AI CapEx Boom Faces Debt Collapse, Skeptic Zitron Warns

质疑者Zitron警告：AI资本支出繁荣或面临债务崩溃 超大规模云厂商的数据中心支出依赖循环融资以及未盈利AI初创企业由债务推动的需求，正在造成比互联网泡沫更严重的系统性市场风险。 科技作家Ed Zitron针对当前的人工智能繁荣提出了严峻且详尽的看空论点，认为其财务根基建立在循环融资、虚高的token年化运行率（run-rate）以及不可持续的单位经济效益之上。讨论将超大规模云厂商极度乐观的预期及高达1万亿美元的基础设施建设，与停滞不前的企业投资回报率（ROI）进行了对比，警告庞大的算力规模正在远远超出实际的终端用户收入。主持人和嘉宾概述了主要系统性风险，包括私人信贷债务敞口、潜在的新型云服务商违约以及知名初创企业的倒闭，并将其与1873年大恐慌等历史市场泡沫进行了对比。 OpenAI 和 Anthropic 消耗了当前约 90% 的 AI 基础设施算力，催生了脱离实际的虚假需求信号，迫使超大规模云服务商投入数千亿美元展开数据中心与 CapEx 军备竞赛。 (6:08) AI 产业陷入循环融资怪圈：以 OpenAI 规模达 1220 亿美元的融资轮为例，其中仅 120 亿美元来自传统 VC，其余 1100 亿美元均由 Nvidia、Amazon 及软银等利益相关供应商以股权与债务形式回输，以此维持虚高的算力收入。 (12:46) 私人 AI 初创公司普遍滥用“ARR 运行率”指标，通过将短期内偶发的 Token 使用量峰值年化来粉饰业绩，掩盖了缺乏真正长期且可预测的 SaaS 订阅收入的现实。 (20:07) 企业界宣传的 AI 效率大幅提升多由高管的 FOMO（错失恐惧症）驱动，实际节省成本往往仅局限于极少数孤立业务，难以支撑大模型开发商到 2030 年高达数千亿美元的算力支出需求。 (30:50) AI 基础设施繁荣潜藏巨大系统性违约风险，其潜在崩塌诱因包括新型云服务商遭遇信贷冻结、超大规模企业发债遇阻、核心 AI 初创公司资不抵债，或是 AI 公司上市遇挫导致软银等巨头流动性陷入困境。 (34:17)

## Background 本期节目由 YouTube 频道 The Compound 于 2026 年 8 月 28 日发布（播客《The Compound and Friends》第 257 期）。主持人 Josh Brown 与 Michael Batnick 特邀知名科技分析师 Ed Zitron 做客，围绕生成式人工智能领域的“终极看空逻辑”，系统拆解英伟达的爆发式增长、OpenAI 与 Anthropic 的商业模式缺陷、超大规模云厂商的资本开支泡沫，以及潜藏在私人信贷与数据中心基建背后的系统性债务危机。 ## TL;DR 节目全面呈现了 Ed Zitron 对当前 AI 产业繁荣本质的尖锐批判，指出以英伟达为核心的千亿美元硬件繁荣严重脱离真实终端需求，全行业近 90% 的算力消耗高度依赖于持续巨额亏损的 OpenAI 和 Anthropic 的资本补贴。所谓强劲的“年化运行率”实质上是截取短期代币消耗峰值的统计障眼法，而超大规模云厂商在缺乏商业闭环的情况下正通过战略投资、回购算力与供应商融资进行“左手倒右手”的资本循环。更为严峻的是，CoreWeave 等新兴云服务商背负着超过 9% 的高息借贷，超大规模数据中心项目深度绑定了缺乏透明度的私人信贷、保险年金与养老金体系。Zitron 警告，与互联网泡沫时期通用光纤基建不同，AI 专用 GPU 无法转为他用且折旧运维极其昂贵，一旦再融资与信贷链条断裂，将引发远超 Dot-com 破裂的系统性金融动荡。 ## Atmosphere & style 整期对话充满激烈的思维对撞与刀刃相见的论辩张力。Ed Zitron 以近乎战斗状态的语速与极其详实的数据猛烈抨击行业乱象，言辞辛辣刻薄却逻辑严密，带着一种目睹灾难即将来临的急迫与愤慨；两位主持人则在多头实战体验与审慎怀疑之间反复平衡，从最初带有调侃色彩的防守反驳，逐渐过渡到面对庞大债务黑洞时的屏息与凝重。整场录制宛如一场在科技狂欢宴席旁悄然召开的金融急诊室会诊，交织着敏锐的智识博弈与令人不寒而栗的现实危机感。 ## Bulletin - [05:16](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=316s) 超大规模云厂商的增长高度依赖极少数初创公司，瑞银等机构预测微软、亚马逊与谷歌约 4400 亿美元的云端增长预期，完全依赖于背负 1.1 万亿美元长期算力承诺却严重亏损的 OpenAI 和 Anthropic。 - [11:22](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=682s) 英伟达营收集中度极高且出现回款放缓迹象，前五大客户贡献了高达 70% 的账目收入，而其应收账款环比骤增 55%，并存在反向租回自售芯片算力的循环关联交易。 - [20:07](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=1207s) AI 数据中心建设与实际商业需求严重脱节，目前规划中的 190 吉瓦（GW）数据中心容量需要每年 1.5 万亿至 3 万亿美元的算力需求来维系，而全市场真实需求估算仅在 220 亿美元左右。 - [01:17:08](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4628s) 私人信贷与保险资金的大规模介入放大了系统性金融风险，约六分之一的保险年金已流入私人信贷，若数据中心因下游初创公司违约而烂尾，其破坏力将远超 2000 年互联网泡沫。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=0s) 科技巨头季报虽然大肆宣扬 AI 投入的回报，但实际上普遍回避披露 AI 业务的真实盈利与真金白银的净利润数据。 - [02:12](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=132s) OpenAI 预计 2030 年营收达 2840 亿美元的财务指引脱离现实，届时仅算力开销就需要超过 2000 亿美元，运营支出规模将超过微软全公司水平。 - [13:41](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=821s) 微软 2026 财年约 343.3 亿美元的 AI 收入中，高达 241 亿美元本质上是 OpenAI 的返购资金，外部商业软件销售实际仅贡献了个位数数十亿美元。 - [16:22](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=982s) 山姆·奥特曼到处签署天价算力承诺，包括向甲骨文承诺超 3000 亿美元及向亚马逊云支付 1380 亿美元，但自身缺乏足够的现金流覆盖这些账单。 - [20:07](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=1207s) AI 初创公司大肆宣传的“年化运行率”（Run Rate）是数字文字游戏，将短期代币（Token）尝鲜消耗强行推算为全年订阅，掩盖了缺乏长期稳定 SaaS 续费合同的真相。 - [22:32](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=1352s) 私人信贷在数据中心投资中尽调极度草率，如蓝猫头鹰资本（Blue Owl）耗资数十亿美元的项目决策被曝仅花费 10 到 15 分钟，完全陷入赌徒心理。 - [27:41](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=1661s) 企业级 AI 的真实落地存在巨大争议，部分自动化客服（如 Airbnb）虽有提效，但整体缺乏足够且普适的商业回报来支撑万亿美元的重资产基建。 - [31:46](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=1906s) 大语言模型频繁换代造成下游开发者严重内耗，且复杂任务下单次推理消耗的 Token 总量激增，抵消了单价名义下降的优势并大幅推高企业账单。 - [34:17](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=2057s) 超大规模云厂商与 AI 初创公司形成了闭环资金游戏：巨头对初创公司注资，初创公司再以算力采购形式将资金返还给投资方，人为做大云业务增速。 - [47:24](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=2844s) 硅谷巨头陷入“腐败经济”（The Rot Economy）与盲目追逐永续增长指标的怪圈，管理层将 AI 视作无需深思、可肆意填充财报故事的魔法黑盒。 - [01:00:25](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3625s) 新兴云算力转租商（Neocloud）如 CoreWeave 面临严重的持续性亏损并承担超过 9.2% 的高昂借贷成本，可能成为引爆算力链条债务危机的“第一骑士”。 - [01:06:30](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3990s) AI 基础设施狂潮在重现 1873 年美国铁路大崩盘与 GE 式激进供应商融资的剧本，过度超前建造的专用算力无法被其他行业有效替代。 - [01:10:11](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4211s) 英伟达在 10-Q 中披露向投资级客户提供长达 3 个月至 1 年的延期付款条款，暴露了硬件供应链依赖信贷垫资维持繁荣的“虚假需求”本质。 - [01:14:01](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4441s) 专用 AI GPU 离开深度学习训练和推理后几乎没有其他转用价值，加之电力与变压器成本高昂，其烂尾代价远超 Dot-com 泡沫时期的通用服务器与光纤网络。 - [01:20:33](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4833s) 英伟达不仅背负 300 亿美元的云算力反向回租协议和 250 亿美元数据中心租赁协议，还伴随应收账款激增，显露出通过关联交易维持账面繁荣的风险。 - [01:22:19](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4939s) 散户投资者在股市中的杠杆率处于历史最高位，一旦 AI 泡沫破裂导致债务连环爆雷，缺乏风险抵御能力的普通大众将承受最严重的财富冲击。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=0s) 从公关到科技批判写作：Ed Zitron 的自学之路与敏锐观察 Ed Zitron 回顾了自己的职业转型经历。在去年之前，他一直经营着一家公关（PR）公司，从事合同制工作。2020 年，他开始独立撰写文章，最初主要关注管理理论、远程办公的好处以及当时抨击远程办公的恶劣舆论。随后他转向加密货币（Crypto）与马斯克（Elon Musk）的相关话题，但长期追踪马斯克的一举一动让他感到身心俱疲、倍感压抑。 转折点发生在 OpenAI 的管理层地震中。当山姆·奥特曼（Sam Altman）被 OpenAI 董事会解雇时，社交媒体（Twitter）上的记者们甚至情绪崩溃，表现得像在崇拜摇滚明星一样为奥特曼流泪、表达狂热的支持。这种离奇的行业崇拜引发了 Ed Zitron 的强烈警觉。随后所有人都在宣扬人工智能是“历史上最伟大、最庞大、不可阻挡的浪潮”，但他深入探究后发现，整个行业几乎找不到任何实在的 AI 营收数据。无论是狂砸数百亿美元资本支出（CapEx）的微软（Microsoft），还是发布季报的谷歌（Google），财报新闻稿每季都在吹嘘“AI 押注获得丰厚回报”，却从未披露 AI 业务究竟赚了多少真金白银。为了拆解这套叙事，他通过阅读大量专业书籍自学经济学与财务知识，开始在其独立平台（基于 Ghost 搭建）上发表深度的科技与财务批判。 ```mermaid flowchart TD accTitle: Ed Zitron 写作关注焦点的演变 accDescr: 展示 Ed Zitron 从 2020 年关注管理与加密货币，到目睹奥特曼事件后转向拆解 AI 财务神话的路径。 subgraph Phase1 ["前期写作探索 (2020-2023)"] A["管理理论与远程办公"] --> B["加密货币与马斯克话题"] B --> C["沉重压抑，寻找新方向"] end subgraph Phase2 ["转折与拆解 (2023 至今)"] C --> D["目睹 Sam Altman 开除风波中媒体的盲目崇拜"] D --> E["深入追查科技巨头实际 AI 财务数据"] E --> F["发现资本支出激增与营收披露隐瞒的断层"] end ``` --- ### [02:12](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=132s) 专业投资人与狂热散户的分歧：AI 巨头不可思议的增长幻象 主持人 Josh Brown 指出，Ed Zitron 的文章在机构投资人圈子中流传甚广。专业投资人即便重仓做多，也懂得兼听则明，愿意了解反方逻辑以管理风险；而散户则常常对质疑者发起人身攻击。Ed Zitron 表示，很多激进的 AI 散户嘲笑他是“公关仔”（PR boy），认为他不配谈论财务，但很多半导体领域的专业多头反而在推特上与他保持友好交流。 Ed Zitron 强调，他不反对技术本身的潜力，也认可那些承认当前存在泡沫的清醒多头，但他无法接受完全脱离现实的狂热预期： - **OpenAI 的夸张财务指引**：OpenAI 预计到 2030 年其营收将达到 2840 亿美元（甚至有传闻称向 3800 亿美元冲击）。若要实现这一目标，OpenAI 必须在短短三年半内体量超越 Meta。 - **失控的运营成本（OpEx）**：为达到该营收目标，OpenAI 在 2030 年仅在算力上的开支就需要超过 2000 亿美元，其整体运营支出将超过 Meta 的两倍，甚至高于微软全公司约 1500 亿美元的运营支出规模。 --- ### [05:16](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=316s) 超大规模云厂商与无底洞需求：被扭曲的 AI 资本支出循环 回顾 2022 年底科技行业的寒冬，当时英伟达（Nvidia）在 2023 财年营收基本持平，微软、谷歌、亚马逊（Amazon）与 Meta 全面面临增长放缓的困境。这些科技巨头通过提价、改变广告竞价机制以及亚马逊发展广告业务来勉强维持业绩，同时 ChatGPT 的爆发成为了救命稻草。各家巨头开始疯狂扫货 GPU，华尔街分析师与媒体随即心照不宣地将随后的所有增长归功于 AI，推升了科技股暴涨。 然而，到了当前阶段，整个产业链陷入了极度危险的脆弱依存关系： | 关键环节 | 现实困境与具体数据 | |:--- |:--- | | **机构预测与超大规模云厂商绑定** | 来自瑞银（UBS）、巴克莱（Barclays）与富国银行（Wells Fargo）的预测显示，谷歌、亚马逊和微软总计约 **4400 亿美元** 的云端增长预期，完全依赖于 OpenAI 和 Anthropic 这两家持续亏损的初创公司。 | | **未兑现的资本承诺与资金缺口** | 这两家初创企业背负着超过 **1.1 万亿美元** 的长期算力采购承诺，但它们当前的实际现金流和客户真实需求根本无法覆盖这些承诺，两家公司现阶段的需求至少需要暴增 10 倍以上才可能维系。 | | **虚假的算力短缺信号** | 市场上充斥着“GPU 极度短缺说明终端需求旺盛”的论调，但实际上全行业 **90% 的 AI 基础设施** 仅被 OpenAI、Anthropic 以及疯狂囤卡却毫无实质回报的 Meta 所吞下。这造成了严重的虚假需求信号（Demand Signal）。 | | **微软 AI 收入的真实底细** | 彭博社数据显示，在 2026 财年微软约 343.3 亿美元的 AI 相关收入中，有 **241 亿美元** 本质上是 OpenAI 贡献的算力回流；微软自研的 AI 软件与第三方 GPU 租赁业务实际规模仅为个位数（数十亿美元），却支撑着超过 **2600 亿美元** 的巨额资本支出。 | | **专用硬件的沉没陷阱** | 与互联网泡沫破灭后铺设的通用光纤不同，AI GPU 属于专用硬件（主要用于 AI、数据分析及 3D 建模等极小众市场），一旦泡沫破裂，巨额折旧与昂贵的电力消耗将无法通过转为其他用途来消化。 | 眼见 Ed Zitron 的犀利剖析全面铺开，主持人 Josh Brown 笑着打断并宣布进入正式节目开场，调侃道必须先给听众留出时间去平掉他们手里的科技股多头头寸。 ### [08:45](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=525s) 赞助口播与开场介绍 节目首先插播了赞助商富达投资（Fidelity Investments）的广告，推介其全新上线的桌面与移动交易平台 Fidelity Trader+，主打跨设备图表偏好同步、定制化多屏监控与专员屏幕共享等功能。 口播结束后，主持人 Josh Brown 热情开场，称本期播客为“全球最伟大的投资播客”，并指出今天将展开一场前所未有的深度对话：探讨人工智能（AI）领域的“终极看空逻辑”（ultra-bear case）。本次特邀嘉宾是知名科技分析师与公关公司 EasyPR 创始人 Ed Zitron。Ed Zitron 同时是播客《Better Offline》（获 Webby 奖）的主持人，并在其时事通讯《Where's Your Ed?》中发表过多篇引爆全网的深度调查，尤其是 2023 年提出的“腐烂经济”（Rot Economy）概念，直指科技巨头牺牲真正创新来盲目追求纸面增长。目前，他是科技界声音最响亮、被引述最多的 AI 怀疑论者之一，长期聚焦于 OpenAI、Anthropic 的单位经济效益缺陷，以及超大规模云服务商（Hyperscalers）资本支出（CapEx）的可持续性危机。 --- ### [11:22](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=682s) 英伟达的惊人业绩与高度集中的客户风险 Josh Brown 开门见山向 Ed Zitron 发难，指出英伟达（Nvidia）最新交出的财报和惊人指引似乎正在“证明 Ed 之前所有的看空言论都是错的”。英伟达不仅打破以往按季度给出业绩指引的惯例，破天荒直接给出了横跨数年的远期指引，预计营收将继续保持高达 70% 的强劲增长，且 GPU 及下一代产品线的订单已被抢购一空。然而，英伟达约 90% 的业务来自数据中心，而数据中心支出的绝大部分又仅由四到五家超大规模云厂商撑起。 ```mermaid flowchart TD accTitle: 英伟达收入高度依赖少数客户的风险流向 accDescr: 展示英伟达芯片收入高度集中在三到五个超大规模云厂商，而终端企业真实需求却无法消化的传导链条。 subgraph Hardware["硬件销售端：英伟达"] NV["英伟达 GPU / 硬件出货"] end subgraph Hyperscalers["买单端：超大规模云厂商"] C1["单一最大客户 (占比 16%)"] C2["前三客户 (半年度占比 44%)"] C3["前五客户 (占比高达 70%)"] end subgraph Reality["终端商业化落地"] EndUsers["终端企业真实需求<br/>(仅贡献个位数数十亿，无法覆盖千亿 CapEx)"] end NV --> C1 NV --> C2 NV --> C3 C1 & C2 & C3 -.->|商业变现受阻| EndUsers ``` 面对主持人的质疑，Ed 澄清自己从未否定英伟达“实打实赚到了大钱”，但繁荣表面下掩盖着极度脆弱的客户结构集中度风险： - **极度依赖少数巨头**：在英伟达最新季度中，仅一家未公开名称的客户就贡献了其总营收的 16%；在 2027 财年上半年，前三名客户合计贡献了 44% 的收入；前五大客户更是直接占到了其全部账目收入的 70%。 - **下游传导断层**：Josh 辩护称这些云巨头并不是给自己买芯片，而是转售给全球 500 强企业、政府与主权机构；但 Ed 反驳指出，终端企业市场根本没有消化如此庞大算力规模的真实商业需求。 --- ### [13:41](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=821s) 微软的 AI 遮羞布：OpenAI 成了无法摆脱的负债 Ed Zitron 用具体数据拆解了号称“软件销售食物链顶端捕食者”的微软（Microsoft）：在刚结束的 2026 财年，微软虽然宣称获得了约 343.3 亿美元的 AI 相关收入，但其中高达 241 亿美元实际上来自 OpenAI 自己的返购与使用。除去这个被 Ed 讽刺为“败家大儿子”（large fail son）的 OpenAI，微软靠着全行业最庞大的数十万渠道商和数万名专业销售，向整个外部市场销售 Copilot 等原生 AI 产品所换来的年化收入，仅仅在个位数数十亿美元徘徊。 ```mermaid flowchart LR accTitle: 微软 2026 财年 AI 收入拆解 accDescr: 微软 34.33 亿美元 AI 收入中，24.1 亿美元来自 OpenAI 返购，仅约 10 亿美元来自庞大的外部商业销售。 MSFT["微软 AI 营收<br/>$34.33B"] --> OAI["OpenAI 返购拉动<br/>$24.1B (占比约 70%)"] MSFT --> Ext["外部企业采购 (Copilot等)<br/>~$10.2B (个位数增长/乏力)"] ``` Ed 进一步抨击微软的 AI 商业模式早已陷入死胡同： - **GitHub Copilot 商业化转向**：作为微软内部唯一称得上成功的 AI 落地产品，GitHub Copilot 已从过去“每月 40 美元随便调用价值 5000 美元 Token”的补贴模式，被迫转向按 Token 用量计费；随着套利空间终结，实际用户用量出现雪崩式萎缩。 - **巨额 CapEx 无法闭环**：在累计高达 2600 亿美元的惊人资本开支支撑下，来自财富 500 强的真正企业级 AI 需求每年仅挤出区区数十亿美元。 - **融资击鼓传花告急**：微软能从 OpenAI 收到真金白银的前提是 OpenAI 能够源源不断在市场上完成巨额股权融资。但风投资本早已望而却步——在 OpenAI 最近一轮 1220 亿美元的融资中，仅有 120 亿美元来自风投和参与数据中心基建的资管机构，其余 1100 亿美元实质上是英伟达（出资 300 亿）、软银（出资 300 亿，此举已让软银财务捉襟见肘）和亚马逊（出资 500 亿）的闭环互保。英伟达甚至已明确放话未来不会再参与这类投资，警告称若继续投钱注资，“糟糕的苦日子还在后头”。 --- ### [16:22](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=982s) 疯狂签单的奥特曼与被推高的数据中心成本墙 Ed 强调，微软与 OpenAI 的绑定关系正演变成一种系统性危机。随着微软放弃与 OpenAI 的排他协议，OpenAI 首席执行官山姆·奥特曼（Sam Altman）开启了令人瞠目的“疯狂签署巨额算力合约”模式，Ed 讽刺奥特曼“唯一的真正天赋就是到处签字画押”： - **OpenAI 的天价算力承诺**：在未来 8 年内，OpenAI 承诺向亚马逊云（AWS）支付 1380 亿美元，向 Cerebras 支付 200 亿美元，向 CoreWeave 支付 224.4 亿美元，向甲骨文（Oracle）承诺超过 3000 亿美元，甚至据瑞银（UBS）分析师 Steven Ju 的报告，OpenAI 每年还要悄悄在谷歌云上花费约 125 亿美元。 - **供应链现实与负债成本**：为了支撑华尔街分析师对英伟达 2028 财年高达 6740 亿美元营收的一致预期，市场不仅需要依赖这几家核心客户继续借入规模空前的高息债务，而且现实世界的基础设施瓶颈也已全面爆发——英伟达芯片自身价格上涨 17%、高带宽内存（HBM）成本飙升、顶级技术人才稀缺，以及电工级硅钢等电力变压器关键原材料严重短缺。 当 Josh Brown 拿出图表反驳，指出英伟达过去 12 个月（TTM）的净利润已经历史性地超越了苹果（Apple）时，Ed 尖锐反问：“这里面到底有多少只是账面上的股权浮盈（Equity Gains）？”他指出，英伟达通过对 Anthropic、OpenAI、CoreWeave、Nebius 和 Lambda 等生态圈客户进行密集的交叉持股，巧妙地将投资增值粉饰到了自己的资产与报表中。 ### [18:42](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=1122s) 曾经的误判与当前市场的最大盲区 Ed Zitron 坦诚地反思了自己在 2024 年做出的错误判断。他表示自己当时的错误纯粹是因为“天真”，以为市场是个讲理的地方，坚信微软、谷歌、亚马逊和 Meta 这些超级大厂不可能毫无缘由地狂砸数千亿美元。主持人打趣问：“所以你认错是因为觉得市场太蠢？”Ed 直言：“不，是我太蠢太天真，低估了这帮人能把事情推到多么极端的境地，也低估了整个债务融资体系居然能承受这种规模的挥霍。” 讨论随即转向当前市场究竟遗漏了什么核心风险。主持人指出，市场上其实许多人都隐约意识到了 OpenAI 是个潜在问题，比如甲骨文（Oracle）的股价即便在其他软件股反弹时依然暴跌 70% 躺在地上爬不起来。然而 Ed 指出，市场远远没有看清危机的本质：算力中心需求的极度中心化。目前全行业大约 60% 的 AI 数据中心需求完全由 OpenAI 一家撑起。如果剔除 OpenAI，市场上根本不存在第二个同等级别的巨额买家；而 Anthropic 或 OpenAI 能如此挥霍，完全依赖于源源不断的风险投资。一旦风投链条断裂，真实的算力需求就会瞬间蒸发。市场之所以还没意识到这一点，仅仅是因为“崩盘还没有在这一两个季度内真正发生”，大家仍在侥幸地盯着眼前的资金流入。 ```mermaid flowchart TD accTitle: AI 资本支出泡沫与真实需求脱节示意图 accDescr: 展现从风投资金与债务流入支撑超大规模开支，到终端缺乏真实订阅需求的恶性闭环流程。 subgraph CapitalInflow["资金输血端（高风险支撑）"] VC["风险投资 (VC) 与私人信贷机构"] Lenders["激进贷款人 / 私人信贷 (如 Blue Owl)"] end subgraph HyperscalersAndStartups["中间消耗端（算力需求高度集中）"] OpenAIAnthropic["OpenAI & Anthropic<br/>(占据约 60% AI 数据中心算力需求)"] DataCenters["超级数据中心项目<br/>(如 1.2 GW Abilene 项目)"] Nvidia["英伟达 Nvidia<br/>(卖铲人：高毛利兑现营收)"] end subgraph RealDemand["真实商业终端（严重脱节）"] EndUsers["终端客户与企业软件"] TokenSpend["非持续的代币消耗 (Token Run-rate)<br/>(真实成本一旦暴露，客户迅速流失)"] end VC -->|"风投资金补贴"| OpenAIAnthropic Lenders -->|"数十亿美元债务融资"| DataCenters OpenAIAnthropic -->|"采购海量算力"| DataCenters DataCenters -->|"芯片采购 (50%+ 成本)"| Nvidia OpenAIAnthropic -.->|"虚高年化运行率"| TokenSpend TokenSpend -.->|"无法覆盖真实算力成本"| EndUsers ``` --- ### [20:07](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=1207s) “年化运行率”的文字游戏与算力供需的惊人缺口 对于 Anthropic 和 OpenAI 号称“营收暴涨”并计划在半年内上市的传闻，Ed 嗤之以鼻，怒斥所谓的“运行率”（Run Rate）是“整个行业最大的骗局”。在传统的 SaaS 商业模式中，ARR 指的是“年度经常性收入”（Annual Recurring Revenue），是由具有法律效力、可预测的长期软件订阅合同构成的真金白银。而现在的 AI 公司却把某个短期周期的代币（Token）消耗量简单乘上 12 或 13 周，强行包装成年化运行率向外吹嘘。甚至连主流财经媒体在报道 Anthropic 获得数十亿美元估值时，也从不对这个运行率的计算口径做任何严格定义。代币消费根本不是经常性订阅——客户本月可能为了尝鲜砸下一千美元，下月转战开源模型就只花五十美元；公司完全可以挑一个刚刚发布新模型、尝鲜涌入高峰的时段截取数据，或者在订阅流失前强行平移统计区间来美化报表。Ed 愤怒地表示，这种数字把戏居然能被华尔街和投资人全盘接受，简直是对投资人智商的侮辱。 尽管主持人辩解称，上市公司在财报电话会议上总在抱怨“算力预算用光了，不得不追加开支”，Ed 解释说，这恰恰是因为 OpenAI 和 Anthropic 抢占了市面上几乎所有现存的基础设施资源，造成了局部的虚假繁荣与短缺。从全局经济账来看，两者存在极其恐怖的鸿沟： - Ed 估算目前全市场真实的有效算力需求仅在 **220 亿美元** 左右。 - 产业研究机构 Sightline Climate 报告显示，仅筹备规划中的数据中心规模就高达 **190 吉瓦（GW）**。 - 要消化支撑如此庞大的装机容量，行业每年需要 **1.5 万亿至 3 万亿美元** 的持续算力需求。真实的商业需求与疯狂的扩张计划之间存在着无法逾越的天文数字缺口。 --- ### [22:32](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=1352s) 吉瓦级“缩微城市”与私人信贷的草率狂欢 当被问及这场狂欢中“谁才是真正的蠢货”——究竟是盲目批准资本支出的 CEO 与董事会，还是二级市场的股票买家？Ed 毫不犹豫地指出：是贷款人（Lenders）。目前投入到 AI 数据中心建设中的债务资金让他“感到深入骨髓的恐惧”。 现代规划的吉瓦级超级数据中心已经演变成极度稠密的“缩微城市”： - **英国布里斯托尔市**：占地约 12.5 亿平方英尺，整座城市的电力负荷约为 800 兆瓦（MW）。 - **Stargate Abilene（甲骨文与 OpenAI 在得州规划的项目）**：占地仅 99.8 万平方英尺，电力需求却高达 1.2 吉瓦（GW / 1200 兆瓦）。 - 这种在千分之一的狭小空间内压缩整座城市电力负荷的怪物项目不仅严重拖延进度，更是直接架设在用“订书钉和胶带勉强粘合”的脆弱老旧电网上。所谓“建数据中心有助于电网升级”完全是科技巨头的自我粉饰。 更令人震惊的是尽职调查的极度缺失。Ed 援引科技媒体 *The Information* 的调查报道指出，私人信贷巨头蓝猫头鹰资本（Blue Owl Capital）在为一个耗资数十亿美元的项目做出出资承诺时，整个决策过程竟然仅仅花了 **10 到 15 分钟**。当主持人难以置信地质疑“这绝不可能是真的”时，Ed 尖锐反驳道，这就是最典型的“赌徒逻辑”（Gambler logic）：参与者根本不做基础审查，完全沉迷在“只要数字一直涨、全球大公司不可能无缘无故花一万亿、我只要抢先建好租出去就能发横财”的盲目狂热与恶意无知（nefariously ignorant）中。 --- ### [25:03](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=1503s) 英伟达的狂欢、补贴破灭与不可承受的债务负担 主持人调侃英伟达 CEO 黄仁勋就像“AI 联邦储备委员会主席”，每隔一段时间就出来安抚市场情绪：财报季通常先由各家超大规模云厂商（Hyperscalers）公布激进的资本开支，在三周的滞后期后，英伟达压轴登场用爆炸性数据“印证”一切狂热都是合理的。数据中心建设成本中芯片占比高达 50% 甚至更多，英伟达凭此斩获了 960 亿美元级别的惊人收入、超过 100% 的同比增速和高达 75% 的毛利率。 主持人和 Ed 达成共识：英伟达收到的支票绝对是真的，但这些采购资金的源头却充满了致命虚妄： - **虚假繁荣全靠补贴**：几乎所有的 AI 初创企业都在不计成本地亏损运营，向终端客户倒贴算力。此前 OpenAI 和 Anthropic 允许用户每月仅付 200 美元就消耗价值数千美元的代币。如果这些公司真有定价权，早就按真实成本收费了，而现实是一旦失去补贴让客户承担真实开销，需求就会瞬间萎缩。 - **全行业被债务绑架**：除了财力极为雄厚的微软，其余大客户（包括马斯克的 SpaceX / xAI 等）都在疯狂背负沉重债务来采购芯片。这些开支根本无法通过自身的自由现金流消化。 - **恶性循环加剧**：科技行业寄希望于虚无缥缈的“未来工业革命”，但在真实终端盈利迟迟无法兑现的背景下，借贷利息和芯片设备成本都在飞速上涨。整个体系正在被逼至承受极限，而金融资本却还在前仆后继地为最荒谬的资产泡沫加杠杆。 ### [27:41](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=1661s) 早期阶段还是皇帝的新衣？关于企业级 AI 生产力的现实争议 讨论从 Anthropic 的收入起点展开。Anthropic 成立于 2021 年，但直到 2023 年才入账第一笔美元营收。Josh Brown 认为当下的 AI 需求才刚刚露出一角，整个市场完全处于极其早期的阶段；然而 Ed Zitron 对此持极为怀疑的态度，指出科技顾问兼工程师 Nick Suresh 曾写过一篇犀利的文章《AI 正在蚕食全球决策力》（*AI is eviscerating global decisionmaking*）。Zitron 用电影《斯大林之死》（*The Death of Stalin*）开头的荒诞场景作比喻：高官们围在斯大林的遗体旁面面相觑，却硬要互相吹捧“他气色看起来真好，非常健康”。很多公司宣称“AI 让我效率提升了 100 倍”，实际上大部分 AI 集成项目根本以失败告终，完全是做样子给外界看。 Josh Brown 随即列举财报季中的企业真实案例进行反驳，尤其是 Airbnb。首席执行官 Brian Chesky 在财报电话会上直言 AI 是公司经历过的最好的技术变革，并给出了具体数据：上个季度 Airbnb 有 45% 的客服咨询完全由 AI 关闭解决，全程无需人工干预，不仅直接削减了运营支出，还显著提振了现金流。Josh Brown 还结合自身财富管理公司的运营日常指出，公司管理着 90 多名员工和 4000 多个客户家庭，在 Salesforce、投资组合核算系统和交易执行之间存在大量日常运维流程。客户根本不在意底层用了什么 AI，但运营团队每周都在尝试用 AI 提效，让公司运转得更快、更低成本。全美各行各业的企业管理层都在向分析师解释为什么业绩能超预期，其关键正是 2023 或 2024 年布局的 AI 投资终于在财报的利润端体现了出来。 ```mermaid flowchart TD accTitle: Enterprise AI Adoption Divide accDescr: Comparison between optimistic corporate operational gains and skeptical evaluations of real AI productivity. subgraph Bull["Josh Brown 的实操视角（多头）"] A["业务运维日常集成\n(Salesforce / 交易系统)"] --> B["具体成本削减\n(如 Airbnb 45% 客服闭环)"] B --> C["企业财报超预期\n与现金流改善"] end subgraph Bear["Ed Zitron 的怀疑视角（空头）"] D["企业管理层焦虑\n『不做显得落后』"] --> E["选择性披露数据\n(如 AT&T 宣称的个别环节省钱)"] E --> F["绝大多数项目失败\n不足以支撑巨额算力投入"] end ``` - **Airbnb 客服闭环**：上季度 45% 的客服支持由 AI 自动化完成，实现零人工介入并切实改善了现金流。 - **AT&T 的案例漏洞**：媒体报道称其使用开源模型节省了 80% 到 90% 的开销，但细看正文会发现仅指极少数个别环节，而 AT&T 接入 AI 的应用场景多达数千个，整体节省效果被严重夸大。 - **管理层跟风心理**：不少公司上马 AI 纯粹是因为害怕落后于同行，或者管理层已经骑虎难下，必须在投资者面前展现 AI 布局。 --- ### [31:46](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=1906s) 基础模型同质化、换代内耗与算力黑洞 讨论转向商业模型的基本面：企业即便是真的在使用 AI，所产生的收入是否足以维系处于严重亏损中的 OpenAI 和 Anthropic？或者企业最终都会倒向自建托管 GPU 结合开源模型？Ed Zitron 指出，OpenAI 和 Anthropic 必须以极其惊人的速度持续扩张，因为它们未来三年半内需要背负高达 4400 亿美元的资本开支承诺。 谈到 Google 的 Gemini 这种闭环生态时，Zitron 直言 Gemini 本质上也就是另一个大语言模型（LLM）。在他看来，市场上绝大多数用户和企业根本不在意底层究竟用的是哪家模型，模型之间的差异度微乎其微。然而，模型频繁迭代给下游开发者带来了巨大的内耗：每当 OpenAI 或其他厂商推出新模型，面向客户的软件公司就觉得自己必须立刻跟进替换，而每次切换都需要重写 Harness 代码、修改 Prompt。更致命的是，即使单个 Token 的名义单价保持不变，新模型在推理时消耗的 Token 总量往往大幅增加——这就像买了一辆百公里油耗相同的车，但每次出行莫名其妙被要求多开 200 英里，实际成本反而更高。 - **高昂的支出缺口**：头部闭源模型公司未来三年半面临着数千亿美元规模的算力与运营成本。 - **Token 消耗膨胀**：新一代复杂模型在思考和输出时往往消耗更多 Token，导致企业实际账单显著上涨。 - **超大规模云厂商的定价博弈**：LLM 目前在经济上很难实现真正盈利，未来超大规模云厂商（Hyperscalers）可能会大幅上调商用 API 的门槛，使之变为极少数特定企业才能负担的昂贵企业级服务。 --- ### [34:17](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=2057s) 大而不能倒？私人信贷预警与云厂商的“循环造血” 面对 AI 泡沫破裂的论调，Josh Brown 提出了一种反向做空陷阱的观点：整个体系已经“大而不能倒”（Too Big to Fail），华尔街和科技巨头牵扯的利益太深，过往历史证明系统往往会被托底硬撑过去。Ed Zitron 则用医疗紧急广播中的隐喻“沃特先生请立刻前往手术室”（Mr. Walter to the operating room）来形容局势的凶险，并直指私人信贷（Private Credit）正在成为巨大的风险引爆点。 Zitron 拆解道，即便资本市场试图对 OpenAI 和 Anthropic 进行救助，也解决不了企业无法赚取数百亿美元利润来填补算力黑洞的客观规律。微软、Google、Meta 和亚马逊等科技巨头自身的核心增长本来就在放缓，这才驱使它们将所有筹码押注在 AI 上。针对 Josh Brown 提出的“本季度除 Meta 外的各大云厂商（AWS、Azure、Google Cloud）营收增速全面加快”的反驳，Zitron 明确指出这恰恰坐实了问题的核心：云业务之所以出现加速反弹，完全是因为 OpenAI 和 Anthropic 在二季度进行了疯狂的“刷 Token”和大规模算力采购，而支撑这两家初创公司付账的资金，相当一部分正是由这些超大规模云厂商自己在上一轮融资中注资提供的闭环资金。这种相互依存的资本游戏一旦脱节，繁荣假象就会瞬间瓦解。 - **微软收入结构关联**：微软 2026 财年约有 7% 的收入直接与 OpenAI 的支出挂钩。 - **私人信贷隐患**：大量非公开债务和私有信贷资金流入了数据中心建设与算力抵押贷款，放大了系统性金融脆弱性。 - **资本闭环运转**：超大规模云厂商对 AI 初创公司进行战略投资，初创公司再将这笔资金作为云服务费全额返还给投资方，推高了云业务的当期营收增速。 ### [36:37](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=2197s) 云计算增长的假象与 AI 无法被救助的规模困境 在审视科技巨头的业绩时，Ed Zitron 指出，Meta 并没有因为 AI 实现爆发式增长，当前真正实现爆炸性收入增长的，唯有直接从 Anthropic 和 OpenAI 手中拿走算力租赁支票的云计算厂商（如微软、亚马逊和谷歌）。Josh Brown 提到 Meta 的看多者逻辑：Meta 目前不做出租算力的生意，而是将算力全部押注在优化 Reels 推荐算法上；但 Meta 随时可以调转方向宣布对外开放数据中心租赁，若是那样，股价必会应声暴涨。然而 Zitron 强调，核心广告业务利润虽然丰厚，但 Meta 已经出现了用户增长下滑的苗头，一旦用户见顶，广告收入终将回落。 ```mermaid flowchart TD accTitle: AI资金循环与依赖链条 accDescr: 展示Anthropic与OpenAI的巨额亏损如何支撑超大规模云厂商的增长神话 subgraph CapitalBurn ["巨额亏损源头"] A["Anthropic & OpenAI<br/>（高额算力预付款 / 巨额亏损）"] end subgraph Hyperscalers ["云巨头收入增长"] B["微软 Azure / 亚马逊 AWS / 谷歌云"] end subgraph Infrastructure ["数据中心投资"] C["6000-7000 亿美元<br/>数据中心新建需求"] end A -->|"资本支出 / 租用算力"| B B -->|"大规模资本开支 (CapEx)"| C ``` 对于市场中存在的救助幻想，Zitron 给予了冷酷的回击：这一次崩盘将是“天文数字级别的严重”，根本无法救助。因为问题不仅在于 OpenAI 或 Anthropic 能否存活，而在于它们必须成长为规模比肩谷歌或微软的庞然大物，并且每年还要花掉远超营收的资金。即使把未来的数据中心建设计划砍掉四分之三，市场上依然需要 6,000 到 7,000 亿美元的算力需求来填补缺口；但所有 AI 初创公司都在巨额亏损，根本无人接盘。微软、亚马逊和谷歌所谓的云业务高速增长，极度依赖 Anthropic 和 OpenAI 的输血，一旦这两家烧不动了，科技巨头的增长神话便将戛然而止。 - Zitron 透露他掌握 OpenAI 内部真实的财务数据：2024 年预计营收虽达 13.07 亿美元，但净亏损高达 209 亿美元，支出扩张速度远超收入增速。 - 整个生成式 AI 领域没有任何一家初创公司能实现盈利，缺乏内生现金流来填补数据中心庞大的折旧与运营成本。 --- ### [38:53](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=2333s) IPO 窗口期抢跑与企业预算见顶的现实 当 Josh Brown 问及 OpenAI 和 Anthropic 是否能通过公开上市（IPO）筹集上千亿美元并推迟危机爆发时，Zitron 承认短期内的确能续命，但无法解决根本的商业模式缺陷。Anthropic 之所以急于推动 IPO，是因为它们在今年前两个季度经历了客户疯狂“烧 Token”（Token maxing）的虚假繁荣。当时企业客户（超过 150 人的公司）从每月 150 至 200 美元的固定套餐被强制切换到按 Token 用量计费，恰逢各企业 CEO 像“脑震荡”一样盲目追捧 AI、下令不计成本投入，带动了收入短期暴增。 然而好景不长，财富 500 强企业和各类大客户在面对暴涨的账单后开始紧急刹车，收紧员工使用 AI 的权限。连 Sam Altman 都在公开活动中承认高昂的价格是客户面临的“巨大问题”。Anthropic 最昂贵、曾被夸大宣传为“因过于强大而被美国政府禁用”的高端模型，根据 Ramp 的企业支出数据显示，其市场份额已停滞在 11% 左右并逐步见顶。此外，扎克伯格虽然传出 Meta 考虑每年向 Anthropic 采购 100 亿美元算力的风声，但 Zitron 提醒大家，扎克伯格曾将公司改名 Meta 却又在 2023 年随手将元宇宙边缘化，“他掌握绝对控制权，随时可以凭一时兴起砍掉项目”，将命运押注在 Meta 身上的供应商风险极高。 - 企业端收费模式从固定订阅费转向按使用量（Token）计费，直接触发了企业财务审计与支出刹车。 - Anthropic 必须赶在第三季度财报显露出增长颓势之前完成上市流程，否则其财务报表将很难在公开市场上讲出动听的故事。 --- ### [41:56](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=2516s) 硅谷炒作周期的轮回：从 Clinkle、Clubhouse 到元宇宙 探讨 Zitron 的观察视角时，Josh Brown 指出 Zitron 的职业生涯始于科技公司的公关与品牌定位，因此他就像一个“看透了太多内幕的人”，亲眼见证了科技圈对待用户的贪婪与虚伪。Zitron 回忆起自己从 2013 年左右就开始发现，主流科技媒体极度不喜欢报道那些真正赚钱的务实企业，反而狂热追逐毫无商业模式的烧钱概念。他列举了曾经名噪一时的荒谬创业项目：号称通过声波实现移动支付的 Clinkle（一年内迅速夭折）以及主打超近距离熟人照片共享的 Color（融了巨资后暴毙），每次泡沫破灭后，VC 和媒体都若无其事地“拍拍屁股转向下一个谎言”。 这种荒唐在疫情期间的 Clubhouse 狂潮中达到了顶峰。当时 Andreessen Horowitz (a16z) 动用各种名流人脉（甚至请来塞缪尔·杰克逊等明星）为其站台，媒体更是宣称 Clubhouse 将彻底颠覆广播行业。Josh Brown 笑着补充，那段时期恰逢 SPAC（特殊目的收购公司）投机狂潮，Clubhouse 上充斥着毫无周期经验的 28 岁年轻人推销 SPAC 壳公司。Josh 分享了自己的一段趣事：他曾受邀参加 Clubhouse 上一档名为《What the SPAC》的语音节目，由于他在早年的经纪人生涯中推销过上百只 SPAC、深知它们最终都会归零，于是他在节目中直言不讳地揭穿真相，导致他从此再也没有被请回该节目——而后来那个房间里讨论的项目果真全军覆没。Zitron 总结道，从全民迷信 Clubhouse 到后来连基本形态都没弄明白、看起来就像粗劣电子游戏的“元宇宙”，整个创投媒体与资本圈始终深陷在制造概念与追逐幻象的恶性循环中。 ### [45:14](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=2714s) 虚假叙事、社会传染与 AI 泡沫的催化 讨论从回顾 2020 年左右的科技荒谬现象切入，嘉宾们笑着提到当年甚至有人在元宇宙中买卖“虚拟房产”（virtual real estate），而主流媒体竟然一本正经地报道，仿佛这是真实有价值的资产。Ed Zitron 指出，从加密货币（Crypto）到 NFT，媒体在每一次泡沫的吹大过程中都扮演了极不光彩的助推角色。尽管 Josh Brown 提醒在现代媒体诞生前历史上就屡见泡沫（例如郁金香泡沫），但 Ed 强调，社交媒体与居家隔离环境彻底放大了这种“社会传染”（social contagion），加速了泡沫的膨胀与恶化。 当 2022 至 2023 年科技行业步入低谷与萧条时，生成式 AI 的横空出世恰逢其时地给了所有人一根救命稻草：每个人突然都有了新由头——人们可以自称 AI 顾问、创办 AI 公司、把 AI 塞进传统软件中兜售，甚至承销 AI 基础设施。投资银行家有了可卖的新故事，整个科技与金融生态都被调动起来。然而，当被问及是否对当下的 AI 产品感到惊艳时，Ed 直言并不感冒，核心原因在于其高昂得不成比例的代价。 ```mermaid flowchart TD accTitle: AI 泡沫与社会传染机制 accDescr: 展示从科技低谷期到媒体助推、全行业兜售 AI 故事的演变链条 A["2022-2023 科技萧条与增长停滞"] --> B["生成式 AI 成为救命叙事"] B --> C["媒体与社交媒体加速'社会传染'"] C --> D["全员变现与资本热潮"] subgraph D["各方抢滩 AI 叙事"] D1["AI 咨询与初创项目"] D2["传统 SaaS 强塞 AI 功能"] D3["投行兜售基础设施概念"] end D --> E["万亿美元资本开支与环境、社会高昂代价"] ``` - 泡沫催化三要素：居家隔离催生的线上情绪孤岛、社交媒体的算法扩散机制，以及主流科技媒体缺乏审慎批判的盲从。 - 巨大的投入与产出失衡：虽然市场上出现了诸如开源/开放权重模型（open-weight models）或蒸馏模型（如 Kimi、K3 等），但追根溯源，它们依然消耗了数十亿至上万亿美元的基础设施资本开支、耗费了海量训练数据与算力。 - 产出可靠性极低：付出万亿美元级别的代价后，得到的技术产物依然无法让人真正信赖，用户仍需每一步进行核查（double check），Ed 认为这根本称不上真正的“智能”。 --- ### [47:24](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=2844s) “腐败经济”与大厂产品体验的系统性崩塌 主持团队将话题引回 Meta 以及科技巨头高层的动机，直接询问 Ed 是否“从骨子里痛恨这些人”。Ed 毫不掩饰地确认了这一态度，并直斥马克·扎克伯格（Mark Zuckerberg）等人根本不在乎用户体验，将 Facebook 和 Instagram 彻底变成了数字地狱。Ed 援引自己曾经曝光过的内部文件揭露，扎克伯格曾硬性要求所有关键指标维持“每年 12% 的永续增长”（12% perpetual growth）。为了追逐指标，团队盲目拉长用户在应用内的停留时间，却根本没意识到核心互动率（engagement）其实在不断下滑。 这种为了指标而牺牲用户体验的现象在硅谷具有普遍性。Ed 指出，这正是他所谓的“腐败经济”（The Rot Economy）——与科利·多克托罗（Cory Doctorow）提出的“平台劣化”（Enshittification）异曲同工。例如在谷歌的反垄断诉讼披露的司法部邮件中显示，当年主管广告业务的副总裁普拉巴卡尔·拉加万（Prabhakar Raghavan）发动了一场内部“政变”，夺取了原搜索负责人本·戈梅斯（Ben Gomes）的职权；戈梅斯曾警告拉加万的做法会极大地损害搜索质量，因为后者纯粹为了增加广告曝光量，故意引导用户反复修改关键词并进行多次搜索。 - 谷歌搜索劣质化的根源：人为降低首轮搜索结果的直接满意度，迫使用户为了获取答案进行更多次查询（queries），从而展示更多广告位。 - 操纵用户心理的恶劣行径：社交巨头不仅取消了默认的时间线倒序排列（chronological feed），更频繁调换按钮布局对用户进行 A/B 测试；《卫报》早年曾曝光 Meta 对 70 万名用户进行秘密的“情绪控制实验”，通过推送负面内容观察其心理反应。 - 传统变现手段耗尽：大厂已经难以通过自然留存榨取新价值，只能频繁对企业与个人用户提高订阅或广告竞价门槛，如微软和谷歌过去几年的多轮强制提价。 --- ### [50:13](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3013s) 增长邪教的终局：为什么说 AI 是大厂的最佳遮羞布 Michael Batnick 与 Josh Brown 提出了市场派的常见辩护：上市公司背负着股东受托责任（fiduciary duty），管理层的首要职责是通过自由现金流增长、股息和股票回购为股东创造回报，而不是充当全人类的“数字保姆”；更何况 Instagram 表面上是免费提供给用户的（由用户注意力变现）。但 Ed 反驳称，资本主义逻辑在被推向极端后会彻底脱轨——当一家公司的管理层被清一色的 MBA 职业经理人（如安迪·贾西、萨蒂亚·纳德拉、孙达尔·皮柴等）接管后，懂产品、在乎用户体验的工匠被驱逐殆尽，整套体制变成了一台只知道追逐无意义财报数字的庞大机器。 当被问及“如果用户真的讨厌这些平台，为什么不直接卸载”时，大家达成了共识：垄断与心理成瘾机制让普通人无法自拔。Josh 坦言自己对 Instagram 既爱又恨，甚至不得不借助实体防沉迷设备（如物理阻断盒 Brick）来强行控制屏幕时间。尤其是在未成年人成瘾问题上，各方一致谴责巨头的无底线扩张。尽管全美 28 个州近期就青少年心理健康问题与 Meta 达成了和解协议（要求对未成年人强制推行每天 2 小时使用上限），但 Ed 指出，分摊到 10 年支付的 120 亿至 170 亿美元罚款，仅占 Meta 市值的 1% 左右，如果这笔钱要求在单季度现金一次性结清，就足以抹平其整个季度的净利润。 ```mermaid flowchart LR accTitle: 腐败经济演化循环 accDescr: 展示从优质产品到过度榨取，最终以 AI 充当盲目扩张载体的全过程 subgraph Phase1["初创期: 体验驱动"] A["打造优质免费服务<br/>吸引海量网络效应"] end subgraph Phase2["榨取期: 腐败经济"] B["管理层换届为纯财务导向"] --> C["暴力插播广告/算法成瘾"] C --> D["自研好产品能力丧失"] end subgraph Phase3["僵局期: 盲目下注"] E["押注 AI 魔法黑盒"] --> F["无需动脑的资本支出扩张"] end Phase1 --> Phase2 Phase2 --> Phase3 ``` Ed 总结指出，脱离道德审判从商业本质来看，这种“增长邪教”的最致命后果在于：企业彻底丧失了研发“能赢得用户真心喜爱的好产品”的能力，只剩下用垄断地位扼住用户咽喉的手段。当管理层陷入“只要盲目砸钱就能换来账面扩张”的思维惰性时，AI 就成了他们眼中最完美的寄生载体——它看起来像一个能回答一切问题的万能黑盒，大厂可以不假思索地把它嵌入任何旧系统，借此向华尔街继续编织下一轮关于无尽增长的宏大谎言。 ### [53:39](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3219s) 大模型是糟糕的产品还是生产力奇迹？ Ed Zitron 直截了当地指出，大型语言模型（LLM）从产品体验角度来看极其可疑。科技巨头们最初幻想着甚至不需要刻意设计产品，“模型本身就会产出产品”，但这一理论彻底落空了。真正注重用户体验的人都会发现，LLM 的输出极其不稳定——有人能从中获得价值，有人完全用不上，而且模型时不时就会崩溃或凭空捏造事实。Zitron 情绪激动地表示：“那根本就是个疯子一样的产品。让人无法理解的是，业内居然还有那么多人为之极力辩解，谎称‘幻觉问题已经被解决了’，这完全是睁眼说瞎话。”他认为所有科技巨头都已经被增长机制死死绑架，高层们以为只要不停狂买 GPU，公司就能持续扩张，而 Meta 等公司的做法显然错得离谱。 听到这里，Michael Batnick 感到难以置信并出言打断：“等等，你觉得这些产品很烂？我很喜欢用 Claude 啊！”Batnick 甚至用“奇迹般的存在（miraculous）”来形容它，称自己整天都在用，而且身边绝非只有他一个人这么觉得。Zitron 承认彼此体验的巨大割裂非常耐人寻味——毕竟这被吹捧为改变时代的全新浪潮，而他本人其实也是个科技爱好者。 Josh Brown 随即加入讨论，以自身创作者的视角剖析 AI 的定位。同为作家的 Brown 表示，他绝对不会让 AI 来替自己写稿，因为读者之所以阅读他的内容，看重的是他独特的行文口吻和鲜明立场。然而，在写作思路顺畅时，他极度抗拒跳出界面去 Google 搜索页面的海量蓝色链接里翻找数据；相反，他希望无论是 Gemini 还是 Claude 都能充当副驾驶（Co-pilot）的角色——“当我在忙着写作时，帮我去复核这组数据，并在我落笔之前把原始信源列出来”。 - Ed Zitron 严厉批判行业对“幻觉”问题的洗白，认为不可控、时常捏造事实的 LLM 根本算不上合格的用户端产品。 - 科技巨头之所以深陷其中，是因为其组织架构完全被增长指标绑架，形成了“买 GPU = 带来增长”的盲目闭环。 - Josh Brown 将 AI 视作写作流中的“副驾驶”，主要用于快速调研、核对数字并检索原始信源，而非代笔创作。 --- ### [55:16](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3316s) 昂贵的搜索与外包思考的危险 Brown 指出，Google 早在一年前就已经意识到这场危机——如果他们不主动自残现有的核心搜索业务并将 AI 概览推到最顶端，随着用户搜索习惯向 ChatGPT 迁移，对 Google 来说就是“灭顶之灾”。然而 Zitron 反驳称，Brown 所描述的这些用途，本质上依然只是“搜索”而已。 Brown 追问，既然 Zitron 也是专业作者，难道不觉得这项能力极其宝贵吗？Zitron 举了彭博终端（Bloomberg Terminal）的例子：过去用户必须手写 BQL（彭博的类 Python 专有查询语言），而现在终端推出了名为“Ask B”的功能，用户只需用自然语言提问，系统就会自动生成 BQL 代码并运行测试，这确实省去了不少麻烦。但 Zitron 尖锐地反驳道：“为了获得稍微好一点的搜索和代码辅助，整个行业砸进去了数千亿美元乃至上万亿美元，这根本不值一提。” 当被问及这是否就是技术终局时，Zitron 预测未来只会有局部的渐进式改善。比如行业正在转向网络安全，或者在编程辅助上继续微调，但这既不能真正替代程序员，甚至越往后越难帮到资深工程师——因为生成的代码会以乘法级的速度变得更繁杂、更难管理。Zitron 强调，他最反感的不仅是成本与回报的倒挂，更是人们逐渐混淆了“外包繁重任务（如检索）”与“外包思考”之间的界限；而 Brown 也深表赞同，认为将独立思考外包给 AI 是极其危险的趋势。Zitron 还吐槽自己曾试图让 AI 帮孩子解决 Minecraft（我的世界）的一个游戏故障，结果 AI 逻辑兜圈子追着自己的尾巴折腾了半个多小时才搞定，中间犯错无数；如果按照每百万 Token 计费，这破事可能得花掉 15 美元。 ```mermaid flowchart TD accTitle: AI投资与实际产出之间的错配 accDescr: 展现巨额CapEx投入与用户仅获得边缘搜索或初级代码生成的对比 subgraph Cost ["沉没成本与资源消耗"] A["巨额资本开支 (数千亿美元)"] B["GPU 折旧与高昂算力成本"] end subgraph Delivery ["实际交付能力"] C["Ask B / 辅助生成检索代码"] D["生成式搜索概览 (替代蓝链)"] E["易错代码排障 (绕圈排查)"] end subgraph Risk ["结构性风险"] F["代码库复杂度乘数级恶化"] G["人类丧失自主思考能力"] end A --> C B --> D B --> E E --> F D --> G ``` --- ### [57:17](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3437s) 巨头的博弈困局与不可持续的商业模式 Batnick 提出不同看法，询问 Zitron 除了彭博终端外，有没有真正利用 AI 自动化某些重复工作流程。Batnick 透露自己虽然不懂技术，但他在 Claude 上创建了专属技能工作流——以往每周五早晨他需要打开 5 个不同网站反复复制粘贴数据，现在只需一句话“周五了，把我要的东西拿来”，Claude 就会自动完成全部抓取与整理，这对他来说简直如同“奇迹”。 Zitron 解释道，如果脱离社会成本、环境损耗和财政资本的语境，LLM 确实是一项有趣的自动化技术；但如果一开始业内把它老老实实叫做“图书馆模型（Library Models）”，资本市场绝对不会产生这种非理性的狂热。科技巨头们之所以陷入疯狂，是因为一种根深蒂固的生存恐慌：如果自己不把数据中心所有的 CPU 拔下来换成 GPU，既有的云业务就会被宣判死刑，AWS 将沦为废墟，而竞争对手将独霸天下。但这种“沉没成本博弈”导致谁都不敢先停手，因为谁率先削减 AI 资本支出（CapEx），市场就会立刻质问：“你为什么停了？你前几年烧的那些钱到底弄出了什么？”更可怕的是，这种开支甚至无法暂停，因为高负荷运转的 GPU 芯片还会自然损耗报废。 随后，Zitron 指出当前整个生态的基础建立在完全不可持续的定价假象之上。他询问 Batnick 目前是如何付费的，Batnick 表示公司目前使用的是类似标准 SaaS 的订阅费模式（如 Gemini、Claude、Hazel 等），并没有按每百万 Token 的 API 计费，而且费用非常便宜。Zitron 随即拿微软旗下的 GitHub Copilot 举例：GitHub Copilot 此前拥有 200 万付费客户，但在今年 6 月 1 日，微软突然掀桌子变相提高门槛并改变规则。之所以出现这种对待用户的疯狂行径，是因为微软之前居然允许用户花着 40 美元的订阅费，却在后端烧掉高达 5000 美元的 Token 算力成本！科技巨头们总是天真地假设算力成本会迅速降下来，但现实中模型为了执行更复杂的任务，单次消耗的 Token 数量反而大幅飙升，所谓“智能成本下降”在更巨大的算力消耗面前被彻底抵消。 - Batnick 分享了利用 Claude 自动化周报跨网站抓取数据的成功工作流，认为对金融分析师而言效率提升巨大。 - 科技巨头被博弈困局套牢：拔掉 CPU 换 GPU 是出于云基础设施被彻底淘汰的恐惧，谁先削减 CapEx 谁就会面临叙事崩盘。 - GitHub Copilot 曾出现向用户收取 40 美元却产生高达 5000 美元 Token 算力消耗的严重倒挂，最终迫使微软强行调整规则。 --- ### [01:00:25](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3625s) AI 启示录的第一骑士：CoreWeave 与私有债务断裂 关于这一轮万亿美元级别 AI 军备竞赛何时落幕，Brown 将其称为华尔街最新的“客厅猜谜游戏（parlor game）”。很多人猜测最可能的剧本是某个超大规模云服务商（Hyperscaler）扛不住压力主动认输，比如 Meta 突然宣布下调资本支出指引，随即引发连锁反应——半导体股票暴跌 30%、公用事业电力股如垃圾股般崩盘、估值倍数严重压缩，最终演变成全面恐慌。然而目前为止，巨头们没有任何退缩迹象。Brown 借此向 Zitron 提问，谁会是这场狂欢落幕时的那声警报？ Zitron 回应称，巨头砍 CapEx 固然是一个显而易见的结果，但他认为真正的“启示录第一骑士”更有可能是新兴云算力转租商（Neocloud）——尤其是 CoreWeave 的债务违约或融资断裂。CoreWeave 的核心业务就是大批量购买 Nvidia GPU，然后将其算力出租给 OpenAI、为 OpenAI 采购的微软、为 OpenAI 提供算力的 Google，以及 Anthropic 和 Meta。虽然表面上核心客户名单极其华丽，但其实际业务陷入了极其严重的亏损。CoreWeave 在过去一年内进行了多达三次数十亿美元规模的巨额融资，其最近一笔借款支付了高达 9.2% 以上的高昂借贷利率。Zitron 警告，这种高度依赖高息私有信贷（Private Credit）支撑的杠杆游戏极为脆弱，一旦信贷市场紧缩导致其资金链断裂，不仅其自身运转会戛然而止，还会瞬间引爆上游 GPU 订单与下游初创企业的整条多米诺骨牌。 - **华尔街的猜顶预期**：普遍认为转折点将由 Meta 等超大规模云厂商下调资本支出指引引爆，从而重创半导体供应链与公用事业板块。 - **CoreWeave 的债务灰犀牛**：作为核心的新兴云服务商（Neocloud），CoreWeave 面临严重的持续性亏损，并背负超过 9.2% 的高昂借贷成本。 - **脆弱的共生链条**：CoreWeave 的客户包括 OpenAI、微软、Google、Anthropic 及 Meta，其算力扩张高度依赖高成本私有信贷，一旦无法按期借新还旧，将成为率先倒下的多米诺骨牌。 ### [01:02:08](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3728s) Neocloud 的债务无底洞与数据中心债券市场的裂痕 Ed Zitron 指出，随着英伟达（Nvidia）GPU 变得愈发昂贵，以 CoreWeave 为代表的所谓“新兴云厂商（Neoclouds）”——包括 Iren、Nebius、Nscale 等——陷入了必须无休止发债融资的死循环。与谷歌拥有坚如磐石的搜索业务来补贴基建不同，这些新兴云服务商根本谈不上拥有真正可持续的商业模式；它们宛如被巨头包养的附属品，每个季度都需要靠注资数十亿美元来续命。一旦其中任何一家试图发债却以失败告终，整个链条就会开始断裂。 与此同时，超大规模云服务商（Hyperscalers）的债券发行市场也初现裂痕。虽然谷歌的一笔债券获得了 4 到 5 倍的超额认购，但近期亚马逊发行的一笔债券仅录得 1.6 倍超额认购。Ed 警告称，如果超额认购倍数进一步降至 1.1 或 1.2 倍这种极度脆弱的水平，舆论与市场情绪将会迅速反转，“数据中心债券需求枯竭”的恐慌便会彻底公开化。过去的盲目乐观寄望于社会道德或巨头自省来踩刹车，但资本市场的现实冷酷得多：决定一切的唯有资金供给。而随着英伟达芯片、内存（RAM）和数据中心建设成本水涨船高，能够持续筹集百亿美元巨资的企业已所剩无几，剩下的大多依靠黄仁勋的注资维持，或者完全寄生于 OpenAI 和 Anthropic 的算力采购。 ```mermaid flowchart TD accTitle: Neoclouds 债务循环与资金断裂链条 accDescr: 展示英伟达芯片高成本导致新兴云服务商不断借债，最终面临流动性枯竭的过程。 subgraph CostPush ["成本持续攀升"] A["Nvidia GPU 与 RAM 涨价"] --> B["超大规模机房建造成本飙升"] end subgraph NeocloudDebt ["新兴云（Neoclouds）生存模式"] B --> C["缺乏自我造血业务 (无搜索/主业现金流)"] C --> D["持续发债与高息融资 (CoreWeave 等)"] end subgraph LiquidityRisk ["流动性危机爆发点"] D --> E["债券认购疲软 (如超额认购降至 1.1x)"] D --> F["评级机构降级 / 债务展期失败"] E --> G["资金链断裂与 AI 资本开支崩塌"] F --> G end ``` - **Neocloud 代表企业**：CoreWeave、Iren、Nebius、Nscale。 - **发债认购反差**：亚马逊数据中心债券仅录得 1.6x 超额认购，而谷歌仍录得 4-5x。 --- ### [01:03:40](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3820s) 甲骨文的财务伪装与 AI 启示录的“第四骑士” Ed Zitron 调侃了甲骨文（Oracle）在社交媒体上的异常举动。每当 OpenAI 传出负面新闻（例如此前传闻英伟达拟向 OpenAI 投资 1000 亿美元却无果而终），甲骨文的官方 Twitter 账号就会急不可耐地公开发文辟谣，声称“我们在 OpenAI 上的财务状况毫无问题”或“星际之门（Stargate）项目进展顺利”。在 Ed 看来，投资者完全可以将甲骨文的声明反向理解。最滑稽的是，甲骨文刚与 OpenAI 签署巨额协议不久，首席执行官萨弗拉·卡兹（Safra Catz）便宣布交权退居幕后，安排了联席 CEO 来接手这个烫手山芋。目前甲骨文仍面临巨大的再融资压力，正在推进大规模股票抛售。 主持人与 Ed 探讨了评级机构的懦弱：尽管甲骨文的信用评级仅比垃圾级高出一档，但评级机构绝无勇气将其打入“堕落天使”（Fallen Angel）；甚至连 CoreWeave 的高风险债务，仅因绑定了超大规模厂商就被评为投资级。随后，嘉宾们系统盘点了可能戳破 AI 泡沫的“第四骑士”（The Pale Horse）： 1. **超大型云服务商（Hyperscaler）单方面退赛**，削减开支； 2. **重度依赖发债的巨头信用评级被降级**（如甲骨文）； 3. **明星 AI 初创公司资金耗尽破产**（如 Perplexity，尽管声量渐消，英伟达仍计划以 300 亿美元估值投资数亿美元，本质上是因为它是少数真金白银购买 GPU 算力的客户）； 4. **CoreWeave 无法完成债务展期（Roll over debt）**，或任何核心参与者发生再融资失败； 5. **标志性 AI 独角兽 IPO 惨遭破发惨败**（尤其是软银在流动性极其紧张、背负巨债的情况下若强推 IPO 失败）。 - **评级机构现状**：面对 CoreWeave 等高杠杆实体缺乏穿透审查，因背靠大厂而给予投资级评级。 - **Perplexity 估值疑云**：英伟达考虑在其 300 亿美元估值轮次中注资数亿，核心动机是维系下游算力采购客户的生存。 --- ### [01:06:30](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3990s) 1873 年铁路危机重演、GE 式供应商融资与需求虚幻 针对 Michael Batnick 提出的“股市并未出现狂热（英伟达远期市盈率仅 18 倍、Meta 仅 16 倍）”的疑问，Ed 反驳称，当前的狂热是“实体活动与建设泡沫”，而不仅是二级市场的估值倍数。甲骨文在剔除通胀后营收已停滞 15 年，近期还解雇了负责核心合同续约的团队，内部管理混乱不堪，却依然被卷入 AI 炒作。 主持人 Josh Brown 提到，重读讲述 1873 年大恐慌（Panic of 1873）的美国铁路大建设与崩盘历史让他彻底转为空头。当前 AI 数据中心的建设狂潮与当年的铁路如出一辙——它们都是极度庞大且昂贵的“重资产物理基建狂潮”，而非单纯的金融游戏。Ed 则进一步补充了通用电气（GE Capital）在杰克·韦尔奇（Jack Welch）时期的先例：当时 GE 旗下的航空租赁部门（GECAS）通过为下游航空公司提供激进的供应商融资（Vendor Financing），靠借钱给客户买自己的飞机来强撑繁荣；如今英伟达通过投资 AI 初创公司让其反哺算力收入，完全是在复制这一套路。 面对 Josh 提出“自动驾驶、人形机器人等通用自动化是否需要这种非线性计算能力”的质疑，Ed 指出核心矛盾在于**体量与投入产出比严重错配**： - 当前规划的数据中心规模要求每年产生 7,000 亿至 1 万亿美元的持续收入才能覆盖成本，但真实商业需求远未达到这个量级； - 特斯拉、Waymo、Zoox 等深耕自动驾驶多年的公司，其技术突破并非单纯堆砌当下的海量通用 GPU 就能解决； - 自动驾驶最大的瓶颈是解决最后“1% 的长尾 corner cases”，这注定是一个需要数十年缓慢演进的渐进工程，根本无法在短期内消化为了大语言模型（LLM）过度超前建造的庞大算力冗余。 ### [01:10:11](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4211s) 虚假需求与英伟达财报里的“异常信号” Ed Zitron 指出，目前整个生成式 AI 产业正深陷一种极度扭曲的增长压力之中。根据英伟达自身的财报逻辑，全球实际上只有三到五家超大规模云服务商（Hyperscalers）拥有足够的财力以当前规模持续买单。然而，OpenAI 和 Anthropic 必须疯狂扩张以维持生存，英伟达也必须展现疯狂的增长数字来支撑高企的估值。Ed Zitron 认为，英伟达在财报沟通中宣称“毛利率可维持在 70% 左右”更像是一种出于绝望的安抚手段——尽管其整体数字极其惊人，但市场此前已经出现过抛售与疑虑，他们迫切需要持续释放刺激信号。 更耐人寻味的是英伟达在最新 10-Q 季报中披露的某些反常细节：他们特别提及部分“投资级合作伙伴”可以延期 3 个月至 1 年付款。主持人 Josh Brown 随即追问，这是否意味着英伟达正在演变成其核心客户的投资人，甚至是债权人？Ed Zitron 直截了当地指出，这正是经济周期中最为典型的“虚假需求”（Illusory Demand）或“人造需求”迹象。当一家硬件供应商需要靠向客户注资或提供信贷额度来让对方购买自己的芯片时，表面的繁荣背后已经埋下了极大的隐患。 ### [01:11:06](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4266s) 是否存在“翻盘”可能？AGI 叙事的破灭与资本开支死结 Josh Brown 提出了观众最渴望听到、也是最具挑战性的问题：“你是否有任何可能判断错了？需要发生什么具体的事情，你才会承认自己看漏了真正核心的变量、承认自己错了？”Ed Zitron 幽默地反击道：“下次你们邀请看多 AI 的嘉宾来节目时，请务必当面问他们一模一样的问题——如果你们错了该怎么办？”他坦言，自己并不需要百分之百正确，局面才会变得极度糟糕。哪怕 OpenAI 和 Anthropic 最终没有彻底归零，而是通过微软、亚马逊或谷歌组建的某种“残局救助联盟”进行复杂的账目粉饰，萎缩成极小规模的特定业务苟延残喘，LLM 作为一个细分行业也已经算实质性破产。 Ed Zitron 断言，根本不会有什么神奇的底层技术突破来挽救这个局面，所谓的“AGI（通用人工智能）叙事”早已破产，就连行业大佬们私下里也在试图淡化这一概念。尽管黄仁勋此前高调宣称 AGI 已经到来，但在业内甚至连 AGI 的标准定义都无人能够给出。更致命的矛盾在于算力基建的天量数字：超大规模企业如果想在明年实现 6700 亿甚至 6900 亿美元级别的硬件销售吸收，亚马逊、谷歌、Meta 等巨头就必须大举发债借贷。尽管亚马逊宣称要采购 200 万张 GPU，但巨头们的自由现金流正在遭受重创，Meta 和亚马逊在三季度给出的指引均不及预期，增长放缓已经初现端倪。GPU 越来越贵，要完成预定规划就必须投入远超以往的资本，Ed Zitron 强调，即便在最疯狂的幻想中，全世界也根本掏不出成千上万亿美元的真金白银来单靠模型调用赚回这笔天量 CapEx。 ```mermaid flowchart TD accTitle: AI 基建资金死循环与现金流承压链条 accDescr: 展示了超大规模巨头举债采购昂贵硬件，但下游缺乏实际造血能力，最终导致资本开支无法闭环的死循环。 subgraph Hardware_Supply["供给端：疯狂增长诉求"] NV["英伟达芯片不断提价<br/>毛利率维持 70% 承压"] Credit["向客户提供垫资/信贷<br/>（虚假需求特征）"] NV --> Credit end subgraph Hyperscalers["中游：超大规模科技巨头"] Capex["背负天量 CapEx 预算<br/>（Meta/亚马逊采购计划）"] Debt["被迫通过发债/消耗现金流买单<br/>（三季度业绩指引放缓）"] Capex --> Debt end subgraph Downstream["下游：LLM 模型公司"] Labs["OpenAI / Anthropic<br/>承诺万亿级别算力开销"] RevenueGap["实际盈利能力极度匮乏<br/>无法消化数千亿算力成本"] Labs --> RevenueGap end Credit -.-> Capex Debt --> NV RevenueGap -.->|无法变现支撑| Capex ``` ### [01:14:01](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4441s) 拒绝参与烟雾弹市场与“毫无价值”的遗留资产 当 Michael Batnick 问及 Ed Zitron 是否真金白银在公开市场下注做空这些公司时，Ed Zitron 明确表示自己没有任何股市仓位。Josh Brown 评价道：“你虽然没有资金下注（Skin in the game），但你已经是 AI 怀疑论的代表面孔，你押上了自己的职业声誉，声誉风险完全不亚于数十亿美元的真金白银。”Ed Zitron 解释，当前的股市根本无法通过基本面分析和审读财务报表来理性参与，现在的二级市场纯粹是靠“解读烟雾信号”。例如此前 OpenAI 仅仅发布了几次毫无实质落地进展的声明，就让 SK 海力士、三星、AMD 和英伟达的股价大幅脉冲上涨，在一个充斥着“虚假公关声明驱动”的市场中下注是极其荒谬的。 面对经常私下向他咨询观点的对冲基金经理，Ed Zitron 始终提醒他们关注底层资产的本质区别。AI 狂潮与当年的互联网泡沫（Dot-com Bubble）有着决定性的不同： - **无法被重用的专用硬件**：Dot-com 泡沫破裂后留下的服务器和网络基建可以低成本平移至后续的新业务（例如亚马逊 AWS 的诞生），而耗资不菲的专用 AI GPU 离开深度学习模型训练与推理后，几乎毫无其他替代用途。 - **运维成本与烂尾代价高昂**：一座尚未竣工的数据中心哪怕搁置 3 年，复工完成它的造价依然分文不减；AI GPU 的电力开销只增不减。很多临时靠燃气轮机发电维持的数据中心，未来将承受极其高昂的气价波动与停运风险。 - **绝无仅有的资本吞噬量**：亚马逊 AWS 从 2003 年创立到 2015 年实现盈利，整整 12 年间整个亚马逊公司的经通胀调整后累计资本开支（包含电商、物流等全部业务）仅为 297 亿美元；而如今 AI 巨头们单季度砸向 GPU 的数字就已远超当年数年总和。 ### [01:17:08](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4628s) 私人信贷暴雷隐忧：超越 Dot-Com 的系统性金融冲击 Ed Zitron 指出，大众对于 OpenAI 这种初创企业某天突然关门其实并不会感到过于意外，真正将给全社会带来毁灭性打击的，是那些耗资数千亿乃至上万亿美元的数据中心因无人使用而沦为空转烂尾楼，随之而来的是规模庞大的数据中心债务出现违约。普通大众对于“数千亿美元、数万亿美元”这种天文数字缺乏直观概念，总盲目笃信“这些硅谷科技巨头汇聚了世界上最聪明的人，他们绝不可能犯傻”。 这一论调让 Josh Brown 深刻意识到事态的严重性：2000 年互联网泡沫破裂时，纳斯达克指数虽然暴跌了 80%，但美国底层实体经济基本未受重创，因为当时主要是公开市场的股票投机者蒙受损失。然而，当下的 AI 军备竞赛通过庞大的私人信贷（Private Credit）体系深度捆绑了商业银行、养老金与保险资金。这些深藏在黑箱之中的非公开债务由于规模过大、透明度极低，一旦发生底层数据中心违约潮，其引发的连锁反应将具备真正摧毁实体金融系统的破坏力，带来远超 Dot-com 时代的一场全经济范围的系统性灾难。 ### [01:18:24](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4704s) 私人信贷与数据中心建设的隐蔽风险 谈话转向了支撑这场 AI 基础设施狂潮背后的庞大资金链条。Ed Zitron 直言，如今市场上充斥着一种极其轻浮的豪赌心态，动辄就是“扔个 100 亿美金无所谓”的做派。以黑石（Blackstone）为代表的巨头正在疯狂砸钱，而这些资金的源头正是保险资金、年金以及蓬勃发展的私人信贷（Private Credit）。Ed 指出，大约六分之一的保险年金已经被配置到了私人信贷领域，再加上各类公共与私人养老金，几乎整个金融体系的底盘都已被裹挟卷入。尽管市场如今才开始对私人信贷领域的风险产生警惕，但 Ed 早在数月前就已撰文预警。诸如 Hunterbrook 的联合创始人 Sam Koppelman 等调查者也在持续揭露其中的问题，但绝大多数市场参与者依然没有把这笔巨额债务与真实的数据中心基本面联系起来。 ```mermaid flowchart TD accTitle: AI基础设施债务链与风险传导 accDescr: 展示从养老金和保险资金流向数据中心，再因需求不足与建设受阻面临违约风险的闭环。 subgraph Capital["资本源头与信贷机构"] P["养老基金与保险年金"] --> PC["私人信贷机构 / 资产管理巨头"] end subgraph Buildout["数据中心与基建承建"] PC --> DC["超大规模数据中心项目"] DC -.-> Risk["成本超支 / 地方反对 / 税收优惠取消"] end subgraph Demand["下游需求与现金流"] DC --> Hyperscalers["OpenAI / Anthropic / 新型云厂商"] Hyperscalers -.-> Shortfall["实际商业化收入无法匹配庞大Capex"] Shortfall --> Default["债务利息无法偿还， senior tranche 风险敞口暴露"] end ``` 在 Ed 看来，即便 OpenAI 和 Anthropic 未来完全不死、甚至营收规模再翻五倍（虽然他认为这根本不可能发生），其产生的终端实际需求也远远填不满目前正在规划和兴建的数据中心容量。更致命的是物理现实：人类历史上几乎从未有过不超预算的宏大工程，而眼下正在上马的正是人类有史以来野心最大、复杂度最高的基础设施建设项目。私人信贷机构习惯了互相眨眼心照不宣的宽松风控，将资金注入这些极其不稳定、复杂的项目之中；与此同时，地方层面的阻力已经开始爆发——亚利桑那州、伊利诺伊州等地正在缩减或取消税收优惠，德克萨斯州和宾夕法尼亚州也出现了强烈的反对声浪。如果这些数据中心无法按时交付，或者完工后缺乏足够的付费客户，那些结构复杂的优先档与劣后档债券将面临巨大灾难，而买入黑石等基建基金的普通机构与个人根本不知道风险究竟潜藏在何处。 - **资金渗透率**：大约六分之一的保险年金资金已流入私人信贷市场，公共和私人养老基金均大比例涉足。 - **地方政策转向**：亚利桑那州、伊利诺伊州削减税收减免，德克萨斯州和宾夕法尼亚州对耗能巨大的数据中心建设展现出明确的地方反对阻力。 --- ### [01:20:33](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4833s) 英伟达的财务迷局：应收账款激增与循环交易质疑 Ed 随后抛出了他整场节目中最尖锐的一个质疑：英伟达（Nvidia）究竟是在何时、以何种方式确认收入的？他指出，在英伟达光鲜亮丽的财报背后，应收账款（Accounts Receivable）正在以令人不安的速度膨胀，环比增长高达 55%，应收账款周转天数（DSO）也在持续拉长。面对主持人和多头常用的“这是因为积压订单（Backlog）太多”的辩解，Ed 反驳道，如果业务真的如此赚钱且供不应求，为何回款周期反而在不断延长？英伟达难道已经触及了客户迅速掏出现金支付的极限？ 更令 Ed 感到警惕的是深层的会计操作与循环关联交易。在现行会计准则下，硬件只要发货就可以计入收入，而实际资金流往往可以延后甚至经过复杂包装。英伟达不仅手握高达 300 亿美元的“云算力协议”（即反向租回自己售出的 GPU 算力），还背负着 250 亿美元的数据中心租赁协议。Ed 质问道：“黄仁勋，你本来是个卖铲人，是要把芯片卖给数据中心，为什么现在你还要反过来花几百亿去租回这些数据中心？你是在自己建数据中心给自己送钱吗？”这种“用债务抽水、左手倒右手把钱喂给自己制造虚假繁荣”的闭环，让 Ed 怀疑这已经演变成一场现代金融炼金术，一旦触碰流动性红线，其崩塌将极其突然和可怕。 - **应收账款异动**：英伟达应收账款环比剧增 55%，DSO 持续延长，显露出回款端的压力。 - **关联承诺敞口**：英伟达披露了约 300 亿美元的云算力回租协议及约 250 亿美元的数据中心租赁协议，形成了复杂的内部循环买卖链条。 --- ### [01:22:19](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4939s) 崩盘担忧与散户杠杆：并非为了“幸灾乐祸” 听到 Ed 情绪激昂地列举了一连串致命隐患，Josh 和 Michael 忍不住打断并问道：“你希望自己是错的吗？你看起来好像非常享受数落这些问题的过程，如果一切真的轰然倒塌，你是不是会在那里狂笑？”Ed 严肃地否认了这种说法，并引用了电影《大空头》（The Big Short）中的经典警醒——失业率每上升 1%，就会有数万人失去生命。他坦言：“如果我被证实是对的，我可能会感到一种智力上的‘满足’（satisfied），但我绝不会为此感到高兴（happy）。” Ed 强调，真正让他感到不寒而栗的是最终为泡沫买单的普通人。当前美股散户投资者的杠杆水平正处于历史最高位，大量散户正在盲目涌入这轮浪潮，类似于此前在散户热潮中遭受洗劫的局面即将重演。他的狂热与激动并非来自恶意，而是源于一种深刻的焦虑与无力感：“我们本可以踩刹车的，但现在已经停不下来了。” 在节目的尾声，Josh 和 Michael 感谢了 Ed 的坦诚直言。虽然频道受众中必然会有人反感这种极度看空的言论甚至留言“别再请这家伙来了”，但 Josh 认为这种反向视角的推演极具价值。如果未来真的发生系统性动荡，至少听众能够理解背后的运转机制。Ed 推荐了阅读其深度调查通讯（edzitron.com）以及收听其主持的科技播客《Better Offline》（betteroffline.com），并在轻松的互侃中结束了本期对谈。 - **散户杠杆风险**：散户投资者的杠杆使用率处于历史顶峰，泡沫破裂时的下行风险将直接波及普通家庭财务。 - **客座与信息源**：Ed Zitron 运营个人科技评论通讯并在各平台主持播客节目《Better Offline》。 ## Notable quotes > [18:42](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=1122s) 不，是我太蠢太天真，低估了这帮人能把事情推到多么极端的境地，也低估了整个债务融资体系居然能承受这种规模的挥霍。 — Ed Zitron > [22:32](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=1352s) 参与者根本不做基础审查，完全沉迷在‘只要数字一直涨、全球大公司不可能无缘无故花一万亿、我只要抢先建好租出去就能发横财’的盲目狂热与恶意无知中。 — Ed Zitron > [53:39](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3219s) 那根本就是个疯子一样的产品。让人无法理解的是，业内居然还有那么多人为之极力辩解，谎称‘幻觉问题已经被解决了’，这完全是睁眼说瞎话。 — Ed Zitron > [53:39](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3219s) 等等，你觉得这些产品很烂？我很喜欢用 Claude 啊！ — Michael Batnick > [53:39](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3219s) 当我在忙着写作时，帮我去复核这组数据，并在我落笔之前把原始信源列出来。 — Josh Brown > [55:16](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3316s) 为了获得稍微好一点的搜索和代码辅助，整个行业砸进去了数千亿美元乃至上万亿美元，这根本不值一提。 — Ed Zitron > [57:17](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=3437s) 你为什么停了？你前几年烧的那些钱到底弄出了什么？ — Ed Zitron > [01:11:06](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4266s) 下次你们邀请看多 AI 的嘉宾来节目时，请务必当面问他们一模一样的问题——如果你们错了该怎么办？ — Ed Zitron > [01:20:33](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4833s) 黄仁勋，你本来是个卖铲人，是要把芯片卖给数据中心，为什么现在你还要反过来花几百亿去租回这些数据中心？你是在自己建数据中心给自己送钱吗？ — Ed Zitron > [01:22:19](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4939s) 如果我被证实是对的，我可能会感到一种智力上的‘满足’，但我绝不会为此感到高兴。 — Ed Zitron > [01:22:19](https://www.youtube.com/watch?v=yoCkR0pn0ns&t=4939s) 我们本可以踩刹车的，但现在已经停不下来了。 — Ed Zitron ## Mentioned Ed Zitron, Josh Brown, Michael Batnick, Sam Altman, Elon Musk, Brian Chesky, Mark Zuckerberg, Prabhakar Raghavan, Ben Gomes, Andy Jassy, Satya Nadella, Sundar Pichai, Jensen Huang, Steven Ju, Nick Suresh, Cory Doctorow, Safra Catz, Jack Welch, Sam Koppelman, OpenAI, Anthropic, Nvidia, Microsoft, Google, Amazon, Meta, Oracle, CoreWeave, Cerebras, Blue Owl Capital, Blackstone, SoftBank, Apple, Airbnb, Salesforce, Bloomberg, AT&T, Nebius, Lambda, Iren, Nscale, Perplexity, Tesla, Waymo, Zoox, SK Hynix, Samsung, AMD, GE Capital, UBS, Barclays, Wells Fargo, Sightline Climate, Ramp, a16z, Fidelity Investments, Ghost, Better Offline, Where's Your Ed?, ChatGPT, GitHub Copilot, Claude, Gemini, The Death of Stalin, The Information, Bloomberg Terminal, Minecraft, The Big Short, Fidelity Trader+, Brick ## Video description On episode 257 of The Compound and Friends, ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Downtown Josh Brown⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ and Michael Batnick are joined ⁠Ed Zitron⁠ to discuss: the ultra-bear case for AI, Nvidia’s explosive growth, the economics of OpenAI and Anthropic, whether AI demand can justify the massive hyperscaler CapEx boom, the data center buildout, CoreWeave and the neoclouds, Oracle’s AI bet, private credit and debt financing, the warning signs that could finally break the AI spending cycle, the “rot economy,” whether AI is actually improving corporate productivity, and much more! This episode is presented by Fidelity Investments and the all-new Fidelity Trader+, Fidelity’s most powerful trading platform yet. Learn more at ⁠http://www.fidelity.com/TraderPlus ►00:00 - Cold Open ►10:00 - Intro ►11:22 - Nvidia ►19:03 - What is The Market getting wrong? ►34:15 - Too Big to Fail? ►41:53 - Meta ►01:00:21 - The Four Horsemen of the AI Apocalypse ►01:11:01 - Is there a positive outcome? 👕 Check out The Compound shop: https://www.idontshop.com 🎙️Listen to our podcasts: TCAF: https://podcasts.thecompoundnews.com/show/TCAF/ Animal Spirits: https://podcasts.thecompoundnews.com/show/animalspirits/ Ask The Compound: https://podcasts.thecompoundnews.com/show/ATC/ The Compound Newsletter: https://www.thecompoundnews.com/subscribe Linkedin: https://www.linkedin.com/company/the-compound-media/ Instagram: https://instagram.com/thecompoundnews Twitter: https://twitter.com/thecompoundnews Tik Tok: https://www.tiktok.com/@thecompoundnews Investing involves the risk of loss. This podcast is for informational purposes only and should not be or regarded as personalized investment advice or relied upon for investment decisions. Josh Brown and Michael Batnick are employees of Ritholtz Wealth Management and may maintain positions in the securities discussed in this video. All opinions expressed by them are solely their own opinion and do not reflect the opinion of Ritholtz Wealth Management. See our disclosures here: https://ritholtzwealth.com/podcast-youtube-disclosures/ The Compound Media, Incorporated, an affiliate of Ritholtz Wealth Management, receives payment from various entities for advertisements in affiliated podcasts, blogs and emails. Inclusion of such advertisements does not constitute or imply endorsement, sponsorship or recommendation thereof, or any affiliation therewith, by the Content Creator or by Ritholtz Wealth Management or any of its employees. For additional advertisement disclaimers see here https://ritholtzwealth.com/advertising-disclaimers. #thecompound #investing #stockmarket #stocks [Source](https://www.youtube.com/watch?v=yoCkR0pn0ns)

[18:42] Ed Zitron: 不，是我太蠢太天真，低估了这帮人能把事情推到多么极端的境地，也低估了整个债务融资体系居然能承受这种规模的挥霍。 [22:32] Ed Zitron: 参与者根本不做基础审查，完全沉迷在‘只要数字一直涨、全球大公司不可能无缘无故花一万亿、我只要抢先建好租出去就能发横财’的盲目狂热与恶意无知中。 [53:39] Ed Zitron: 那根本就是个疯子一样的产品。让人无法理解的是，业内居然还有那么多人为之极力辩解，谎称‘幻觉问题已经被解决了’，这完全是睁眼说瞎话。 [53:39] Michael Batnick: 等等，你觉得这些产品很烂？我很喜欢用 Claude 啊！ [53:39] Josh Brown: 当我在忙着写作时，帮我去复核这组数据，并在我落笔之前把原始信源列出来。 [55:16] Ed Zitron: 为了获得稍微好一点的搜索和代码辅助，整个行业砸进去了数千亿美元乃至上万亿美元，这根本不值一提。 [57:17] Ed Zitron: 你为什么停了？你前几年烧的那些钱到底弄出了什么？ [71:06] Ed Zitron: 下次你们邀请看多 AI 的嘉宾来节目时，请务必当面问他们一模一样的问题——如果你们错了该怎么办？ [80:33] Ed Zitron: 黄仁勋，你本来是个卖铲人，是要把芯片卖给数据中心，为什么现在你还要反过来花几百亿去租回这些数据中心？你是在自己建数据中心给自己送钱吗？ [82:19] Ed Zitron: 如果我被证实是对的，我可能会感到一种智力上的‘满足’，但我绝不会为此感到高兴。 [82:19] Ed Zitron: 我们本可以踩刹车的，但现在已经停不下来了。

[3:36] Ed Zitron: OpenAI is going to become bigger than Meta in 3 and a half years if you believe their projections... and they're going to do that while spending more than twice of Meta's OPEX. [18:20] Ed Zitron: I was wrong because I was naive about how the world works... I didn't think that the debt system would support it. [20:07] Ed Zitron: Run rate could mean four weeks times 12... They are annualizing token spend, which is not a recurring software subscription. A customer could spend a thousand this month and then spend 50 next month. [52:16] Downtown Josh Brown: I will never have AI write my words because I think the only reason people read me is because I have a very specific way of talking. [75:37] Ed Zitron: Between 2003 when it was created and 2015 when it became profitable, Amazon's total CapEx was $29.7 billion adjusted for inflation... This costs way more. That was all of Amazon's CapEx, not just AWS. [80:33] Ed Zitron: You sell to the data center, Jensen. Why are you also renting the GPUs back? Are you just building a data center to feed yourself money? This is eventually what they wanted to create—where the only thing was handing yourself money, but the money increased by draining debt.

[0:00] Like I ran a PR firm until like last [0:01] year and I was like doing that as kind [0:03] of contract labor thing and then I [0:06] started writing this in 2020 and I just [0:08] kept going and going and going and going [0:10] and then initially I was writing about [0:11] like management theory stuff and how [0:13] much I liked remote work and how evil [0:15] the articles were about it. Then I wrote [0:17] about crypto and I wrote about Elon and [0:19] by the end of those two things I was so [0:21] thoroughly depressed with both. I was [0:22] just like what am I going to write [0:24] about? Well, with the topics or with [0:25] writing? [0:25] >> With the topics. The topics were so [0:27] depress like every week following [0:29] whating Elon Musk does and just and then [0:33] the metaverse happened. That was kind of [0:34] fun to write about, but that was before [0:35] that. But then OpenAI got like um sorry, [0:39] Sam Orman got fired from OpenAI. And I [0:42] saw on Twitter that there were [0:43] journalists who were like, "Oh my god, I [0:45] hope they bring him back." They were [0:46] like talking about him like a rock star. [0:49] It was like, "Oh, then everyone's there [0:50] was someone like, "Oh, I'm crying [0:52] because how beautiful it is that Sam [0:54] that everyone's showing support for Sam [0:55] Orman." [0:56] >> Josh cried. [0:56] >> Journalists. [0:57] >> Yeah, journalist. [0:59] >> Crazy. So, um, I saw this and I just I [1:01] didn't even judge them for it. I was [1:02] like, "Fine, whatever. Feel." But I was [1:04] like, "This is weird. [1:05] >> Why are you acting?" Like I was like, [1:07] "What does this mean? Why are you acting [1:09] this way?" And so, I kept looking really [1:12] intently and I just immediately saw this [1:14] thing where everyone was saying, "Ah, [1:15] it's the biggest, most hugest, best [1:17] thing ever. It's inevitable." blah blah [1:18] blah blah, but I couldn't find any [1:20] revenues. Like I couldn't find and every [1:22] time I was like, "Okay, great. But how [1:24] much does it make?" And everyone was [1:25] like, "I I couldn't possibly say, "How [1:27] much does Microsoft? They're spending [1:28] billions of dollars on capex. How much [1:30] does this make?" Can't couldn't tell [1:32] you. No one wants to tell you, right? [1:33] >> But you kept having these articles every [1:35] earnings season saying Microsoft's AI [1:37] bet pays off. Google's AI bet pays off [1:39] even though they never mention the AI [1:40] revenue. And I just kind of taught [1:43] myself economics from there. Like a lot [1:45] of this like everything's selftaught [1:47] which to the checkin of my many haters [1:50] is just like stuff I learned. [1:52] >> Yeah. [1:52] >> But you can learn a lot by reading books [1:54] and such. [1:55] >> Okay. So you all right so your I guess [1:59] is it's a substack or [2:01] >> ghost. Okay. So what you're doing is [2:05] being read everywhere. people are [2:07] sharing it and bulls on AI are reading [2:10] it because I think [2:12] >> so my view of the market is most [2:14] professional investors are intelligent [2:16] enough to listen to both sides [2:18] >> and they don't have to agree with [2:20] everything they read in order to want to [2:22] keep reading it [2:22] >> right [2:23] >> I don't think that's the same is true [2:24] for retail investors but we're not going [2:26] to concern ourselves with that I think [2:29] professionals do like to hear both sides [2:31] especially if they have a lot of money [2:32] on the line [2:33] >> that's actually been my experience [2:34] talking to like some of the banks and [2:35] the institutions like they they want to [2:37] even the ones who are very pro. [2:38] >> I agree with that. [2:39] >> Yeah. [2:39] >> All right. We're the same way. So, we [2:41] we've been doing this show [2:43] >> uh we tape every Thursday going back to [2:46] the summer of 2021 and for 5 years [2:49] >> the majority of our guests are bullish [2:51] and so far so good for the most part. [2:54] >> Yeah. [2:55] >> Okay. Um we have had people who are [2:57] bearish but they are non-specifically [2:59] bearish. So, we've had people like [3:02] Jeremy Granthm who's become famous for [3:05] he he takes issue with the Perma Bear [3:07] label, but he's almost always bearish. [3:09] So, we've had people like that on the [3:10] show and um I always find that [3:13] fascinating, but we haven't really had [3:15] specifically an AI bear, [3:18] >> somebody that like spec, you know, most [3:20] of the people that come on portfolio [3:21] managers or they're analysts covering [3:24] the space [3:25] >> and you know all the incentives to see [3:28] the brighter side. So, this will be an [3:30] interesting test of all of the things [3:32] that we've done on the show. Yeah. To [3:33] hear the other side. So, we're really [3:34] excited about it. I'm pumped. Okay. [3:36] Awesome. [3:36] >> Do bulls think you're an idiot or do [3:38] they generally like like what you're [3:39] saying to them? [3:40] >> So, it's got to a weird point. So, if [3:42] we're mixing in every kind of bull here, [3:44] so I imagine a lot of them are retail [3:45] bags. [3:47] They are driven like cuz I used to be in [3:49] PR. They're like, "He's just a PR boy. [3:51] He's just a PR guy. He shouldn't talk [3:52] about stop talking to him. He could say [3:54] something." As if financial experts have [3:57] never got anything wrong before. [3:59] There are some of like what's weird is [4:00] like there are some of them who are [4:02] quite bullish on semiconductors like who [4:05] follow me on Twitter like bubble boy [4:06] who's really great and very very smart [4:08] on semis I'm not sure he agrees with me [4:10] on everything but very kind and friendly [4:12] and like there are lots of them who are [4:13] some of them are vile and horrible but [4:16] so say many such cases and so the thing [4:20] is with the bulls is the ones who are [4:23] like I love this tech I have some crazy [4:26] dream what it'll [4:28] But there's a bubble. I respect those. [4:30] If you can like love a I still think [4:32] there are massive environmental and [4:33] social problems that they're not [4:35] thinking of, but at the very least [4:37] you're living in reality. I think the [4:39] ones who are like anthropic and open AI [4:41] are going to grow to $284 billion in [4:44] revenue. That's what OpenAI is [4:45] projecting in 2030. The people who read [4:47] that and they're like, "Yeah, sounds [4:48] good to me." Those people are not living [4:50] in reality. And I don't need a degree of [4:52] >> the number or [4:54] >> OpenAI is going to become the going to [4:56] become bigger than Meta in 3 and a half [4:58] years if you believe their projections. [4:59] And they're going to do that while [5:01] spending more than twice of Meta's OPEX. [5:04] I think I think Microsoft's OPEX is like [5:06] 150 billion and um OpenAI is going to [5:09] spend 200 something billion or more on [5:11] compute in 2030 [5:12] >> to get to 380 billion in [5:15] >> 284. 280 billion. [5:16] >> So are you outright bearish or you think [5:18] the bulls are smoking crack or am I [5:19] saying the same thing? uh possibly the [5:21] same thing. I think that [5:23] after 2022 when everything got really [5:26] grim in tech, I'm don't know if you [5:28] remember like Nvidia had a flat year, [5:30] fiscal 23 course, [5:31] >> really rough time and everyone was kind [5:33] of blowing hard. Microsoft, Google, [5:35] Amazon, Meta, cranked up prices, changed [5:37] ad auction stuff. Amazon actually made a [5:39] surprisingly healthy ad business. Like [5:41] they had to find ways to grow revenue [5:43] and also chat GPT happened. So they [5:44] went, "Oh, we'll buy a bunch of GPUs." [5:46] And the market immediately was like all [5:48] of this revenue growth is coming from [5:49] AI. And because the companies never [5:52] disclosed their AI revenues and [5:53] journalists don't have the teeth to [5:55] bother them or analysts especially, [5:57] people were like, "Oh, it's AI. Every [5:59] time they spend this money, it's AI." [6:00] And this was really good. It helped [6:01] their stocks pop crazy style. Like the [6:03] fair one fair criticism of my work is if [6:06] you traded off of my work, which I've [6:08] never said to do, I do not give [6:09] financial advice. Yeah, you probably [6:12] >> Well, but if you end up being right [6:13] about this, they're going to put a gun [6:15] to your head and say start a hedge fund. [6:16] >> I mean, oh Christ, I don't even know how [6:18] to do that. But [6:19] >> I'm just telling you that's [6:20] >> That sounds fine. But the point is I'm I [6:23] like writing a lot. I really enjoy [6:24] actually getting into it's genuinely [6:26] intellectually fascinating. But the [6:28] point I'm making is up until about 2025, [6:30] it was mostly just buying GPUs and [6:32] hoping OpenAI and Anthropic grew into [6:34] them and also hoping they become [6:35] profitable. But then what happened is [6:37] the hyperscalers have now become [6:40] financially dependent on the growth from [6:42] open AAI and anthropic. Analyst [6:44] expectations uh from like UBS, Barkclays [6:46] and uh Wells Fargo have $440 billion of [6:50] cloud revenue across Google, Amazon and [6:52] Microsoft coming just from open air and [6:54] anthropic to unprofitable startups who [6:57] need to constantly raise money and also [7:00] the big companies need to build the data [7:01] centers to make the money. So there's [7:03] all of these very improbable, if not [7:05] impossible things that need to occur. [7:07] And also, OpenAI and Anthropic need to [7:09] have 10 times the demand they have right [7:11] now. It's not even, however you may feel [7:13] about these companies, they are not big [7:15] enough. They are not generating enough [7:17] cash right now to even get close to [7:20] covering their 1.1 or more trillion [7:22] dollars in commitments. And on top of [7:24] that, they're taking up 90% of AI [7:27] infrastructure. This is creating an [7:29] illusory demand signal because right now [7:31] it's difficult to get GPUs. People are [7:33] saying, "Oh, oh, it's because there's so [7:35] much demand for AI." There's a big Barry [7:37] Bonds asterisk at the top of that. [7:39] There's so much demand from AI from two [7:42] companies pretty much meta as well, but [7:44] they're not doing anything. They're just [7:45] they're just rolling in their filth. But [7:48] so what's happening is everyone's going, [7:49] "Oh, there's tons of demand for AI. I'm [7:51] going to build a bunch of data centers." [7:53] Even though when you so in fiscal year [7:55] 26 Microsoft they made corn to Bloomberg [7:58] about $ 34.33 billion on AI $24.1 [8:01] billion of that is open AI which means [8:04] that open uh Microsoft software and GPU [8:07] rental services are singledigit billion [8:09] dollar businesses and they have 260 plus [8:12] billion in capex in this. It's a [8:15] disaster and it's not something that [8:17] gets easily I don't I don't think it [8:19] gets fixed at all. On top of it, AIG [8:22] GPUs are pretty specialist. They're [8:24] useful for like AI and data analytics [8:26] and 3D modeling, which is a very small [8:28] business. There's no com bubble thing [8:29] after this. The electricity will be just [8:31] as expensive, if not more. [8:33] >> Oh, wait. You're on a roll. Um, I want [8:35] to give people the chance to close out [8:37] their long positions and we're going to [8:39] introduce the show properly and then [8:40] we're going to dive right back in. Who's [8:43] uh who's doing my countdown? Miss [8:45] Nicole. All right. [9:01] This message is brought to you by [9:02] Fidelity Investments. When timing is [9:04] everything, you need powerful tools and [9:06] research that can meet you in the [9:07] moment. With the allnew Fidelity Trader [9:09] Plus platform, your charts and [9:11] preferences show up consistently, synced [9:13] up across all your devices, so you can [9:15] act fast whenever and wherever you're [9:17] trading. [9:18] You can save an order on your desktop at [9:20] home, get a mobile alert when you're at [9:23] work, and complete the trade in the [9:24] Fidelity app without starting over. And [9:27] with the downloadable Fidelity Trader [9:29] Plus desktop platform, you have more [9:32] control with multimonitor views, [9:34] enhanced tools, and customization [9:36] options, and integrated screen sharing. [9:39] With Fidelity Trading Specialists, try [9:41] Fidelity's most powerful trading [9:43] platform yet at fidelity.com/trader. [9:46] US Fidelity Investments and the Compound [9:48] are not affiliated. Views, opinions, [9:50] products, services, and strategies [9:51] discussed are not endorsed or promoted [9:53] by Fidelity Investments. Fidelity [9:55] Brokerage Services LLC member NYC. [10:01] >> All right, ladies and gentlemen, welcome [10:03] to the world's greatest investing [10:04] podcast. I say that without a trace of [10:07] irony. I really mean it. [10:10] Today, you're in for a very special [10:11] treat. We are having a conversation that [10:14] we have not really had here before. What [10:16] is the ultra bare case on AI? Why are [10:20] people clapping? It's crazy. So, [10:22] retirement on the line. What's going on, [10:26] guys? Compound. All right. Uh, joining [10:29] us today, Ed Zitron. Ed is the founder [10:32] and CEO of EasyPR, a research and media [10:35] analysis firm he launched in 2013. [10:40] He also hosts the Webbby award-winning [10:42] better offline podcast with iHeart Radio [10:45] and Cool Zone Media. Ed writes a [10:47] newsletter called Where's Your Ed? that [10:50] he launched in 2020 where his 2023 Rock [10:54] Economy essay, I'm going to get into [10:56] that, arguing tech firms have chased [10:59] growth at the expense of innovation and [11:02] users has gone viral. Today he's one of [11:05] the loudest and most cited AI skeptics [11:09] focused on the unit economics of open AI [11:12] anthropic and the hyperscaler capex [11:15] buildout. Ed, thank you for being here. [11:17] We appreciate it. [11:18] >> Thanks for having me. [11:19] >> All right. [11:21] >> All right. So, we have to start with [11:24] Nvidia. [11:24] >> Yeah. [11:26] >> I feel like they're they're proving [11:27] everything that you've said to be not [11:29] correct. [11:30] >> I know, but just [11:31] >> No, no, no, no. Hear me. Hear me out. [11:34] They just they just came out here. Here [11:36] here's what they did last night. [11:37] >> Um and Michael will jump into the [11:39] numbers and then we're going to give you [11:40] a chance to comment on it. [11:42] >> But they just said for the first time [11:43] ever we're actually going to give a full [11:45] year's worth of guidance. Um they were [11:47] going quarter by quarter. They actually [11:49] gave guidance through 2028, fiscal 2028, [11:52] which is not calendar 28, but um [11:53] >> it's February February 2027 off in the [11:55] woods. [11:56] >> So they're saying like we're going to [11:58] grow revenue by about 70%. [12:01] We've got the sales locked in for GPUs [12:04] and uh Vera CPUs and etc etc. We know [12:08] that like 90% of their business is [12:10] selling to data centers. A huge chunk of [12:12] that is selling to the same four or five [12:14] customers, the hyperscalers. They are [12:16] talking about broadening that out. [12:18] They're talking more about now uh [12:20] automation, robotics, but the data [12:22] center business is the business. Okay. [12:24] So, [12:26] is he going to be completely wrong in [12:28] his own forecast? Or will they make [12:31] their number, but that's part of the [12:33] problem because of who the buyers are [12:36] and what their motivations are? Like, [12:37] what? Let's give people the nuance take [12:40] on exactly what you're saying because [12:42] Nvidia really is at the heart of the [12:45] entirety of the AI story. [12:46] >> I just want to be clear as well. Nvidia [12:48] made a bunch of money. Like, that's I'm [12:49] not questioning that. That'd be crazy. [12:50] >> Of course. However, 16% of their latest [12:53] quarter revenue was one customer. 44% of [12:56] their first half of fiscal 2027 was [12:59] three customers. Five customers make up [13:02] 70% of their accounts. [13:03] >> Can I jump in right there? [13:04] >> They're good customers. [13:05] >> Well, can I jump in right there? [13:06] >> Why? Because they don't name them. [13:08] >> Well, well, we know who they are, [13:09] >> do we? [13:10] >> Well, we might not know who's 16 and [13:12] who's 12, but we know the [13:14] >> But if one of them is, [13:17] >> but hear me out. [13:19] That's their immediate customer, but [13:22] those customers have millions of [13:24] customers. And that's what this in other [13:27] words, [13:27] >> you don't have a hyperscaler spending on [13:29] Nvidia chips to serve itself. Except in [13:32] the case of Meta, [13:34] >> right? They're serving Fortune 500 [13:36] businesses, governments, sovereigns, uh, [13:39] etc. That those are their customers. [13:41] >> But I just the point I made about [13:42] Microsoft's AI revenue is where this [13:44] comes in. So fiscal 26 which just ended [13:47] $ 34.33 billion of AI revenue $24.1 [13:50] billion of that is open AI that means [13:53] that selling GPUs to anyone else other [13:56] than their large fail son is a [13:58] singledigit billion business it's [14:00] Microsoft is the apex predator of [14:02] software sales they have hundreds of [14:04] thousands of resellers they have tens of [14:05] thousands of salespeople and they can't [14:07] scrape together more than singledigit [14:09] billion dollars of sales for co-pilot [14:11] and all the stuff and this is on top of [14:13] the fact that GitHub copilot But one of [14:15] their only successful AI products has [14:17] now gone tokenbased billing, which means [14:19] people were before able to spend 40 [14:20] bucks a month, spend $5,000 of tokens. [14:23] Now the token based billing, [14:24] >> now it's usage gone. That business is [14:26] dead. [14:27] >> And on top of all of that, wow, $260 [14:29] billion in capex and all of that Fortune [14:31] 500 demand drew on like whatever it may [14:34] be is single is what $10 billion a year. [14:36] >> Adela would say, G, give us a couple of [14:38] years. Why do I have to give sorry I'm [14:40] but the point is is like we've given him [14:43] a few years and this is the best he's [14:45] got for us. We have a li open AI is a [14:48] liability now like that's they are [14:50] material enough that Microsoft actually [14:52] has to I think what's like 27% they own [14:54] of them it's actually a deeply low [14:56] bearing [14:57] >> they have to consolidate their [14:58] financials onto their their own income [15:00] statement. The other thing is is [15:01] Microsoft's only getting paid if OpenAI [15:04] can keep raising money. And OpenAI it [15:06] has a shrinking amount of people to [15:08] raise it from. Their last round 122 [15:10] billion only 12 billion of that came [15:13] from venture capital and the very same [15:15] asset managers who were funding data [15:17] centers. 110 billion was Nvidia 30, Soft [15:21] Bank 30 which put them into fire [15:22] financial direates and they're still [15:24] there and uh Amazon 50 billion. That's [15:28] not going to happen multiple times or if [15:30] it if Nvidia invests again, they said [15:32] they wouldn't. They literally said the [15:33] words, "If they invest again, it's bad [15:35] times ahead." So, Microsoft's growth is [15:38] dependent on OpenAI and now anthropic [15:40] spending. Outside of that, they don't [15:42] have significant demand. This isn't even [15:45] an opinion. This is a Bloomberg article [15:47] that spelled this out. And if you go and [15:49] look at [15:49] >> Was that a Bloomberg opinion? [15:50] >> It No, it was it was Brody Ford, I [15:52] think. But it was and it was also based [15:54] on the actual actual earnings documents. [15:56] Microsoft doesn't have the demand for [15:58] its own AI native products from its own [16:00] customers [16:01] >> at the scale that they would need to [16:02] justify like I'm not saying there's zero [16:04] demand they have found but this is also [16:06] after them harassing their customers and [16:08] massive sale I must be clear Microsoft [16:11] sales teams are there's only one level [16:13] higher of more annoying and that's [16:15] Oracle and Oracle is a whole other [16:17] problem but the point is is if this was [16:20] a case where [16:22] 30% of their revenue was open AI I would [16:24] be like all right maybe my kind of [16:26] falling But 70% of their AI revenue [16:28] being open AI is existentially bad. They [16:31] have $250 billion of Azure spend [16:34] committed to that's so much that's [16:36] Microsoft was uh 331 billion I think [16:39] last fiscal year [16:40] >> but now they've divorced themselves from [16:43] the exclusivity agreement with open AAI [16:45] >> which only makes it worse because now [16:47] OpenAI needs to spend 138 billion over [16:50] eight years with Amazon 20 billion with [16:51] Cabus 22.4 4 billion with core. We have [16:54] 300 billion plus with Oracle. U they're [16:57] spending, this is the greatest thing, [16:58] Steven Ju from uh UBS, he had an analyst [17:00] note, they're spending like 12 a.5 [17:02] billion a year at Google. I didn't know [17:04] this until very recently. It's weird [17:07] that it's not being reported very well, [17:08] but the thing is is OpenAI Sam Alman has [17:11] one real talent and that's signing his [17:13] name. My man loves signing contracts. [17:15] And the thing is when the problem that [17:17] everyone has right now is you can only [17:19] really think a quarter two quarters in [17:21] the future. You can't because even like [17:23] oh it's going to happen in 2028. Nvidia [17:25] will get to to 600 something 674 billion [17:29] I guess it would be based on consensus [17:31] for fiscal 28. So Nvidia will get there [17:34] because right now Nvidia makes lots of [17:36] money. It's not a problem. Now to get [17:38] there Nvidia is going to have to [17:40] basically take three to five customers. [17:43] They're going to have to find way more [17:45] debt. Like way more because the price of [17:48] debt is increasing. But also, Nvidia [17:50] just bumped up prices 17%. The cost of [17:52] memory skyrocketing. And on top of that, [17:54] you have a shortage of talent, shortage [17:56] of electrical grade steel. There are all [17:59] sorts of things stacked against Nvidia [18:01] here. They could pull it off, but I [18:03] don't know how they're going. The the [18:05] actual money that needs to get raised is [18:08] astronomical. But they so Daniel chart [18:10] five Ed if if somebody said to you two [18:14] years ago that this is going to happen [18:16] you would be saying the same things that [18:17] you are right now about how implausible [18:19] it would be. So what we're looking at [18:20] for people that are listening is the [18:22] trailing 12 month net income for Nvidia [18:25] just screamed past Apple. [18:27] >> Uhhuh. But how much of that is equity [18:29] gains though? [18:30] >> What does that mean? as in their [18:32] operating anthropic, Anthropic, Open [18:34] AAI, uh they've invested in like [18:36] Coreweave, Nebius, Iron, they have a [18:39] bunch of equity investments that help [18:40] bumping that up. To be clear, I was [18:42] wrong back in 2024. And why? Cuz I was [18:44] naive. I was like, the market's [18:46] a sensible place. Microsoft and Google, [18:48] Amazon, Meta wouldn't spend hundreds of [18:49] billions of dollars for no goddamn [18:50] reason. [18:50] >> Wait, so you were wrong cuz the market's [18:52] stupid? [18:52] >> I was wrong because I was stupid. I was [18:54] wrong cuz I was naive about how the [18:56] world works and I've had to take a hard [18:58] >> You didn't think they would take it as [18:59] far as they [19:00] >> Exactly. and I didn't think that the [19:01] debt system would support it. So yeah, I [19:03] absolutely [19:04] >> So, so what do you think the market is [19:05] getting wrong right now? Because I think [19:06] I think everyone knows that OpenAI is a [19:08] potential problem. Look at Oracle stock. [19:10] It's down 70% still on the mat even as [19:13] as uh software stocks are bouncing. [19:15] Yeah, Oracle has caught no bounce. So [19:17] what do you think the market has [19:18] misunderstanding that you're that you're [19:20] seeing? [19:20] >> The thing I said about the [19:21] centralization of data center demand, [19:23] open AI, I think is like 60% of all AI [19:27] data center demand. Without OpenAI, [19:29] there is not another OpenAI sized [19:31] spender. And the only reason there is an [19:33] anthropic or OpenAI sized spender on AI [19:36] compute is because of the availability [19:37] of venture capital dollars. So without [19:40] that, you can't like the demand isn't [19:43] there. The market isn't seeing that [19:46] because it hasn't happened yet. the [19:48] quarter two quarters think they're doing [19:50] where they're like well the money is [19:51] still coming in because what I kind of [19:53] said this already but the amount that [19:55] Anthropic and Open AI has to spend has [19:57] to increase by like 10x [19:59] >> well their revenues are increasing and [20:01] there those are real revenue dollars and [20:04] they're both going public in the next [20:06] call it 6 months [20:07] >> wait you just made a face anthropic [20:09] doing 60 billion is that not [20:11] >> run rate go rate it that's this is the [20:14] biggest scam of them all what a rate [20:17] goes to less quarter. That's the first [20:18] part of the scam. So, ARR used to mean [20:21] annual recurring revenue referring to if [20:23] I have 10 100 million contact contracts, [20:26] I've I've got like a billion dollars a [20:28] year. Easy peasy. That's annual [20:30] recurring revenue. Wow, how amazing. [20:32] >> That's SAS. [20:33] >> SAS. I probably up the math [20:34] there, I realized, but whatever. Moving [20:35] on. Um, [20:36] >> and now they're booking their revenue [20:37] probably don't like. [20:38] >> Well, now they're calling it run rate. [20:39] And run rate could have Anthropic has [20:41] never defined this. Even Bloomberg in [20:43] the 65 billion story didn't define it. [20:45] It can mean four weeks x 12 or 4 weeks x [20:48] 13. The problem is is anthropic and [20:51] openai are both annualizing token spend [20:54] which is not a recur. It's not a [20:56] software subscription. It's not [20:57] recurring. A customer could spend a [20:59] thousand this month and then not spend [21:01] then spend 50 next month especially if [21:03] they're moving to open source models. [21:04] >> Sorry that's 100% true. But every [21:09] publicly traded corporation that says [21:11] anything to Wall Street about its spend [21:14] >> Mhm. [21:15] >> the only thing you ever hear is we [21:17] already ran out of our comput spend and [21:19] need to spend more. [21:20] >> Yeah. That's cuz anthropic and open AI [21:22] are taking up most of the compute [21:24] infrastructure so there's not enough. [21:26] There is more demand than there is right [21:27] now. The question is how much more [21:30] because I don't think that I mean I kind [21:33] of estimated in an est in a newsletter [21:35] week ago I think there's about $22 [21:37] billion of compute demand. Sighteline [21:40] climate said back in February that there [21:42] was 190 gawatt in planning. That's about [21:46] $1.5 to3 trillion a year in compute [21:48] demand they'd need. There is a complete [21:51] economic mismatch and it's [21:54] stark. You don't but you don't like ARR [21:56] run rate because it's just a [21:58] extrapolation [21:59] >> of a it's a snapshot of a period in time [22:02] they plug it out to like a year. [22:04] >> Yeah. So it's we also don't know what [22:05] the period we don't even know what it [22:06] means. They never define it. Anthropic [22:08] has used it several times in their own [22:09] announcements. They never define it. It [22:11] could refer to hey we just released a [22:13] model and everyone's trying it. Hey we [22:15] just did a new feature and everyone's [22:17] buying a new subscription. Oh, on this [22:19] day we lost a bunch of subscriptions, so [22:21] we'll pull the period over here so we [22:23] look bigger. It is not a trustworthy [22:27] measure of a business and the fact it [22:29] gets accepted is an insult to investors [22:31] intelligence. [22:32] >> So, who's the idiot here? Is it the CEOs [22:34] and the boards for all of this capex? Is [22:37] it the equity investors? Is it the [22:39] people loaning the money to build these [22:40] data centers? Like, who is not seeing [22:43] past their nose right now? [22:44] >> Lenders. Uh all of the investment in AI [22:47] data centers terrifies me to my core. [22:49] It's so bad. These things are like [22:52] miniature cities. They condense city of [22:54] Bristol in England using as an example [22:56] about 1.25 billion square ft takes about [22:59] 800 megawatts of power. Stargate Abene [23:02] Open AAI data center 1.2 gawatt of power [23:05] in 998,000 ft. So they're condensing a [23:08] city's power into a thousandth the [23:11] space. And that thing is way behind [23:12] schedule. Each one of these, even the [23:15] smaller ones are like, [23:16] >> is that the Oracle one in Texas? [23:17] >> Yeah, one one of the is one of the ones [23:19] in Texas and um all of these data [23:22] centers are very ambitious and the [23:24] bigger they are, the more ridiculous [23:25] they are. Gigawatt data centers are a [23:27] new idea. Also, a gigawatt data center [23:29] is a campus of smaller buildings, just [23:30] to be clear. But they're new. They're [23:32] brand new and they're trying to build so [23:34] many of them in a power grid kind of put [23:36] together with like staples and tape and [23:38] such, which is I mean, and the thing [23:39] they always say is, "Oh, it's helping us [23:41] upgrade the power grid." No, it's not. [23:42] It's strategically placing power in a [23:44] place that you can maybe rent it to [23:46] someone. And the problem is is that for [23:49] everything to be [23:50] >> But is it nefarious or do they just not [23:53] understand the demand as well as you do? [23:55] >> Little little here's the thing [23:57] >> because it seems like they're the ones [23:58] taking the risk. [23:59] >> It's nefariously ignorant. I think that [24:01] their ignorance is so harmful. Stargate [24:04] Abene the information had a report about [24:05] this that Blue Owl agreed to invest in [24:08] 10 minutes. [24:10] That's how that's how the due diligence [24:11] is going, boys. We're just 10 minutes. [24:14] >> Come on. [24:15] >> It's just like and that is for a project [24:17] that will cost tens of billions of [24:19] dollars. [24:20] >> And it's just Why is no one putting the [24:22] time in? [24:23] >> Wait, hold on. There's no way that's [24:24] true. [24:24] >> It's I will I will glad it was 15 [24:27] minutes, but it was maybe it was 15. No, [24:29] no, I can get I can get you the article [24:31] right now if you'd like. [24:32] >> I believe that the article is real, but [24:34] I just don't believe that that is [24:35] actually real. [24:36] >> Have you read anything about Blue Owl? I [24:37] don't believe it in 2 seconds man like [24:40] I'm sorry like having seen what blue owl [24:42] does I yeah I I would believe that and [24:45] also this is a mania and number go up [24:48] hey the biggest the justification makes [24:51] sense if you don't think for a second [24:52] the biggest companies in the world [24:54] wouldn't spend a trillion dollars for no [24:56] reason I got to build one of these [24:57] things I will make so much money it's [25:00] it's gambler logic [25:01] >> let's go through let's go through some [25:03] of these numbers uh Daniel you want to [25:05] do chart one So, back on Nvidia just [25:08] because this is like well I call Jensen [25:11] Wang the in the uh the chairman of the [25:14] AI Federal Reserve basically basically [25:17] he's the person that comes out and makes [25:19] everybody feel better [25:20] >> pretty much every quarter we get the [25:22] hyperscaler earnings [25:24] >> and then there's a 3-week lag [25:26] >> and then Nvidia comes out at the end of [25:27] the earning season and ratifies [25:30] everything that we Okay so just like on [25:34] its face 96 billion in revenue. We know [25:37] that the cost of a data center is [25:39] approximately 50% on chips. [25:41] >> Yeah. More now thanks to memory, but [25:43] yeah. [25:44] >> Okay. That'sund That's 106% revenue [25:47] growth at 75% gross margins. [25:50] >> Yeah. [25:51] >> Um Nvidia right now, I mean, this is [25:53] real money. [25:54] >> Yeah. [25:54] >> So, your issue is not whether or not the [25:56] spending is real. Your issue is the [25:59] source of where the capital is coming [26:01] from to buy from Nvidia is really what's [26:04] in question. [26:05] >> Yes. And just to be clear, the source of [26:07] the revenue and the money into all the [26:08] this stuff real the the money going into [26:11] AI software, which this is meant to prop [26:14] up very small when you remove the [26:16] subsidized stuff. Every AI startup is [26:19] subsidized. Every single one loses [26:21] money. Like uh up until fairly recently, [26:23] OpenAI and Anthropic let you burn [26:25] thousands of dollars of tokens for 200 [26:27] bucks a month. I think you still can, [26:28] but not as much. These companies when [26:31] their customers are exposed to the real [26:33] costs they shrivele away and just very [26:35] simply if they thought they could charge [26:37] the real cost of their services they [26:39] charge the real cost of their services. [26:40] They wouldn't d this weird dance. So the [26:43] problem isn't finding well I mean [26:45] actually it is becoming one now. This a [26:47] lot of this was I think that that the um [26:50] the one big customer the 16% one I think [26:52] it was SpaceX. So remember SpaceX, bunch [26:55] of debt. All of this bunch of debt. [26:57] Other than Microsoft at this point, [26:58] they're all having to take on debt. And [27:00] this is the thing. Everything it it's [27:03] not actually about real demand because [27:05] they can't pay this out of cash flow. [27:06] They're clearly not making enough money [27:07] to pay this out of cash flow. And they [27:09] would say, "Oh, it's the new industrial [27:10] revolution or some such wank." Well, [27:13] it's clearly not going very well. Oh, it [27:15] will work out in the future. When [27:17] where's the real demand? Where is it? [27:19] And so they're incre and because [27:21] everything's getting more expensive, [27:23] both the debt and the things the debt's [27:24] buying. They're going to need more and [27:25] more and more money at a time when the [27:27] system's kind of straining itself, but [27:29] the only thing the system wants to [27:31] invest in is the dumbest thing possible. [27:33] It it's actually more spelled out the [27:35] worry against. [27:36] >> When did Enthropic first turn their [27:37] revenue? I think it was March 2023. [27:39] >> What when was when did they what [27:41] >> when did they first book their dollar [27:43] first dollar of revenue for Enthropic? [27:44] It was sometime in 2023. [27:45] >> Yeah, maybe. I they they founded in [27:48] 2021, but I can't remember when they [27:49] actually. [27:50] >> So, let's But you you say that like [27:51] you're sick and tired of of saying where [27:54] when it's going to come. You don't you [27:56] don't have any uh open minds in this to [27:58] the fact that it genuinely is early. How [28:01] much of the demand that we're going to [28:03] see are we seeing now? Because I think [28:04] it's a tiny tiny fraction. [28:06] >> Well, will people all of a sudden stop [28:08] spending on this customers? So there's a [28:10] guy called Nick Sesh who has a really [28:11] good blog called uh AI is eviscerating [28:13] global decisionmaking and he's talked to [28:15] he's a tech consultant gifted software [28:17] engineer talks to a lot of CEOs. A lot [28:19] of this is you ever see the death of [28:21] Stalin? [28:22] >> Okay so a lot of people doing make [28:24] >> on Netflix. [28:25] >> No well maybe actually um but [28:28] nevertheless there's a bit at the [28:29] beginning where they're sitting around [28:30] Stalin's corpse and going he's so he [28:31] looks great he looks very healthy. Oh [28:33] yes. A lot of this is people saying, [28:36] "Oh, AI is making me 100x more pro um [28:39] productive or whatever." It's not true. [28:40] Most AI integrations fail. [28:42] >> So, all right. So, I want to give you a [28:44] counter to that though because [28:46] >> I listen to [28:47] >> I'm going to say 20 calls every earning [28:50] season and usually the stocks I own, but [28:52] sometimes just stocks that make a big [28:54] move and I just want to learn like why [28:56] >> why all of a sudden are people really [28:58] bullish or really bearish. So, one of [29:00] the stocks, and I happen to been talking [29:01] about this today, so it's fresh in my [29:03] mind, is Airbnb. [29:05] >> Uhuh. [29:05] >> And Chesy comes out and says, "AI is the [29:07] best thing that ever happened to us." [29:09] And then proceeds to site spec specific [29:12] examples. One of which is that last [29:15] quarter [29:16] >> 45% of all their customer support calls [29:18] were closed by AI with no human [29:21] involvement whatsoever. [29:22] >> Okay. [29:23] >> Um, it boosted cash flows. It reduced [29:26] expenses. [29:26] >> Did they say by how much? [29:28] >> They gave the numbers. Yeah, [29:30] >> that's good. [29:30] >> So, [29:31] >> they also moved to open source. [29:33] >> They very Chesy very specifically [29:35] >> not paying not paying the open [29:37] >> one of the early ones. [29:38] >> So, for me, that's the thing that I'm [29:39] most bearish about, but and we'll get [29:41] into that a second. I'm not So, here's [29:43] what I here's where I don't agree with [29:44] you. [29:45] >> Yeah. [29:45] >> I really don't think this is going to [29:47] stop on a dime as an AI [29:49] >> user [29:51] at a company. Every week we're finding [29:54] new things that we can plug into uh AI [29:58] and do better than we did the week [29:59] before. [30:00] >> I'm curious. [30:01] >> We have to we have to manage uh 90 some [30:04] odd employees, 4,000 client households. [30:07] >> There's a constant um orchestration of [30:11] operations activity that has to take [30:12] place between Salesforce and our [30:15] portfolio accounting software and the [30:16] trading that we're doing. And the [30:18] general idea from our standpoint is the [30:21] customers don't give a about our AI [30:23] use. This is operations. How do we run [30:26] the firm better, right? How do we do [30:28] things faster, more efficiently? And uh [30:31] that's what I think every company in [30:34] America right now is trying to do. And [30:36] we are increasingly hearing from them [30:38] one by one on Q&As with analysts saying, [30:42] "Why did we have an upside surprise?" [30:44] Well, glad you asked. We invested in AI [30:47] probably in 23 or 24 [30:50] >> and now all of a sudden we're able to [30:51] tell you specific things that we're [30:53] doing with AI that are improving our [30:55] earnings. I don't think that part's [30:57] going to stop. [30:58] >> So I think it will stop happening as [31:00] much because I think a lot of it is make [31:02] believe. AT&T being the worst one. I [31:04] don't [31:04] >> if they're lying then right [31:06] >> with AT&T if here's the clever thing. So [31:08] they had this article in the journal [31:09] saying 80% 90% uh savings on using open [31:13] source. Then you go down the article, it [31:15] says 80 to 90% savings in some fe some [31:19] functions. And then the next paragraph [31:21] is AT&T uses has thousands of things it [31:24] has plugged into AI. So it's like oh so [31:26] some things in some things may have [31:28] saved you some money. This is the [31:29] problem. A lot of this the reason I [31:32] brought up Nick Suresh's article is [31:33] because he makes a point where a lot of [31:36] people are doing this because they're [31:37] afraid of how it will look if they don't [31:39] or indeed the management is so cooked [31:41] that they must see AI because it's the [31:43] thing they're agree with that. I [31:44] definitely agree that there's a lot of [31:45] that going on. [31:46] >> And here's the thing on there is a point [31:48] we can come to which is even with that [31:50] being the case is that going to be the [31:52] revenue sufficient to keep open and [31:54] anthropic alive who are both [31:55] unprofitable or is that going to move on [31:57] to I don't know a hosted GPU system with [32:01] an open- source model. There are so many [32:03] way like anthropic and open AI must keep [32:05] growing and they must keep growing so [32:07] much faster. This is not even close to [32:10] where they need to be. they need to be [32:11] able to afford $440 billion in the next [32:14] three and a half years. [32:15] >> So, a lot of your bare case um rests on [32:18] that idea that the two most well-known [32:21] foundation models um are being relied [32:24] upon by the S&P 10 to deliver their no. [32:28] Okay, I I don't dispute that. [32:31] >> Um but what about Gemini then? What do [32:33] you do with the Alphabet Gemini [32:35] situation where they do not they do not [32:38] require Open AI or or Anthropic to spend [32:41] with them? [32:42] >> I think Sund Pashai is asking that [32:43] question based on all the press I've [32:44] seen around Gemini. Gemini is just [32:46] another LLM company. The problem is is [32:48] that as much as people love to pretend, [32:51] oh, OpenAI, Anthropic Special, maybe [32:53] they like Soul, whatever. For the most [32:55] part, people don't really give a [32:56] which model they use. And each model, [32:58] the differentiation between them is kind [32:59] of, uh, why do I need this? And then [33:01] there's a new better one. You try this [33:02] and every time you move it, there's a [33:04] bunch of harness changes and prompt [33:05] changes. The actual functionality of [33:07] this stuff is so malleable and volatile [33:10] that I think it's exhausting to engage [33:12] with on the regular. And actually [33:14] talking to pro AAI people, the few that [33:15] I talk to, they experience this. There's [33:17] an exhaustion around it. You constantly [33:19] have to keep up and they kind of [33:20] rationalize it like any bad relation [33:23] like you're saying um OpenAI releases [33:25] the newest model. [33:26] >> I must integrate this now. [33:28] >> Yeah. Yeah. And I must and especially if [33:30] they run a company that has models [33:31] facing the customer, they got to have [33:33] that immediately which bumps up the cost [33:35] because the thing is the cost of the [33:36] tokens may stay around the same but the [33:39] amount of tokens being burned changes [33:41] arbitrarily between models. So if a [33:43] model's the same price, but it's burning [33:45] more tokens. It's like having a car that [33:47] miles per gallon is the same, but you're [33:49] driving 200 extra miles. So my thing is [33:53] I could see a world where there are LLMs [33:55] in the future, but I don't know how [33:57] profitable it is to well I don't know if [33:59] it's possible to make these things [34:00] profitable even with custom silicon. I [34:02] don't know if there's a future for these [34:04] companies at all or indeed whether [34:06] hyperscalers will still offer these [34:08] services at the current prices. They may [34:10] offer them as a really expensive thing [34:11] for certain businesses. They may get [34:13] that. That might happen. I think that's [34:14] realistic. Gemini, for example, could [34:16] become that. [34:17] >> This might make you more bearish. Okay, [34:19] >> there is too much these things are too [34:21] big to fail. There is too much riding on [34:24] the success of maybe not just these two [34:26] companies, but this whole idea not [34:28] imploding and bearish investors have [34:30] learned over and over and over again. [34:32] The system is rigged to work out and to [34:35] go up. Now, that might make you more [34:37] bearish, but I'm curious to hear what [34:38] your take would be on that. [34:39] >> So, the problem with this is people talk [34:42] about a bailout and I actually think [34:44] private credit is the warning sign [34:46] there, Mr. Walter to the operating room [34:49] like it it just like uh lot of bad [34:51] things happening with private credit. [34:53] With AI and data centers, you've got [34:55] multiple problems. You have open AI and [34:57] anthropic. You bail these out. Okay. But [34:59] that doesn't fix the fact that they [35:00] won't be able to afford to pay $440 [35:02] billion in three and a half years. It [35:04] doesn't fix the fact that it's very [35:06] clear that Microsoft, Google, Meta, and [35:07] Amazon, their overall growth is slowing. [35:10] Otherwise, they would have never done [35:11] this. They are slowing down. and m 7% of [35:14] Microsoft's fiscal year 26 revenue was [35:16] from OpenAI. [35:17] >> I have to challenge that. Their cloud [35:19] their cloud revenues are accelerating [35:21] this quarter [35:22] >> based on anthropic and OpenAI spend. [35:24] >> But you're saying they're slowing but [35:26] they materially are not [35:27] >> when you remove the Open AI and [35:28] anthropic. [35:29] >> But why do you have to why would you [35:30] remove them? [35:32] >> Sorry, you're right. [35:33] >> Okay. [35:33] >> In the future, you can't count on it. I [35:35] should have put it like that. You can't [35:37] count on it in the future. When I say [35:38] that 440 number that is exponentially [35:40] larger because yeah they are growing [35:41] right now because open AI and anthropic [35:43] their checks are clearing this [35:45] >> well all of the hyperscalers all of them [35:48] reported way better than expected growth [35:50] numbers for cloud in the quarter they [35:52] they just reported second quarter [35:54] numbers like [35:54] >> other than meta [35:56] >> other so meta is a whole other basket [35:58] >> case this is actually a point though [36:00] >> okay all right but it is true that AWS [36:03] and Azure and um Google cloud are seeing [36:06] accelerating growth You're not [36:08] questioning that, you're saying there's [36:11] a big asterisk, it's two companies that [36:14] are ramping up all that spend. [36:15] >> So what also happened in Q2, the massive [36:18] token maxing, the massive spend on cloud [36:21] compute, it times directly with them. [36:23] Their numbers went up when Anthropic and [36:25] OpenAI and that money is real. Some of [36:27] that money came directly from the [36:29] hyperscalers themselves. The reason I [36:30] brought up meta was not not to be [36:32] argumentative is you'll notice that Meta [36:35] didn't grow like gang busters. Oracle [36:37] did, open AI, um Microsoft did. [36:39] >> Meta is not in that business. [36:40] >> But that's the point I'm making. You'll [36:42] notice that AI is not making meta grow [36:44] massively. It's just the ones who get [36:45] the money from anthropic and open AI. [36:47] And this is the problem. The problem is [36:50] is that for now everything you look a [36:52] quarter, two quarters in the future. [36:53] Wow. Money not money good money here. [36:55] Well, meadow meadows the Metabulls and [36:58] I'm not one of them, but the Metabulls [37:00] would argue they're not rent and [37:03] compute. They think they want to [37:05] prioritize it to improve reels. [37:07] >> Uhuh. [37:07] >> Um, they could at any time do an about [37:10] face and say, "Oh, yeah, by the way, now [37:12] we're a data center. Um, and we're going [37:14] to rent compute." And if they do that, I [37:17] think you would agree with me that day [37:18] at least, the stock goes up. [37:20] >> Yeah. [37:20] >> Okay. They haven't done it yet. [37:21] >> They rumored though. [37:22] >> They haven't done it yet. But their core [37:24] business, which is serving ads, [37:27] >> is in better shape than it's ever been. [37:29] >> That's not They had their first decline [37:30] in users. [37:31] >> No, no, no, no, no. the core ad business [37:34] like the profitability of their ad [37:36] business [37:36] >> right if the users all the spending [37:38] >> the users are declining though that will [37:40] decline too again it's I understand what [37:42] you're saying I'm saying for now it [37:44] looks good and this is the eternal [37:46] problem I've had it's like [37:48] >> at some point this goes wrong and the [37:49] reason I'm no longer giving timelines [37:51] cuz I was wrong [37:52] >> at some point everything goes wrong [37:54] >> but this is the way this goes wrong is [37:56] astronomically bad because it's not just [37:58] getting back to the bailout thing it's [38:00] not just oh open AI and anthropic have [38:02] to survive. They have to become 10. They [38:04] basically have to become the size of [38:05] Google or Microsoft and spend more than [38:07] they make every year just on those two [38:10] two or three companies. So, you can't [38:12] bail that out. On top of that, if we [38:14] build, let's say they cut back the the [38:16] process and they only build a quarter of [38:18] the data centers. We're still talking [38:20] 600700 billion of data center demand [38:22] that needs to exist that won't because [38:24] most AI startups are I mean, all AI [38:27] startups are unprofitable. So, where's [38:28] the money going to come from? Do you [38:30] bail out the data centers? Do you bail [38:32] out all of them? What happens to them [38:33] afterwards? What how do you bail out the [38:36] fact that the point I'm making around [38:38] cloud growth, which is most of the [38:40] growth for Amazon, Google, and [38:41] Microsoft, though Amazon has a [38:42] surprisingly good ad business comes from [38:45] cloud and most of that growth comes from [38:47] anthropic and open AI. Without that, [38:50] their growth story ends and they I [38:52] actually think will eventually [38:53] >> do you think do you think they'll be [38:54] able to come public? [38:56] >> I think anthropic will. I think OpenAI [38:59] is going to have the problem of [39:00] anthropic going public and having I [39:03] think their economics are going to be [39:04] bad. I've seen OpenAI's I've seen their [39:07] actual numbers. They're bad bad and [39:10] >> bad meaning yes there's a lot of revenue [39:12] growth but the expenses are ramping up [39:13] even faster. [39:14] >> Yep. 13.07 billion of revenue loss of [39:17] $20.9 billion [39:18] >> if it does okay SpaceX losing a lot of [39:20] money too. Yeah. If it does manage to go [39:22] public and get a reasonably warm [39:25] reception on Wall Street, either or both [39:27] of those companies and they do raise [39:30] what will probably be record setting IPO [39:34] uh you know uh uh share sales, they do [39:36] raise that kind of money, does that [39:38] alleviate some of your concern on the [39:41] sources of all the data center demand [39:42] going forward? [39:44] >> I mean in the short term because hundred [39:46] billion dollars [39:47] >> So it does push it out though. It pushes [39:49] it out [39:49] >> cuz they raise capital and they can keep [39:51] spending. [39:51] >> But the thing is with these cloud [39:53] compute agreements, you have to do [39:55] prepayments. So that money is going out [39:57] the door. Anthropic is rushing this IPO [40:00] because they had this first two quarter [40:02] window where token maxing was happening. [40:04] Everyone was saying burn as many tokens [40:06] as possible. And they'd moved all of [40:07] their enterprise customers, so [40:09] businesses over 150 people from paying [40:12] 200 bucks a month or 150 a month to [40:14] paying per token. So the revenue [40:15] exploded. It was great. And this was at [40:18] a time when the CEOs had all got [40:19] concussions. So they're like spend as [40:20] much as possible. So Anthropic needs to [40:22] go public ideally before they have to [40:26] really have the books for Q3, but they [40:28] can still show the growth there even [40:29] though most companies are now cutting [40:31] back. They're talking about [40:32] >> you think the data centers customers, [40:34] the Fortune 500 will actively pull back [40:38] on their AI spend now that the pricing [40:41] has changed and they'll look at it and [40:43] say, "Whoa, whoa, whoa, whoa, whoa. We [40:45] need to have a more of a handle of the [40:47] budget on what we're letting our people [40:49] do on AI. [40:50] >> Yeah. And I think that they've hit a [40:51] ceiling. I think there are some that [40:53] will now I think they've already hit a I [40:55] mean Sam Orman said on stage he was [40:57] saying yeah it's a huge issue for our [40:59] customers. He's so good at this. Uh [41:01] >> the price and anthropics fable [41:04] >> it's an issue for our customers. [41:05] >> It's a huge issue for our customers. [41:07] It's just like when you need a guy to [41:08] reassure you just going out there. Um [41:11] but the thing is fable so the expensive [41:14] an the most expensive anthropic model [41:16] the one that was so that had the [41:18] marketing of being banned by the US [41:20] government for being too powerful not [41:21] true but still they had it has kind of [41:24] peted out according to ramp at 11% of [41:26] market share and that's bad cuz that's [41:29] the only way you really grow revenue now [41:31] because it's clear that business there [41:32] are some businesses that will spend a [41:34] lot of money met has talked about [41:35] spending 10 billion if they're spending [41:37] 10 billion a year on anthropic the story [41:39] came out today That's also really bad [41:41] because Mark Zuckerberg changed the name [41:44] of his company to Meta mere years ago [41:46] and then just dumped that in 2023. [41:48] Like this man does not care. He has [41:50] complete control. He will stop doing [41:52] something on a whim. That's a really bad [41:55] situation. [41:56] >> Could we do a little bit of meta stuff? [41:58] >> Yeah, sure. [41:59] >> Okay. So, I know this is Let me Let me [42:02] ask this question. [42:04] >> My understanding of your story is that [42:06] you started out representing technology [42:09] companies. Yeah. [42:11] >> PR, helping them with messaging. [42:14] >> And while you're doing that, you start [42:16] to become disillusioned with their [42:18] intentions, the way they think about [42:20] their users, the um the rapacity with [42:24] which they are willing to monetize their [42:27] user B. So that's like a lot of your [42:29] origin story. You're almost like the guy [42:32] who saw too much. [42:33] >> Yeah. Uh, and I would say I was lucky in [42:35] that a lot of my clients because I would [42:37] I've been saying this since like 2013 [42:39] like journalists don't like profitable [42:41] companies and I've tried to work with [42:42] profitable companies just a really easy [42:44] start where if I work with unprofitable [42:46] ones I need to hear a very convincing [42:48] tangible story as to why don't do [42:50] anymore but still that was the thing but [42:51] yeah when I would see the companies that [42:53] got press and they were just these big [42:55] pieces of like clinkle clinkle you [42:57] ever hear clinkle [42:58] >> I'm still bullish on clinkle [42:59] >> I'm I'm I'm a clinkle head no but that [43:01] was one where it was like oh it's going [43:02] to use sonic wave gives to send payments [43:04] died within a year. It was a big loss. [43:06] Color, an app where you could share [43:08] photos with the people you're closest [43:09] to, dead. All these things kept dying. [43:11] They kept raising a bunch of money, [43:12] dying. They didn't have a business [43:13] model. And everyone just went, "Ah, [43:15] Bob's perfect. Moving on." And I saw [43:17] this again and again and again. And [43:20] really looking, what actually really got [43:22] me was Clubhouse. So, I don't know if [43:25] >> Oh, I remember that one. [43:26] >> Yeah. Because everyone was gooning out [43:27] over Clubhouse. They were like, "We're [43:29] all stuck in our houses and looking to [43:31] talk people." A16Z was big in that, [43:32] right? [43:33] >> Yeah. And the A16Z did the the classic [43:35] conif. They got all the people they [43:36] knew, didn't it? Samuel Jackson or [43:38] something. It was so strange. But [43:39] everyone was at home and they were like, [43:40] "This is going to replace radio." I'm [43:42] like, "Holy crap. Do you not you don't [43:44] respect your radio hosts?" Cuz these [43:47] people, yeah, so um yeah, I'm here um [43:52] from startup. Um the But everyone's [43:55] like, "This is the biggest thing. [43:55] Everyone needs to have a clubhouse [43:57] strategy." I'm like, "You people, you [43:59] >> that was weird. Do you know that that do [44:00] you know that that era coincided with [44:03] the spa boom? And so Clubhouse and [44:05] selling spaxs to people. [44:07] >> Oh yeah. [44:07] >> It became synonymous. All of the [44:09] financial conversations on Clubhouse [44:11] >> were like, [44:13] >> you know, a 28-year-old who just managed [44:16] to launch a spa. [44:17] >> So cool. [44:17] >> Who's who's never experienced the cycle [44:19] or anything. There was actually show [44:21] called What the Spack [44:23] >> and they invited me on it. And it was a [44:25] huge mistake for them to have done that. [44:27] >> Because in a prior career as a retail [44:29] stock broker, I sold a hundred spacks. [44:32] So I knew they were all going to zero. [44:34] >> Oh god. [44:34] >> And I told everybody in clubhouse and I [44:36] was never invited back to what this [44:38] back. [44:39] >> All of the people involved in what this [44:40] back all went to zero. Everyone. [44:42] >> Not to alleviate everybody from their [44:44] responsibility, but that was a very [44:45] bizarre I think all of our brains were a [44:47] little bit broken. [44:48] >> I sure. [44:50] I think it was more the media that [44:51] really got jokified because I just saw [44:53] people being like Clubhouse is going to [44:55] be the new thing. Everyone's just going [44:56] to listen to Clubhouse. Everyone's going [44:57] to do Clubhouse going to do all. And I [44:58] was like, are you people? And then the [45:00] metaverse, [45:01] >> they thought it was Instagram. They [45:02] thought it was like Snap or Instagram [45:03] where it was just going to take off. [45:04] >> Then the metaverse happened. I'm like, [45:06] "Holy You're either [45:07] describing something that doesn't exist [45:09] or video games. What are you What do you [45:11] mean? This looks like this VR thing [45:13] looks like What are you doing?" [45:14] And people were like, [45:15] >> we were laughing at that in real time. [45:16] >> And that's the thing. I Oh, yeah. Of [45:17] course, 2020. was selling virtual land. [45:20] I said I [45:20] >> no but virtual real estate virtual real [45:23] estate. I can't have this. [45:24] >> And I saw it being written about like it [45:26] was real. I'm like [45:27] >> oh my god what it is like having the [45:29] glass system they live. Jesus [45:30] Christ. And so every week I would see a [45:32] new article. I'm like Jesus. And then [45:34] crypto happened and NFTTS happened and I [45:36] would read things in the newspaper [45:38] telling me about crypto and NFT and I [45:40] went there's a reason people keep being [45:43] there's a reason now why bubbles keep [45:45] happening and it is the media. It's a [45:47] mixture of other things. They're not the [45:48] only problem. [45:49] >> There was bubbles before the media. [45:50] >> Yes, but the media still helped them. [45:52] Social media, [45:53] >> the Fed, [45:54] >> I'm Let me finish. [45:56] >> The point is [45:56] >> everybody's in. [45:57] >> But everything is a social contagion [46:00] ultimately. And social media, partly [46:02] from being at home to your point, [46:03] >> allowed it to exacerbate these bubbles [46:05] and speed them up. And AI was special [46:07] because of that 2022 2023 depression. [46:10] Plus, it gave everyone something to do. [46:13] You could become an AI consultant. You [46:14] could have an AI business. You could put [46:16] AI in your software. You could sell AI [46:18] infrastructure. Everyone the bankers had [46:20] something to sell. We had new things. [46:21] >> Are you not at all impressed with the [46:22] products? [46:23] >> Not for the amount of cost. [46:25] >> Okay. So if they were so an openweight [46:27] model that you pay almost nothing for [46:30] and it gets close enough. [46:31] >> I mean the cost to get here. So a [46:32] trillion dollars infrastructure cost [46:34] >> infrastructure and uh the environmental [46:36] cost and the social the horrifying [46:40] things the media has done scaring people [46:41] about jobs. Disgusting. [46:43] >> I agree with that. But the thing is you [46:45] can't divorce the technology from that. [46:48] You cannot what it does today is [46:50] unremarkable compared to its cost and [46:51] it's about what I would expect and [46:54] remember this just even if you take out [46:56] the infrastructure this still costs tens [46:57] of billions of dollars. the Kimmy K3s of [46:59] the world they still distilled from they [47:02] used the same training data they still [47:03] used a bunch of comput they still used [47:05] innovation from other people that and I [47:06] don't care about anyone copying anyone [47:08] in AI that's kind of the point but the [47:10] thing is there is this massive cost to [47:12] make this thing that great it's kind of [47:13] helpful I guess somewhere in a specific [47:15] thing but you can't really rely on it [47:17] you can't really trust it you have to [47:18] look at everything it does or you have [47:20] to double check it didn't muff something [47:21] up which doesn't sound particularly so [47:23] >> intelligent but so back to meta so the [47:25] reason I went into your backstory is cuz [47:27] I want to ask Do you hate these people? [47:29] >> Sounds like a [47:30] >> I absolutely hate them. [47:31] >> You personally [47:32] >> I think Mark Zuckerberg is a scumbag. I [47:35] think he's an piece of He is He [47:37] cares. He does not give a rat about his [47:39] customers. He has turned Facebook and [47:41] Instagram into a hell hole. He has [47:44] deliberately I've seen documents. I [47:46] published them where he they Mark [47:48] Zuckerberg requested 12% perpetual [47:51] growth of all the numbers. They at one [47:53] point for years were growing I think [47:55] time spent on app not realizing that [47:58] engagement was dropping because they [47:59] were measuring the wrong thing. They [48:01] have made that product like all of these [48:02] companies worse to increase growth. [48:05] Google search. Google search worse [48:06] because Prabagar Ragavan, the VP of ads [48:08] at the time, did a coup to take over [48:11] from Ben Gomes, who in emails with the [48:14] Department of Justice, published these [48:15] as part of the antitrust trial, said [48:17] like, "Hey, this is going to make the [48:18] user experience worse cuz you just want [48:19] to increase queries, which just means [48:21] people searching more times to keep them [48:23] to show them more ads." This is why [48:25] Google sucks now. And Microsoft, [48:27] >> you've called this the rot economy. [48:29] Exactly. This is this so you wrote this [48:31] big piece in 2023. Corey Doctoro talked [48:34] about in shitification. Yeah. [48:36] >> Your version of that it seems like is [48:37] like this idea where all of these [48:40] products start out good with like the [48:42] right intentions. Let's let's create a [48:45] searchable index of the entire internet. [48:47] Let's connect everybody's family members [48:49] with photos. Like it's a great idea. And [48:52] also probably if we get enough people [48:54] using it, there's some money. [48:56] >> And then your point about the rot [48:57] economy is like five years goes by, the [49:01] numbers aren't numbering. Yeah. [49:03] >> How now? All right. We have to shear the [49:05] sheep now. The sheep have gotten fat. [49:07] They're all on the platform. We charge [49:09] them nothing. We We subsidized this. Now [49:12] we're at now we're at this like critical [49:14] mass of people. Let's turn them upside [49:17] down and dump their pockets out. [49:18] >> And what's crazy is for a few [49:20] >> But that's how you see it. That's how [49:21] you see it. Okay? [49:21] >> Because they had ads on Google search. [49:23] They had ads on Facebook and Instagram. [49:25] Now it's the whole first page. And now [49:26] it's everything. And you can't have a [49:28] default chronological feed and [49:30] everything. It's like Ganesha painting [49:31] trying to find stuff. They move the [49:33] buttons around because they're [49:34] experimenting on their users. There was [49:36] a there was a Guardian article eight [49:38] years about it where they took 700,000 [49:40] people and they just started emotionally [49:42] abusing them with the notifications they [49:43] gave them, the stories they've fed them [49:44] to see what would happen. That should be [49:46] illegal. But these companies, it's not [49:48] just they want to grow. It's that they [49:50] have to grow perpetually and they will [49:52] do anything. It's why the honestly [49:55] they've run out of ways to squeeze other [49:57] than price increases at this point. And [49:59] that's why Microsoft has done multiple [50:01] price increases the last few years. Same [50:03] with Google, Meta. I think it was Google [50:04] and Meta changed how ads work around how [50:07] the first bid thing. I think that they [50:08] basically have a way of producing more [50:10] money out of it and raising the bid. The [50:12] point is [50:13] >> it's sort of capitalism though. Also, [50:15] >> it is, but at some point it eventually [50:17] goes completely haywire and explodes. [50:19] Like that is what it that is the [50:20] ultimate thing. You eventually get to a [50:22] point where you've chased out anyone at [50:24] the top who cares about product or good [50:27] stuff or happy customers and all they're [50:29] there is for growth. Andy Jesse MBA uh [50:32] Saji Nadella NBA Pashide Mackenzie and [50:35] NBA Mark Zuckerberg not NBA Cheryl [50:38] Samberg NBA and the original growth team [50:41] I know had another one. They had [50:42] Schmath. They had Naomi Gl. They had [50:44] Javier Olivan who's now the CEO. [50:45] >> So but like philosophically these [50:47] companies have shareholders. The [50:49] shareholders are there because they [50:50] expect the cash flows to grow, some [50:53] portion of that to be returned to them [50:54] in the form of dividends, buybacks or [50:56] both. They want to see the earnings [50:58] growth. That's like sort of the job of [51:00] the people running these things. It's [51:02] not to um it's not necessarily to be [51:06] caretakers for the population of users. [51:08] It's [51:08] >> you mean you mean provide a good service [51:09] at a fair price? [51:11] >> Yeah. [51:12] >> So that's the thing. [51:12] >> Well, Instagram would tell So Meta would [51:15] say, "Well, here's Instagram. How much [51:17] do you pay for it? Zero. [51:19] >> They make 55 bucks. [51:21] >> You pay for it in attention. That's the [51:23] That's what they did. [51:24] >> Sure, but there is a way of doing that [51:25] without making it horrible, which they [51:27] have. The point I'm making is however [51:29] you feel about this ethically, it has a [51:31] natural endpoint where you you change [51:34] the company from making a product that [51:36] you sell and you turn it into a growth [51:37] engine. You turn it into something [51:39] that's only about expressing growth. You [51:42] hire people with that explicit intent. [51:44] You direct them with an explicit intent. [51:47] >> What's So, okay. So, what's stopping the [51:49] users of Google, Amazon, and Meta from [51:52] saying, "I no longer like this and will [51:54] not use it." [51:55] >> It's addictive. So, Monopolies. [51:56] >> No, I'm with Ed. I'm with you. I think I [51:58] think what you're describing is how a [52:00] lot of us feel. I hate Instagram. I love [52:03] it so much. I'm addicted. I have to use [52:05] a physical brick device to stop myself [52:08] >> from using it. I think we all feel that [52:10] these these places are toxic and [52:11] disgusting. I don't like the way that I [52:13] don't like the way that children were [52:15] onboarded. [52:16] >> I uh [52:17] >> that's so that's where I probably you [52:19] and I are probably agree the most. [52:21] >> There's just to me there's something [52:23] different about a 12-year-old and a 42y [52:25] old. I'm not worried about a 42y old. [52:27] >> Yeah. [52:28] >> You want to sit there and scroll your [52:29] phone for 9 hours? It's not my problem. [52:31] >> I don't like the addiction of the And [52:34] there was a settlement this week. Um [52:36] >> Yeah. But over 10 years, [52:38] >> how much money was it? [52:38] >> It's 12 billion and then maybe 17 [52:41] billion. like 1% of Meta's market cap [52:43] basically. [52:44] >> It's I thought it was going to be an [52:45] upfront payment and the the settlement [52:47] people are always [52:48] >> 28 states all settled at once. [52:50] >> Yeah, apparently. [52:51] >> Yeah. But so what will this do? So Meta [52:53] will enforce Meta will enforce the [52:55] 2-hour daily limit on Instagram and [52:56] Facebook for children as part of a [52:58] record claim. [52:59] >> That's great news. [53:00] >> That's awesome. Now I love those [53:01] changes, but the they should have made [53:02] them pay up front because that would [53:04] have been most of their net income from [53:05] last quarter. But there is a point to [53:06] this as well. taking away the ethical [53:08] concerns. You eventually engineer your [53:11] company around growth and you stop [53:13] learning how to build new things that [53:14] customers like. You find ways to twist [53:16] customers arms, but you don't really [53:18] find ways to win them over. You [53:19] eventually get into a mindset where [53:21] you're like, "What if I could just spend [53:24] money and get money back without really [53:26] thinking about it?" AI is the perfect [53:29] vehicle for that. In theory, if you were [53:32] a company that hasn't really thought [53:33] about user experience, you got this [53:34] magic box you can ask questions to. You [53:36] could detach that to anything and you [53:38] could answer any question. you [53:39] don't even need to come up with a [53:40] product. It will come up with a product, [53:42] right? This is the theory they had. It [53:44] didn't work. But when you're not really [53:46] thinking user friend, anyone who's [53:48] thinking, "Wow, I want to make a user [53:49] friendly experience." LLMs areing [53:51] insanely questionable product. Some [53:54] people get value out of them. Some [53:55] people get no value. Sometimes it just [53:57] goes wrong. Sometimes it makes stuff up. [53:59] That's a crazying product. I'm sorry. [54:01] That's insane. The fact that people are [54:03] so defensive and they say hallucinations [54:04] have been solved when they haven't is [54:06] insane. But all of these companies have [54:08] become engineered around growth. And so [54:10] they went, "Wow, I can just buy these [54:12] GPUs and I'll keep growing." Right? Meta [54:15] was horribly wrong, of course. [54:17] >> Wait, you think the products suck? I [54:18] love I love using Claude. [54:20] >> I mean, why? What do you mean why? I [54:22] find it I find it to be miraculous. I [54:24] use it all day. And I'm not I'm not I'm [54:26] not the only one. [54:27] >> But this is the thing. Like the fact we [54:29] have such divergent experiences is so [54:30] interesting to me because this is meant [54:33] to be the new big thing. I I to be clear [54:35] love technology. I [54:37] >> really So here's the word for it. I'm a [54:39] writer like you. [54:39] >> Yeah. [54:40] >> I will never have AI write my words cuz [54:43] I think the only reason people read me [54:45] is because I have a very specific way of [54:47] talking. Yeah. [54:48] >> And a very specific point of view. Um, [54:51] however, I don't necessarily want to use [54:53] the blue links on Google's front page to [54:56] go hunt down information when I'm in the [54:59] flow of writing. I'd much prefer to [55:02] have, whether it's Gemini or Claude, [55:04] have a co-pilot where I say, "While I'm [55:07] busy writing this, go doublech checkck [55:10] these numbers for me and cite the [55:11] sources before I actually [55:14] >> I use it as research." [55:16] >> You use it. That's describing. [55:17] >> Search, search, and research. And now [55:19] what Gemini has done or what Google has [55:21] done I should say Google saw this a year [55:24] and a half ago was extinction level [55:26] event if they didn't cannibalize their [55:28] own search and put the AI result up top. [55:32] >> I don't Yeah, I know. I know they [55:34] thought that way. I don't know how I [55:35] agree with [55:35] >> Well, I don't know. I don't know what [55:36] else they could have done if people's [55:37] searches started to move to um to chat [55:40] GPT [55:42] sit there. [55:42] >> But the thing is you're describing [55:44] search. [55:44] >> But do you not find that you're a great [55:45] writer, you must find that valuable. [55:47] >> Here's the thing. I use so my terminal [55:49] for example BQL not having to write BQL [55:52] ising great I could I can't do that [55:53] there I don't know what that is so with [55:55] the Bloomberg terminal they have this [55:56] thing called ask B so if you need to [55:58] look something up it would run basically [56:00] Python their coding language it's BQL [56:03] and instead of doing that you can go [56:04] look up this and it will generate the [56:06] code and run it so you can actually try [56:07] it take it test it yourself the thing is [56:10] >> that doesn't impress you [56:12] >> doesn't impress me for the cost for like [56:14] a trillion dollars and tens of billions [56:15] of dollars to get to a better search [56:18] >> but You think this is the end state? You [56:19] don't think that you do? [56:20] >> I think that we're going to get [56:21] iterations from here. There's a reason [56:22] they move to cyber security. They've [56:24] kind of they're going to iteratively get [56:26] better at code which doesn't necessarily [56:28] replace anyone or even necessarily help [56:30] software engineers past the certain [56:32] point because it gets more multipatively [56:34] I could say that word correctly more [56:36] complex and more hard to manage. And I [56:39] think the big problem with LLM I have is [56:41] the distinction between when you are [56:42] outsourcing work like searching for [56:45] something and thinking. [56:46] >> Oh, I totally agree. And that's and that [56:48] the thing is you tons of people are [56:50] doing the thinking one which is a [56:51] problem [56:52] >> and no I'm not that impressed because I [56:54] used it to fix a thing with my kids [56:56] Minecraft game recently. I because I [56:58] think ethically I should try and use [57:00] this. I should and I have and it's like [57:02] it took half an hour of chasing its [57:04] tail. It eventually got it done and I [57:06] probably couldn't have done that or [57:07] maybe it would have taken but it's just [57:09] kind of like okay how many mistakes did [57:11] it make? If I was paying on a per token [57:13] per million token basis, I'd be pissed [57:15] because I'd have probably spent $15. [57:17] >> Are there workflows that you've been [57:18] able to automate utilizing AI where [57:21] you've said, "Okay, I still hate these [57:24] people, but this is literally saving me [57:26] time from some of my repetitive tasks [57:28] >> outside of the Bloomberg terminal." No. [57:30] >> Okay. So, we're on in the other boat. [57:33] Like there are skills that I've been [57:35] able to create on Claude where I used to [57:37] have to go to these five websites, copy [57:40] paste, do this whole thing. And now I [57:42] can say to it, "All right, it's Friday [57:44] morning. You know what I need? Go get [57:46] it." [57:47] >> I find that to be like [57:49] >> So I'm not a tech person. I'm a finance [57:51] person. But for me, it's like [57:53] miraculous. [57:53] >> I also want to be clear about something. [57:55] LLM's in a vacuum when you remove the [57:57] social cost, the environmental cost, the [57:59] fiscal cost, all this interesting [58:00] technology. I like I think they if [58:03] they'd have called them like library [58:04] models no one would be getting rock hard [58:06] about it but it's just like they are [58:08] clouds they're automation like and like [58:11] fine great I don't know why we have to [58:13] talk about it I don't mean here but it's [58:15] like I don't know why everyone's sinking [58:16] a trillion dollars into this that's [58:18] crazy because well there is an there is [58:20] also there is a thing where if they [58:22] think if they don't you've probably [58:23] heard this a million times [58:25] >> like the existential threat to their own [58:27] companies like if we miss this wave [58:30] our existing data center business is [58:32] worthless. If we don't rip out all these [58:34] CPUs and replace them with GPUs and be [58:37] able to do these calculations for these [58:39] models, then AWS will go to zero and [58:44] Google Cloud will win. They feel that [58:46] way. I get that. And I think what's [58:48] happened is maybe they've realized, but [58:49] they now realize they have to double [58:51] down because everyone else is doubling [58:52] down. And I think when one of them pulls [58:53] back [58:54] >> some cost and they already spent too [58:55] much. [58:55] >> And that's the thing. [58:56] >> Yeah. when they stop spending on AI [58:59] capex, everyone's gonna go, great, [59:01] you're done with this. What? Why did you [59:03] do it? Like that's the thing. The moment [59:04] they do it, it's just like everyone's [59:05] gonna go like brilliant. Um what were [59:07] you doing the last few years? Have you [59:09] got enough AI now? What are you how are [59:11] you [59:11] >> and the spending literally can't stop [59:13] because the chips burn out. [59:15] >> Yeah. Well, that that one it's like [59:17] again they obiscate all this data. But I [59:19] actually had a question around your [59:20] workflows. Do you pay on a per million [59:22] token basis? [59:23] >> No. Right now, everything that I'm able [59:25] to do is very cheap because I'm not Jane [59:27] Street and I'm not running a billion [59:30] calculations a day. We are working on [59:33] Gemini. We're working on CL working on [59:36] Hazel. [59:37] >> Yeah, we're pay but we're paying SAS [59:38] like we're paying a SAS-like fee at this [59:41] point. Yes. But you're not paying and I [59:42] know it could change. [59:43] >> That's and that's kind of the point I'm [59:44] getting to. It's like also this is an [59:46] insane way to run a business. GitHub [59:49] Copilot had two million customers and on [59:50] June 1st of this year, Microsoft just [59:52] went pulled the rug out from under them. [59:55] That is an insane thing to do to [59:56] customers. That kind of it's the most [59:58] economy. [59:59] >> Why did they Why did they do that? [60:00] >> Because they were allowing them to burn [60:01] $5,000 of tokens for $40. I But the [60:04] thing is, why did you do that? And they [60:06] would argue, "Oh, we thought it would [60:08] get cheaper." Why? Why'd you get that? [60:10] Because it's never got nothing has got [60:12] cheaper. The cost of intelligence has [60:15] come down, but the models spend more [60:16] tokens. And the models may be able to do [60:18] more stuff. They should have to [60:20] considering the tens of billions. Like [60:22] I'm not saying they haven't improved, [60:23] but it's like what will be the first [60:25] sign do you think if you had to guess? [60:28] And I know you're not timing it and I [60:29] know you're not telling people to trade [60:30] on this. What would be the the first [60:33] sign of the apocalypse for all of this [60:36] spending and like the comeuppance or or [60:38] like like when everyone goes uhoh? [60:42] Because I here's why I ask you that [60:44] question. I genuinely [60:46] know enough to know that I will never be [60:49] the person that spots it. [60:50] >> But this is the new parlor game. [60:52] >> Yeah. [60:53] >> People say one of these days Meta is [60:57] going to throw in the towel. [60:58] >> Yeah. [60:59] >> And say that capex projection we gave [61:01] you last quarter, don't rely on that. [61:03] It's actually going to be lower. And [61:05] then that will set off a chain reaction [61:07] where the semiconductor uh stocks [61:09] instantly lose 30% of their value. the [61:12] utilities fall like penny stocks um and [61:15] all of a sudden like multiples compress [61:17] other companies pull back that creates a [61:20] panic. Okay, so that would be like my [61:22] guess is it has to come from one of the [61:24] hyperscalers. [61:25] >> So none of them is showing any signs [61:28] that they have any intention of pulling [61:29] back yet. But is that the thing that [61:31] you're looking for? [61:32] >> That's an obvious one. I agree. Like it [61:34] would be a capex pullback. Been saying [61:35] that for years. Who knows when they they [61:37] eventually cry uncle. I actually think [61:39] there are a few different pale horses. [61:41] One of them is if coreweave cannot raise [61:43] debt. Coreweave as a neocloud so they [61:45] rent AIGU compute and their customers [61:48] are open AI, Microsoft for open AI, [61:51] Google for open AI, Anthropic and Meta [61:53] >> and they buy a lot of GPUs [61:54] >> and they are horribly unprofitable and [61:56] they need I think they've raised [61:58] billions of dollars three times this [61:59] year. They need to do that. Last time [62:01] they paid over like 9.2 or something [62:03] like some ridiculous rate on it. And [62:05] hey, I don't think things are getting [62:06] cheaper in that realm. And also the [62:08] chips more expensive. So corewave needs [62:10] to perpetually raise debt. In fact, [62:11] everyone does. Iron, Nebius, Nscale, who [62:15] maybe I [62:16] >> all the Neoclouds, [62:16] >> they all need to raise debt. And [62:18] >> so they don't have a Google search like [62:19] business to fund this. [62:20] >> They don't have a business at all. They [62:22] don't they just have daddy like the Find [62:24] that calls them and asks them for $3 [62:26] billion every quarter. So one of those [62:29] companies doesn't raise money. It's [62:31] going to be they try and they fail. I [62:33] could also see a hyperscala bond sale [62:35] that gets even interest or like just [62:38] above even. We just had an Amazon one [62:40] that only got 1.6x [62:42] overs subscribed. If that comes to like [62:44] 1.1 1.2 if it gets like real thin, [62:46] people will write articles like the [62:48] enthusiasm for data center bonds are [62:51] drying up. [62:51] >> And I thought that was happening because [62:52] the Amazon one happened then Google did [62:54] one that was four or 5x over subscribed. [62:56] So I'm like okay. But the thing is those [62:58] are the kind of signs because the big [63:00] naivity I had was like that some sort of [63:02] ethical or social concern would change [63:04] things or that the companies wouldn't [63:06] keep doing something because they all [63:08] knew that they had to and if one of them [63:10] stops all of them can stop but the first [63:11] one has to first no it comes down to [63:13] money. Will the money be available? And [63:16] that's the thing because Nvidia's GPUs [63:19] are so expensive and they just got more [63:21] expensive and the cost of RAM and [63:23] construction are going up. you [63:24] necessitate being able to raise billions [63:27] of dollars. That's actually a fairly [63:29] small amount of companies that can do [63:31] so. Most of them are bankrolled by [63:34] Jensen Huang or dependent on OpenAI or [63:37] anthropic. So, it's the reason you ever [63:40] do you look at Oracle's Twitter account? [63:41] I know this is weird. I'm going [63:42] somewhere. [63:43] >> No. [63:43] >> So, every time something goes wrong in [63:45] open, I haven't seen it. [63:47] >> So, Oracle when I that story came out [63:50] cuz Nvidia was going to invest 100 [63:52] billion in open AI. It didn't happen. [63:53] When that came out, Oracle did a tweet [63:55] that was that was saying something along [63:57] the lines of, "We have no concerns over [63:58] our financial situation with OpenAI." [64:01] Oracle will just randomly post up be [64:03] like, "Starate Abene's going great." And [64:05] you can basically just take the opposite [64:06] of whatever it's saying. That's because [64:09] >> who do you think runs out of that's not [64:11] Saffricat City? [64:12] >> God, no. Saffric Saffricat's funniest [64:14] thing I've ever seen. Oracle signs this [64:15] massive deal with OpenAI. Like a week [64:18] later, she's like, "See you. Yeah, we're [64:20] you're going to have two CEOs now. Let [64:22] these guys deal with this. I'm going to [64:23] go sit back here, [64:24] >> right? [64:25] >> And so Oracle also has to keep raising [64:29] money. Oracle is still, I think, in the [64:31] process of doing their mass share sale. [64:33] >> You But you must So, but you must be [64:35] heartened by the fact that the stock [64:37] market is treating Oracle differently [64:39] than it's treating Alphabet. [64:41] >> Yeah. [64:41] >> And the and the bond ratings agencies [64:43] are certainly looking at. [64:44] >> They don't have the stones to downgrade. [64:46] They don't have the stones to to [64:47] >> It's that junk. It's a notch above. [64:49] >> They're notch above. They don't. Will [64:51] they have the balls to make them fall an [64:52] angel? Hell no. That the ratings [64:54] agencies rated coreweave debt investment [64:57] grade cuz it was connected to a [64:58] hyperscaler. Unbelievable. But that's [65:00] the thing. It comes down to when the [65:02] money runs out. If open AI fails to [65:04] raise, if a major AI startup goes [65:06] insolvent. So, have you heard about [65:08] Plexity? [65:09] >> What's going on? [65:10] >> So, [65:11] >> I haven't heard much about them in a [65:12] while. Actually, [65:12] >> they speak to a fade of the way. [65:13] >> That's actually really good to know [65:15] because Nvidia is considering investing [65:17] billions at a $30 billion valuation. [65:19] Why? Because Perplexity is one of the [65:21] few companies that actually spends any [65:22] real money on GPU compute. Perplexity's [65:25] death would hundreds of millions, but [65:27] still in this industry that's a lot when [65:30] no one else is spending much more than [65:31] millions or tens of millions. So Nvidia [65:34] is keeping them alive because if they [65:36] fail, well, everyone knows. [65:37] >> So the pale horse could either be a [65:40] major hyperscaler decides it wants out [65:43] of the race. [65:45] >> A major hyperscaler has its debt [65:48] downgraded. Allah Oracle. [65:50] >> Yes. [65:50] >> Probably the most likely [65:52] >> more than certain [65:53] >> major credit a major startup [65:57] >> uh runs out of money [65:58] >> that like everyone associates with the [66:00] AI trade like a perplexity. Okay. [66:02] >> Um or what was the fourth one? If [66:05] coreweave [66:06] >> and coreweave can't roll its debt [66:08] >> basically anyone who can't raise [66:09] morning. Oh, also [66:10] >> or an IPO bombing. [66:11] >> IPO bombing of course. [66:13] >> That would be a big one. If an IPO did [66:15] not go well, [66:16] >> if SoftBank can't take a company public, [66:19] Soft Bank's liquidity situation is [66:21] really bad. They are having to raise 20 [66:24] billion in bonds cuz they raised 40 [66:25] billion. [66:26] >> Dude, that guy's crazy. [66:27] >> I love him. I love his goose. His goose [66:29] math rocks. [66:30] >> Ed, where's the mania though? Because I [66:31] feel like at least in the equity market, [66:33] a lot of investors are not treating [66:35] these companies as if they're Teflon [66:37] because Nvidia's 70% uh profit margins, [66:40] stocks trading 18 times forward [66:41] earnings. Uh Meta 16 times like where's [66:44] the mania? And the mania is in the data [66:46] center construction and the companies [66:48] themselves. [66:48] >> And it's an activity bubble as much as [66:50] it is a financial bubble. Like the [66:52] amount of these deals, [66:53] >> yeah, there is momentum that still [66:55] exists that keeps everyone thinking this [66:56] is normal. Oracle shouldn't trade at [66:58] whatever. If it's anything above $40, it [67:00] should Oracle's revenue has been flat [67:01] for 15 years when you adjust for [67:03] inflation. Like what? Like this dog of a [67:06] company. I'm surpris they are the only [67:08] one of them I'm certain would get bailed [67:09] out. [67:09] >> Stock's down like 65%. I mean [67:12] >> it should go lower. It suck. Oracle is a [67:14] horrible company. They also fired a [67:15] bunch of people and I know this from [67:17] sources that deal with contract renewals [67:19] which is like the number one most [67:21] important thing at Oracle. And if [67:22] they're firing those people, they're [67:24] just kind of disorganized. [67:25] >> The data center thing that makes me the [67:27] most bearish. I made the huge mistake of [67:29] rereading a book called 1873. [67:32] >> Oh. [67:33] >> About the rail the build build out the [67:35] the boom and bust of the railroads. I [67:37] don't think that this looks like [67:40] and I don't think that this looks like [67:41] the great financial crisis. I think that [67:44] is the best analog because of how [67:46] physical the demand is. the build of the [67:49] railroads was like extraordinarily [67:52] physical in the way that whatever size [67:54] whatever megawatt data center you just [67:57] described like [67:58] >> it's it's that type of a uh [68:00] infrastructure buildout bubble to me. [68:03] >> Yeah. [68:04] >> And I should not have reread that book [68:05] because made me want to sell everything. [68:07] >> GE Capital is the other comparison I'd [68:09] make with Nvidia. Nvidia is acting a lot [68:11] like Welch. They're doing some [68:13] >> vendor financing. vendor financing, but [68:15] also with the uh was it GCAS, the [68:17] airline side over there. Yeah, they kept [68:20] airlines alive constantly. They were a [68:22] massive leaser. They were America's [68:24] largest equipment leaser at one point to [68:26] their customers. There were airlines [68:27] that could only survive as long as GE [68:29] could lease them stuff. But the another [68:32] thing is as well is this is a bit of [68:34] everything because the data the [68:36] difference and the problem with data [68:38] sense is these things are not really [68:40] useful for other things. GPU compute is [68:42] not really that useful outside of this [68:43] and it's very clear that the only people [68:45] that will pay a lot for it are [68:47] unprofitable startups if real businesses [68:49] are not buying it at the scale they need [68:50] to and there doesn't seem to be a reason [68:52] in fact there seems to be more reasons [68:53] to buy less open source models but [68:56] you've got Nvidia [68:58] >> sorry uh self self-driving cars [69:00] automation in general humanoid robots [69:03] we're saying that nonlinear processing [69:05] is not useful for those things [69:06] >> I got to go to the bathroom carry on [69:09] >> not not at the scale they build them. If [69:11] they were building 30 billion of the $30 [69:13] billion worth of these things, very [69:16] different. [69:16] >> Okay, [69:17] >> but they're building what will require [69:19] 700 to a trillion dollar a year in [69:21] income for something that doesn't need [69:23] that. Look, we've had Tesla and Volvo [69:25] and all sorts of other companies that [69:26] have self-driving cut uh Zuks and um [69:29] what is it? Whimo. You have all of these [69:31] things and they didn't need all these [69:32] GPUs and indeed having more of these [69:34] GPUs has not changed like there hasn't [69:36] seemed to be a breakthrough in them. [69:38] They're getting better. I love [69:39] autonomous cars. I think they're [69:40] fascinating. I love getting in a way. I [69:42] think that there are social problems [69:44] we're going to have as a result of taxi [69:45] drivers. I think that there's going to [69:47] be real problems there. [69:48] >> Yeah. [69:48] >> I wish we actually do some socialism and [69:50] actually help people. [69:51] >> Not exactly sure what all those drivers [69:53] are going to do. Maybe the only get out [69:55] of jail free card is this could take [69:56] place over 30 years. Not [69:58] >> I think it's decades and decades and [70:00] decades because the problem with [70:01] autonomous cars is not the 99%, it's the [70:03] 1% issue and that means you have to [70:05] expand slowly. But the point is although [70:07] those GPUs might be useful for that, but [70:10] we've built so much of this. We're [70:11] building so very much of this and per [70:14] Nvidia's own earnings, we're at a point [70:16] where there are like three to five [70:18] companies that can afford this at any [70:20] kind of scale and Nvidia has to make so [70:23] much more money. Like so everyone is in [70:26] this situation. Open AI and Anthropic [70:27] have to grow like crazy. Nvidia has to [70:29] grow like crazy. I think Nvidia may have [70:32] said the 70% thing because they're [70:35] desperate. And I know that sounds crazy [70:36] because they're doing mad numbers, but [70:38] remember they've been down. People have [70:40] been a little bit worried about Nvidia. [70:41] They needed something exciting. Also, [70:43] they got some concerning in that in [70:45] that 10Q. There's some really weird [70:47] stuff. That statement they made about [70:48] investment grade partners that can pay [70:50] between 3 months and one year. It's [70:52] really weird. [70:52] >> So, you don't like seeing them become [70:54] investors and in some cases creditors to [70:57] of their own customers? [70:58] >> It seems like a it's the classic sign of [71:02] illusory demand or artificial demand, I [71:04] should say. [71:04] >> All right. So to close so to close this [71:06] out, this is the question my at this [71:08] point my entire audience would be [71:10] screaming this at us. [71:11] >> Yeah. [71:12] >> This is the thing they would most want [71:13] to hear from you in closing. [71:16] >> Is there any chance that you're wrong [71:18] about how this turns out? [71:20] >> And if so, what would have to happen for [71:23] you to change your mind and say, "I [71:26] probably got a lot right, but I missed [71:28] the thing that actually mattered and um [71:32] I I was I was wrong." like what what are [71:34] those concrete things or cuz it's not [71:36] just cinnamon chef just promise me the [71:38] next time you have an AI bull you ask [71:40] them the same question like next week [71:41] please please do please ask them the [71:42] exact same question what if you're wrong [71:44] but so I don't have to be 100% right for [71:47] things to be really bad I agree with you [71:49] really like open AI and anthropic could [71:52] become much smaller businesses that hang [71:54] in there through some messy accounting [71:56] and some wing of Microsoft or Amazon and [71:58] Google some consortium and the LLM [72:00] industry could be much smaller but not [72:01] disappear those are feasible [72:04] There is no breakthrough coming that's [72:05] going to magically change this. The AGI [72:07] story is dead. Even they are trying to [72:08] move away from it. Even though Jensen [72:09] Wong [72:10] >> nobody can even define it. [72:11] >> Exactly. Jensen Wong said it was here [72:12] and it's just like shut up Jensen. Come [72:14] on mate. But the thing is the hypers [72:18] scale growth story is so difficult for [72:20] them to meet. It's so crazy. Like even [72:23] what Jensen Jensen if anyone can pull it [72:24] off it's Jensen. But to do $670, $690 [72:28] billion next year probably from the same [72:31] customers means that Amazon, Google, [72:35] maybe Microsoft, definitely Meta are [72:36] going to have to take out a bunch of [72:37] debt. SpaceX 100%. Amazon says they're [72:40] buying 2 million GPUs. Really? How how [72:43] much money are you going to have to [72:44] raise your cash flow negative? All these [72:46] companies are going to see their [72:47] economics and or I think Meta and Amazon [72:51] both guided lower than expected in Q3. [72:54] So, we're already seeing some slowdown. [72:55] How interesting. But the point is, they [72:58] are going to have to spend more and more [73:00] and more and more money and more. Even [73:02] if they're buying the same amount of [73:04] GPUs they plan to 6 months ago, they're [73:06] going to have to spend more cuz they're [73:07] more expensive now. So, [73:08] >> so there's no chance you're wrong is [73:10] what I'm hearing. [73:11] >> I don't see how this goes on as long as [73:13] they they expect to be making hundreds [73:16] of billions of dollars in compute [73:17] revenue just from anthropic and open AI. [73:19] I don't even in the dumbest craziest [73:22] world I don't see how the money exists [73:24] to do that because that requires them to [73:26] likely spend half a trillion to a [73:28] trillion dollars in capex and have those [73:30] data centers actually answer that [73:32] question how would you be how would you [73:34] know for sure that ed ends up uh being [73:37] wrong [73:39] >> um [73:39] >> what would be like would it be the [73:40] profitability of anthropic and open AI [73:44] happens in 28 [73:45] >> even then are they profitable enough to [73:48] be able to afford the compute they've [73:49] committed to 1 1.1. [73:51] >> So even that wouldn't be enough to [73:53] change. [73:54] >> 2 trillion is so 100. [73:55] >> Why would you What would you say? [73:56] >> I don't know. I don't know how you could [73:57] disprove that or prove that he's wrong. [73:59] Are you betting against any of these [74:00] companies? Do you have any money on the [74:01] line? [74:01] >> No, I don't. Here's the thing. I don't [74:03] have cash in the market. People go, "Oh, [74:05] he doesn't have skin in the game." I I [74:07] have emotional skin in the game. I don't [74:08] know what to say. I [74:09] >> Oh, no. I agree. You have made You have [74:11] become the face of AI skepticism. you [74:14] have as much in uh reputational risk as [74:18] other people have in dollar risk. [74:20] >> And but the other thing is is the market [74:22] is not you don't invest in the market [74:24] now based on good sense and reading [74:26] about companies. You invest based on [74:28] reading the smoke signals. And last year [74:31] in September there were like three or [74:32] four OpenAI announcements that [74:34] manipulated SKH High, Samsung, uh AMD [74:37] and Nvidia. Four announcements where [74:39] nothing happened and the stocks popped. [74:41] I can't play in a market like that. I [74:43] can't play in a market where companies [74:44] do fake announcements. [74:46] >> Do you talk to hedge fund managers that [74:47] are making active bets against these [74:49] companies or they're long but they want [74:52] to hear like your your views? [74:54] >> I occasionally do [74:55] >> without like naming people like what [74:56] kinds of conversations are you having [74:58] similar to this? [74:59] >> Yes, but I never get near the actual [75:01] trades. No one ever talks investing [75:03] which is great. [75:03] >> But no, but they might want to hear your [75:05] train of thought like why why shouldn't [75:08] I have a billion dollars in [75:10] semiconductor stocks right now? [75:11] >> Yes. And and that's the thing when it [75:13] comes to just equities. I'm kind of like [75:15] if you are willing to ride the crazy [75:16] train, go ahead. Like if you want to if [75:18] you want to read the smoke signals, if [75:20] you're capable of seeing where the [75:21] market will go and what will influence [75:22] the market, bully for you. This is not [75:23] my expertise. But if it comes to a [75:25] long-term thing, especially if it [75:26] involves investing in like the debt [75:28] underlying a data center, I will just [75:30] explain how the demand is a losery. [75:33] There's not no demand, but it's a teeny [75:35] tiny amount compared to how much we're [75:36] building. [75:37] >> It's a hype. It's definitely a hype [75:38] cycle. Even the bulls would have to [75:40] admit there's a lot of hype. There's [75:41] nothing after this. Nothing. There's no [75:44] dot comb bubble style fixer here. AIGPUs [75:47] are not useful for other stuff. And [75:49] also, if a data center is left unc [75:51] incomplete, unbuilt, it's going to cost [75:53] just as much in 3 years to finish. Data [75:55] center GPU AI GPUs are going to cost [75:57] just as much to run in electricity, if [75:59] not more. And if the power isn't [76:01] finished, if the power never made it, or [76:03] if you were running it using gas [76:04] turbines, so nothing permanent, it's [76:06] going to be even more expensive [76:08] depending on the cost of gas at the [76:09] time. And there's just not a with a.com [76:12] bubble there's a bunch of server [76:13] hardware that went out that was useful [76:15] for building the future with like Amazon [76:16] web services for example Amazon web [76:18] services [76:18] >> dark the dark fiber was useful later [76:20] like that [76:21] >> but energizing that fiber using that [76:22] fiber wasn't anywhere near as expensive [76:25] Amazon web services by the way between [76:27] 2003 when it was created and 2015 when [76:29] it became profitable Amazon's total [76:31] capex was $ 29.7 billion adjusted for [76:34] inflation [76:34] >> point and that's the thing that I try [76:37] and tell people because oh Amazon web [76:39] services cost a lot money. Yeah, this [76:42] costs way more. That was all of Amazon's [76:43] capex, not just AWS. This is nothing [76:47] people and I do the same thing. People [76:49] want to compare things, people want to [76:50] smart and they can look at and say this, [76:52] this looks like this and that and the [76:53] other. [76:54] >> I just challenge people to say stop [76:57] relying on that. Don't if you don't [76:59] agree with me then at least don't rely [77:01] on there being a happy ending here and [77:03] prepare accordingly because I think that [77:08] the day people may expect open AI to [77:10] die. They may not be surprised when it [77:12] dies. I think some will but they won't [77:14] be like there will be some people went [77:15] yeah I kind of saw that coming. the data [77:18] centers not being used and the data [77:20] center debt not being paid. That is [77:23] going to be what shocks people because [77:25] people because we're talking hundreds of [77:27] billions, trillions of dollars, regular [77:29] people can't think of that. Like that's [77:31] an unknowable amount of money. It's [77:32] almost to the point that stops meaning [77:34] and you come to this logic of these are [77:36] the smartest people in the world. Why do [77:38] you think they're smart? [77:39] >> So the All right. So the implications [77:41] though of what you're saying, you [77:43] realize we're not talking about a.com [77:46] meltdown where the NASDAQ declines 80%. [77:50] But the rest of the economy is fairly [77:52] unscathed, which is actually what [77:53] happened. [77:54] >> Yeah. [77:55] >> What you're describing [77:57] because the numbers are so much bigger [77:59] and because the private credit people [78:01] are all in. Yeah. [78:02] >> And they are not NASDAQ stocks. These [78:05] are banks and insurance companies and [78:08] All right. So what you're describing [78:10] does have the ability to literally take [78:13] down the economy in the way the dot [78:16] capex blowoff top could not have. And [78:18] the problem is it's hard to quantify how [78:20] bad it will be because of the private [78:21] part of private credit. We don't [78:23] >> there's also stuff we don't even know [78:24] that people are doing. [78:25] >> And that's the thing there are these [78:27] kind of like guys just like yeah 10 [78:29] billion whatever I don't give a [78:31] Blackstone pissing money up the wall and [78:32] they're funding it with insurance money, [78:34] private credit money. Everyone's talking [78:35] about this Mark Wall situation. Cool. I [78:37] wrote about this three months ago with [78:39] the private credit situation. Like I [78:41] think a sixth of insurance annuities are [78:44] private credit funded. Uh we've got [78:46] public pension funds. We've got private [78:48] pension funds. We've got private [78:49] everybody's in. [78:50] >> Everyone's in. Yes. [78:51] >> I don't know how widespread the damage [78:53] is. I truly don't I don't want to be [78:55] like it's going to [78:56] >> but it seems really bad and people are [78:58] alarmed about it now. And I mean Pablo's [79:02] done a great job and Nick Neth has done [79:03] an amazing job pulling out. Yeah. Nick's [79:06] wrong. Oh, it's part of this world. [79:07] Nick's really great. Have done a good [79:10] job really illuminating this. And um Sam [79:12] Coppelman over at Hunter Brook legend. [79:13] >> We know we know [79:14] >> and um they've done a really good of [79:16] elevating this. Still people don't [79:18] connect that to data centers like they [79:20] should. And they don't connect it to the [79:21] fact that it openai and anthropic don't [79:24] even need to die. They don't even need [79:27] to stop growing. Like they could grow, [79:28] they won't, but like they could grow [79:29] like five times as big and they still [79:31] won't have the demand for these things. [79:32] And what happens if the thing doesn't [79:34] get built or runs over? When have you [79:36] heard of a building project that didn't [79:37] run over budget? And these are the most [79:40] ambitious [79:42] uh infrastructure projects of all time. [79:44] And so you've got private credit with [79:46] their wonderful underwriting quality [79:49] where most most just people winking at [79:51] each other. And you've got them funding [79:53] these big unstable difficult complex [79:56] projects that can get locals getting in [79:58] the way. They're cutting the tax breaks [80:00] in Arizona, Illinois, and somewhere [80:01] else. I can't remember. Texas, [80:04] Pennsylvania both showing opposition to [80:06] them and they need these things to [80:08] finish and the customer to exist to make [80:11] any money on the debt and with all the [80:13] weird senior tranches and all that who [80:15] knows where they'll actually be at the [80:16] end of it and they don't know because [80:17] they're investing in a Blackstone or [80:19] what have you Blackstone infrastructure [80:21] fund and it's just very dangerous in a [80:24] way that makes me really angry because [80:26] it could have been avoided. It could [80:29] have, but everyone got a might bit too [80:31] greedy. And Jensen as well has done so [80:33] well on this. But here's my real Nvidia [80:36] question, and I've been waiting to bring [80:37] this up. It feels like the right time. [80:39] How and when does Nvidia realize [80:41] revenue? That is a very interesting [80:44] question. When you actually look at [80:45] their earnings, those accounts [80:46] receivable have grown. Nobody special [80:48] found this. 55% sequentially. Like [80:50] what's going on there? Why are your [80:51] accounts receivable? Why are your daily [80:53] sales outstanding growing? [80:57] Like why is that? [80:57] >> We like to think of them as backlog. [80:59] >> But that's the thing though. Why is this [81:01] extending as you're making more money? [81:04] Is that because you're reaching the edge [81:05] of when you can get money quickly or at [81:07] all? When does that how much more [81:09] affordance are you going to have to make [81:11] your customers to grow to $690 billion? [81:15] Also, technically on an accounting [81:17] level, you can ship something and you [81:18] can count the revenue and no money can [81:20] pass. I don't know. Maybe you're feeling [81:22] generous with corewave. You really have [81:23] to disclose that. You might eventually, [81:25] but Collet Crest might just not send it [81:27] out as quickly. They're going to push [81:29] because we don't have an SEC right now. [81:32] They're going to push the absolute [81:34] limits of accounting shenanigans here. [81:36] It's It's going to It's already looking [81:38] really weird. Inventory is growing as [81:40] well. Nvidia has $30 billion of cloud [81:43] compute agreements, as in to rent back [81:45] their GPUs. They have $25 billion of [81:48] data center lease agreements. Why are [81:50] they you sell to the data center Jensen? [81:53] Why are you also renting the you just [81:55] are you just building a data center to [81:57] feed yourself money and this is the [81:59] ultimate point of the economy? This is [82:01] eventually they wanted to create [82:02] something where the only thing was just [82:04] handing yourself money but the money [82:05] increased by draining debt [82:07] >> right [82:07] >> and there are limits and I don't know [82:10] when we hit them but we will and I worry [82:13] that when we do it's just going to be so [82:15] abrupt and horrifying and the era [82:18] afterwards because [82:19] >> do you want to be wrong? Do you hope [82:21] you're wrong or are you [82:22] >> Yeah, because you do see like you do [82:23] seem to I don't want to say enjoy [82:26] reeling off all of these problems that [82:28] you have with it, but it it does seem [82:30] like you're enthusiastic to see this [82:32] thing. [82:32] >> I feel like you're going to be cackling [82:34] if this all comes down. [82:34] >> No, I So there's the was the bit in the [82:36] big short where it's like unemployment [82:38] drops 1% 30,000 people die, right? [82:40] >> I'm not going to I might be satisfied. [82:41] I'm not going to be happy. like I I [82:43] don't know if it's come across [82:43] >> but you're not going to take personal [82:45] pleasure in seeing [82:46] >> you but you will get some intellectual [82:48] satisfaction because out of having put [82:51] all your personal capital on the line so [82:54] to speak and and been right [82:56] >> but the thing is is like [82:58] >> re regular people are going to suffer on [83:00] this one retail [83:01] >> in that'll never that'll never change [83:02] >> retail investors though are barreling [83:04] into this what we're at the highest [83:05] point of retail leverage in history like [83:07] we're seeing a lot of people look at [83:08] what's happening with the Cosby it's [83:09] going to be like that in American like [83:11] it's there's something very scary about [83:12] that. That chills me. And I don't know. [83:15] I know I sound enthusiastic, but I am [83:16] like this comes from a place of like [83:18] worry because we could have stopped [83:21] this. We can't now. And now it's just [83:23] how big is the system going to let [83:25] itself grow before the system itself [83:27] runs out of cash. Cuz there's enough [83:29] cash. There's enough. But there are [83:32] limits to how quickly you can mobilize [83:33] it and where you can put it. And [83:34] everyone's asking for more. Everyone is [83:37] asking for more and more and more and [83:39] more. On that uplifting note, did you [83:42] did you at least have fun on the show [83:43] today? [83:43] >> I would love to come back. This is so [83:45] most of them are out of [83:46] >> No, listen. We we feel that these [83:48] conversations need to be had and there [83:50] will be people that are fans of our show [83:52] that will listen to this and say [83:54] >> don't ever have that guy on again. But [83:57] what if not, you're right and the world [84:00] comes to an end, but what if there is a [84:01] minor catastrophe and people listening [84:03] to the show were able to kind of [84:06] understand why it's happening in real [84:08] time because they listen to it. Yeah. [84:10] So, we do put bear. We've had Sam [84:12] Compliment on the show. I wouldn't call [84:13] him a bear, but he is [84:15] >> he did he did the Bloom Energy piece [84:17] back [84:18] >> and I've had Nick on on uh on the [84:20] YouTube live stream. So, I I think these [84:22] conversations are important to have. I [84:23] don't have to agree with everything you [84:24] say, [84:25] >> but I am learning from you and I hope [84:27] you've learned from us. [84:28] >> I have. [84:28] >> Okay. All right. You learned how how to [84:30] do a proper podcast. Minimal. All right, [84:33] >> dude. This has been this has been fun. [84:34] Such pleasure. [84:35] >> Fun might be the wrong [84:37] >> This has been intellectually [84:38] stimulating. [84:39] >> There we go. Yeah. Nice. Nice. [84:41] >> Guys, thank you so much for watching. [84:43] Thank you for listening. Where can [84:44] people go to learn more uh from you, Ed? [84:46] Tell people the URL. Where's your ed. at [84:49] for the newsletter, subscribe to the [84:50] premium, and of course, better [84:51] offline.com for links to the podcast. [84:54] >> Dude, you are a virtuoso at this. You [84:56] You crushed it. So, thank you. [84:57] >> I love podcasting. [84:58] >> All right. Thank you. Thank you guys. [85:00] We'll see you soon.