneur and the author of the newsletter Exponential View. He has a more optimistic take. 00:00 Intro 01:58 The Case That There’s an A.I. Bubble 07:20 Azeem’s Dashboard View 09:12 Tokens and Pricing Pressure 16:37 Compute vs. Model Labs 20:15 What Counts as a Pop? 22:01 OpenAI and Anthropic’s I.P.O.s 31:49 Why Is There No Apple A.I.? 37:34 Debt-Fueled AI Buildout 44:20 What A.I. Adoption Looks Like 46:32 Repricing Jobs and Labor 50:16 Backlash and Regulation Risk 57:59 We’re Not Ready for A.I. Bots 01:01:57 What Would Change Their Minds? 01:04:03 Book Recommendations Read the full transcript here: https://www.nytimes.com/2026/10/16/opinion/ezra-klein-podcast-azeem-azhar-paul-kedrosky.html Watch more on @EzraKleinShow Thoughts? Guest suggestions? Email us at ezrakleinshow@nytimes.com. You can find transcripts (posted midday) and more episodes of “The Ezra Klein Show” at nytimes.com/ezra-klein-podcast. Book recommendations from all our guests are listed at https://www.nytimes.com/article/ezra-klein-show-book-recs.html [Source](https://www.youtube.com/watch?v=9GLdGc9OX-A)

[49:02] Paul Kedrosky: 这是历史上普及速度最快的技术，而它的核心功能就是给认知劳动力重新定价。 [5:08] Paul Kedrosky: As long as you have a supply of capital and a race dynamic you're guaranteed over capacity much like whether it's canals, railroads, electrification, take your pick. [12:49] Paul Kedrosky: 永远不要把你的早期客户误认为是未来真实市场的样子。 [1:28] Paul Kedrosky: The bubble is this vast oversupply of something largely fueled now by credit which in turn will lead to massive business failures and write down of assets and all of that credit has spread through the economy. [55:37] Ezra Klein: We don't want to be replaced by robots, which actually when you say it aloud like that seems like a fairly straightforward thing and probably the kind of thing that people are going to express in a democracy.

[5:08] Mark Zuckerberg: If we end up misspending a couple hundred billion dollars, that would be unfortunate, but the risk of not acting aggressively enough is considerably higher. [18:51] Azeem Azhar: I think we may end up in a world where there are grindingly low and embarrassing returns for credit funds and for hyperscalers because of the investments they've made. That degree of pain is not a bubble bursting. [46:20] Azeem Azhar: It's just not the case that you get the best model from OpenAI and say fix my company and it fixes it. [31:49] Azeem Azhar: Apple really cares about detail and they care about control. And large language models are powerful in many ways, but they’re not about control.

[0:00] Right now, AI is the driving force [0:02] behind the gains in the US stock market, [0:04] behind US economic growth. It is sucking [0:06] up so [music] many construction [0:08] resources, it now appears to be crowding [0:09] out housing, manufacturing facilities, [0:13] like basically everything else you might [0:14] need to build. And the question, even [0:16] just from a financial perspective, is is [0:18] this all worth it? Is all of this [0:21] investment going to pay off or is AI a [0:23] bubble and one of these companies is [0:25] going to miss an earnings target or get [0:27] downgraded and the whole thing is going [0:29] to pop taking all of our retirement [0:31] portfolios with it. So I [music] wanted [0:33] to have two different perspectives on [0:35] this. Dr. Paul Kadoski is a partner at [0:37] the venture capital firm SK Ventures. [0:39] [music] He's a fellow at MIT's [0:40] initiative on the digital economy. It's [0:43] popular newsletter at paulcadoski.com [0:44] and he's been very vocal that he thinks [0:46] we are in an AI bubble. Azimar is a [0:50] British technology entrepreneur, an [0:51] author and an analyst. He is a creator [0:53] of the exponential view which is both a [0:55] great substack that I read often and a [0:58] research platform where among other [0:59] things they've been creating these [1:00] dashboards tracking [music] [1:02] whether or not there's financial stress [1:03] in the eye industry and his view is [1:05] somewhat more optimistic I think it is [1:07] fair to say than Paul's. So I want to [1:09] put the two of them into conversation [1:11] together to try to see why you might [1:13] think [music] and why you might not [1:14] think we are in an AI bubble. [1:23] Paul Kadoski Azimar, welcome to the [1:25] show. Hey, good to be here. It's great [1:27] to be here. Thanks. [1:28] >> Paul, make the case to me this is a [1:31] bubble. I mean, that's a big it's a big [1:33] claim, right? You're saying that all [1:36] these investors, all this money pouring [1:38] in, all these sophisticated analysts at [1:41] Black Rockck and Goldman and wherever [1:43] and you know the the hyperscalers like [1:45] Google and Meta, [1:46] >> but they're all getting this wrong on [1:48] some level and you're not. So why do you [1:50] think it's a bubble and they're all [1:51] idiots? [1:52] >> Well, I wouldn't go quite as far as [1:53] [laughter] they're they're all idiots. [1:55] That may be true though. I'll just I'll [1:57] step back on that one. So what's unusual [1:59] about this? Let's just talk about really [2:01] quickly what a bubble is. the bubbles. [2:02] Whenever asset prices detach from any [2:05] semblance of economic reality, we have [2:07] one of the largest capex competitive [2:11] capex spending waves in history. So in [2:13] western economic history, there have [2:14] been like eight or nine of these moments [2:16] where we build out uh some giant general [2:19] purpose technology, a GPT, not not the C [2:22] the chat GPT kind, but a general purpose [2:24] technology. And so, and we're building [2:26] it out like we were the railroads, like [2:28] we were electrification, like we were [2:30] the canals, like we were the interstate [2:32] highways. And every one of those moments [2:35] has led to overbuild. There isn't a [2:38] major economic capex episode in Western [2:41] economic history that didn't end in a [2:44] rolling series of crises, vastly [2:46] inflated asset prices, and vast overupp [2:49] of capacity of whatever the thing was, [2:51] right? Railroads, fiber, take your pick. [2:54] And so that's the bubble. The bubble is [2:56] once again we're vastly overupplying [2:59] something because of the availability of [3:01] capital and the eagerness of the [3:03] suppliers of the capital to continue [3:05] building out in anticipation of future [3:07] demand. Now none of this is to say that [3:09] like AI is frivolous or it doesn't work [3:12] or that there aren't incredibly useful [3:14] cases for it. That's not the point [3:16] anymore than that was the point for [3:17] railroads, right? Railroads are [3:19] incredibly use didn't still ended up [3:20] where something like 50% of the track [3:22] miles ever laid turned out to be [3:23] useless. Right? So the bubble is this [3:26] vast over supply of something largely [3:28] fueled now by credit which in turn will [3:30] lead to massive business failures and [3:32] write down of assets and all of that [3:35] credit has spread through the economy. [3:37] It's inside of insurers, inside of [3:39] private credit, it's inside of sovereign [3:40] wealth funds and we're on the order of [3:42] you know trillions of dollars now and [3:44] it's probably going to hit 10 trillion [3:46] before it's done. So this all of this [3:48] investment is at risk and it's across [3:51] the entire economy and sitting inside of [3:53] all kinds of I mean it's even sitting [3:54] inside of you know your individual you [3:56] know your 401ks right I mean we've all [3:57] seen the data that something like the [3:59] top 10 stocks in the S&P 500 are like [4:02] 40% of the market capitalization of the [4:04] index. So if you're holding an S&P 500 [4:06] index my friend you are very very long [4:09] AI right portfolio Paul like yeah get [4:12] that crap out of my portfolio. So you're [4:14] very very long AI and this is the [4:15] problem is the the expression you hear [4:18] all the time it's a one factor world [4:20] there's everything is leveraged to AI [4:22] across every asset class across all of [4:24] retail investing and then the last stat [4:26] I'll give you and the reason why that [4:28] matters is we're at record percentages [4:31] in terms of the percentage of people's [4:32] net worth that are tied to equity 50% of [4:36] people's net worth is now in equities. [4:38] So follow that math through. 50% of net [4:40] worth is in equities and that's driven [4:43] by the AI trade. People are even more [4:46] overlevered to this single trade than [4:48] they ever were to dos. And that creates [4:51] this incredible fragility. So all of [4:53] this this this sort of nexus of of [4:56] factors is what creates this bubbish [4:58] moment which has nothing to do with [4:59] whether or not AI is useful but has [5:01] everything to do with why the other side [5:03] of this will see you know a rolling [5:05] series of crises that could potentially [5:06] become really catastrophic. I I do I [5:08] want to focus for a minute on the [5:10] rationality of the core players, your [5:13] metas, your alphabets, your open AIs and [5:16] anthropics, [5:18] >> all of whom believe that there's only [5:21] going to be a couple AI winners. I mean, [5:24] Zuckerberg gave this amazing quote where [5:25] he said, [5:26] >> "If we end up misspending a couple [5:28] hundred billion dollars, that would be [5:30] unfortunate, [5:31] >> but the risk of not acting aggressively [5:33] enough is considerably higher." And so [5:37] there is this dynamic where I think they [5:38] believe this is going to be a highly [5:39] concentrated or oligopolistic market [5:42] where maybe it's one model that rules [5:44] them all or three models that rule them [5:45] all. And so there is this race to be one [5:48] of the three. But that creates a kind of [5:52] pressure to invest way beyond what the [5:56] market can collectively sustain. [5:58] >> Right? That's the classic competitive [6:00] dynamic that's existed in every one of [6:01] these capex buildouts. It existed with [6:03] the railroads. you could not afford to [6:06] be left behind. You had to invest in it. [6:08] And that's why we ended up with a [6:09] duplicate roots between cities. I mean, [6:11] that's what that's the essence of the [6:13] competitive dynamic crossed with a kind [6:15] of, you know, prisoners dilemma that I I [6:18] would like to slow down, but I can't [6:19] because no one else will. And so, we [6:21] have to we all continue on with this [6:22] frantic competitive dynamic. And so, [6:25] that's the what you're seeing incarnate [6:26] there. And we see that even whenever I [6:28] don't know a couple of weeks ago Dario [6:29] Emmed at Enthropic was suggesting that [6:31] we need to and this is a ridiculous [6:33] expression but we need to pace the [6:34] frontier which sounds vaguely like we're [6:36] training for the Boston Marathon but [6:38] it's this idea that we need to slow [6:40] down. Well of course everyone in a in a [6:42] classic prisoners dilemma thing said go [6:44] [snorts] for it you slow down I'm not [6:46] slowing down and even worse they said [6:48] even if we all agree here in the US [6:50] frontier companies to slow down China [6:52] won't. And then you get into this [6:53] terrible discussion, well the what about [6:55] with respect to what China will do. And [6:57] so this is the problem is there's a a [6:59] race dynamic that's been created that if [7:02] you stop there's a fear that either my [7:04] competitors will gain an advantage or my [7:06] shareholders will punish me. And so as a [7:08] result the race dynamic continues and [7:10] that's what guarantees the overbuild. As [7:12] long as you have a supply of capital and [7:14] a race dynamic you're guaranteed over [7:16] capacity much like whether it's canals, [7:18] railroads, electrification, take your [7:20] pick. Azim, you you and your team over [7:22] the past year, year and a half have been [7:25] building out this dashboard trying to [7:27] track how much stress there is in the AI [7:30] industry financially, trying to think [7:32] through what it looks like for this to [7:35] be on a healthy financial path and what [7:36] it looks like to be a a bubble. So, why [7:39] don't you walk me through sort of your [7:41] perspective, the the indicators you [7:43] watch most closely and why you're a [7:44] little bit more optimistic right now [7:46] than Paulus. Well, Ezra, when we're [7:49] looking at a technology boom, a [7:53] technology cycle, we've got to keep into [7:55] consideration two different things. The [7:57] first is, are customers finding it [7:59] useful? Are they paying for it? Are they [8:01] increasingly paying for it? In in other [8:04] words, is demand real? And is demand [8:06] there? And when you look back at bubbles [8:09] at burst, it is always almost always [8:11] connected to a problem in the financing [8:14] mechanism. The refinancing of a shorter [8:17] term debt. It runs into a problem and [8:20] capital goes away and you see that all [8:22] of the things that we we associate with [8:24] with bubbles. The question is where are [8:26] we today? And where we are today is that [8:28] demand is very strong from American and [8:30] global enterprises. There is not an [8:34] overbuild yet. There is vast vast under [8:37] supply of the compute needed for AI and [8:40] the race seems to be can we get enough [8:43] chips and power enough chips over the [8:45] next two or three years to keep up with [8:46] the demand and on the financing side we [8:49] are certainly moving off the cash that [8:51] the hyperscalers have earned over the [8:53] last 20 years or so and we're bringing [8:55] debt on board but it's not yet uh at the [8:58] levels at which you would say this is a [9:00] historically high and complex uh [9:02] position. So for me today it's you know [9:05] it's in that productive boom with the [9:07] hints of maybe getting for forgive my [9:10] French a little bit stinky. [9:12] >> So uh I I want to focus on this supply [9:14] versus demand right at this very simple [9:17] level. What Paul's basically saying is [9:18] that there is a strong incentive to [9:20] build too much supply and what you are [9:23] saying is that there's actually plenty [9:24] of demand. So I want to decompose what [9:27] we mean when we talk about demand. So [9:29] this is demand for AI usage which is [9:32] tokens which again for people who are [9:34] not as used to this are you know when [9:36] you ask a question the um AI model is [9:41] consuming tokens to answer it. When you [9:43] ask the AI coder to code you a program [9:46] it is consuming tokens to to code the [9:49] program. And so functionally what is [9:50] happening with a company like OpenAI or [9:52] Anthropic is people are buying token use [9:56] from them right they don't sell you [9:58] clawed they sell you tokens at the high [10:00] levels so that's the reason tokens are [10:01] an important concept here and so there's [10:04] first this question of how many tokens [10:08] are people going to want and how strong [10:09] is the demand for tokens going to be and [10:11] then there is a secondary question of [10:13] how much will they pay for them because [10:15] the price per token is plummeting. [10:18] So there's this question of even if [10:20] there is a lot of demand, is it going to [10:22] be demand for the kinds of very [10:25] expensive models that the American [10:26] companies are investing in or or not? So [10:28] I'd like you to walk me through how that [10:31] demand picture looks to you. [10:33] >> Well, we have to start a layer below the [10:36] the tokens. The asset we're investing in [10:39] is data centers, 65% of the cost of [10:42] which is the chip, the GPU or the AI [10:45] accelerator. So what what I think is [10:47] important to to look at is what is the [10:50] demand for those chips to do the [10:53] computing that in some cases most cases [10:56] produces tokens and and who are the [10:58] customers of those GPUs. Uh and when we [11:02] when we look at this we see these [11:04] hyperscalers Amazon and Microsoft and [11:06] and Google and then new labs called [11:09] Neoclouds. These people buy billions of [11:12] these GPUs. They put them together in uh [11:14] data centers that now will start to cost [11:16] billions of dollars and they rent those [11:19] GPUs out. And the thing that that that I [11:21] spend a lot of time thinking about is is [11:23] that capital expenditure on the actual [11:26] productive asset turning into revenue [11:31] with just enough of return on investment [11:33] to to pay itself. And right now, the [11:35] reason we know we're in a computed [11:37] environment is because people are making [11:39] extremely good returns in 18 to 24 [11:41] months on on an asset because they're [11:44] able to charge essentially Uber Surge [11:47] pricing for their for their GPUs. And [11:50] about 50% of their revenues go to the [11:53] two big AI companies, Anthropic and and [11:56] OpenAI, for their spending. The other [11:58] 50% is a a mix of other AI companies and [12:03] firms companies directly running those [12:06] open source models that you alluded to [12:07] or openweight models on the hyperscalers [12:10] infrastructure and and then I think we [12:12] get to your your other question which is [12:13] well if 50% of this is coming from [12:16] anthropic and open AI what of that is [12:18] actually real customer spending and and [12:20] what I'd say is if this if those ratios [12:23] stayed the same for the next three or [12:25] four years we'd very much be in a world [12:27] where you know this is someone [12:29] effectively paying themselves around the [12:30] round around the back door. But what has [12:32] been happening is that that ratio has [12:33] been declining because essentially the [12:36] hyperscalers are getting more and more [12:37] business from real customers. And all [12:40] all of that for me is a sign of a market [12:42] that is maturing week after week, [12:44] quarter after quarter. [12:45] >> Paul, what do you how do you see that [12:47] differently? [12:49] >> I hardly know where to start. Um, so let [12:52] me just start like a lot of the the the [12:54] discussion and I use that word advisedly [12:56] with respect to how people talk about [12:58] this stuff reminds me of that old joke [12:59] about like a mother comes downstairs in [13:02] the morning and finds her daughter under [13:03] the Christmas tree and there's a giant [13:04] pile of horse manure and the daughter's [13:06] digging in the pile of horse manure and [13:07] the mother's like but what are you [13:09] doing? What are you looking Well giant [13:10] pile of horse manure? There must be a [13:11] pony in here somewhere. And it's for me [13:14] a lot of this discussion about giant [13:16] scale with respect to tokens is a lot [13:18] like this digging in a pile of manure [13:20] because there must be a pony in there [13:21] somewhere. Uh because it's this this [13:24] confusion of scale, massive volume of [13:26] tokens with value. So I'll take Azim's [13:29] point with respect to um this prodigious [13:32] use of tokens. Tokens alone don't [13:34] matter, right? It's this pile of manure [13:35] under the Christmas tree. What matters [13:37] is tokens times price. And you you you [13:39] mentioned this earlier that we're seeing [13:41] over the last few years roughly a [13:43] 13-fold annual decrease in the price of [13:46] tokens. Now that's performance adjusted. [13:48] So the answer of the frontier model [13:50] companies will be say hey sure the old [13:53] models are falling like that but not the [13:54] new models. The new models are the ones [13:56] that can break into banks. Those ones [13:58] are still much more expensive. So so [14:00] those are holding on to their value. But [14:01] the problem is of course for most people [14:04] for most of what they do they don't need [14:06] the much more expensive models. And as a [14:08] matter of fact, most normals most normal [14:10] people get by perfectly adequately on [14:12] relatively low performance models. A [14:15] good example is this incredible uptake [14:17] in Meta's Muse product which has [14:19] dominated the app charts and everything [14:21] else which uses a relatively primitive [14:23] model. It's not using a leading edge [14:24] model and nobody cares. It just works [14:27] perfectly fine for what most people want [14:28] to do. Which take lets me sort of take [14:30] it back to the beginning. One of the [14:33] reasons why people really [14:35] mischaracterize this entire market, and [14:37] this is like an old venture capital kind [14:39] of adage, is never confuse your early [14:42] customers for what the real market will [14:44] look like. An awful lot of what's [14:46] happened with frontier models is people [14:48] confusing the early customers of [14:50] frontier models for what you and I want. [14:52] So, who were the early customers of AI [14:54] models, right? Who who really drove [14:56] this? Coders, right? something like 80% [14:59] of the token usage and the revenues from [15:02] the largest a the largest frontier model [15:04] companies are coders are developers and [15:06] developers are really idiosyncratic for [15:08] at least two reasons why one they want [15:11] the leading edge models right they're [15:12] willing to pay up for them because they [15:14] have really complicated problems two [15:17] they are what I like to think of as kind [15:19] of expansive in terms of their token [15:20] usage meaning that I enter a small [15:22] prompt and say you know bro me a fitness [15:24] app and all of a sudden seven words [15:26] turns into a million lines of coat [15:28] that's not anything like what you and I [15:30] do for the most part. A lot of what [15:32] normal people and normal businesses do [15:34] is closer to kind of a onetoone [15:36] relationship or even compressive like [15:38] I'll say I got 25 emails from Goldman [15:41] Sachs babbling about something or [15:42] another are any of those things worth [15:44] reading and that's compressive right [15:45] because I'm taking a whole lot of stuff [15:47] and producing and turning into something [15:48] small and that feels efficient to me or [15:50] similarly some other email or a calendar [15:52] or whatever. So the point being that [15:54] normals, the people that these frontier [15:56] models want to sell to now are not [15:59] really well represented by what people [16:01] in the early days of the market were [16:03] doing because the coders are really [16:05] unusual. And you're already seeing some [16:07] of that with uh Meta cutting in half the [16:10] number of developers in house who are [16:11] allowed to use external frontier tools. [16:13] Microsoft the same thing. You know the [16:16] days of token maxing where you like had [16:18] you know charts internally that said you [16:19] know who's winning at using the most [16:21] tokens. Those are long over. Normal [16:23] companies can't tolerate that. People [16:24] are blowing their entire budgets in like [16:26] months. So my point is is we're already [16:29] moving away from this early days of [16:31] proflegate token use and everybody [16:33] looking for ponies inside of giant piles [16:35] of manure. [16:36] >> Azim. [16:37] >> Yeah. In his giant pile of manure, Paul [16:39] also had some slight of hand. Uh what [16:42] what Paul talked about was the frontier [16:44] models and and tokens. uh and what I [16:47] talked about was the capital expenditure [16:49] in the capacity we've been talking about [16:51] which is the GPUs and I think that they [16:52] are two different things that merit two [16:55] different discussions I mean what [16:56] happens to the frontier model labs open [16:59] AI and anthropic with this business [17:01] model of in ever more powerful AI [17:04] systems that have a smaller and smaller [17:07] sort of set of users is quite different [17:09] to the question of what's going to [17:11] happen with the trillion plus dollars or [17:13] so of capital that will go into building [17:16] data centers upon which American and [17:18] global businesses will start to run [17:21] increasing amounts of AI. And I and I [17:23] think we need to separate those two out. [17:25] One is really about like the fundamental [17:27] viability longer term of of anthropic or [17:30] open AI and what they might do in the [17:32] face of the pressures that that Paul has [17:34] articulated. And I think the other is [17:37] whether the uh the the the GPUs and this [17:40] new digital infrastructure that's being [17:42] built out will find uses amongst [17:45] predominantly enterprises but of course [17:47] uh as we rightly heard uh consumers [17:50] through to through things like um you [17:52] know Muse and other agents. I think [17:54] those are different questions. I'm [17:55] confused by you separating these [17:56] questions and and I recognize that maybe [17:58] this is an important point of confusion [18:00] to to untangle because people can mean [18:03] different things when they talk about [18:04] whether or not there's a bubble and what [18:05] it would mean for the bubble to pop or [18:07] to deflate. [18:09] as far as I can tell, if it turns out [18:11] that OpenAI or Anthropic or for that [18:14] matter, you know, Alphabet or whoever, [18:18] that their business model here doesn't [18:20] work. That in fact, people are perfectly [18:23] happy to use DeepSeek from China for a [18:25] lot less money and Anthropic like [18:27] collapses in on itself. That is going to [18:30] pop the economy. Like that that looks to [18:32] me like it would become a contagion in [18:35] the stock market. I mean, I don't [18:37] necessarily think it'll throw us into a [18:38] depression, but I think that would have [18:40] a very very very big effect on [18:42] valuations on um you know, what people [18:46] are willing to lend in the future. Do [18:48] you not think that? [18:49] >> Well, I think what you have to do is is [18:51] zoom back out and ask what is the the [18:53] fundamental risk that emerges with a uh [18:56] with a with a an infrastructure [18:58] buildout, which is that you don't have [18:59] enough customers who are demanding a [19:01] product and paying enough for that [19:04] underlying asset. And if people can [19:06] move, enterprises can move to uh [19:10] openweight models that are being run [19:11] directly by a hyperscaler that is just [19:14] the demand moving. It may hurt [19:16] Anthropic's margins, but it still [19:18] monetizes the underlying capital uh [19:21] expenditure. And then what one of the [19:22] things [19:23] >> which are these chips basically the data [19:25] center? [19:26] >> Anthropic might might collapse, but [19:28] Nvidia will be fine. [19:29] >> But but even but even so, an anthropic [19:32] has got many many uh degrees of freedom. [19:34] I mean in the last uh 24 hours or 48 [19:36] hours since before we recorded this uh [19:39] it it cut prices and released even [19:41] cheaper more efficient lightweight [19:42] models. These are still models around [19:44] which they have a pretty decent margin. [19:47] They will be doing the things that [19:49] software companies have done for for [19:51] decades which is trying to construct [19:54] reasons for you to use their software [19:56] beyond the core capability. That's why [19:57] they have claude co-work and they have [20:00] claude for lawyers and that's why I [20:01] think this question of if anthropic [20:03] disappeared tomorrow overnight well I [20:05] can't quite imagine the circumstances [20:07] under which that would would happen but [20:09] what I what I'm saying is that you're [20:11] seeing the end customer demand and [20:12] that's the thing without which this will [20:14] always end horribly c can I just ask you [20:16] also to because I want to actually ask [20:18] you both this question of what for you [20:21] counts [20:23] as the bubble popping because there's a [20:27] world here where you know you have say a [20:29] giant stock market correction but the [20:32] data centers are still being used and it [20:35] sounds to me like you're saying that's [20:36] not a bubble popping [20:38] that's just a correction or something. [20:40] >> Well historically uh bubbles when they [20:43] they've popped have resulted in in two [20:45] things. one is a huge decline in the [20:47] level of investment in that particular [20:50] asset and they have been accompanied [20:53] with a significant crash in a relevant [20:56] asset market whether it's the stock [20:58] market or or or something else and that [21:00] is the test that that I would use. I [21:03] think we may end up in a world where [21:05] there are grindingly low and [21:08] embarrassing returns for credit funds [21:11] and for hyperscalers because of the [21:13] investments they've made. that that [21:16] degree of pain is not a bubble bursting. [21:19] That is just somebody made some bad [21:21] investments and they bought high uh when [21:23] they should have bought at a lower at a [21:24] lower price. I mean I think I would use [21:26] a very very classic historical [21:28] definition which is that we need to see [21:29] the level of investment drop by 50% and [21:31] we need to see a really really [21:33] substantial and typically it's been 50% [21:35] or more uh correction in whatever the [21:37] relevant market is. [21:39] >> Paul same question to you. [21:43] >> I Okay. So I I I sort disagree on almost [21:47] every point that Azim is making. These [21:49] are arbitrary definitions and I think [21:51] let's just let's put it this way. Okay, [21:53] so back to my original point. Something [21:54] like 50% of people's current net worth [21:56] in the United States is tied into [21:58] equities. Most of equities is currently [21:59] tech. Most of tech is currently AI. [22:01] That's going to only get worse as these [22:04] two monster IPOs come to market over the [22:06] next, I don't know, two months or [22:08] whatever. [22:08] >> Open AI and anthropic. [22:09] >> Open eye and anthropic. And just to put [22:11] that in context, by the way, Open Eye, [22:13] Anthropic, and SpaceX, if you want to [22:15] include it, combined, those three IPOs [22:18] are larger than all post World War II [22:20] IPOs combined. Like, this is a really [22:22] unusual moment. [22:23] >> That's a crazy That's a crazy stat. [22:24] >> It's a crazy stat. And not only that, [22:27] the combined value of those companies is [22:30] larger than the value of the NASDAQ [22:32] market capitalization at its peak. Like, [22:34] the bubble is sitting right in front of [22:36] people and they want to say, "Well, I [22:38] expect you to come with blonde hair, not [22:39] brown hair." Dude, just because it [22:41] doesn't look the way you want it to look [22:42] doesn't make it any less fragile. So you [22:45] could imagine, let's say, just for the [22:46] sake of argument, the anthropic IPO gets [22:48] pushed out. It doesn't happen in [22:50] November. It happens next year. There [22:52] will be a sharp revaluation in their [22:55] ability to raise funds as well as in the [22:58] companies that are supplying them [22:59] because everything is levered off the [23:02] valuation of these companies that are [23:03] currently sitting in private markets. So [23:05] I I like to say this is the failure here [23:07] is overdetermined. There's so many ways [23:09] it can all break and break really [23:11] catastrophically because I'll give you [23:14] another example. The the what's called [23:16] the c the Schiller uh cycllically [23:18] adjusted price earnings multiple which [23:20] is a sort of a longer term measure of [23:22] where US equities sit. We're sitting at [23:24] the levels we sat just before the crash [23:26] of 29. We're sitting at absolutely [23:29] historically unprecedented levels in [23:31] terms of marketwide valuation. And this [23:33] is the problem is people lack [23:35] imagination and simply refuse to engage [23:37] with the broader data about what's [23:39] actually happening in the world because [23:40] they say, "Well, this doesn't look like [23:41] the last bubble, so we must be fine." [23:43] >> Well, let me but I think I'm going to [23:44] I'm just going to come in directly as I [23:46] mean I I think that it's very easy to [23:48] use um very colorful language. I'm [23:51] always mindful of [23:52] >> Well, this is data. It's not colorful [23:54] language. It's actually numbers. numbers [23:56] like what's what's your response to like [23:58] we're sitting at historically [23:59] unprecedented numbers in terms of like [24:01] the cyclical uh price adjusted earnings [24:03] is good is that good in your book [24:05] >> one of one of the uh one of the [24:07] responses about the Cape Schiller is [24:09] that Cape Schiller doesn't say anything [24:11] about timing [24:12] >> uh because it's because it's a it's an [24:14] adjusted long period horizon historical [24:17] it's a historical 10-year look back that [24:20] can sometimes understate underlying [24:23] changes in the regime of the the the [24:25] economy [24:26] and the market. But there's there's [24:27] something else that's going on, which is [24:29] I think a lot of what you're describing, [24:30] Paul, is already priced by equity [24:33] investors and it's being priced by uh [24:35] the credit markets. We've seen it just [24:37] in the last few weeks. We've seen uh an [24:40] IPO for a data center company Australia [24:42] get pulled because investors weren't [24:45] happy with what they were being [24:46] promised. [24:47] >> That's not why it was pulled. [24:49] >> Well, why was it pulled? It's pulled [24:50] because of this unprecedented moment [24:53] where two trillion dollar IPOs are in [24:56] orbit around us waiting to come to [24:58] market and are sucking all the risk [25:00] capital to the market as people say, [25:02] "Yeah, this looks really cute, this IPO [25:04] we have, but there's we don't have any [25:05] capacity to do that right now." [25:06] >> Okay. But I think that in a way actually [25:08] goes to Azim's point a bit. When you [25:10] talk about these massive IPOs that are [25:12] circling and we're talking about OpenAI [25:14] and Anthropic and Open AI has gotten [25:16] pushed back a little bit for when they [25:17] seem like they're going to launch [25:18] actually into the into the public [25:20] markets. The thing that is interesting [25:22] about that from I think the perspective [25:25] of it's not all a bubble is that [25:28] sophisticated investors in mass numbers [25:31] who really really do have skin in the [25:33] game here are looking at these and [25:35] saying they make sense. So when you hear [25:38] this from other people, Paul, and you're [25:40] an investor and you think about these [25:41] things, what mistake do you think all [25:44] these people are making? When you say [25:45] it's just math, it's clearly not just [25:47] math because everybody else can do math. [25:49] It is your interpretation of the math. [25:50] So where does your interpretation of the [25:52] math differ from, you know, the market [25:54] >> because Sure. So you have to think about [25:56] why why these kinds of IPOs happen and [25:58] why people are piling into them. It's [26:00] partly in anticipation of what they [26:02] think could be, you know, a fantastic [26:04] business with a long-term franchise, da [26:05] da da da da. But it's also because they [26:08] think that when it comes time when the [26:09] lockup expires, they can sell the stock [26:11] to someone else who thinks that, right? [26:12] This is this is the essence of going [26:14] back to the very beginning of why do [26:16] IPOs exist? IPOs used to exist to allow [26:19] you to raise money for a company who [26:21] needed to grow. These are not that, [26:23] right? These companies have done their [26:25] growing. They're probably actually going [26:26] to be, I don't know, hundred billion [26:29] dollars in run rate revenues by the time [26:30] they come out or 70 billion or something [26:32] like this. So, these are exit events. [26:34] These are people inside the company [26:36] saying, "I don't need my stock anymore. [26:38] You can have it." So, they're selling [26:40] their stock to people on the outside [26:42] because they at this valuation, they'd [26:44] rather have cash than anthropic stock or [26:46] cash than OpenAI stock. [26:48] >> At that valuation, I'd like to have [26:49] cash, too. [26:50] >> Yeah, sure. Absolutely. Hand it over. [26:52] But the point being that like this is [26:53] not a this is not a a gesture where [26:56] they're saying you know I really like [26:58] the our prospects for the company. [26:59] There's all there's going to be selling [27:01] shareholders across the board because [27:02] this is an exit event not a [27:04] >> final but it is a gesture where other [27:05] people are saying I do like your [27:07] prospects as a company and I want to own [27:10] this. [27:11] >> I the same thing was said in [27:12] anticipation of the SpaceX IPO and we've [27:14] seen what happened subsequent to that [27:15] that all of this forever before it [27:17] turned into a you know kind of a a sort [27:20] of a light roller coaster decline. So my [27:22] expectation is you'll see the exact same [27:24] thing because these are in market terms [27:26] uh exit events. You're going to be a [27:28] wall of selling shareholders showing up [27:31] post lockup selling it to anyone who [27:33] will show up and want to and want to [27:34] take the shares from them. And that's [27:36] why this needs to be understood for what [27:38] it is an exit event. People are buying [27:40] in anticipation of selling to you. If [27:42] you want it's like the old line about [27:44] the poker table. If you don't know who [27:45] the greater fool at the table is, it's [27:46] you. This is that. And I don't [27:48] understand why people don't see it that [27:49] way because that's what makes this so [27:51] unprecedented. [27:52] >> I I'll just I'll just come in on that. I [27:54] mean, I think what Paul describes is [27:55] really what many IPOs have been like for [27:58] for many many years. And when we take a [28:00] look at anthropic and whether it can be [28:02] a sustainable business, we probably need [28:04] to look at what it will show in its S1 [28:08] when it's the S1 is the document. It's a [28:10] document essentially for what this [28:12] company is when you're going to buy it [28:14] at at IPO and look at what's happening [28:16] with its revenue trajectory and its [28:18] margin trajectory and the breadth of its [28:20] customer base and whether customers are [28:23] sticking with it, whether those accounts [28:24] are growing beyond what has been frankly [28:27] a lot of patchy reporting and leaks, [28:30] some of which is accurate and some of [28:32] which is is inaccurate, but not much of [28:34] which gives us a cleaner picture of of [28:37] the company. And I think that draws one [28:39] of the really big issues we've all had [28:40] to deal with over the last 2 or 3 years, [28:43] which is these companies have got very [28:45] big, but they are still private [28:46] companies. So they don't really need to [28:48] disclose much, if anything at all. And [28:51] what comes out is a smattering of rumor, [28:54] scuttlebutt, and the occasional [28:56] wellsourced leak. And because they have [28:59] essentially been structural to the AI [29:02] trade so far and they have been as a [29:04] consequence the thing that has spun the [29:07] valuations uh of the of the wider market [29:10] whatever it is now 38% uh tech but we [29:14] have not been able to inspect the way we [29:17] normally can with public companies [29:19] what's really really going on. And I [29:20] think what happens with the IPO is that [29:23] we we won't get every answer, but we get [29:26] some of the answers that we need about [29:27] these these businesses. You have a [29:29] personal answer now though, just on that [29:31] point though, which I think is amazing [29:32] because it came from, [29:34] >> you know, what's that thing with the [29:35] horror movies with the killers inside [29:36] the house. This is a little bit of one [29:37] of those where Anthropic accidentally [29:39] leaked who their that who one of their [29:42] largest customers was. They said we [29:43] needed to have what was it called? A [29:44] mega account executive or something like [29:46] this. when they were trying to hire one [29:48] and then they quickly realized that by [29:50] saying who their largest customer was [29:51] you could triangulate and say this was [29:53] for Meta so this was their account [29:55] executive to handle sales to a single [29:56] company [29:57] >> that it worked out that uh Meta [30:00] something like 10% of their revenues and [30:02] so this goes back to [30:03] >> Meta is 10% of anthropics revenue that's [30:05] right but that was in 2025 right so [30:08] >> and let's just be clear let's be clear [30:10] but revenues no but why it matters is [30:13] that anthropics revenue in the first [30:14] quarter of this year was bigger than the [30:16] entire ity of last calendar year. So [30:18] what we're doing is we're looking at a [30:20] rear view mirror and saying we hit a [30:21] squirrel 700 yard ago ago. What we need [30:25] to look is look out of the [laughter] [30:26] the windscreen. This is why these [30:28] companies are so unusual. We have never [30:31] seen companies on this size go public [30:33] with such a small number of customers. [30:35] The concentration of their customer [30:37] list, it might be eight companies that [30:39] make up 80% of their revenues. That [30:42] doesn't happen. And so as a I was an [30:44] equity analyst years ago. This would be [30:45] a colossal source of risk for me. I'm [30:47] looking at it and I'm saying the [30:48] dependency on such a small number of [30:50] customers who are mostly using it for [30:52] the same thing is historically [30:53] unprecedented to have such customer [30:56] concentration. So what they accidentally [30:58] leaked was the fragility of the business [31:00] in terms of its exposure not just to a [31:02] single customer but to a single customer [31:04] who was doing that same old thing of [31:06] writing code. And this is this this is [31:08] the point I want to make. That's what [31:09] makes these so unusual. The beauty of [31:11] what what Paul has just said which is [31:13] sort of eight customers 80% of the [31:14] business is that that's 10% per customer [31:16] that would have to be disclosed in the [31:17] S1. So we have actually a very very [31:19] falsifiable prediction that we will be [31:21] able to check when the S1 comes out if [31:23] anthropic does successfully go public [31:25] and say was it really 80% and eight [31:27] companies representing 80% of the [31:29] business and that's really helpful [31:30] because if Paul's right that is a bit [31:33] that's a bit exotic [music] right it's [31:34] it's a bit spicy. [31:36] If you're not a subscriber to the New [31:38] York Times, we have [music] some news [31:40] for you. You can now explore the Times [31:42] for free without any payw walls at all [31:43] during [music] your first month in the [31:45] New York Times app. [31:49] I think Apple is a really interesting [31:51] company here because unlike Alphabet, [31:54] Google, unlike Meta, obviously Open and [31:56] Anthropic are AI companies, Microsoft [31:58] was a major OpenAI investor and is now [32:00] also trying to make its own frontier AI. [32:03] Oracle is heavily, you know, invested [32:05] here and is like now just like right [32:07] above junk bond status. [32:10] Apple just decided to not compete in [32:13] making a frontier AI. While their [32:15] competitors are now spending, you know, [32:17] 100 to20 billion in capex every year. [32:21] Apple spending what 15ish billion in [32:23] capex per year, their stock this year [32:25] has performed better than any of these [32:27] other companies stocks. [32:30] Is there does it imply anything that [32:33] Apple and being the one major player [32:35] here to not try to compete in the [32:37] frontier model Olympics seems to be [32:40] winning. [32:42] >> Let me let me start with here. [32:44] >> Yeah. Yeah, sure. Yeah. I mean, Apple [32:46] has a business model that means they [32:48] don't need to run a frontier model. We [32:50] buy their phones, we buy their laptops, [32:52] and then we might subscribe to the odd [32:54] movie or or something like that. So, [32:56] they make money regardless of of the AI. [32:59] But I also think there's a really [33:00] important cultural dynamic which is uh [33:04] Apple cares a lot about detail. I [33:06] remember this story of uh Steve Jobs uh [33:09] when uh he was rolling out the iPhone [33:11] many many years ago phones the then [33:14] senior vice president at Google who's [33:16] responsible for the partnership and [33:18] blasts him out because they'd used the [33:20] wrong color yellow uh on the Google uh [33:23] then Google logo. Apple really cares [33:25] about detail and they care about [33:27] control. And large language models are [33:30] powerful in many ways, but they're not [33:31] about control. They can get very very [33:33] hapless. And and I I think that one [33:35] cultural reason why Apple would have [33:38] held back even though it's so well [33:40] positioned to, you know, be the personal [33:42] agent. It's got the device. It's got the [33:44] privacy. is that it it won't it won't [33:47] want the mess uh that a large language [33:50] model will ultimately deliver if you've [33:52] got a billion people using it, which is [33:54] what Apple's estate looks like. [33:57] >> So yeah, I I think I don't think Apple [33:59] was very smart about this. I think they [34:00] mostly screwed it up and got lucky. Um [34:03] so they tried various times to make [34:05] foray into AI. I mean they they they had [34:08] they had Teams, they disbanded Teams, [34:09] they were going to integrate it with [34:11] Siri, they weren't going to integrate it [34:12] with Siri, and they didn't ever get [34:13] anywhere. And what they benefited from [34:15] though was by their inaction, they were [34:17] a classic free rider, right? They got to [34:19] free ride on everyone else's gigantic [34:22] expenditures and they come out the other [34:24] side smelling like roses because they [34:26] didn't actually have to make the [34:27] expenditures. They can use what's being [34:28] built by others and sit on top of it. [34:30] But I think writ large we're going to [34:32] see a lot of that because people are [34:34] burning money on a bonfire which can [34:36] benefit lots of other companies, you [34:37] know. So Apple sitting on top of it, not [34:39] having to build models. Uh look at [34:41] people like Cloudflare for example. I [34:43] mean here's a company that largely [34:45] forgotten was kind of basic internet [34:47] infrastructure stock has tripled in the [34:49] last what 18 months because suddenly [34:51] people like wait a minute wait a minute [34:52] wait a minute if we have all these bots [34:54] ranging around the internet that's kind [34:56] of like a new form of attack that we [34:58] need to make sure that we have who who's [34:59] out there that can deal with all of this [35:01] new traffic and so a lot of this [35:03] investment that we've seen out there [35:04] will turn out to be beneficiary to a [35:06] whole host of companies whose names are [35:08] not OpenAI Anthropic Grock or whoever [35:11] whomever else and I think that's a [35:12] really important point and that goes [35:14] back to the you know the days of of [35:16] electrification in the 1920s like who [35:18] were the primary beneficiaries once the [35:19] utilities all began collapsing in the [35:21] crash of 29. Well of course it was the [35:23] people building things that required [35:24] electricity air conditioner [35:26] manufacturers and others and so yeah for [35:27] sure we'll see a lot of this you know [35:29] accidental piggybacking and free riding [35:31] on top of this. It's it's but it's not [35:33] accidental piggybacking and free riding [35:35] as it's called complimentary services [35:38] that and complimentary businesses that's [35:40] what economists when they're you know [35:42] being very mod moderate with their [35:44] language you describe in the process [35:46] which is a general purpose technology [35:47] shows up it continuously improves and [35:50] then it starts to create the need for [35:53] complimentary services whether it's [35:54] maintenance or whether it is uh you know [35:57] repair or or whether it's something you [35:58] can now do with the technology uh this [36:01] is the the normal evolution of a general [36:04] purpose technology as it finds its way [36:05] in and in some cases it's a bit of free [36:07] riding. In other cases, what's happened [36:09] with with Cloudflare is that we've been [36:11] able to across the economy automate so [36:15] many bots and AI agents that businesses [36:18] on the internet are turning to [36:19] Cloudflare to try to protect them if [36:21] they if they can. Let let me ask [36:23] something actually about these big [36:25] companies. It has come up a lot in the [36:27] bubble debates I have followed which is [36:31] the the major companies we are talking [36:33] about here your alphabets your metas [36:35] etc. they just have a lot of money and [36:38] so they're doing this from a very [36:40] healthy balance sheet which was not the [36:42] case sometimes for these railroad [36:43] companies etc. When when Zuckerberg [36:45] says, you know, it would be sad if we [36:48] misspent hundreds of billions of [36:50] dollars, what he's also saying is we can [36:52] misspend hundreds of billions of [36:53] dollars. It'll be fine. And and one [36:56] argument I've seen is that one reason [36:58] you don't really need to worry about [36:59] this is that it is being driven by [37:03] companies that can absorb losses on [37:06] their own balance sheet without [37:07] contagion in the rest of the economy. [37:09] you're not dealing with something like, [37:11] you know, derivatives in the 2008 [37:13] financial crisis. Um, Paul, you've [37:16] thought a lot about the private credit [37:17] dimensions. I mean, there has been a [37:19] movement towards debt in some of the the [37:20] way this is being financed now. How do [37:23] you think about the question of are we [37:26] protected from this? We being broader [37:29] society because these companies have [37:31] ungodly amounts of money to lose if they [37:33] so choose. [37:35] >> No, we're not. Um, so this is a funny [37:38] thing. So, like a year ago, people told [37:40] me to stop babbling about this stuff [37:42] because these are smart companies. They [37:44] have predigious cash flows. Who are you [37:46] to tell them what to do? And I was like, [37:47] "Oh, shucks. You've got me there. So, [37:49] I'll stop saying this." And then, of [37:51] course, we're now a year later. What [37:52] does the data show? Spending is moving [37:55] away from cash flow to dad. So, I, you [37:57] know, you put up your hand and you say, [37:58] "Hey, wait a minute, wait a minute. [37:59] Remember that thing you said last year? [38:00] It's no longer true. So, should I worry [38:02] now?" And it's like, "No, you don't need [38:03] to worry now because they have pristine [38:05] balance sheets. These are great [38:07] credits." And I'm like, "Oh crap, I [38:09] might just my timing is wrong twice on [38:11] this." And then, of course, to be [38:12] serious, what happens whenever you have [38:14] a pristine balance sheet? What Wall [38:15] Street is very very good at is making it [38:18] less pristine because they'll they'll [38:20] continue to raise money for you and [38:21] raise and raise and raise for you [38:23] because the fees are tremendous. And [38:24] what and as as they do so, your credit [38:27] rating declines. And we're seeing that [38:28] with Oracle. We're seeing that we're on [38:30] the cusp of that with Microsoft. We're [38:32] on the cusp of that with some of the [38:33] other hyperscalers because that's what [38:35] Wall Street does. They see a pristine [38:37] balance sheet, an unlevered company in [38:39] the term of of art of the industry and [38:41] they'll continue to burden it with [38:42] credits. And so the way that these [38:44] companies try to get around it, Meta is [38:46] probably the best example is they use [38:48] somewhat more exotic structures. So the [38:50] debt doesn't actually sit on their [38:51] balance sheet. They use things called [38:52] special purpose vehicles. They use [38:54] things that are off balance sheet [38:55] financing. Say, "Hey, hey, that's not my [38:58] debt. I know I control it, but it's over [39:00] there with that funny named thing that [39:01] sounds like a sort of a New Orleans [39:03] bagel. um that's somebody else's debt. [39:05] And so this is all just a shell game. [39:07] And the reality is it rolled up. We've [39:10] now at a point where tech whose main [39:13] advantage historically was fast growth, [39:15] no debt, right? Tech is now the largest [39:18] issuer of investment grade bonds in the [39:19] United States. This is an industry whose [39:21] main claim to fame other than growth was [39:23] that they were relatively clean. They're [39:25] now the exact opposite. These are hugely [39:27] debt-enccumbered companies increasingly [39:29] and that are dominating the issuance [39:32] marketplace to the point that it's even [39:33] having an impact on sovereigns. Right? [39:35] So one of the reasons why we had the the [39:38] bond market freakout over the last 2 3 [39:40] weeks was because of the role of [39:43] hyperscalers issuing credit at the same [39:46] duration as the 10-year in the United [39:49] States which at the margin was actually [39:51] competing the US which caused yields to [39:55] have to go up on the 10ear. That's how [39:57] crazy things have become. So, yeah, no, [39:59] it's completely transitioned, which is [40:00] going to have, you know, huge [40:02] consequences down the road in terms of [40:04] this debt overhang that starts to look [40:06] increasingly like, you know, Japan circa [40:08] 1990. This kind of all these [40:10] corporations with giant balance sheets [40:12] loaded with debt that are trying to work [40:13] it off. [40:14] >> Well, the other side of that, Azim, is [40:16] that it's not just debt that is being [40:18] crowded out where these, you know, hyper [40:20] scalar companies are competing with [40:23] everybody else for for money. It's also [40:26] construction. It's also real materials. [40:28] I mean, I know I'm sure you've seen [40:29] these charts going around [40:31] >> that imply there's real crowd out [40:33] happening that yes, we are building. We [40:35] are pumping money into AI capex, [40:38] >> but home building in a country that does [40:40] not have enough homes is going down. [40:42] Spending on manufacturing plants and [40:44] factories is going down or at least like [40:46] the the sort of share. And so there [40:48] there does appear to be a movement from [40:50] investing in other parts of the economy [40:53] at a construction and capex level into [40:55] investing in AI. I guess if AI works out [40:58] great, that's fine. But if it doesn't or [41:01] even if you just think these other [41:02] things are, [41:04] >> you know, underrated as say maybe I [41:06] think for homes, homes are a more stable [41:09] >> investment. They don't return give you [41:10] returns like the like investing in an [41:12] Nvidia does, but they're pretty [41:14] important. And I have like a cleaner way [41:17] of explaining why more homes would make [41:18] more people's lives better. [41:20] >> How do you think about the the crowd out [41:21] question here? [41:22] >> Um, when we look at uh investment rates, [41:25] a year ago, back in the summer of 2025, [41:27] I wrote about how I assessed whether [41:30] we're in a boom or a bubble. And one of [41:31] my observations was that at the moment [41:33] the hyperscalers are spending out of [41:35] their own cash flows and their own [41:36] balance sheets. As soon as that starts [41:38] to change and they go to external [41:39] finance, you introduce lots more risk. [41:42] and my threshold for more risk was about [41:45] 30% uh of the uh uh uh 30% which is [41:49] about the level that we we are [41:51] approaching now at the end of 2026. [41:53] These things are are not set in stone. [41:56] The crowding out problem is a is a real [41:59] problem and it's it's it's really uh a [42:02] complicated and thorny issue. I mean, I [42:04] think one place to to think about was um [42:06] what happened in in the Netherlands in [42:09] the 1970s when they discovered a lot of [42:11] uh natural gas off the North Sea and the [42:14] entire economy oriented around gas [42:17] extraction and you got this thing called [42:19] Dutch disease where effectively that [42:22] part of the economy was sucking up all [42:24] of the activities and there you know [42:26] there could be a risk that that will [42:27] happen uh in the US if it isn't already [42:29] happening right now and and I think then [42:31] the question is what do you do about it [42:33] And how do you actually uh uh you know [42:37] adapt to that? And I'm afraid my only [42:38] answer is that it requires something of [42:41] policy something of a policy stimulus to [42:44] make sure that those sectors get the [42:46] activity that they need because I'm not [42:48] persuaded that market forces on their [42:50] own uh would would do that. [42:52] >> That seems like a tough uh prescription [42:54] at a time when I'd say both we have very [42:56] erratic leadership but our borrowing [42:58] costs are going pretty high amidst [43:01] incredible levels of debt. [43:03] Well, look, your your politics is your [43:05] politics and you know, you get to vote [43:07] every four years and and make the [43:08] choices that that that you make. Um the [43:11] >> and then live with them, [laughter] [43:12] >> right? And then live with them. But but [43:14] it doesn't have to involve uh subsidies, [43:16] right? It can involve other types of uh [43:19] you know policy interventions. So, I'm [43:20] not it's it's a really difficult problem [43:22] and it will continue to happen so long [43:25] as the returns on the other investments [43:27] are are higher and so long that that's [43:30] that that is creating a uh a a supply [43:33] shock on electricians and construction [43:35] workers uh and HVAC technicians and all [43:38] the other things that are needed uh in [43:40] data centers. But, but and there is [43:42] there is a a but to all of this which is [43:45] that of course there's a lot of [43:47] upskilling going on. There are many many [43:49] electricians being trained. There are [43:50] many HVAC engineers being trained. There [43:52] are many construction workers being [43:53] trained. There are construction firms [43:55] who are growing and who will start to [43:58] have spare capacity. And I don't know [44:00] how that ends up, but in a year or two [44:02] or 3 years, that capacity, those skills [44:06] will be in the economy [music] [44:08] and able to be used. [44:14] >> [music] [44:20] >> So I want to ask this not in terms of [44:23] stock markets and financial flows. I [44:26] want to try to ask what does this [44:27] economy look like in which [44:31] companies and people and governments for [44:33] that matter are using enough AI to make [44:37] these investments writ large pay off [44:40] even if you know some companies may not [44:42] end up being in the winner circle here [44:43] and may may fail. So you spend a lot of [44:46] time thinking about this. If you were to [44:49] try to paint me a picture of what things [44:52] look like at normal companies, your GES, [44:55] your Coca-Cas, [44:57] um, in this world where the AI bet pays [45:00] off, [45:01] what does it look like? What is [45:03] different? What are we not what is not [45:05] being spent on if so much money is being [45:07] spent on AI tokens? Like what is [45:08] happening here? [45:11] You know, when I talk to to bosses and I [45:13] talk to them across Europe and the and [45:14] the US, dozens every every year, the the [45:17] the I I hear this the same sort of [45:20] thing, which is um we're making progress [45:22] with AI. It's harder than we thought. [45:24] We're going to keep doing it and we're [45:26] starting to get results. and the more [45:27] confident ones say we're getting more [45:29] and more results. But a large part of it [45:32] is coming now I think will will come [45:35] unless businesses grow very very quickly [45:37] from just a slow reduction in in [45:40] replacement of of workers that companies [45:44] that are getting more productive with AI [45:46] will be delivering more dollars per [45:48] employee and if they're not growing [45:50] faster than that productivity growth [45:52] then their their headcounts will will [45:53] start to to to fall. But we should just [45:56] add that you know there are very few AI [45:59] native companies there are very when we [46:01] think about electricity the electricity [46:03] native companies were people like uh the [46:05] Ford Motor Company with its endto-end [46:08] production line but most of America [46:10] didn't get fully electrified for another [46:12] 15 or 20 years after that and I don't [46:15] think I can really point to a main [46:18] street firm that is not outside of [46:20] Silicon Valley that you could say this [46:23] company has become AI native what you [46:25] can see are bits of the business that [46:28] look like they're using a lot of AI. [46:31] >> Okay. But if if the AI native company [46:32] that you're talking about is a company [46:34] where you've managed to replace people [46:36] with AIS. [46:37] >> Yeah. [46:38] >> So now you seem to be really quickly [46:40] tipping into the much worried about [46:42] world in which AI is net negative for [46:45] jobs across the economy unless you [46:48] believe there are all these new [46:49] companies starting that are somehow very [46:51] human labor intensive. Even though in [46:53] theory the new companies starting are [46:54] going to be more AI native so less human [46:57] labor intensive. [46:59] So is that actually a good world? And [47:01] then if that's the world, who's spending [47:03] more money in this world where people [47:05] are losing jobs? [47:06] >> Well, I I think there are there are a [47:07] number of different things going on. You [47:09] know, one is that uh there are go there [47:11] are still things that the AIS won't be [47:13] able to do and those will be layers of [47:16] human humans doing verification, doing [47:18] the last sign off, doing the things that [47:20] the models cannot do well. And what [47:23] would typically happen is that those [47:24] people will end up being paid more and [47:25] more uh you know better and better as a [47:27] consequence of of sitting at that choke [47:30] point. I think that the second thing [47:32] that uh you will start to to see is a [47:37] whole class of roles that we can't [47:40] imagine within within companies you know [47:42] whether it's things like the web master [47:44] which didn't exist in you know 1993 and [47:46] and and then I think the third thing is [47:48] this is really the the cultural [47:49] dimension of a general purpose [47:51] technology. This is where, you know, [47:52] Paul was describing the GPT is that they [47:55] tend to construct entire new classes of [47:58] of work and behavior that we hadn't been [48:01] able to put our fingers on. And you [48:02] know, if you just think about Henry [48:04] Ford, well, because of Henry Ford, we [48:06] ended up with motorcars. Because we got [48:08] motorcars, we ended up with people [48:09] having suburbs and suburbs to vacations [48:12] and vacations to all the things that the [48:14] services that sit around around there. [48:17] Now, I'm not blas about that. I mean I [48:18] think the the challenge we may find is [48:22] that it takes longer to create those new [48:24] roles than anything else that that that [48:26] happens. But the the other observation I [48:28] would make is that you know my my time [48:30] horizon for this runs in 5 7 10 years [48:35] not one or two not two or three because [48:38] I do think it just takes time for [48:41] management teams to absorb what it is [48:43] they have to do with this technology. [48:44] And it's it's just not the case that you [48:46] get the best model from OpenAI and say [48:50] fix my company and it fixes it. I mean I [48:52] just these these things take time and [48:53] that's what I see when I talk to to [48:55] bosses. However fast they want to move [48:58] it just takes time. Paul, how do you see [49:00] this? [49:02] >> Uh we're basically repricing labor, [49:04] right? I mean this is the fastest [49:06] adopting technology in in history and [49:09] its fundamental function is to repric [49:11] cognitive labor. And I mean this is the [49:14] deep struggle because the article of [49:15] faith the west coast ideology on this [49:18] stuff is technology always creates more [49:20] jobs than it destroys. And so it becomes [49:23] just a thing that people repeat like a [49:25] mantra that they that they hope will [49:26] become true. And the reality is we've [49:28] never seen a general purpose technology [49:30] that had such broad applicability was [49:32] adopted so quickly and was repricing [49:34] labor at the same time because and when [49:36] I say repricing labor I mean it's making [49:38] it vastly cheaper to do things like call [49:40] centers for example. And as a result, [49:43] these things, this change is going to be [49:45] destabilizing long before we figure out [49:47] how to turn them all into I don't know [49:48] what the equivalent of web masters is. [49:50] And so this is the problem, right? Is [49:52] that it's happening so much faster than [49:55] any prior GPT that these sorts of, you [49:57] know, armwaving about creating more jobs [49:59] than it destroys belies the lived [50:02] experience of people who are affected by [50:04] this stuff. And so it be can quickly [50:06] become just some kind of strange mantra [50:09] that it's like you know don't worry [50:10] you're pretty little head everything [50:11] will be fine. So no I think the [50:13] disruption will vastly outpace our [50:15] ability to create new positions. This [50:17] has also struck me as another form of [50:18] economic risk for not just workers but [50:21] but actually AI and for you know some [50:23] form of this bubble popping or or [50:25] deflating if it if it is a bubble [50:27] >> because you can imagine a world where [50:28] actually the problem is not whether or [50:31] not companies are willing to adopt AI [50:33] [snorts] that actually particularly [50:35] under more economic strain interest [50:36] rates keep going higher maybe there's a [50:38] recession of some sort and companies [50:40] begin moving very aggressively to [50:42] replace labor human labor which you know [50:45] needs healthcare and unionizes and has [50:48] opinions with AI and then there's a huge [50:52] backlash. There's already a huge AI [50:54] backlash already people organizing [50:56] against building data centers. They [50:57] don't want those where they live. But [51:00] you know if you know anything about the [51:02] history of occupational licensing or [51:03] building homes or anything else it's [51:06] actually not that difficult for say [51:09] teachers to organize and say no AI in [51:12] the classroom. Actually education is off [51:14] limits for AI. doctors all of a sudden [51:16] to organize maybe there's a couple of [51:18] bad events where people are getting bad [51:19] diagnoses off of AI and so it's sort of [51:21] like locked out of hospitals. Anthropic [51:23] in their risks to their IPO says [51:26] existential risk to humanity is possibly [51:28] one of the risks of this IPO. If there's [51:31] some major control incidents where [51:32] instead of hacking hugging face, OpenAI [51:36] hacks Wells Fargo, [51:37] >> it hacks something in China. [51:39] >> Yeah. [51:39] >> And all of a sudden regulation comes in, [51:42] not a little bit, but all at once. [51:45] >> All of those things seem like real risk. [51:48] the the the possibility of discontinuous [51:52] regulation [51:53] >> seems higher to me than elsewhere and [51:55] also seems I don't want to say [51:57] inevitable to me but but quite likely to [51:59] me. So I'm curious how you think about [52:01] that. Uh this has been my complaint from [52:03] the very beginning of the process by [52:06] which these companies have launched [52:08] models that they've essentially been [52:10] conducting unlicensed drug trials on the [52:12] public in the sense that these are [52:13] hugely consequential technologies that [52:15] are launched in a way that were they a [52:17] drug company that you would immediately [52:19] have the FDA on your back and yet [52:20] because of their technology companies [52:22] you can do these mass trials. Oh by the [52:24] way we're going to transform all of [52:26] education. We're going to launch these [52:27] things that you know can penetrate um [52:29] government websites and other things. [52:31] This idea that you can do this with [52:33] impunity in technology because we live [52:35] in the timeline where so far these [52:37] things haven't killed us is just [52:39] ridiculous and indefensible. And I think [52:42] we're inevitably headed towards a world [52:43] of where there's going to be after some [52:45] kind of terrible event. There's going to [52:47] be some kind of post hawk um regulation [52:51] imposed, but it'll only be after [52:52] consequential damage. And I'll give you [52:53] a specific this is the way I often try [52:55] to characterize it to people because [52:56] they'll they'll say well what's going to [52:58] cause that is if like you know you can [52:59] just come up with this stuff like if you [53:01] think about it there's [snorts] probably [53:03] 5 10 15 different sorts of scenarios we [53:06] can come up with that might lead to [53:07] regulation none of them probably having [53:10] a probability of like let's say less [53:11] maybe more than 5%. So let's say I have [53:14] six events each of which have a 5% [53:16] chance of happening what's the [53:18] likelihood that any one of them happens [53:20] it's like 20%. If I have a dozen events [53:23] like possible scenarios, right, that are [53:25] all bad, that each of which have no more [53:27] than a 5% chance of happening, what's [53:29] the likelihood that one of those [53:30] happens? It's more than 50%. We have way [53:34] more than a dozen ways that these things [53:36] could cause trouble and end up requiring [53:37] regulation. So looking at it that way, [53:40] it's almost inevitable. So it's like a [53:42] 50 60% chance that something happens [53:44] that drives regulation. And yet people [53:46] continue to sit around and say, "Well, [53:49] you know, this isn't that likely or that [53:51] isn't that likely." So, no regulation [53:52] here is inevitable. It'll take the form [53:54] of something like the FDA, maybe some [53:56] kind of atomic energy board because it's [53:59] just as consequential and because there [54:00] are so many different ways for them to [54:02] have broader consequences in society. [54:04] And it's amazing to me that people sort [54:06] of globally pass that on. [54:07] >> Azim, how do you see this political risk [54:09] dimension? Well, you know, I think when [54:11] you look at industries, uh, industries [54:14] grow quite well when there's a little [54:17] bit of regulation as the pharmaceutical [54:18] industry is, uh, you know, is a great [54:20] great example. And I think when we think [54:23] about who's going to pay for this, the [54:26] AI companies from the model labs through [54:29] to the hyperscalers, uh, are going to be [54:31] making a lot of money. So it's it's [54:33] fairly clear that there is money in the [54:35] system and there are people who firms [54:37] who can cover some part of of the bill [54:40] and I think one of the things that's [54:41] really exciting about the data center [54:44] protests that you're seeing particularly [54:45] in the US is that actually it is people [54:49] expressing their sense of unease about [54:52] this technology the way it's manifesting [54:55] itself and and the way to look at that [54:57] is that that is a de democratic [54:59] expression right that is an expression [55:00] of will that is expression of [55:01] uncertainty [55:02] And what you would hope to come out of [55:04] something like that would be a better [55:08] social license which people are are more [55:10] and more comfortable with. Now you know [55:13] I I suspect and I'm that that within the [55:15] US because you have state level uh you [55:17] know governance power that some of this [55:20] may be done on a statewide basis. what's [55:22] happening in the UK is that that there [55:24] is a an attempt to get people developing [55:27] data centers to reinvest annuities into [55:30] the local community because one of the [55:32] things that we you know that we don't [55:34] really talk about much is that a huge [55:35] data center might employ very very few [55:37] people indeed. I mean, it may be [55:39] billions of dollars, but it's a handful [55:41] of jobs. And that's not crazy. Like two [55:43] to six people for a multi-billion dollar [55:45] facility. [55:45] >> Well, and if you do the math in terms of [55:47] job uh costs per jobs created, it's [55:49] often I was doing looking at one in New [55:50] Mexico recently, it's $4 million in [55:53] terms of subsidies per job created, [55:54] leaving aside the cost of the center, if [55:56] you look at the tax abatements and the [55:58] required improvements to, you know, [56:00] local utilities. The numbers are insane. [56:02] >> So, so, so none of that none of that [56:04] works long term. uh but the the useful [56:06] thing is that there's money in the [56:07] system because there are profits being [56:08] made up and down the stack and it you [56:11] know one would feel that because of that [56:13] there's an obvious place for that bill [56:15] to land and that bill can land through [56:18] uh you know I don't think [56:19] self-regulation works and I would agree [56:21] that we will need to find whether it's [56:23] local national or international real [56:26] rules about about how these systems work [56:28] and the harms and the damages that they [56:29] might cause. Well, I the but the thing [56:32] to me that's interesting in that of [56:33] course there's going to be regulation. [56:34] That that's clear to me [56:37] >> the range of how that could fall all the [56:40] way from pharmaceutical regulation where [56:42] what we want is for pharmaceutical [56:43] companies to produce safe drugs although [56:46] people maybe argue you know we made that [56:47] too slow also [56:49] >> all the way to say nuclear power [56:51] regulation where fundamentally what [56:54] we've ended up wanting is no new nuclear [56:56] power plants [56:57] and depending on what happens with AI [57:00] and given how negative people are on it [57:02] already one one of the possibilities [57:03] here is that it could swing such that [57:06] the answer on a bunch of things is [57:07] simply no. We don't want to be replaced [57:10] by robots, which actually when you say [57:12] it aloud like that seems like a fairly [57:13] straightforward thing [snorts] and [57:15] probably the kind of thing that people [57:16] are going to express in a democracy. Or [57:19] no, we don't want robot agents crawling [57:22] all over the economy. No, we don't want [57:24] to be called endlessly as a restaurant [57:26] by your muse agent [57:28] >> trying to get reservations everywhere [57:30] because now you've just made it [57:31] functionally impossible for a restaurant [57:33] to ever answer the phone again with a [57:35] human being. And that the the sort of [57:37] like collective expansion of all this. I [57:39] mean, I know people who, you know, are [57:41] are managers at different firms and [57:42] they're they're just saying, I'm just [57:45] drowning [57:46] in AI produced reports all of a sudden. [57:49] my email is just full of obviously AI [57:53] written inquiries which is making my [57:55] responsiveness all across the board go [57:58] down. I'm not just saying will there [58:00] will there not be regulation but I think [58:02] one risk is that people decide maybe [58:05] they don't want this in actually fairly [58:06] large swasts of the economy or used in [58:09] some of the ways people are expecting. [58:11] There's a very good piece by John Herman [58:12] in New York Magazine about the just sort [58:15] of beginning agony of companies that all [58:19] of a sudden are absorbing tremendous [58:22] amounts of muse agent inquiry. And like [58:25] we're not prepared for everybody to have [58:27] an agent calling you as a robot trying [58:29] to negotiate down the price or get a [58:31] reservation or sell a concert ticket [58:33] >> or just whatever it might be. And people [58:36] may people may actually not like it. [58:38] >> Yeah. [58:38] >> They they might not like it. Uh the [58:40] trouble is there appears to be some [58:42] revealed preference uh here and that [58:44] picture is is not even. So only about 2 [58:46] and a half% of Americans pay for an AI [58:49] subscription which is considering it's [58:51] they've been available for 3 years is a [58:52] very slow slow take. [58:54] >> You had this I love sorry I love this [58:56] stat in one of your newsletters recently [58:57] that more people pay for GLP1s a larger [59:00] number of Americans paid for a GLP1 than [59:03] for AI. Well, yeah, because one is [59:06] really quite useful and you feel the [59:08] impact within a matter of a few a few [59:10] weeks, but hundreds of millions of [59:12] people in America and around the world [59:13] do use AI through the the free version. [59:16] And when companies like Meta start to [59:18] prod put these tools in front of you, [59:21] I'm sure people will end up using them. [59:24] Even if they vote with their uh [59:26] complaints and with their their social [59:27] media posts saying they don't want any [59:29] AI, either it'll become inescapable or [59:31] they will still have it on their phone [59:32] and they'll be using it. I think we have [59:34] to just contend with that that [59:37] possibility. I think we we also have to [59:39] that you know realize and this is where [59:41] we think about technology as being this [59:43] gigantic unpermissioned experiment which [59:46] it it has been for 40 or 50 years that [59:49] what you're describing about robo calls [59:52] from AI agents into hotels and [59:54] restaurants or as my my uh uh orthopedic [59:58] surgeon was telling me he just gets long [60:01] thousandword 10,000word emails from [60:04] patients who've clearly chat GPT their [60:06] medical notes This is unfortunately [60:08] going to happen and that cost has now [60:10] been pushed into the system and we're [60:13] going to have to figure out how to deal [60:14] with it. I now pay, it's not a lot, I [60:17] pay about a buck a month uh to my uh [60:20] phone operator to screen out spoof phone [60:24] calls that are essentially roocalls and [60:27] that that's what's going to happen and [60:28] we're going to call that GDP [60:30] those additional costs. I think the one [60:33] of the things that's your your comment [60:35] Ezra really reminded me of this is that [60:37] societ we we have this idea in sort of [60:39] western capitalism that it's always [60:40] about efficiency in a weird way [60:42] societies require friction [60:44] >> and when you take away frictions even [60:46] even frictions that you didn't realize [60:48] existed all of a sudden it has [60:50] consequences you know there's been [60:51] examples of people using large language [60:53] models u to launch claims in small [60:56] claims court for example do we want that [60:57] to be more to have less friction do we [60:59] want more of that do we this, you know, [61:02] your example of this of the of the [61:04] restaurants and we have all of these [61:06] frictions that exist in society and [61:08] blindly removing all of them is kind of [61:11] creating in in tech parliament almost a [61:13] denial of service attack on modernity [61:15] where all of a sudden we're having stuff [61:17] launching at us from every angle almost [61:20] like you know people you know Russian [61:22] bots attacking a website except it's [61:24] coming from all angles because of the [61:25] increasing you know um numbers of bots [61:28] and other things and I think that's what [61:29] people are responding to is it's this [61:31] kind of outofcontrol feeling as all of [61:33] the frictions drop away and suddenly you [61:36] know I can't get access to things or [61:38] nothing works the way I'd expect. So [61:39] simply dealing with the reduction in [61:41] frictions has huge consequences across [61:43] society and I think people are reacting [61:44] to that [61:45] >> and what's going to happen is that it's [61:46] going to be these two AIs bouncing [61:47] things back and forth and [61:49] >> right [61:49] >> and it's just going to be a more and [61:52] more and more unpleasant economy which [61:54] is why like I do think there going to be [61:55] large areas for people to say maybe not. [61:58] So I want to end on on this question [62:00] which is first for for you Azim what [62:02] would you need to see to start thinking [62:06] oh no this is getting pretty bubbly and [62:09] then and then to you Paul the reverse [62:11] what would you need to see to think oh [62:13] maybe I do have this wrong and I [62:14] shouldn't be these investors have it [62:16] right and there's a lot of money to be [62:18] made here [62:19] >> well there there are so many risks that [62:22] we have talked about and there's all the [62:23] complexity and fragility in in the [62:25] financing the the thing that keeps uh [62:29] this on the other side of healthy is [62:31] whether real customers are spending real [62:33] money and getting real results. So the [62:36] thing that would make me uh uneasy would [62:38] be if I saw a couple of quarters where [62:41] the amount that American enterprises [62:44] were spending on AI was uh the growth [62:47] rate was h havinging say two quarters in [62:49] a row. that would suggest that they [62:51] they've filled their appetite and [62:53] there's not going to be the revenue to [62:55] to keep this on the right side of of [62:57] healthy Paul. [62:59] >> Uh there's so many I mean I I'm [63:02] convinced I'm wrong about most things so [63:04] I'm always happy to find more things I'm [63:05] wrong about. So So I I just look at [63:07] economic data. So if we've began to see [63:10] I don't know productivity growth go well [63:13] beyond trend that would be striking, [63:15] right? I mean so far we're sitting at [63:16] around trend data with respect to US [63:18] productivity growth. there's no sign of, [63:19] you know, AI improving productivity. If [63:21] we saw participation rates sharply [63:23] increasing, that would be that would be [63:25] something that would help to tell me, [63:26] okay, AI is an inclusive technology and [63:28] bringing people back into the economy, [63:30] that would really help. If we saw GDP [63:32] growth going back above trend, we're [63:34] still sitting around the levels of, you [63:35] know, sort of the postWorld War II trend [63:37] at the low end of it. Any of those [63:39] things would tell me that something big [63:41] was going on, like a true productivity [63:43] shock in the economy that warranted the [63:45] kind of expenditures. So, you know, if [63:48] any of those and probably the three of [63:49] them at once happen, that would tell me [63:50] I was wrong. You know, if those [63:52] productivity numbers come through, then [63:54] fine, everything changes. But if it [63:56] doesn't, you're looking at something [63:57] that looks more like the 1930s than it [63:59] does like, I don't know, 2000s. [64:01] >> All right, I think that's a good place [64:02] to end. So, as our final question, what [64:04] are three books you'd recommend to the [64:05] audience? Uh, Paul, why don't we begin [64:07] with you? [64:08] >> Sure. Um, Middle March is sort of my [64:11] go-to on thinking about normal people [64:15] being swept along by kind of the winds [64:17] of change as a major technology [64:19] revolution's taking place. It was, you [64:20] know, it took place just as the [64:22] industrial revolution was taking was was [64:24] sort of taking hold. And I think it's a [64:26] tremendous uh a tremendous book about [64:28] that. So the other one that I like to [64:29] recommend is Liberty's Dawn by Emma [64:31] Griffin which is um a collection of the [64:34] average workers biographies and and [64:36] notes from during the industrial [64:38] revolution and it's a just a wonderful [64:41] wonderful look at what it was like in [64:43] the mid in the middle of that wrenching [64:46] transition and as you know people are [64:48] talk about economic data and everything [64:49] else what it was like to live through it [64:50] what did it feel like what was the pain [64:52] I felt and you know what were my hopes [64:55] for the future it's just just like a [64:57] wonderful book about that and the last [64:58] one um that I I was quoted as Samuel [65:02] Inel largely forgotten figure. He was [65:04] the probably one of the most important [65:06] entrepreneurs of the 1920s whose empire [65:09] and electricity shows what happens that [65:11] you can build building an empire on top [65:13] of a GPT and even doing that [65:16] successfully can still end up in in in [65:18] collapse and calamity but it's a it's a [65:20] wonderful story. So insult the rise and [65:22] fall of a billionaire utility tycoon by [65:24] uh Forest McDonald Azim. [65:28] >> Uh so my my three books uh the first is [65:30] uh the day the universe changed. It's [65:32] written by a BBC science journalist back [65:34] in the early 80s. I read it as a as a [65:36] teenager and I've just reread it and it [65:38] is all about how uh as we develop new [65:41] types of knowledge we start to rethink [65:44] our role in the world, what matters to [65:46] us and how unsettling that is. and he [65:48] walks us through uh a thousand years [65:50] plus uh of that through through western [65:53] and Asian history. Uh the the second [65:56] book is um uh fathers and sons by Tony [66:00] which is a book written in the 19th [66:02] century. It's at that moment where uh [66:05] the ideas of science and modernity are [66:07] making their way into Russia and you [66:09] have the generational conflict uh [66:12] between the more uh traditional [66:13] generation and the the younger [66:15] generation and and what uh stays as a as [66:18] a bond between them. What even the uh [66:21] the most recent breakthroughs in science [66:23] can't tell you about what it is to be [66:26] human. I think very uh relevant. And [66:28] then the the final one which I'm I'm [66:30] just rereading again is uh Condied by [66:33] Voltater [66:35] Paulski. Thank you very much. Thanks. [66:37] >> Thank you. [66:45] [music]