SpaceX Leases xAI Data Center to Anthropic as AI ARR Reaches $44B

SpaceX 向 Anthropic 出租 xAI 数据中心，AI ARR 达 440 亿美元 Elon Musk 与 Anthropic 的算力合作推动了这家 AI 初创公司前所未有的营收增长，同时让 Elon Web Services 成为传统超大规模云服务商的直接威胁。 主持人深入剖析了 Elon Musk 旗下 xAI/SpaceX 与 Anthropic 之间里程碑式的基础设施交易，分析了其对 Anthropic 惊人的 ARR 增长以及 SpaceX 估值的影响。随后讨论演变为一场激烈辩论：Anthropic 究竟是在利用安全言论谋求监管俘获并逐渐成为危险的垄断者，还是 AI 市场依然竞争激烈。此外，嘉宾们还评估了白宫的 AI 安全提案、硅谷的公关失误、宏观环境的“金发姑娘”（温和平衡）状态，以及企业在 AI token 上的支出何时才能转化为实际的企业利润率提升。 SpaceX 与 Anthropic 达成重大算力合作协议，Anthropic 整体租用 xAI 的 Colossus 1 设施（配备 22 万多块 GPU 及逾 300 兆瓦电力），由此催生出直接叫板传统云巨头的“Elon Web Services (EWS)”。 (4:38) 受代码开发需求的爆发式驱动，Anthropic 的 ARR 在 2026 年 4 月底跃升至 440 亿美元，预计年底逼近 1000 亿美元；David Sacks 预警未来可能出现一骑绝尘的“Mag 1”垄断格局，而 Brad Gerstner 则指出电力与基建瓶颈仍存限制。 (13:00) 针对白宫可能设立“AI 版 FDA”实行模型预审批的传闻与恐慌，嘉宾澄清白宫办公厅主任 Susie Wiles 已明确否认此类监管机构，并倡导通过高能力模型 KYC 身份验证机制在保障国家网络安全的同时避免扼杀创新。 (35:20) 三大云服务巨头（AWS、Azure、GCP）财报表现强劲，单季度合计新增约 300 亿美元 ARR，以强劲的企业级 AI 采用率有力回击了市场对 AI 泡沫与纯投机论调的担忧。 (60:04)

The All-In hosts examine how massive infrastructure deals, surging model revenues, and regulatory flashpoints are reshaping artificial intelligence, highlighted by Anthropic leasing Elon Musk's Colossus 1 compute facility. While David Sacks and Brad Gerstner emphasize unprecedented expansion and enterprise adoption, Chamath Palihapitiya warns of unresolved supply constraints, public backlash, and an approaching return-on-investment reckoning for enterprise software buyers. Compute Bottlenecks and 'Elon Web Services' Anthropic secured more than 220,000 Nvidia GPUs and over 300 megawatts of power by leasing xAI’s Colossus 1 data center from Elon Musk, immediately doubling rate limits for Claude Code and lifting Opus API caps. Meanwhile, xAI shifted its internal frontier model training pipelines to Colossus 2. Chamath Palihapitiya asserts that revenue growth at top AI labs is strictly supply-constrained by electrical power and data center availability rather than enterprise demand, noting that roughly half of the 9 gigawatts of planned data center additions this year face regulatory or activist resistance. The panel views the transaction as the functional debut of 'Elon Web Services' (EWS) competing against AWS, Microsoft Azure, and Google Cloud. Brad Gerstner estimates this compute leasing strategy could generate up to $45 billion in annual revenue for SpaceX, providing high-margin cash flow to subsidize ongoing training expenditures. Furthermore, the hosts note that distributed compute initiatives—such as pairing residential developments with micro-data center clusters—and Tesla’s battery and solar manufacturing give Musk unique structural advantages over legacy cloud hyperscalers. Anthropic's Growth and the Monopoly Debate David Sacks highlights Anthropic’s rapid commercial acceleration, reporting that its annualized run rate expanded from $10 billion on January 1 to $30 billion by March 31, and reached $44 billion by late April. Sacks projects Anthropic could finish the year near $100 billion ARR and potentially reach $1 trillion by 2027. Because the combined revenue of the 'Magnificent 7' tech giants is approximately $2.3 trillion to $2.35 trillion, Palihapitiya notes that a single firm generating $1 trillion would establish an unprecedented 'Mag 1' market concentration. The software development market is seen as the primary driver, where Jason Calacanis estimates $1 trillion in annual global developer spending could expand to $2 trillion through AI-driven coding gains. Sacks warns that Anthropic could secure an unassailable monopoly within 18 months, comparing its emphasis on safety regulations to John D. Rockefeller branding kerosene as 'Safe Oil' to secure regulatory capture and block competition. He points to Anthropic's restrictions on third-party model wrappers like OpenClaw as an anti-competitive indicator. Gerstner strongly pushes back, arguing that Anthropic and OpenAI maintain roughly equal GAAP run rates and remain fragile startups competing against entrenched, cash-rich incumbents in a vibrant market that requires no preemptive government intervention. The 'FDA for AI' Policy Panic (35:20) Reports that the White House is considering an executive order to form an AI working group sparked debate over a prospective pre-release review agency for frontier models. Media reports linked the administration's scrutiny to a leak of Anthropic’s advanced cyber-warfare model, *Mythos*. Despite public comments by National Economic Council Director Kevin Hassett analogizing oversight to FDA drug reviews, Gerstner reports speaking directly with Hassett, who clarified that the administration does not endorse a mandatory pre-approval board. Sacks adds that Chief of Staff Susie Wiles explicitly dismissed the concept, pointing instead to the administration's March 20, 2026 legislative framework prioritizing narrow statutory rules over executive overreach. Addressing cyber risks, Sacks contends that offensive and defensive capabilities are technically symmetrical and that rival domestic and foreign labs will reach parity with *Mythos* within three to six months. Rather than establishing bureaucratic bottlenecks, he advocates implementing Know Your Customer (KYC) identity verification on developer APIs and releasing frontier models to commercial cybersecurity vendors like CrowdStrike and Palo Alto Networks. Meanwhile, Palihapitiya warns that Silicon Valley faces mounting political vulnerability, grading tech leadership's public messaging a 'D-minus' for failing to demonstrate broad economic benefits while Main Street worries about job displacement and wealth concentration. Enterprise ROI and Macro Market Valuations (60:04) Cloud infrastructure spending showed sharp acceleration, with AWS reaching a $150 billion annual run rate (28% growth), Azure hitting $108 billion (39% growth), and Google Cloud generating $80 billion (63% growth). Gerstner argues the broader market remains fundamentally sound rather than bubble-inflated, citing moderate valuations among hyperscalers (Meta at 17x GAAP earnings, Nvidia at 19x, Microsoft at 20x, Alphabet at 24x) and low multiples for memory suppliers such as SK Hynix (5x), Samsung (6x), and Micron (7x). He also highlights that S&P 500 operating margins widened from 11% in 2023 to 13% in 2026, while Big Tech headcount grew only 3% over three years. Conversely, Palihapitiya cautions that enterprise software spend faces a reckoning within roughly 500 days. He argues that outside of tech infrastructure suppliers, there is minimal evidence that token expenditures have lifted the operating margins of broader S&P 500 corporations, meaning enterprises must eventually demonstrate measurable cost reductions or top-line gains. The hosts also split on labor market impacts: Sacks cites low 4.2% unemployment and improving job placements for recent graduates as evidence against AI displacement, whereas Calacanis points to a depressed labor force participation rate of 61.9% as an indicator that headline employment statistics mask labor market friction. While the panel agrees that capital expenditure and cloud consumption are driving historic top-line growth for infrastructure suppliers and frontier labs, a fundamental divide remains unresolved: whether enterprise AI investments will translate into tangible corporate profitability across the broader economy before compute constraints, public skepticism, or regulatory interventions slow the sector's momentum.

本期节目聚焦Anthropic与SpaceX达成的大规模算力租赁协议，探讨了“埃隆网络服务”重塑云基础设施格局、前沿AI实验室营收爆发式增长背后的垄断隐忧，以及围绕“AI版FDA”的监管恐慌与宏观投资回报大考。 算力基建交易与埃隆网络服务崛起 Jason Calacanis披露，Elon Musk将xAI的Colossus 1数据中心全量租予Anthropic，为其提供超22万张英伟达GPU及300兆瓦以上电力支持。Anthropic随即宣布Claude Code速率限制提高一倍并解除付费用户高峰限制，而xAI则将训练迁移至Colossus 2。这一交易催生了可与AWS、Azure和谷歌云直面竞争的“埃隆网络服务”（EWS）。 Chamath Palihapitiya分析认为，前沿AI实验室的收入瓶颈完全受制于数据中心和电力供给而非终端需求，目前年内计划上线的9吉瓦数据中心容量中近半数遭遇政治抗议或阻滞。Brad Gerstner估算，出租地面算力每年可为SpaceX带来高达450亿美元的增量收入，不仅能对冲天基数据中心的不确定性，更实质性补贴了xAI高昂的研发支出，推动SpaceX在IPO前夕获得40至50倍的高额远期市销率估值溢价。 超级增长假说与单一巨头垄断之辩 (16:22) David Sacks指出，Anthropic的年化经常性收入（ARR）呈现惊人飞跃，从年初的100亿美元跃升至3月底的300亿美元，4月下旬进一步升至440亿美元，预计年底将接近1000亿美元，并提出其有望在2027年实现1万亿美元ARR并形成空前垄断。Chamath则列举2025年“美股七巨头”总计约2.3万亿美元的收入基数，指出若Sacks预测成真，市场将由“七巨头”演化为前所未见的“单一巨头”（Mag 1）。 围绕这一前景，Sacks将Anthropic对安全监管的倡导比作标准石油公司的垄断策略，批评其限制OpenClaw等第三方接入的行为是意图打造监管护城河。Gerstner则坚决反对垄断定性，强调Anthropic与OpenAI营收规模目前旗鼓相当且均处于初创脆弱期；面对全球每年约1万亿美元的软件工程师薪酬市场，谷歌、OpenAI及xAI等正全力反扑，过早出台反垄断干预将严重扼杀充满活力的本土创新。 AI版FDA监管恐慌与国家安全防线 (35:20) 媒体此前报道白宫正考虑颁布行政令建立“AI版FDA”以应对Anthropic“Mythos”模型外泄引发的网络攻击担忧。Gerstner澄清其与国家经济委员会主任Kevin Hassett的沟通，确认政府仅是将FDA作为漏洞协同的借喻，绝无推行前置审批官僚体制的计划；Sacks亦指出白宫办公厅主任Susie Wiles已公开否认设立此类机构，重申白宫更倾向于针对儿童保护与深度伪造的精准立法。 针对前沿模型的网络武器风险，Sacks主张网络攻防具备双重属性，一味压制发布只会延误防御部署；解决之道在于推行强制性的“了解你的客户”（KYC）身份核验机制，并加速将工具交付给CrowdStrike等本土安全厂商。Chamath则警示，科技界若无法向公众阐明AI带来的普惠收益，反而过度渲染末日风险与财富分化，必将引发大众抵触并招致严厉监管。 云计算繁荣与企业端ROI商业大考 (60:04) 巨头财报展现出强劲的AI需求拉动：AWS年化运行率达到1500亿美元（同比增长28%），Azure达1080亿美元（增长39%），谷歌云增长63%至800亿美元，仅三大厂商单季便新增约300亿美元ARR。同时，SK海力士、三星与美光等芯片厂商仅以5至7倍市盈率交易，Meta与英伟达市盈率亦落在17至19倍区间，Gerstner据此认为当前市场绝非投机泡沫，而是受业绩驱动的“金发女孩”宏观环境。 然而，Chamath发出了清醒的冷思考，强调截至目前尚无确凿证据证明AI有效提升了标普500成分股的整体营业利润率。他预测在未来约500天内，企业将迎来投资回报率（ROI）大考：巨额采购算力的传统企业必须证明AI能够实质性削减运营成本或推动营收增长。尽管Gerstner以标普500营业利润率升至13%及代码生成工具替代外包为例反驳，企业能否跨越实验性支出转化为真实利润仍是市场后续发展的核心检验标尺。 本期节目梳理了算力基建扩张与宏观科技资本的强劲动能，但针对超前垄断的定性争议、安全监管与创新的博弈，以及企业端Token开支能否在500天内兑现为实质性财务利润率，构成了当前AI繁荣期悬而未决的核心命题。

## Background 本期内容由知名科技投资播客《All-In Podcast》于 2026 年 5 月 27 日在 YouTube 平台发布。节目简介重点导读了洛杉矶市长选举选情、马斯克旗下 xAI 将 Colossus 1 算力中心租给 Anthropic 的世纪大单与“埃隆云服务（EWS）”崛起、Anthropic 是否成为下一个超级垄断组织、白宫关于“AI 版 FDA”审批制的恐慌与真相，以及当下 AI 繁荣对宏观经济与交易市场的影响。 ## TL;DR 节目围绕马斯克将庞大算力集群租给竞争对手 Anthropic 展开深度推演，剖析此举不仅化解了 xAI 的天量资本开支、奠定 EWS 算力服务地位，更助力 Anthropic 年化经常性收入（ARR）暴涨至 440 亿美元，引发了其是否成为“人类史上最强垄断组织”的激烈辩论。随后，嘉宾们澄清了白宫内部所谓“AI 领域的 FDA”监管传闻，指出行政层面绝无搞事先准入牌照的意图，核心挑战在于网络安全攻防与企业自律防御。最后，几位投资人深入复盘宏观经济与科技巨头财报，针对 AI 到底是否已经实质性拉高标普 500 企业的真实营业利润率，以及未来 500 天后终端企业商业投资回报率（ROI）是否面临清算时刻爆发了针锋相对的交锋。 ## Atmosphere & style 全场氛围兼具顶尖科技投资人的敏锐博弈与老友拌嘴式的辛辣幽默，既有像“洛克菲勒把标准石油包装成安全石油”这种一针见血的犀利讽刺，也有围绕市盈率、GAAP 盈利和算力电网指标展开的硬核推演。随着关于 AI 究竟是“既成事实的生产力革命”还是“缺乏宏观利润证据的资本狂欢”的争辩层层深入，演播室内的对抗张力一度拉满，随后又在互相调侃拟任“开司米与红酒部长”的插科打诨中轻松收场，让听众在密集的认知轰炸与硅谷内幕爆料中获得极具代入感的畅快体验。 ## Bulletin - [04:38](https://www.youtube.com/watch?v=10MdOvK-aG4&t=278s) 马斯克旗下 xAI 将整个 Colossus 1 超算中心（超 22 万张英伟达 GPU、逾 300 MW 电力）全面租赁给 Anthropic，立即使 Claude Code 限额翻倍，并标志着“埃隆云服务”（EWS）正式进军超大规模云市场。 - [14:09](https://www.youtube.com/watch?v=10MdOvK-aG4&t=849s) Anthropic 展现出科技史上罕见的指数级爆发，其 ARR 在短短数月内从 100 亿美元暴增至 440 亿美元，被嘉宾戏称为正在走向“人类历史上最强大的垄断组织”。 - [43:02](https://www.youtube.com/watch?v=10MdOvK-aG4&t=2582s) 针对外界对白宫将设立“AI 版 FDA”进行事前审批的恐慌，嘉宾澄清这属于媒体炒作与稻草人谬论，白宫主张针对具体安全漏洞施策而非设立全局官僚准入机构。 - [01:00:05](https://www.youtube.com/watch?v=10MdOvK-aG4&t=3605s) 三大超大规模云服务商交出亮眼财报，单季度合计新增近 300 亿美元 ARR，且标普 500 指数营业利润率跃升至 13.0%，但关于 AI 是否已真正为终端企业创造实质净利润仍引发激烈交锋。 ## Key takeaways - [04:38](https://www.youtube.com/watch?v=10MdOvK-aG4&t=278s) xAI 与 Anthropic 达成重磅算力租赁协议，不仅缓解了 Anthropic 的算力饥渴，也让马斯克将庞大的前期 CapEx 转化为可预测的高利润现金流，并为 SpaceX IPO 筑牢估值底座。 - [06:00](https://www.youtube.com/watch?v=10MdOvK-aG4&t=360s) 大模型厂商的营收天花板不在于终端需求，而完全受制于物理层面的电力供给与数据中心建设瓶颈，全美约 50% 计划投产的算力电力正遭遇阻力。 - [08:53](https://www.youtube.com/watch?v=10MdOvK-aG4&t=533s) SpaceX 正在形成集火箭发射、Starlink、EWS 算力服务、太空数据中心及上层应用为一体的五层业务架构，今年算力租赁预计带来 40 亿至 50 亿美元增量收入。 - [17:48](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1068s) Anthropic 依靠深耕代码生成这一核心企业变现通道，ARR 闪电跃升至 440 亿美元，相较于分散探索各类多模态玩具模型的竞争对手筑起了极高的壁垒。 - [20:25](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1225s) 马斯克展现了将边缘储能、Tesla 车辆计算、住宅断路器面板及星链串联的野心，有望催生规避地面繁重监管与能源约束的分布式及太空算力体系。 - [26:48](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1608s) 大卫·萨克斯提出洛克菲勒“安全石油”思想实验，警惕头部 AI 企业借“AI 安全”之名游说监管准入标准，以此阻击中小竞争者并构筑行政垄断护城河。 - [30:01](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1801s) 针对垄断定性，布拉德·格斯特纳认为按照 GAAP 会计准则前沿实验室依然脆弱，谷歌和亚马逊等巨头拥有海量自由现金流，政府过早选定赢家输家将扼杀创新活力。 - [39:55](https://www.youtube.com/watch?v=10MdOvK-aG4&t=2395s) 科技界遭遇公众舆论危机与“氛围剧变”，其根源在于未能向全美民众清晰阐释 AI 的普惠红利，反而让公众滋生了赢者通吃和极端贫富差距的抗体心态。 - [45:06](https://www.youtube.com/watch?v=10MdOvK-aG4&t=2706s) 针对 Mythos 等具备强大渗透能力的军民两用攻防模型，行业应在封闭预览期落实严格的 KYC 身份验证与快速补丁分发，而非诉诸低效的中心化官僚审查。 - [52:01](https://www.youtube.com/watch?v=10MdOvK-aG4&t=3121s) 为扭转全美公众的 AI 焦虑，嘉宾建议推行全民 IPO 红利共享（如将 1%-5% 股份注入儿童投资账户）、聚焦健康长寿与教育普惠，以切实利好换取社会认同。 - [56:27](https://www.youtube.com/watch?v=10MdOvK-aG4&t=3387s) 民意调查显示 AI 仅在选民关注的 39 个议题中排名第 29 位，而数据中心建设掀起的实体蓝领繁荣正在为建筑工人和整体经济注入强劲动力。 - [01:01:32](https://www.youtube.com/watch?v=10MdOvK-aG4&t=3692s) 超大规模云服务商 AWS、Azure 和 GCP 均维持 28% 至 63% 的高速增长，纳斯达克核心标的估值并未过度透支，宏观环境整体呈现平稳的“金发女孩”格局。 - [01:07:07](https://www.youtube.com/watch?v=10MdOvK-aG4&t=4027s) 查马斯警告市场可能面临“500 天大考”：目前尚无直接证据表明 AI 普遍提升了标普 500 终端企业的营业利润率，未来 2 到 3 年企业必须在财报中证明真实的商业 ROI。 - [01:16:15](https://www.youtube.com/watch?v=10MdOvK-aG4&t=4575s) 全美失业率维持在 4.2% 的历史低位且年轻大学毕业生就业状况呈现改善，但劳动参与率相较疫情前依然存在缺口，AI 对劳动力市场的全貌冲击仍在演进中。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=10MdOvK-aG4&t=0s) 闲聊开场：最佳好友齐聚与洛杉矶市长选情热议 节目在轻松的互相调侃中开场。杰森·卡拉卡尼斯（Jason Calacanis）测试着自己的音频和出镜状态，查马斯·帕利哈皮蒂亚（Chamath Palihapitiya）打趣道他现在的状态“长相适合做广播”，不过比起前几周看起来没那么疲惫了。杰森笑称观众在评论区对他眼袋的攻击十分无情，甚至感叹“还好老子有钱”。本期节目中，大卫·弗里德伯格（David Friedberg）因病缺席，被戏称为得了“社会主义相关的流感”，预计下周将携两场重量级专访回归；“第五位好友”阿尔特米尔资本（Altimeter Capital）创始人布拉德·格斯特纳（Brad Gerstner）受邀补位，与查马斯、大卫·萨克斯（David Sacks）一同坐镇。 随后几人迅速转向当下的热点政治议题——洛杉矶市长选举。布拉德由衷赞叹竞选广告的制作水准，直言这是他见过最顶尖的政治广告之一，背后的社交媒体与视频制作团队堪称“次世代级”。萨克斯指出，不仅广告吸睛，候选人在辩论现场的表现同样极具杀伤力。面对极左翼的现任市长凯伦·巴斯（Karen Bass）以及立场更为激进、被萨克斯调侃为“印度版菲德尔·卡斯特罗”的市议员尼蒂亚·拉曼（Nithya Raman），候选人针对洛杉矶的无家可归危机进行了犀利反击。 拉曼试图将流浪汉危机的责任推给市长，却被当场揭穿她自己正是负责这些项目的主管官员。他在辩论中一针见血地指出，洛杉矶的核心问题根本不是“缺少住房”，而是严重的毒瘾和精神疾病问题，甚至直言如果拉曼一个人走上街头，在像贫民区（Skid Row）那种地方恐怕会被人拿刀刺伤颈部。导播切出一段在网络上疯传的辩论切片，面对拉曼指责那是“MAGA 共和党人眼中的洛杉矶幻想”，他只是耸肩摊手、无奈晃头，这一极具喜剧效果的临场反应让主持人们大笑不已。查马斯提到，加州即将进行涉及退休保护与储蓄法案的公投，极有可能高票通过并以此封杀财富税。如果加州能通过宪法修正案保护个人资产并选出务实的领导层，将向全美释放极其强烈的非共识积极信号；而萨克斯则幽默地表示，在愿景成真之前，自己还是老老实实待在得克萨斯州。 ### [03:40](https://www.youtube.com/watch?v=10MdOvK-aG4&t=220s) 治安恶化与富豪迁徙：肯·格里芬撤离纽约 话题随后延伸到美国大城市的治理恶化与针对商业领袖的激进政治风向。几周前纽约左翼政治人物佐兰·曼达尼（Zohran Mamdani）专门跑到城堡投资（Citadel）创始人肯·格里芬（Ken Griffin）的私人豪宅门前录制攻击视频并公开发布。在近期联合健康（UnitedHealth）高管因意识形态动机遭枪击、山姆·奥特曼（Sam Altman）住宅遭针对等一系列极端事件的阴影下，这种公然公开富豪住址并进行政治煽动的行为极其危险。 肯·格里芬在随后的采访中明确表态，出于对自身及家人人身安全的严重担忧，他将彻底退出纽约，并将未来的发展重心全盘转移至佛罗里达州。这一幕正如当年格里芬因治安和税负问题将城堡总部从芝加哥搬迁一样。而曼达尼在面对质询时仅仅给出了支支吾吾的敷衍回应，毫无歉意甚至变本加厉。主持人们感慨，若这种针对企业家的敌视与治安恶化持续蔓延，纽约也正在沦为被高净值人群抛弃的“飞越之城”。 ### [04:38](https://www.youtube.com/watch?v=10MdOvK-aG4&t=278s) 重磅交易：马斯克将 Colossus 1 租给 Anthropic 与 EWS 崛起 杰森抛出了本期的重磅科技头条：埃隆·马斯克（Elon Musk）旗下的 xAI 已将整个超级计算中心 Colossus 1 全面租赁给了 Anthropic。查马斯五天前曾在节目中预言马斯克与 Anthropic CEO 达里奥·阿莫代伊（Dario Amodei）应该立刻展开合作，如今这一戏剧性联盟迅速落地。由于 Anthropic 一直面临极度严苛的算力短缺，这笔交易直接为其注入了超过 22 万张英伟达 GPU 以及逾 300 兆瓦（MW）的电力供应。 ```mermaid flowchart TD accTitle: SpaceX 与 xAI 的算力分层与商业化架构 accDescr: 展示马斯克通过 Colossus 数据中心向外租赁算力并形成 EWS 超级云的结构 subgraph Infrastructure["算力基础设施层"] C1["Colossus 1 算力中心<br/>(220,000+ Nvidia GPUs / 300+ MW)"] C2["Colossus 2 算力中心<br/>(聚焦下一代 Grok 训练)"] end subgraph Commercialization["EWS (Elon Web Services)"] Lease["外部算力商业租赁"] Anthropic["Anthropic (Claude 模型)"] Cursor["Cursor 及其他 AI 开发者"] end subgraph InternalTech["内部模型迭代"] Grok["xAI 自研 Grok 模型"] end C1 --> Lease Lease --> Anthropic Lease --> Cursor C2 --> Grok ``` 这笔交易的影响立竿见影： - **Claude Code 速率翻倍**：此前由于算力吃紧频频遭遇限流的 Claude Code，调用速率上限即刻提升一倍。 - **取消高峰使用限制**：付费用户的峰值使用限制被直接移除。 - **提升旗舰 API 吞吐**：大幅增加了面向企业与开发者的 Opus 模型 API 并发额度。 - **xAI 战略重心转移**：xAI 自身的主力模型训练已全面迁移至更为庞大的 Colossus 2。 继上周与 AI 编程工具 Cursor 达成合作之后，马斯克在此次交易中展现出惊人的算力基建落地速度。杰森将其正式命名为“埃隆云服务”（EWS, Elon Web Services），这意味着马斯克不仅自研模型，更正式以超大规模云厂商（Hyperscaler）的姿态，与亚马逊 AWS、微软 Azure 以及谷歌云展开正面交锋。 ### [06:00](https://www.youtube.com/watch?v=10MdOvK-aG4&t=360s) 查马斯拆解逻辑：AI 算力供给瓶颈与 SpaceX 的估值支撑 查马斯深入剖析了这笔交易背后的底层商业逻辑。他指出，市场目前过度关注 Anthropic 或 OpenAI 的短期季度营收预测，但实际上这两家头部大模型厂商的业绩表现“与需求毫无关系，需求是无限的，其营收完全受限于数据中心和电力的供给瓶颈”。如果拥有无限电力，它们的收入曲线将呈现更加夸张的抛物线式爆发。 ```mermaid flowchart LR accTitle: AI 数据中心供给约束与马斯克基建套利 accDescr: 分析 AI 厂商因电力短缺陷入瓶颈而马斯克凭借现成电力成为造王者 subgraph Bottleneck["行业面临的严峻瓶颈"] A["今年原定投产 9 GW 算力电力"] --> B["社区抵制与政治审查<br/>(近 50% 项目遭遇阻力)"] B --> C["电力供给极端短缺<br/>算力扩张严重受限"] end subgraph Advantage["马斯克的套利护城河"] D["提前锁网并建成 Colossus 规模基建"] --> E["掌控稀缺电力与 GPU 现货"] E --> F["对冲太空数据中心不确定性<br/>强化 SpaceX IPO 估值底座"] end C -.-> E ``` 与此同时，整个行业正遭遇严峻的现实挑战： - **公共关系与领导力危机**：主流 AI 企业的领导层在公众与监管层眼中显得缺乏信任度且过于功利，导致舆论与社区抵触情绪高涨。 - **电力落地严重受阻**：全美今年原计划上线的约 9 吉瓦（GW）数据中心电力指标中，有将近 50% 正在遭遇地方社区的抗议与阻挠，按历史规律大部分最终可能会被叫停，令算力短缺进一步恶化。 在这样的宏观背景下，马斯克的远见展现出了巨大的战略价值。查马斯分析，在 SpaceX 筹备 IPO 的路演过程中，空头或红队审视其高估值时，最大的质疑点往往在于所谓“轨道太空数据中心”的远期叙事能否兑现。而马斯克如今通过抢先锁网、光速建成的地面数据中心直接拿下巨额现货客户，建立了一个极其坚固且高利润的地面基建业务。即便太空数据中心的研发推迟数个季度，地面 EWS 业务产生的充沛现金流也能从容反哺并补贴 Grok 模型的巨额训练成本，使马斯克在算力战局中扮演起左右逢源的“造王者”（King-maker）。 ### [08:53](https://www.youtube.com/watch?v=10MdOvK-aG4&t=533s) 布拉德·格斯特纳点评：五层蛋糕架构与百亿美元级增量收入 布拉德·格斯特纳对查马斯的观点表示高度赞同，并感叹“在这个星球上，没有人比埃隆更擅长把电子（Electrons）转化为代币（Tokens）”。他援引红杉资本合伙人肖恩·马奎尔（Shaun Maguire）对 SpaceX 业务架构的比喻，指出 SpaceX 实际上由五层精密业务叠加而成： 1. **火箭发射**（Launch） 2. **星链网络连接**（Connectivity） 3. **算力超大规模服务商**（Compute Hyperscaler，即 EWS） 4. **轨道太空数据中心**（Space Data Centers） 5. **上层应用与基础模型及其他前沿押注**（Applications, Models & Other bets） 布拉德分析，在此次路演中，外界此前最大的困惑在于：尽管 xAI 投入了海量资本开支，但其自研模型的商业化收入规模并未达到 OpenAI 或 Anthropic 的爆发速度。而马斯克亮出的这一张底牌——出租底层算力成为 EWS，瞬间化解了这一结构性矛盾。据布拉德初步测算，在华尔街分析师对 SpaceX 预期的约 250 亿美元基准营收之上，仅今年这一项算力租赁业务就将为马斯克带来高达 40 亿至 50 亿美元的超预期增量收入。这笔巨大的现金流不仅能完全覆盖前期的重资产投入，也彻底夯实了整体业务的财务弹性。 ### [09:56](https://www.youtube.com/watch?v=10MdOvK-aG4&t=596s) 算力租赁的双赢棋局与数据中心抗议的真相 马斯克通过将其算力资产商业化，巧妙地化解了巨额前期投资带来的资金压力。马斯克旗下布局了三大超级计算中心：Colossus、Macrohard 以及 Macroharder，其中后两者配备了高达 1.2 吉瓦（GW）的英伟达 Blackwell 芯片。在此次合作中，他将互联带宽要求相对较低、但极其适合做推理（Inference）任务的 H100 集群转租给了 Anthropic。这笔交易不仅让马斯克获得了巨额变现，更为 Anthropic 注入了急需的算力燃料。最关键的是，它解决了外界对 xAI 估值模型中最大的担忧——即资本开支（CapEx）严重超前于实际收入的倒挂问题，使 xAI 不再背负短期内必须拿出亮眼营收的巨大压力，一跃成为超大规模算力租赁市场的有力竞争者。尽管外界此前普遍唱衰电力短缺、算力告急、商业化不及预期，但正如嘉宾所言，“美国资本主义的混乱总能以某种方式找到出路”，Anthropic 团队与马斯克达成的这一“破冰停火”对美国维持前沿模型竞争力至关重要。 与此同时，针对全美多地反对修建数据中心的抗议浪潮，嘉宾直指其背后并非自发的本地社区诉求，而是高度组织化、在全美流窜挑起事端的激进环保团体，其操作手法与三十年前摧毁美国核电产业的策略如出一辙。当年激进活动导致美国整整一代人几乎没有新建任何核反应堆，反观中国如今已投产和在建上百座核电站。有关数据中心“耗尽地下水”和“推高民用电价”的指控纯属虚假信息：现实中电费飞涨的正是没有新增电网供给的纽约州和加州，而在全美建设数据中心最密集的得克萨斯州，电力成本反而呈现下降趋势。 ```mermaid flowchart TD accTitle: xAI与Anthropic算力合作协同架构 accDescr: 展示xAI与Anthropic在算力、资金和模型层面的互利交换逻辑 subgraph MuskEcosystem ["马斯克生态体系 (xAI / SpaceX)"] A["算力基础设施<br/>Colossus / Macrohard (H100/Blackwell)"] B["短期痛点：巨额CapEx缺乏软件营收支撑"] C["化解方案：转租闲置/推理算力至第三方变现"] end subgraph AnthropicNeed ["Anthropic 前沿模型体系"] D["核心瓶颈：爆发式增长遭遇算力极度短缺"] E["承接算力：获得大规模H100集群用于推理"] F["商业飞轮：支撑编码代理爆发，释放指数级ARR"] end A -->|"转租推理算力"| E F -->|"支付巨额租赁收入"| C C -->|"填补资产负债表亏损"| B ``` --- ### [12:31](https://www.youtube.com/watch?v=10MdOvK-aG4&t=751s) xAI 与 Anthropic 的互补解题：萨克斯祝贺达里奥“赢得 AI 竞赛” 大卫·萨克斯（David Sacks）深入拆解了这桩交易在财务层面的极度互补性。SpaceX 旗下的航天发射与 Starlink 卫星通信业务盈利丰厚，而并入其中的 xAI 此前却面临着惊人的亏损。究其根本，前沿大模型的训练需要消耗天价算力集群，但在模型真正跻身第一梯队前几乎没有任何进账；雪上加霜的是，当前整个 AI 市场绝大多数的企业端商业化收入都被“编程（Coding）”这一核心场景所垄断。虽然 xAI 刚刚与 Cursor 达成合作试图弯道超车，但自身仍缺乏杀手级的编程落地产品，陷入了“承担全部基础设施成本，却无法分享分毫营收蛋糕”的被动局面。将过剩算力租赁给 Anthropic，让马斯克既保留了追逐前沿模型的资格，又成功剥离了沉重的未兑现资本开支，极大修复了资产负债表。 反观 Anthropic，算力一直是卡住其高速狂飙的唯一脖子。这一合作促使萨克斯在节目中公开表态：“让我成为第一个祝贺达里奥·阿莫代伊（Dario Amodei）赢得本轮 AI 竞赛的人。”此言一出，主持人立刻打趣萨克斯过去在节目和社交平台 X 上对 Anthropic 及达里奥的尖锐批评——当时萨克斯常讽刺达里奥是个“110% 的末日论安全狂（P-doomer）”。面对如今两大巨头握手言和的“AI 中东和平时刻”，萨克斯坦言必须尊重客观商业现实：Anthropic 正在展现科技史上前所未有的统治力。 - SpaceX 现金流充裕，但单靠其无法长期无底线补贴 xAI 庞大的芯片与电力研发亏损。 - 编程代码生成目前是企业级 AI 市场变现的核心通道，xAI 在此领域缺乏直接抗衡的产品。 - 算力租赁在不动摇马斯克底层掌控权的前提下，将闲置 GPU 产能转化为可预测的现金流。 --- ### [14:09](https://www.youtube.com/watch?v=10MdOvK-aG4&t=849s) 史无前例的指数级爆发：从 100 亿到 440 亿 ARR 的“人类最强垄断” 萨克斯详细列举了 Anthropic 震撼硅谷的爆炸式财务数据。在过去三年里，Anthropic 始终保持着每年 10 倍（10x）的增长速度。原本外界普遍认为，随着业务基数急剧膨胀，这种复合增速在物理上绝不可能持续，然而今年前四个月的数据彻底打破了所有传统认知： - 第一季度（1月1日至3月31日），其年度经常性收入（ARR）从约 100 亿美元直接三倍跳跃至 300 亿美元； - 4 月单月增长不仅未放缓，反而进一步加速，ARR 闪电冲上 440 亿美元。 硅谷从业者虽然习惯了指数级增长，但从未在如此巨大的体量下目睹过这种速度。算力短缺曾是其前行路上唯一的物理阻碍，随着与马斯克及其他供应商的大规模算力协议落地，Anthropic 年底以 1000 亿美元 ARR 收官已几乎成为定局，甚至引发了其能否在 2027 年突破 1 万亿美元营收大关的激辩。萨克斯断言：“如果按照当前的势头发展下去，达里奥称之为 AGI，而我称之为人类历史上最强大的垄断组织。” 查马斯（Chamath Palihapitiya）随即列出 2025 年美股科技七巨头（Mag 7）的年营收数据进行直观对比：苹果 4200 亿美元、微软 3000 亿美元、Alphabet 3900 亿美元、亚马逊 7000 亿美元、英伟达 1900 亿美元、Meta 1850 亿美元、特斯拉 1100 亿美元，七巨头总营收合计约为 2.3 万亿至 2.35 万亿美元。换言之，若 Anthropic 真能单体斩获万亿美元年收入，它将不再是跻身什么“Mag 8”，而是让整个科技格局沦为 Anthropic 一家独大的“Mag 1”。不过布拉德（Brad）也提出审慎视角：达里奥此前在 Dwarkesh Patel 的播客中预计，到 2029 年整个 AI 领先梯队的合计营收才可能达到 1 万亿美元；且从底层物理电力推算，Anthropic 计划在今年底获得 5 GW 算力支持、明年底达到 10 GW，单家企业在短时间内能否吸纳并转化如此规模的电力与芯片产能仍需观察。 --- ### [17:48](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1068s) 刺猬与狐狸的对决：聚焦编程代币与万亿美元 TAM 的竞争新局 讨论聚焦于市场总容量（TAM）能否支撑起这一万亿级庞然大物。据测算，全球仅在软件开发工程师薪资及软件工程构建相关的现存支出每年就高达 1 万亿美元；而当高效低廉的代码代币（Coding Tokens）将软件生产效率和业务价值放大十倍甚至百倍时，整个软件市场的实际规模完全有望翻倍至 2 万亿美元，足以容纳万亿美元规模的超大型企业。 面对巨大的蛋糕，大厂之间的战略定位差异成为胜负分水岭。萨克斯引用了经典的“刺猬与狐狸”哲学隐喻： - **刺猬（Anthropic）：** 奉行一元论，自始至终死磕编程底座这一终极杀手级场景，并在此基础上衍生出 Co-work、企业级 Agent 等高价值应用。 - **狐狸（其他前沿模型实验室与巨头）：** 试图面面俱到，把大量算力与工程资源耗费在诸如 Nano Banana 模型、Sora 视频生成、图像创作以及虚拟人角色陪伴聊天等分散方向上。 事后看来，那些五花八门的尝试在创造坚实企业现金流方面收效甚微。真正的商业主战场已被证明就是代码生成及其驱动的自动化生态。尽管 OpenAI 等同行已经警醒并迅速调头重新聚焦编程场景，但在这场关乎万亿美元企业级入口的争夺战中，全速狂奔的 Anthropic 已经筑起了极高的竞争壁垒，对手能否夺回流失的市场份额依然充满变数。 ### [19:29](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1169s) AI 赛道的竞争动态：OpenAI 与 Google 的反扑，以及 Anthropic 的先发优势 围绕底层大模型与编程工具的较量正在加剧。Sacks 提到，基于下一代基座模型“Spud”的 GPT-5.5 已经让其 Codex 代码能力备受好评，且增长曲线重新出现加速势头。OpenAI 显然具备夺回市场份额的硬实力，而 Google 凭借自身顶级工程团队在代码领域的深厚积累，也绝不会在这场竞争中掉队。再加上 Elon Musk 近期与 Cursor 达成合作，多方围剿之势愈发明显。 然而，从硅谷的发展历史来看，抢占第一名位置至关重要。处于领跑位置的公司只需“维持原有的惯性（maintain inertia）”向前冲刺，而落后对手则必须做出重大改变才可能重回战局。正因如此，Sacks 半开玩笑地指出：“祝贺 Dario 赢得了这场 AI 竞赛”，虽然不能说 Anthropic 已经稳操胜券，但 Dario Amodei 领导的团队目前确实牢牢占据着领跑者的身位。 ```mermaid flowchart TD accTitle: AI 核心模型竞争格局 accDescr: 展示以 Anthropic 为领跑者，OpenAI、Google、Elon 等多方力量追赶与突围的竞争生态 subgraph Leaders["领跑集团"] Anthropic["Anthropic (Claude)<br/>维持既有高增长惯性，领跑赛道"] end subgraph Challengers["追赶力量"] OpenAI["OpenAI<br/>GPT-5.5 / Spud 基座<br/>Codex 增速反弹"] Google["Google / DeepMind<br/>顶尖工程团队，深耕代码能力"] ElonGroup["Elon 阵营<br/>与 Cursor 结盟<br/>xAI 协同布局"] end Challengers -.->|"必须打破现有市场惯性"| Anthropic ``` --- ### [20:25](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1225s) “Elon Web Services”构想：从工商业储能到太空分布式算力 Jason 分析了 Elon Musk 打造全产业链闭环的底层逻辑。当前 SpaceX 的发射服务与星链（Starlink）预计年营收已达 200 亿美元左右，但对比 AWS、Azure 和 GCP 构筑的规模——合计约 3,000 亿美元营收、如果独立拆分将值 4 到 5 万亿美元市值的庞大生态——传统云巨头的地盘依然巨大。Elon 在工业制造领域的核心壁垒就是“建工厂”，而超大规模数据中心本质上就是巨型计算工厂；在能源端，他又有 Tesla 的储能电池 Megapack 和早年并购 SolarCity 沉淀的太阳能方案支撑。 如果 Elon 借此整合出价值 50 亿美元增量收入的全新“新一代云（Neo-Cloud）”服务体系，他的布局不仅限于地面数据中心，更可能演化出极其激进的分布式计算架构： - **家庭与汽车的边缘算力**：将芯片植入 Tesla 车辆与 Powerwall 家用储能系统，甚至反向向家庭补贴电费，构建遍布千家万户的分布式算力网络。 - **星链与空天算力**：借助 Starlink 传输数据，甚至将算力直接送入太空，避开地面监管与能源阻力。这与 Anthropic 发布的合作声明中双方对“太空计算”的兴趣完全契合。 --- ### [22:25](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1345s) 分布式家用算力潮与 SpaceX 的超级估值乘数 这种将算力沉降到建筑与住宅侧的趋势已经演变为行业现实。Chamath 举例提到，美国头部建筑商 PulteGroup 刚与智能配电面板企业 Span 达成合作，直接在新建住宅旁部署配有 Nvidia GPU 集群的微型数据中心。Jason 补充道，Span 最早起步于用手机 App 管理断路器的智能电箱，如今果断转型接入 AI 算力硬件；与此同时，Zach Dell 创办的 Base Power 也在工业端推行类似的分布式能源与算力部署。 针对这些跨界业务整合，Brad 预测 SpaceX 即将到来的 IPO 可能会以惊人的 40 到 50 倍市销率（P/S）定价。如果次年营收达到 400 亿至 500 亿美元，其上市估值有望直接冲击 2 万亿美元关口，远超以约 25 倍市盈率交易的传统科技巨头。这种超高估值溢价来自 Elon 对整个科技底层创新的整合能力。 --- ### [23:45](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1425s) 估值溢价之争：突破性创新还是传统大厂的平庸退化？ 四位主持人就 Elon 旗下企业的“估值溢价（Elon Premium）”爆发了激烈的言语交锋。Jason 与 Chamath 认为市场给予 Tesla 和 SpaceX 高估值，是因为公众清晰感知到 Elon 展现的未来蓝图，而传统科技巨头则因缺乏 Steve Jobs 式的愿景领袖而陷入“停滞性平庸”： - **创新停滞与“挤牙膏”**：Chamath 犀利指出，资本市场正在向 Big Tech 传递明确警讯——靠更好的广告定向“从石头里挤血”，对社会福利已经毫无净收益。当 Tesla 与 SpaceX 最终走向合并、诞生统一的“Elon Corp”时，整个华尔街的认知框架将再度被颠覆。 - **互怼与语义争论**：Jason 吐槽 Tim Cook 治下最大的所谓创新竟然是跟风做流媒体的 Apple TV，而关键的汽车和前沿 AI 探索却屡遭腰斩，因此理应承受估值折价。 - **溢价还是折价？** Brad 随即反驳，强调 Google、Meta、Apple 和 Amazon 当前仍然享有强劲的绝对估值倍数，并未遭到市场实质性“惩罚”；核心逻辑是“Elon 凭借硬核创新享受了超额溢价”，双方为了究竟是“大厂被折价”还是“Elon 得溢价”争执不休，被戏谑为纯粹的语义游戏。 --- ### [26:48](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1608s) 思想实验：当垄断者掌握了公关——“安全石油”与 Anthropic 的安全叙事 插科打诨间，Jason 调侃刚发表了两万字长文、自诩“辩论队长”的 Sacks，追问他频繁暗示 Anthropic 走向垄断，是否意味着 FTC 等监管部门应该进场干预。Sacks 随即严肃抛出自己的行业研判：科技行业历来具有赢者通吃、走向垄断或双寡头的自然规律。当前市场上唯有 OpenAI（年化 3-4 倍增长）和 Anthropic（年化 10 倍指数级暴增）在生成真实且庞大的现金流。若 Anthropic 保持这一增速仅需再跑 18 个月，就将跃升为人类历史上最具价值的企业，进而掌握对 AI 这一时代最关键技术的绝对支配权。 ```mermaid flowchart LR accTitle: 标准石油与 AI 安全垄断策略对比 accDescr: 洛克菲勒标准石油的强势扩张与当下以“安全”为旗帜的 AI 监管与护城河构建对比 subgraph Historical["历史镜像：洛克菲勒"] SO["标准石油 Standard Oil<br/>初代产品：煤油 Kerosene"] Ruthless["公关拙劣，被公众视为无情垄断者<br/>《血色将至》式的野蛮掠夺"] end subgraph ModernAnalogy["思想实验：安全石油 / Safe Oil"] PRHypo["若改名为『安全石油 Safe Oil』<br/>理由：煤油易燃、极度危险"] RegCapture["呼吁政府设立严苛的『安全许可标准』<br/>打着保护公众名义排除中小竞争者"] end Historical -.-> ModernAnalogy ``` 为了剖析这种垄断潜在的伪装形态，Sacks 提出了一个关于约翰·D·洛克菲勒（John D. Rockefeller）的思想实验： - 洛克菲勒作为美国历史上最高效也是最冷血的垄断巨头，曾因拙劣的公关形象遭受公众唾弃（正如电影《血色将至》所刻画的那样）。 - “但想象一下，如果洛克菲勒当年精通现代 PR 话术，他绝不会把公司叫做‘标准石油’（Standard Oil），而会把它包装成‘安全石油’（Safe Oil）。” - 洛克菲勒完全可以凭借“煤油极易燃烧且充满危险”这一冠冕堂皇的公众安全借口，大肆游说监管机构，通过设立严苛的“安全标准”扫清所有竞争对手，以“安全”之名筑起无人能破的行政垄断高墙。这正是 Sacks 对眼下以“AI 安全”为名构筑行业壁垒的深层警惕。 ### [28:51](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1731s) 洛克菲勒煤油寓言：安全辩论掩盖下的垄断 大卫·萨克斯（David Sacks）打了一个辛辣的讽刺比喻：煤油既可以用来为家庭照亮房间，也可以将房子付之一炬；若落入坏人之手，它能烧毁一座城市，甚至被用来制造炸弹。按此逻辑，约翰·D·洛克菲勒（John D. Rockefeller）当年就应该呼吁成立一个新的政府监管机构来把控产品安全——进行严苛的测试、发放牌照、实施“常识性监管”。届时大众会陷入对安全标准的激烈辩论：“灯芯应该多厚才合规？”以及“我们是否该允许那些危险的独立炼油厂存在？” 萨克斯指出，人们会被“什么是安全石油、什么是安全煤油”的伪议题彻底带偏，从而忽略背后真正发生的大事——洛克菲勒正在打造人类历史上最庞大、最富有的垄断帝国。公众甚至可能会把洛克菲勒奉为“有效利他主义者”（Effective Altruist），只因他对产品安全“如此关切”。这番犀利的独白引得全场哄堂大笑，杰森·卡拉卡尼斯（Jason Calacanis）调侃这是萨克斯幕后写手纽曼（Newman）的手笔，萨克斯则坚称完全是自己原创。 ```mermaid flowchart TD accTitle: 洛克菲勒垄断与AI安全监管的讽刺类比 accDescr: 展示以安全监管为名行垄断护城河之实的策略路径 subgraph Pretext["安全话语假象"] A["高危技术双刃剑<br/>（煤油可照明亦可纵火 / AI可赋能亦可造弹）"] --> B["呼吁设立政府监管机构<br/>（严格测试、牌照准入、行业准则）"] B --> C["公众陷入安全标准细节争论<br/>（如灯芯厚度、独立厂商资格）"] end subgraph Reality["商业本质与终局"] C --> D["构筑极高准入门槛与护城河<br/>（清除独立竞争对手）"] D --> E["确立超级市场垄断地位<br/>（标准石油帝国 / 前沿AI巨头）"] E --> F["巨头披上‘有效利他主义’外衣"] end ``` --- ### [30:01](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1801s) 垄断与竞争之争：布拉德·格斯特纳与萨克斯交锋 布拉德·格斯特纳（Brad Gerstner）对萨克斯的垄断论调极为不满，认为这一论调在当前阶段极为荒谬。格斯特纳强调，虽然大家都在讨论年化收入跑率（run-rate），但按照 GAAP 会计准则，Anthropic 在 3 月份的实际营收与 OpenAI 大致相当，此时定性垄断为时尚早。仅仅 5 个月前，所有人还坚信 OpenAI 将独霸天下；与此同时，谷歌在 AI 领域的营收极其庞大，谷歌和亚马逊等巨头坐拥数百亿美元的自由现金流来支持增量投资，而 OpenAI 和 Anthropic 充其量只是在整个格局中依然脆弱的新兴创业公司。美国拥有全球最具活力的 AI 竞争环境，这也是美国能够领跑前沿、横扫全球的原因。华盛顿最不该做的就是插手干预、给竞争设卡。 卡拉卡尼斯随即调侃道，格斯特纳的意思翻译过来就是“别动老子的账面投资”（Don't mess with my paper），因为格斯特纳重仓了这些公司。萨克斯回应称，没有人比他更希望看到激烈竞争，这正是去年他们制定行动计划的核心初衷，美国要赢下 AI 竞赛全靠五大前沿实验室的拼杀。然而萨克斯警示，历史规律表明华盛顿总是在巨头拿下 80% 市场份额后才后知后觉。按照 Anthropic 当前的发展势头，只需再持续 18 个月，就会形成前所未有的超级垄断地位。那些打着“安全”旗号的政策，本质上是在寻求监管俘虏（regulatory capture），为正在成型的垄断或双寡头构筑更厚的护城河。 萨克斯进一步抛出具体指控：真正的垄断者最热衷于扼杀竞争，比如直接禁止竞争对手使用其模型，“Anthropic 封禁 OpenClaw 使用其模型，究竟有何站得住脚的理由？这难道不是赤裸裸的反竞争行为吗？”格斯特纳则坚持认为，行业尚未冲出起跑线，政府过早下场挑选赢家和输家将是一场灾难。 - **GAAP 营收对比**：在 GAAP 会计准则下，Anthropic 与 OpenAI 在 3 月份的实际营收规模基本处于同一梯队。 - **巨头现金流支撑**：科技巨头（谷歌、亚马逊等）拥有数百亿美元自由现金流支撑 AI 军备竞赛，初创公司地位远未固化。 - **市场窗口预警**：萨克斯认为，如果当前爆发式增长势头再持续 18 个月，前沿阵营将彻底蜕变为不可撼动的垄断体。 - **反竞争操作质疑**：Anthropic 针对竞争对手（如 OpenClaw）调用自身模型实施封禁限制，被萨克斯直斥为反竞争行为。 --- ### [34:25](https://www.youtube.com/watch?v=10MdOvK-aG4&t=2065s) 峰会插曲与白宫“AI 版 FDA”监管风波 在两人争执不下之际，卡拉卡尼斯打圆场切入 All-In 峰会的售票进展，透露洛杉矶峰会（9 月 13 日至 15 日）的席位正在光速售罄，候补名单已彻底关闭，团队正在为大会开发专属社交软件。随后，讨论迅速切入重磅爆料：据《纽约时报》披露，白宫正在考虑出台行政命令设立一个“AI 工作组”，由科技高管与政府官员共同组成，研究潜在的审查程序，其中包括对新 AI 模型的审查机制——外界普遍将其恐慌性地解读为“针对 AI 的 FDA（食品药品监督管理局）审批制”。 报道指出，这一动向的直接导火索正是 Anthropic 最新发布的模型，其强大能力引发了白宫内部的高度警觉与恐慌。拜登/特朗普政府的核心诉求是“政治免责”（CYA），极力避免未来发生灾难性的 AI 赋能网络攻击时承担连带政治责任。白宫国家经济委员会（NEC）主任凯文·哈塞特（Kevin Hassett）在福克斯商业频道受访时公开证实了该报道，甚至直言后续 AI 在进入真实世界之前，应如同“FDA 审批新药”一样证明其安全性；而候任财长斯科特·贝森特（Scott Bessent）在受访时则表达得较为折中，强调政府既要支持 AI 公司持续创新，又要守住国家安全的底线。 格斯特纳随即透露，他在该电视片段播出当晚便直接联系了哈塞特。哈塞特向他澄清，提出“FDA”这个类比只是想表达政府希望 AI 公司能主动展示前沿模型以便协同应对，确保政府系统加固、情报机构跟上技术节奏，哈塞特本人以及右翼阵营内部，绝无任何人支持建立真正的“模型事前审批准入机制”（approval regime）。在格斯特纳看来，要求每个模型都必须经由华盛顿官僚机构盖章预审的设想，纯粹是一场行业灾难。 ### [38:14](https://www.youtube.com/watch?v=10MdOvK-aG4&t=2294s) AI 监管争议：“AI 领域的 FDA”构想与华盛顿的审查瓶颈 讨论首先聚焦于政府监管可能带来的结构性危害。如果在模型领域让华盛顿扮演裁定“输赢家”的角色，将带来极其糟糕的后果：美国目前在 AI 竞赛中骑在领跑的“赢家骏马”上，遥遥领先于全球其他国家，此时临阵换马或彻底改变治理体系毫无意义，更不能用繁文缛节与过度民主化的官僚程序来拖垮创新。然而，对于 Mistral、Spud 等前 AGI（pre-AGI）尖端模型的涌现，产业界与政府之间的协同正逐步加深，但现有框架仍待进化。最核心的堵点在于政府自身的审查能力——目前进行网络安全审查（cyber review）的时间拖延太久，必须建立严格的时限机制，确保政府能在限定窗口内给出反馈，而不是无限期卡死发布流程。 针对政界部分人士提议设立“AI 领域的 FDA”（像食品药品监督管理局审查新药一样审批前沿模型）的论调，主讲人强调内阁核心成员（如 Kevin 与 Scott Bessent）对此危害心知肚明，绝不会在华盛顿设立类似机构。然而，Jason 随即提出质疑：公众与理性人士普遍认同 AI 不可能处于完全放任的丛林状态，必须设立安全护栏。这究竟是一个糟糕的类比、媒体炒作，还是某些游说者趁着 David Sacks 离开白宫时趁虚而入、吹“枕边风”试图左右特朗普的决策？ ```mermaid flowchart TD accTitle: AI Review Mechanism Comparison accDescr: 展现繁重的中心化准入审批与敏捷高效的限时安全协同之间的机制对比。 subgraph Trap["中心化官僚准入陷阱 (AI 版 FDA)"] A1["前沿模型预训练完成"] --> B1["提交华盛顿审查机构"] B1 --> C1["无休止的审批与合规拖延"] C1 --> D1["扼杀美国创新领先优势"] end subgraph Agile["敏捷协同与能力增强路线"] A2["前沿模型预训练完成"] --> B2["行业与政府既定协同对接"] B2 --> C2["有限时限的网络安全快速审查"] C2 --> D2["安全部署与持续攻防加固"] end ``` ### [39:55](https://www.youtube.com/watch?v=10MdOvK-aG4&t=2395s) 硅谷的舆论危机与“氛围剧变”：缺乏共赢叙事的恶果 Chamath 深刻剖析了监管风声背后的根本成因。他认为这并非单纯的宫斗或个别说客的进言，而是整个社会层面正在发生一场深刻的“氛围剧变”（vibe shift）。从主街（普通民众）到华盛顿，公众对科技寡头、硅谷以及 AI 的敌意与戒备正迅速渗透。无论未来是民主党还是共和党执政，某种形式的强力监管都不可避免，差异仅在于民主党治下手段可能更严苛，但共和党也绝不会听之任之。 之所以演变至此，症结在于科技界犯下的致命错误： 1. **糟糕至极的舆论公关**：整个舆论场被 AI 的负面情绪主导。科技巨头和领袖们舍得挥霍资本，却不肯花精力和资金去清晰勾勒 AI 带来的普惠红利和巨大上升空间，以获取跨阶层的广泛民意支持。 2. **财富极度集中引发的群体抗体**：行业所描绘的前景，是少数赢家通吃、千万人成为潜在受害者的残酷画面。科技领袖们既没有联合起来对全美社会进行大规模的反哺与再投资，也没有用行动打消民众疑虑，导致社会机体产生针对科技行业的“免疫抗体”。Chamath 直言如果给科技领袖目前的沟通与作为打分，只能打“D- 且正在向 F 滑落”。如果最终结局只是让两三个巨头坐拥万亿美元身家并独揽底层关键钥匙，公众自然会反问并支持严厉监管。 Jason 补充道，人类的大脑进化本就带有强烈的风险防御机制与安全偏见，天生就会去警惕深伪漏洞（deepfakes）、杀人机器人、自动驾驶夺走蓝领饭碗以及生物武器等至暗威胁，而整个行业却几乎无人能有力地传达光明的一面。 ### [43:02](https://www.youtube.com/watch?v=10MdOvK-aG4&t=2582s) 白宫内部动态与虚假叙事：破除“AI 版 FDA”的迷雾 David Sacks 随后直接打破外界的所谓“白宫宫廷秘史”猜测，指出整场恐慌很大程度上是媒体制造的虚假新闻（fake news）。不论是 Brad 还是 DeepMind 的 Demis Hassabis，没有哪位核心高管或政府高层认同搞一个“AI 领域的 FDA”；特朗普总统本就是历史上最支持技术创新的总统，而白宫幕僚长 Susie Wiles 前一晚刚刚公开发表声明，彻底击碎了设立该类审批机构的传闻。这一概念最初是由 Andrew Ross Sorkin 等评论员在《纽约时报》等媒体上率先炒作并不断发酵，导致硅谷产生了极其强烈的本能抗拒。 Sacks 进一步拆解了建制派舆论构建的另一层稻草人谬论：反对者试图将特朗普政府第一年的 AI 政策抹黑为毫无护栏、彻头彻尾的“自由放任”（laissez-faire）。事实上，白宫在 3 月 20 日就发布了全美 AI 监管框架立法建议报告（包含一份四页纸的立法支持清单）。政府倡导的是针对具体风险对症下药（specific solutions to specific problems），而不是通过华盛顿的大一统权力扩张来碾压创新。 - **媒体定调偏差**：将特定评论员的推论当作政府既定方针大肆宣扬。 - **白宫澄清**：Susie Wiles 声明明确了支持创新、拒绝设立 AI 官僚审批机构的立场。 - **政策实质**：3 月 20 日白宫政策框架主张针对具体安全威胁精准施策，拒绝华盛顿借机进行全局性的权力寻租。 ### [45:06](https://www.youtube.com/watch?v=10MdOvK-aG4&t=2706s) 网络空间攻防战：Mythos 能力爆发与反制策略 针对眼下最紧迫的实质性风险，Sacks 指向了以 Mythos 为代表的高级网络攻防模型。这不仅限于个别前沿模型，OpenAI 目前也研发出了具备同等强大网络渗透能力的模型，并且在未来 3 到 6 个月内，所有主流前沿实验室乃至中国的模型、开源社区的模型都会达到同等网络能力水准（开源社区目前已具备 80% 至 90% 的能力）。网络安全工具本身具有极强的军民两用性，用于安全防御的底层机制与用于黑客攻击的技术完全重合。 面对这场不可避免的攻防竞赛，防御策略不能依赖于“一禁了之”，而必须依靠以下协同手段： - **政企高效协同防御**：调动美国庞大的网络安全生态，依靠 Palo Alto Networks、CrowdStrike 等行业巨头，以及数百家利用 AI 进行自动化渗透测试（pen testing）的新兴安全初创公司。 - **工具即武器赋能**：将 Mythos 和 OpenAI 的尖端网络分析工具作为力量倍增器（force multiplier），以最快速度分发给网络安全厂商，用于对企业关键代码库进行全自动化扫描与漏洞挖掘，赶在黑客行动之前完成底层系统的快速打补丁和系统加固。 - **探索身份认证与 KYC 机制**：针对调用此类超级网络能力的前沿模型，Jason 与 Sacks 探讨了引入类似于金融领域的“了解你的客户”（KYC, Know Your Customer）准入封装机制，要求使用者完成身份核验，尽可能拦截具有敌意的国家级黑客组织或恶意行动者。Sacks 强调，Anthropic 和 OpenAI 在发布超强模型时均表现出了极高的自律性，这也从反面证明，强推事前牌照审批既脱离现实，也完全低估了领军企业的合规自觉。 ### [47:54](https://www.youtube.com/watch?v=10MdOvK-aG4&t=2874s) 针对前沿模型的 API 访问与预览期 KYC 监管讨论 嘉宾们针对强力前沿 AI 模型的安全交付机制展开讨论。David Sacks 与 Chamath Palihapitiya 认为，建立事前审批制往往是在“解决一个根本不存在的问题”——因为 Anthropic 和 OpenAI 等实验室深知这些模型的威力与可能面临的法律诉讼风险，本身就在高度自律地克制发布。 ```mermaid flowchart TD accTitle: 前沿 AI 模型的安全与分发链条 accDescr: 展示预览期严格 KYC、模型滥用监控与全面公开发布后的监管演变过程 subgraph PreviewPhase ["预览测试阶段 (Preview Phase)"] A["前沿模型部署<br/>(例如 Mythos / OpenAI 对应产品)"] --> B["严格身份认证 (KYC)<br/>确认交付对象为'好人'"] B --> C["API 行为追踪与日志记录<br/>防蒸馏与异常调用告警"] end subgraph GovCoordination ["政府联动"] C --> D["向政府共享异常情报<br/>跨部门安全协同"] end subgraph GeneralRelease ["全面公开发布阶段 (General Release)"] D --> E["大规模普惠交付"] E --> F["受制于隐私法规限制<br/>KYC 与全面深度日志逐步弱化"] end ``` 然而在技术验证与封闭预览阶段，Chamath 强调必须实施严格的 KYC（了解你的客户）身份验证机制：在将超级强大的模型 API 交付出去前，绝不能提供给不可知、不透明的机构或个人。当务之急是迅速将工具交给经过审核的“正面力量”（good guys），而 KYC 正是确认其身份的前提。Jason Calacanis（J-Cal）随即发问：如果预览阶段对前沿模型进行全方位的身份识别与日志审计以防范安全漏洞，一旦面向全社会公开发布，是否会因公众隐私保护问题而无法维持同等监控力度？ - 预测市场 Polymarket 上关于“特朗普在 5 月 31 日前签署行政命令要求对 AI 模型进行联邦级全面审查”的概率当时显示为 21%。 - J-Cal 开玩笑质疑作为主持人自己是否有内部消息优势、能否借此下注赚钱，Chamath 随即严厉劝阻：“J-Cal，千万别下注。” - 几位嘉宾一致认为，对于早期内测预览期，KYC 是必不可少的标准配置；一旦模型转入全面公开发布（General Release），由于受众极广，传统严格的 KYC 可能不再适用，重点转向通用防范。 - 实际上，所有一线大模型实验室目前均在持续追踪 API 使用情况，严密防御模型蒸馏行为（anti-distillation），并已与政府建立了高频的情报互通与安全协同机制，任何可疑活动都会被标记甚至联动国家防范力量。 --- ### [50:27](https://www.youtube.com/watch?v=10MdOvK-aG4&t=3027s) AI 末日论者的监管算盘与华盛顿的“FDA 陷阱” David Sacks 深入分析了华盛顿政策圈背后的推手动机。他指出，目前确实存在网络安全（Cyber）层面的真实紧迫挑战，各大系统与基础设施必须在未来 3 到 6 个月内完成安全加固，但这一技术问题终究会被克服。真正危险的是一帮被称为“AI 意识形态狂热者”或“AI 末日论者”（AI doomers）的人，正在遵循经典的“绝不要浪费任何一场危机”策略，试图借题发挥，在华盛顿凭空建立起永久性的庞大监管行政官僚机构。 Sacks 指出，虽然白宫现任政府的本意并非如此，但在社交媒体、各大智库以及部分政客（如伯尼·桑德斯 Bernie Sanders）的推波助澜下，监管呼声被无限放大。J-Cal 模仿起桑德斯的标志性口吻：“这是我头一次认同本届政府，他们终于理解了对前 1% 的 1% 征税！AI 完全失控了，它不仅要抢走所有人的工作，还要夺走我的避暑别墅！”Sacks 补充道，桑德斯等人本质上完全全盘接受了末日叙事，甚至主张禁止新建数据中心，其青睐类似于“AI 版 FDA”的设想，根本目的就是要彻底给科技创新踩刹车。 --- ### [52:01](https://www.youtube.com/watch?v=10MdOvK-aG4&t=3121s) 扭转公众 AI 焦虑：财富共享机制、大额慈善与 Invest America 设想 为了从根本上扭转全社会对 AI 的负面情绪（bad vibes），J-Cal 提出了多维度的社会救济与科技红利共享策略。他认为，当前最需要科技富豪与顶级科技公司用实质行动来改变大众认知，而不是继续玩弄虚伪的公关花招： - **全民 IPO 红利共享（IPO for Kids）**：像 Nvidia、未来上市的 SpaceX、Anthropic 乃至可能的 OpenAI 等核心巨头，在公开上市（IPO）时，应直接将 1% 到 5% 的股份捐赠给全体美国公民，注入到 Brad Gerstner 发起的 Invest America 儿童投资账户中，让底层儿童与公民直接成为 AI 繁荣的股东。 - **唾弃虚伪的“捐赠承诺”（Giving Pledge）**：J-Cal 批评现行的巴菲特/盖茨式“捐赠承诺”纯属美化形象的道德姿态（virtue signaling），仅承诺在临终前捐款毫无实质约束力；他主张真正的财富所有者应在未来 20 年内，每年拿出 1% 的股票直接投入社会基金，“反正无论对贝佐斯还是谁，这笔钱这辈子都花不完”。 - **聚焦健康与科学普及**：科技界应大幅强化 AI 在医疗健康、生命延长、基础科学研究以及降低教育成本方面的落地贡献，让公众切身体会到痛苦的减轻与福利的增加。 --- ### [53:35](https://www.youtube.com/watch?v=10MdOvK-aG4&t=3215s) 消费驱动型经济与最低工资、全民医保的激烈激辩 J-Cal 进一步将讨论从 AI 焦虑推向美国底层的核心生存痛点：普通民众之所以抵触 AI，根源在于对失去收入与失去医疗保障的深层恐慌。他旗帜鲜明地呼吁资本家们必须正视最低工资与全民医保议题： | 讨论议题 | Jason Calacanis 的观点与立场 | 传统资本主义/反对派视角的反驳（Sacks 等） | |:--- |:--- |:--- | | **最低工资调整** | 敦促企业主动自愿或阶梯式提高最低工资。参考新西兰、瑞典、瑞士、澳大利亚的经验，底层人群拿到工资后不会储蓄而是直接消费，能立即反哺由 60%-80% 消费拉动的美国经济。 | 资本主义内部历来对强制提升最低工资存疑；Sacks 困惑于讨论为何从 AI 技术突兀转向了最低工资。 | | **全民医疗保障** | 医疗焦虑是普通家庭的头号压力源。企业若不再背负雇主医保包袱，公众将拥有更多可支配收入用于消费（如订阅 Netflix、在亚马逊购物）。 | 核心障碍并非大家不愿提供医保，而是天量财政成本。Sacks 援引讽刺作家 P.J. O'Rourke 的名言反讽道：“如果你现在觉得医疗昂贵，那就等等看全民免费后的账单吧。”去除了市场激励机制只会让医疗体系崩溃。 | 面对 Sacks 的调侃与不解，J-Cal 坚持认为，AI 带来的社会极化与财富分化若无法通过制度性福利缓冲，公众对技术革命的抵触只会愈演愈烈。Sacks 则极力把话题拉回现实，强调尽管各类民调显示大众对 AI 确实存在广泛反感，但最新的关键民意数据却揭示了更为复杂的舆论切片。 ### [56:27](https://www.youtube.com/watch?v=10MdOvK-aG4&t=3387s) AI 的民意与经济现实：通缩红利与蓝领繁荣 嘉宾们深入探讨了选民关心的议题优先级与公众对 AI 态度的巨大落差。民意调查数据显示，在选民最关心的 39 个议题中，AI 仅排在第 29 位，根本未进入前十甚至前二十名；选民真正置于首位的是生活成本（Cost of Living），第二位则是整体经济。大卫·萨克斯（David Sacks）指出，现实中 AI 本身具有极强的通缩属性，既能压降生活成本，又正在掀起一场全国性的经济繁荣——AI 相关的投资与建设贡献了第一季度 75% 的 GDP 增长。 这种增长红利绝不仅限于硅谷的科技初创企业，而是已经迅速外溢到实体经济：全国正掀起数据中心建设潮，推动蓝领工人和建筑工人迎来景气期，建筑工人工资甚至暴涨了 25% 到 30%。布拉德·格斯特纳（Brad Gerstner）随即补充，图表中甚至连医疗健康领域也开始受益。萨克斯强调，虽然 AI 在舆论场上可能并不讨喜，但它带来的经济效果其实极受欢迎；媒体如果能客观报道事实，公众就会意识到 AI 正在切实驱动当下的经济繁荣。 - **选民关注排名**：在全部 39 项议题中，AI 排第 29 名，完全不在选民关心的前列。 - **选民核心痛点**：排在第一位的是生活成本，第二位是整体经济。 - **GDP 贡献率**：AI 相关产业链拉动了第一季度 75% 的 GDP 增长。 - **蓝领红利**：数据中心基础设施建设带动建筑行业蓝领工人工资上涨 25% 至 30%。 --- ### [57:27](https://www.youtube.com/watch?v=10MdOvK-aG4&t=3447s) 监管之害 vs. 市场红利：如何让全美共享 AI 蛋糕 格斯特纳对萨克斯的分析表示高度赞同，并尖锐批评了以伯尼·桑德斯（Bernie Sanders）为代表呼吁设立暂停期（Moratorium）、关停所有数据中心的强监管思潮。他警示称，若真按此执行，今年美国 GDP 将立刻陷入负增长，股市会暴跌 15% 到 20%，失业率也将大幅飙升。他援引历史教训指出，1929 年大萧条前夕失业率仅为 4%，但短短三年后因政府过度介入管制、扼杀正在正常运转的经济机制，失业率飙升至 23%。如今 AI 在控制失业率、推动经济增长和生产力飞跃方面带来了巨大的净收益，过度行政管制是当前最大的威胁。 为了彻底扭转公众偏见，科技行业必须主动向全民兑现净收益。格斯特纳呼应了查马斯（Chamath Palihapitiya）此前的构想，主张让每个美国人都拥有一个能够随 AI 红利复合增长的投资账户。此外，他还提出了极具实操性的方案：例如若在得克萨斯州阿比林（Abilene）建设巨型数据中心，就应顺带为该地全体居民提供免费用电。让民众直接拿到看得见摸得着的实惠，社会对 AI 的悲观情绪才能真正跨越谷底。 针对杰森·卡拉卡尼斯（Jason Calacanis）关于最低工资与资本主义创新关系的追问，格斯特纳明确表态，美国如今背负着高达 38 万亿美元的国债，政府自上而下的包揽干预从来解决不了问题。观察各类通胀图表可以发现，医疗、高等教育等成本飙升最严重的领域，恰恰是政府过度干涉的地方。与其依赖官僚规制，不如依靠自由市场催生“丰饶时代（Age of Abundance）”，届时人人都能拥有高质量的 AI 医疗和教育顾问。卡拉卡尼斯也深表赞同，直言许多初创企业和 VC 避开住房、教育和医疗，正是因为这些领域被政府过度监管视作“氪石（Kryptonite）”，唯有拆除管制的阻碍，创新者才能放手施展。 --- ### [01:00:05](https://www.youtube.com/watch?v=10MdOvK-aG4&t=3605s) 超大规模云厂商暴赚与强劲的宏观基本面 随着话题切入资本市场，卡拉卡尼斯指出当下科技与消费板块表现出令人瞠目的强度。几大超大规模云厂商（Hyperscalers）交出的营收答卷直接引爆了市场。尽管各家统计口径略有差异——微软在 Azure 中计入部分软件产品，谷歌在 Google Cloud（GCP）中涵盖 Google Workspace 等办公套件，但底层势头极为凶猛： - **亚马逊 AWS**：作为最纯粹的云基础设施业务，在极其庞大的体量基数上维持着 28% 的高速增长，年化营收跑道（Run Rate）达到惊人的 1,500 亿美元。 - **微软 Azure**：以 39% 的增速推进，年化营收达到 1,080 亿美元。 - **谷歌 GCP**：增长速度更是震撼华尔街，交出了 63% 的亮眼增速，年化营收升至 800 亿美元。 这三大云厂商单季度合计净增近 300 亿美元的 ARR（谷歌增加约 100 亿，AWS 增加 100 亿，Azure 增加 95 亿）。与此同时，美股大盘持续处于历史高位，七巨头（Mag 7）势头正猛，而作为耐用消费与服务风向标的 Uber 和迪士尼（Disney）财报同样全线爆表，显示底层消费极其强韧，科技基础设施更是在全速快跑。 ```mermaid flowchart TD accTitle: Big Three Cloud Hyperscalers Revenue and Growth accDescr: Diagram illustrating annual run rate and growth metrics for AWS, Azure, and Google Cloud along with incremental revenue. subgraph Hyperscalers["三大超大规模云服务商年化规模与增速"] AWS["AWS<br/>年化规模: $1,500 亿<br/>增速: 28%<br/>增量: +$100 亿 ARR"] Azure["Azure<br/>年化规模: $1,080 亿<br/>增速: 39%<br/>增量: +$95 亿 ARR"] GCP["Google Cloud<br/>年化规模: $800 亿<br/>增速: 63%<br/>增量: +$100 亿 ARR"] end Hyperscalers --> Total["三巨头季度合计新增<br/>~$300 亿美元 ARR"] ``` --- ### [01:01:32](https://www.youtube.com/watch?v=10MdOvK-aG4&t=3692s) 从审慎到“All In”：AI 算力估值与金发女孩经济 格斯特纳对宏观环境进行了全局复盘。回顾过去一段时间，市场充斥着悲观叙事：有人预言关税政策会导致恶性通胀并摧毁 GDP，有人担忧委内瑞拉和伊朗的地缘冲突将引发能源震荡。然而现实情况却是美国 GDP 加速上行、十年期美债收益率平稳运行在 4.3% 左右、通胀完全受控，美国在算力与 AI 领域继续扩大全球领先优势。更为关键的是，标普 500 指数年内仅上涨 8%，市场绝非处于泡沫阶段： - **Meta**：基于全额计税 GAAP 盈利的市盈率仅为 17 倍。 - **英伟达（NVIDIA）**：市盈率为 19 倍。 - **微软（Microsoft）**：市盈率为 20 倍。 - **谷歌（Google）**：市盈率为 24 倍。 - **半导体存储板块**：格斯特纳透露其投资组合有 25% 仓位押注于 SK 海力士（SK Hynix，仅 5 倍 PE）、三星（Samsung，6 倍 PE）和美光（Micron，7 倍 PE）。 | 公司 / 标的 | 全额计税 GAAP 市盈率 (P/E) | 备注说明 | |:--- |:--- |:--- | | **Meta** | 17x | 科技巨头估值依然理性 | | **Nvidia** | 19x | AI 算力龙头估值并未透支 | | **Microsoft** | 20x | 云与企业级 AI 核心龙头 | | **Alphabet / Google** | 24x | 搜索与 GCP 驱动全面回升 | | **SK 海力士 (SK Hynix)** | 5x | 基金重点重仓，占组合 25% | | **三星 (Samsung)** | 6x | 存储芯片关键供应链 | | **美光 (Micron)** | 7x | 存储供应链关键标的 | 格斯特纳解释了他自身立场的转变：去年底由于涨幅过大他一度持谨慎悲观态度，当时笼罩在市场上空的最大疑云是“AI 基础设施投资能否真正转化为下游收入”。若 OpenAI 和 Anthropic 的营收无法兑现、云厂商没有重新加速，市场本会回调 10% 到 15%。然而到了 12 月和 1 月，真实营收全面爆发——OpenAI 与 Anthropic 的合计年化收入在短短四个月内从 300 亿美元跃升至 800 亿美元。确认投资回报率（ROI）成立后，他果断将风险敞口从中等直接拉升至重仓（80% 以上配置于 AI、算力及存储）。 卡拉卡尼斯虽指出若没有数千亿美元的军事冲突开销和被判违宪的关税折腾，经济可能跑得更快，但格斯特纳强调，当前美国所处的是难以想象的“金发女孩（Goldilocks）”完美局面。随着全球贴现率回落与债市稳定，一级市场孕育出多家数万亿美元级别的巨无霸（如即将上市的 SpaceX，以及 OpenAI、Anthropic），全美科技优势无可撼动。正如全场所达成的高度共识，全世界没有任何一个国家愿意放弃当下的位置而不去交换美国的国运，这是由 SpaceX、Google 等卓越企业所彰显的“美国例外论（American Exceptionalism）”的集中体现。 ### [01:05:20](https://www.youtube.com/watch?v=10MdOvK-aG4&t=3920s) 特朗普政策与能源释放推动的 AI 繁荣 在讨论宏观经济与执政表现时，主持人 Jason 指出，尽管现任政府展现了对商业的友好态度，但在关税执行以及陷入当前冲突这两点上犯了严重错误，必须尽快找到解决方案。随后他请 David Sacks 对整体经济局势发表看法。 Sacks 旗帜鲜明地将当前的 AI 繁荣归功于唐纳德·特朗普总统的政策。他回顾道，特朗普在上任首周便废除了拜登时期关于芯片与模型的监管框架：在原有制度下，达到一定 FLOPs 训练门槛的模型必须送往华盛顿特区审批，全球每一笔 GPU 销售除非获得极狭窄的豁免，否则都必须获得华盛顿的许可。特朗普废除了这一套“所有事情都必须经过华盛顿批准”的审批机制，明确宣示美国必须赢得 AI 竞赛，全面为科技企业松绑。 此外，Sacks 强调了能源层面的决定性作用。早在十年前，特朗普就提出了“Drill, baby, drill（尽情开采吧，宝贝）”的口号，坚持美国能源是整个经济的基础，也是 AI 的底座。特朗普政府允许 AI 企业成为自身的“能源公司”，将自带发电设备接入数据中心，而不是从现有公共电网抽电、与普通民众争夺民用用电。Sacks 认为，正是这种政策催生了当前为 AI 基础设施提供动力的蓝领建设繁荣；相比之下，像伯尼·桑德斯（Bernie Sanders）这类政客主张的则是“全面禁止数据中心”，后果将不堪设想。 - **监管转变**：特朗普政府废除了拜登政府针对高算力模型训练（FLOPs 审批制）和 GPU 全球销售的许可审批体制。 - **能源自给自足**：支持数据中心自主发电，避免占用公共电网资源并激化民生用电矛盾。 - **政策对比**：Sacks 指出激进派政客曾提出禁止建设数据中心，若采取该方案将扼杀整个产业的扩张。 --- ### [01:07:07](https://www.youtube.com/watch?v=10MdOvK-aG4&t=4027s) 繁荣背后的“500天大考”：AI 到底有没有提升真实利润？ Jason 邀请 Chamath Palihapitiya 对整体经济做总结发言。Chamath 开玩笑道：“我认为市场还会继续涨一阵子，然后到了某个节点，就会跌下来。”面对这番废话式预测，Jason 现场模仿起“记笔记”表情符号调侃道：“太棒了，欢迎收看 All-In 播客，获取最顶级的市场操作建议——行情会先涨后跌，大家据此操作就好！” 在随后的严肃讨论中，Chamath 展开了更深层的逻辑推演：短期内，制造“新事物”的企业确实会获得极高估值并证明其价值，包括英伟达（NVIDIA）、存储芯片厂商、Anthropic、SpaceX 以及 OpenAI。然而，经济规律最终会清算一切——你不能光为市场制造工具，而买单的终端客户却无法从中获得可衡量的实质收益。Chamath 语气尖锐地直言：“坦白讲，现在字面意义上没有任何一丝证据表明 AI 帮助 S&P 500 指数成分股提升了营业利润率（Operating Margins）。” Chamath 预测，在未来两到三年内，市场将迎来一个分水岭式的关键抉择点： 1. **路径一：压缩运营支出（OpEx）**。通过裁减劳动力、缩减运营开支占收入的比例，从而实现利润率提升； 2. **路径二：驱动营收增长**。在运营支出持平甚至略微增加的情况下，通过业务规模的大幅扩张带动利润率上升。 他认为，目前的窗口期大约还剩 500 天左右，在这几百天里投资者依然可以净做多（net long）。但 500 天之后，一场必须面对的清算时刻即将来临：为所有这些 Token 付费的企业，必须从自己的财务报表上看到真实的利润回报。 ```mermaid flowchart TD accTitle: AI投资分水岭演进路径 accDescr: 展现从当前硬件与模型繁荣过渡到500天后终端企业ROI商业验证分水岭的逻辑图。 subgraph Phase1["当前阶段：短期资本盛宴（~500天）"] A["底层软硬件赋能者<br/>(NVIDIA / 存储厂商 / Anthropic / OpenAI)"] --> B["企业大额采购 Token 与算力<br/>(进行广泛试验与自研探索)"] end subgraph Fork["未来2-3年的商业化分水岭"] B --> C{"终端客户能否兑现 ROI？"} C -->|"路径 A：降本模式"| D["缩减 OpEx 与雇员规模<br/>营业利润率被动提升"] C -->|"路径 B：增收模式"| E["开拓全新业务增长点<br/>营收暴增带动利润率扩张"] C -->|"未见成效"| F["预算紧缩与估值回撤<br/>迎来市场清算时刻"] end ``` --- ### [01:09:42](https://www.youtube.com/watch?v=10MdOvK-aG4&t=4182s) 企业级 ROI 争论：既成事实还是纸上谈兵？ 针对 Chamath 的质疑，Sacks 梳理了逻辑链条：年初 Brad Gerstner 曾指出，进入 2026 年时市场对庞大的基础设施资本支出（CapEx）充满担忧，不确定模型收入能否支撑起巨额投资；然而这一担忧已被证伪，模型公司的营收实现了爆发。如今 Chamath 提出的则是下一环节的考验——买 Token 的企业到底能不能算过来账？Sacks 认为，虽然这尚未在全市场宏观数据上得到证实，但他非常乐观：如果企业客户看不到 ROI，根本不可能月复一月地疯狂加购编程 Token（Coding Tokens）。这些用于定制软件开发的 Token 正在释放前所未有的生产力浪潮，ROI 正在顺着“基础设施 $\to$ 基础模型 $\to$ 软件应用 $\to$ 最终用户”的链条层层下传。 Jason 完全赞同 Sacks 的观点，感叹这犹如 PC、互联网、云计算和移动革命的重演：“过去每次革命我们都在纠结：该不该做个 App？该不该建个网站？该不该上云还是留在本地？这些问号最终都变成了感叹号。”Jason 拿自己的投资机构举例：原本一家 22 人的小型投资机构应该购买标准 SaaS 软件，但现在仅凭 3 个人借助 AI Agent 和代码生成工具，就能在内部源源不断地开发并上线各种产品与界面。在他看来，初创企业通过 AI Token 获得了巨大的单位生产力，可以用更少的资金走得更远、雇佣少一半的员工，ROI 在创业圈已经是“既成事实（fait accompli）”。 Brad Gerstner 与 Chamath 则对此表达了强烈反对，认为这绝非既成事实。Chamath 回击道，虽然像 80/90 这类咨询和技术实施业务做得很好，但“利润不是单凭主观意愿就能凭空变出来的”： - 百威英博（Anheuser-Busch）归根结底得卖出更多的啤酒； - 耐克（Nike）归根结底得卖出更多的运动鞋； - 医疗器械公司归根结底得卖出更多的人工髋关节与膝关节。 Chamath 指出，当前充其量只是一场极具建设性和创造性的“广泛试验”，绝大部分并没有证明自身的最终商业价值。经济飞轮加速的前提是企业能算清一笔明确的账——“我投入了 $X$ 成本，赚回了 $Y$ 收益，且 $Y > X$ 并切实拉高了利润率”。现在行业仅仅完成了前半段方程（花掉了 $X$），后半段（收获 $Y$）在宏观层面根本没有显现：全球 GDP 未见异常加速，全球生产力数据未见暴增，S&P 500 的整体利润率也没有看到由 AI 驱动的跨越式跃升。 面对质疑，Jason 再次试图拉 Brad 进场辩护，坚称自己看到大量企业在成本控制上迎来了暴跌、广告投放效率大幅提高，同时普遍放缓了招聘甚至停设新增岗位。双方就 AI 降本增效能否真正撬动真实商业利润展开了激烈的交锋。 ### [01:13:56](https://www.youtube.com/watch?v=10MdOvK-aG4&t=4436s) 企业端 AI 降本增效与标普 500 利润率扩张 Jason 结合投资组合中的真实案例，阐述了生成式 AI 在企业端落地的震撼效果。过去耐克（Nike）等大品牌为应用制作宣传图和物料时，必须耗费巨资雇佣摄影团队进行实地拍摄，而如今初创公司利用 AI 生成图像，帮助品牌大幅削减了拍摄预算；类似的情况也发生在 DoorDash 上，餐馆的餐品展示图以往全靠专业摄影师逐一拍摄，现在已全面转向 AI 生成。广告创意的制作成本不仅腰斩，投放转化效率反而实现了两位数的百分比增长。面对 AI 究竟是否已经体现在财报中的疑问，Brad Gerstner 列举了亮眼的企业盈利与宏观数据：本季度微软 Azure 实现了 39% 的强劲增长，Google Cloud 更是暴涨 63%；过去三年里，科技巨头 Mag 5 的合计员工总数仅增长了约 3%，经营利润率全线大幅扩张。放大到整个标普 500 指数，2023 年企业的营业利润率约为 11%，2024 年第一季度提升至 11.8%，而今年更是攀升至 13%，整整实现了 200 个基点的惊人跃升。 面对利润率的激增，几位合伙人展开了激烈辨析：这究竟是 AI 带来的真实生产力红利，还是企业在后疫情时期的惯性财务工程？Chamath 直言，所谓的利润改善大概率与 AI 无关，依然是过去十年推高每股收益的经典财务工程手段；从 2022 至 2023 年，硅谷企业经历了从“过剩时代”（age of excess）向“精简健身时代”（age of fitness）的急刹车，许多大厂只是借着 AI 的幌子裁撤疫情期间过度招聘的冗员。Brad 承认这种质疑具有合理性，市场对于利润率在未来两年持续扩张的预期仍待检验，但他强调从微观层面观察，越来越多的企业正在以极低的员工增速支撑顶线营收（Top line）的加速狂飙，二者的脱钩已是不争的事实。 - **云业务增长指标**：微软 Azure 季度增速达 39%，谷歌云（Google Cloud）季度增速达 63%。 - **巨头团队控制**：Mag 5 科技巨头过去三年合计员工人数（Headcount）仅微增约 3%。 - **标普 500 营业利润率飞跃**：从 2023 年的 11.0% 上升至 2024 年第一季度的 11.8%，并在今年达到 13.0%（提升 200 个基点）。 ### [01:16:15](https://www.youtube.com/watch?v=10MdOvK-aG4&t=4575s) 就业数据交锋：失业率真相与应届生就业走势 David Sacks 随后接过话题，直面关于 AI 导致大规模失业的社会恐慌。他指出，在全美企业生产效率大幅提升的同时，失业率却持续维持在 4.2% 左右的历史低位，依然完全处于经济学界公认的 4% 到 5% 的充分就业区间内。更重要的是，最新媒体报道显示，年轻大学毕业生的失业率不升反降，打破了此前“入门级白领岗位将被 AI 赶尽杀绝”的悲观叙事。《华尔街日报》甚至刊发头条指出大学毕业生终于在就业市场中迎来转机，Sacks 推测这或许是因为年轻一代本身就是“AI 原生代”（AI natives），掌握了更娴熟的 AI 工具使用技能。他强调，迄今为止没有任何客观数据能够支撑“AI 引发灾难性失业”的理论推演，相反，早期显现出来的全部是巨大的生产力跃升。 Jason 对此提出了针锋相对的质疑，认为单纯看失业率掩盖了劳动力流失的残酷现实。他援引关键数据指出，全美劳动参与率（Labor Force Participation Rate）在 3 月份仅为 61.9%，远低于疫情前的 63.3%；当求职者彻底放弃寻找工作而退出统计口径时，官方失业率就会失真。此外，大学毕业生的实际境遇高度分化，文理工各专业的就业冷热截然不同，现在下定论还“为时过早”。Sacks 随即反唇相讥，调侃 Jason 每当遇到反驳其叙事的数据时就搬出“为时过早”的说辞；两人围绕宏观数据真实性、废除美联储（Fed）的争议言论以及“特朗普综合征”展开了一轮经典的拌嘴，Jason 笑着宣布大家重新回到了熟悉的“辩论社模式”。 - **失业率与就业基准**：当前全美失业率维持在约 4.2% 的低位，处于 4% 至 5% 的充分就业区间。 - **劳动参与率缺口**：3 月份全美劳动参与率为 61.9%，相比疫情前（COVID 前）的 63.3% 仍有明显差距。 - **大学应届生就业改善**：《华尔街日报》以头条确认大学毕业生在当下的就业市场正迎来显著转机。 ### [01:18:44](https://www.youtube.com/watch?v=10MdOvK-aG4&t=4724s) 尾声调侃与全美 AI 创新生态总结 在节目尾声的欢快氛围中，主持人开始互相打趣。Jason 戏谑地将刚刚达成重磅算力合作的两位科技巨头形容为“马斯克与达里奥·D·洛克菲勒”（暗指 Anthropic 首席执行官 Dario Amodei 展现出的垄断野心），并借机向与 Dario 熟识的 Brad 施压，要求其下周务必邀请 Dario 登上 All-In 节目现场对质。Brad 则由衷感叹，美国拥有全球无可匹敌的技术创新框架，无论是马斯克（Elon Musk）、Anthropic、OpenAI，还是谷歌（Google）与亚马逊（Amazon），都在同一片土地上激烈竞逐并全力推出顶级模型。正如刚刚结束的肯塔基赛马会（Kentucky Derby），美国正稳稳骑在赢面最大的“冠军之马”上，必须坚定不移地保持政策连贯性与技术领先优势。 演播室随后演变成一场内阁职位的搞笑封赏。大家打趣 Brad 若竞选公职，必须先摘掉标志性的红色眼镜并脱掉灯芯绒短裤，Brad 笑称自己更青睐“财政部长”（Treasury Secretary）一职；Jason 顺水推舟封 Sacks 为“国务卿”（Secretary of State），而 Chamath 则被戏封为掌管奢华享受的“开司米羊绒与红酒部长”（Secretary of Cashmere and Wine）。Sacks 总结道，主流媒体对微型劫掠（micro looting）等社会负面充斥着荒谬报道，却刻意忽视科技创新与经济基本面带来的巨大利好。在全场互道“爱你兄弟”（Love you boys）与经典的片尾神曲串烧中，本期播客圆满落下帷幕。 ## Notable quotes > [06:00](https://www.youtube.com/watch?v=10MdOvK-aG4&t=360s) 与需求毫无关系，需求是无限的，其营收完全受限于数据中心和电力的供给瓶颈 — Chamath Palihapitiya > [08:53](https://www.youtube.com/watch?v=10MdOvK-aG4&t=533s) 在这个星球上，没有人比埃隆更擅长把电子（Electrons）转化为代币（Tokens） — Brad Gerstner > [09:56](https://www.youtube.com/watch?v=10MdOvK-aG4&t=596s) 美国资本主义的混乱总能以某种方式找到出路 — David Sacks > [12:31](https://www.youtube.com/watch?v=10MdOvK-aG4&t=751s) 让我成为第一个祝贺达里奥·阿莫代伊（Dario Amodei）赢得本轮 AI 竞赛的人 — David Sacks > [14:09](https://www.youtube.com/watch?v=10MdOvK-aG4&t=849s) 如果按照当前的势头发展下去，达里奥称之为 AGI，而我称之为人类历史上最强大的垄断组织 — David Sacks > [26:48](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1608s) 但想象一下，如果洛克菲勒当年精通现代 PR 话术，他绝不会把公司叫做‘标准石油’（Standard Oil），而会把它包装成‘安全石油’（Safe Oil） — David Sacks > [30:01](https://www.youtube.com/watch?v=10MdOvK-aG4&t=1801s) 别动老子的账面投资（Don't mess with my paper） — Jason Calacanis > [01:07:07](https://www.youtube.com/watch?v=10MdOvK-aG4&t=4027s) 坦白讲，现在字面意义上没有任何一丝证据表明 AI 帮助 S&P 500 指数成分股提升了营业利润率（Operating Margins） — Chamath Palihapitiya > [01:09:42](https://www.youtube.com/watch?v=10MdOvK-aG4&t=4182s) 过去每次革命我们都在纠结：该不该做个 App？该不该建个网站？该不该上云还是留在本地？这些问号最终都变成了感叹号 — Jason Calacanis > [53:35](https://www.youtube.com/watch?v=10MdOvK-aG4&t=3215s) 如果你现在觉得医疗昂贵，那就等等看全民免费后的账单吧 — David Sacks ## Mentioned Jason Calacanis, Chamath Palihapitiya, David Sacks, Brad Gerstner, David Friedberg, Elon Musk, Dario Amodei, Sam Altman, Ken Griffin, Zohran Mamdani, Karen Bass, Nithya Raman, Shaun Maguire, Demis Hassabis, Kevin Hassett, Scott Bessent, Susie Wiles, Bernie Sanders, Donald Trump, Joe Biden, John D. Rockefeller, P.J. O'Rourke, Anthropic, Claude, Claude Code, Opus, xAI, Grok, Colossus 1, Colossus 2, Macrohard, Macroharder, SpaceX, Starlink, Tesla, Megapack, Powerwall, OpenAI, GPT-5.5, Spud, Codex, Sora, Cursor, Google, Google Cloud, Google DeepMind, Amazon, AWS, Microsoft, Azure, Nvidia, Meta, Apple, SK Hynix, Samsung, Micron, PulteGroup, Span, Base Power, CrowdStrike, Palo Alto Networks, DoorDash, Nike, Uber, Disney, Anheuser-Busch, Citadel, UnitedHealth, Altimeter Capital, The New York Times, The Wall Street Journal, FDA, Polymarket, Invest America ## Video description (0:00) Bestie intros! Thoughts on the LA mayor election (4:38) SpaceX-Anthropic deal, Elon Web Services, SpaceX IPO valuation, Anthropic's insane growth trajectory (26:48) Is Anthropic the next great monopoly? Early signals or major overreaction? (35:21) "FDA for AI" freakout, how the White House thinks about AI safety (52:01) Flipping AI's negative perception: Giving, healthcare and education innovation (1:00:04) Trading the AI market, state of the economy Apply for Summit 2026: https://allin.com/events Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg https://x.com/altcap Follow on X: https://x.com/theallinpod Follow on Instagram: https://www.instagram.com/theallinpod Follow on TikTok: https://www.tiktok.com/@theallinpod Follow on LinkedIn: https://www.linkedin.com/company/allinpod Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg Intro Video Credit: https://x.com/TheZachEffect Referenced in the show: https://x.com/WallStreetApes/status/2052413443739951366 https://www.anthropic.com/news/higher-limits-spacex https://x.com/elonmusk/status/2052073463029055926 https://x.com/shaunmmaguire/status/2052075296002625942 https://x.ai/news/anthropic-compute-partnership https://www.realtor.com/news/trends/nvidia-pultegroup-span-date-center-backyard https://www.gradesaver.com/there-will-be-blood/study-guide/a-brief-history-of-standard-oil https://www.nytimes.com/2026/05/04/technology/trump-ai-models.html https://x.com/SusieWiles47/status/2052192419718783246 https://www.nytimes.com/2026/05/05/business/dealbook/trump-ai-regulation.html https://www.whitehouse.gov/wp-content/uploads/2026/03/03.20.26-National-Policy-Framework-for-Artificial-Intelligence-Legislative-Recommendations.pdf https://x.com/ahall_research/status/2052042535661691282 https://x.com/SenSanders/status/2052116733683470556 #allin #tech #news [Source](https://www.youtube.com/watch?v=10MdOvK-aG4)

[6:00] Chamath Palihapitiya: 与需求毫无关系，需求是无限的，其营收完全受限于数据中心和电力的供给瓶颈 [8:53] Brad Gerstner: 在这个星球上，没有人比埃隆更擅长把电子（Electrons）转化为代币（Tokens） [9:56] David Sacks: 美国资本主义的混乱总能以某种方式找到出路 [12:31] David Sacks: 让我成为第一个祝贺达里奥·阿莫代伊（Dario Amodei）赢得本轮 AI 竞赛的人 [14:09] David Sacks: 如果按照当前的势头发展下去，达里奥称之为 AGI，而我称之为人类历史上最强大的垄断组织 [26:48] David Sacks: 但想象一下，如果洛克菲勒当年精通现代 PR 话术，他绝不会把公司叫做‘标准石油’（Standard Oil），而会把它包装成‘安全石油’（Safe Oil） [30:01] Jason Calacanis: 别动老子的账面投资（Don't mess with my paper） [67:07] Chamath Palihapitiya: 坦白讲，现在字面意义上没有任何一丝证据表明 AI 帮助 S&P 500 指数成分股提升了营业利润率（Operating Margins） [69:42] Jason Calacanis: 过去每次革命我们都在纠结：该不该做个 App？该不该建个网站？该不该上云还是留在本地？这些问号最终都变成了感叹号 [53:35] David Sacks: 如果你现在觉得医疗昂贵，那就等等看全民免费后的账单吧

[6:00] Chamath Palihapitiya: Anthropic and OpenAI's revenue performance has nothing to do with demand. Zero. It is entirely to do with the supply constraints that exist in data centers and specifically in power. [26:45] David Sacks: Imagine if John D. Rockefeller was way better at public relations. And instead of calling his company Standard Oil, he called it Safe Oil... Standard Oil's first big product was kerosene. Kerosene can light your house, or it can burn it down... So John D. should have called for the creation of a new government agency to regulate the safety of his product... People would have gotten so wrapped up in this debate over what constituted safe kerosene that they would have missed what was really going on—which is that Rockefeller was building the richest, most powerful monopoly of all time. [35:20] Kevin Hassett: We're studying possibly an executive order to give a clear roadmap to everybody about how this is going to go, and how future AIs that also potentially create vulnerabilities should go through a process so that, you know, they're released into the wild after they've been proven safe—just like an FDA drug. [35:20] Scott Bessent: What we had in the past month was a step change in the power of one large language model, but we're going to see it from the other AI companies. What we are determined to do is work with our AI companies to allow them to continue to innovate, but our charge at the US government is maintaining safety. [39:55] Chamath Palihapitiya: The response from the tech community... should be the leadership of the tech world coming together and actually reinvesting in America writ large. They're not doing that in enough of a scale that blunts this... So the response is: 'Hey hold on, we're going to give three guys trillion-dollar net worths and allow them to control the keys?' That's why this is happening. [78:36] Brad Gerstner: I thank our lucky stars that we have Elon, that we have Anthropic, that we have OpenAI, that we have Google, that we have Amazon all innovating in this country.

[0:00] How do I sound? [0:00] >> You sound perfect. [0:01] >> great. [0:02] >> I look? [0:02] >> Yeah, you sound great. Better than you [0:04] look. [0:05] >> You got that face made for radio. [0:07] >> You [laughter] don't look as tired as [0:09] you have in recent weeks. [0:10] >> That's true. Yeah. [0:11] >> Yeah. [0:11] >> Somebody was slagging me for the bags [0:13] under my eyes. I mean, this audience is [0:16] brutal. [music] [0:17] >> They're brutal. They're brutal. [0:19] >> It's a good thing I'm rich. [0:20] >> [laughter] [0:24] [music] [0:24] >> You let your winners ride. [0:27] Rain Man David Sacks [0:30] >> [music] [0:32] >> We open sourced it to the fans and [0:33] they've just gone crazy [music] with it. [0:39] >> All right, everybody. Welcome back to [0:41] the number one podcast in the world. [0:42] It's the All-In podcast. With us today [0:46] Chamath Palihapitiya, David Sacks, and [0:49] our fifth bestie, Mr. Brad Gerstner is [0:52] here. I think David Friedberg is [0:55] suffering from some [0:56] socialist-related flu. [0:58] >> He's very sick of reading about [1:00] >> socialists, but he'll be back next week [1:02] with two incredible, incredible [1:04] interviews. [1:05] >> You guys see those Spencer Pratt ads? [1:07] >> Wow. [1:08] >> It's one of the best political ads I've [1:09] ever ever seen. [1:10] >> Oh, there's like three or four of them. [1:12] >> There are multiples. [1:13] >> Yeah. [1:13] >> Whoever that social media team is is on [1:15] fire. [1:15] >> If you get a good social media team and [1:17] you get a good ad production team, I [1:19] think it's next gen cuz these things go [1:21] crazy. And Spencer Pratt, if he wins [1:23] this election, which I think he's going [1:24] to in Los Angeles, the reason is what [1:27] Brad said. Those ads are incredible. [1:30] >> Well, he's also quite a good debater. [1:31] Did you see clips from this debate? [1:34] >> Absolutely. He's so funny. He's so [1:37] chill. [1:38] >> Yeah, well, he's up against Karen Bass, [1:39] who's the mayor, who is basically [1:41] extremely left-wing. And then there's [1:43] someone who's a city councilwoman, who's [1:45] even further to the left of Karen Bass. [1:48] I mean, she's [1:48] >> Yeah, yeah. [1:49] >> often like Fidel Castro territory. [1:51] >> She's an Indian Fidel Castro. [1:53] >> Raman. So, she was basically, I guess, [1:56] criticizing the mayor for the [1:58] homelessness problem. [2:00] And then Pratt pointed out that this [2:02] councilwoman is actually in charge of [2:04] all these homeless programs already. [2:07] >> [laughter] [2:07] >> He eviscerated her. [2:08] >> He eviscerated her. And he basically [2:10] made the key point, which is, "Look, the [2:12] problem here is not lack of housing, [2:13] it's an addiction issue, and it's a [2:15] mental illness issue." And he said, [2:17] "Look, if you [2:18] >> said [2:18] he said if she went to the street, she'd [2:20] get stabbed in the neck. [2:21] >> Yeah. [2:22] >> [laughter] [2:23] >> Which is pretty accurate if you've been [2:24] to Skid Row. I mean, you would not want [2:26] to walk through there. [2:28] It was like the Spider-Man photo. [2:30] >> got the clip. Play the clip, Nick. [2:32] >> Oh, god. [2:32] >> This clip is brutal. [2:33] >> a different clip, but this one super [2:34] viral. It reminded me of Trump a little [2:36] bit. [2:37] >> Let's see. [2:38] >> I'm not sure how to respond to that [2:39] vision of Los Angeles. This is a MAGA [2:41] Republican's idea of what Los Angeles [2:44] looks like. This is This is really not [2:47] >> [laughter] [2:49] >> Unbelievable. [2:51] >> For those of you listening, he put his [2:52] hands up and wiggled his head like, "Oh [2:54] my god." [2:55] >> [laughter] [2:55] >> Hey, hey, Sachs. You know, Stephen Pratt [2:58] wins mayor, [2:59] >> Spencer Pratt. [2:59] >> Spencer [3:00] >> the you know, the ballot initiative, the [3:02] retirement protection and savings act, [3:04] right? It's going to pass. It's going to [3:05] pass with big numbers. This is the, you [3:07] know, referendum that effectively is [3:09] going to knock out the wealth tax. Can [3:10] you imagine if California effectively [3:13] passes a constitutional amendment [3:15] protecting retirement savings and [3:17] personal assets and banning the wealth [3:19] tax and Pratt gets elected, the message [3:21] that would send to the country? That's [3:23] That's a very non-consensus view that [3:26] I'm becoming increasingly optimistic [3:28] about. [3:29] >> Well, from your lips to God's ears. [3:32] But until that message actually is sent, [3:34] I think I'm going to be uh [3:36] >> in Texas. [3:37] >> In Texas. [3:38] >> [laughter] [3:40] >> Well, I mean, this is also in the face [3:41] of I don't know, just a follow-up story [3:43] here, but Mondami did like an attack [3:45] video on Ken Griffin's house. We talked [3:47] about it on the pod a couple weeks ago. [3:49] And like, literally stood in front of [3:50] his house, pointed at it. and this is in [3:52] the face of like a CEO getting shot for [3:56] ideological reasons, Sam Altman's house [3:58] being targeted. This is like a really [3:59] dangerous thing for Mondami to do. [4:02] And Ken Griffin came out today in or [4:04] yesterday in an interview and said, [4:05] "Hey, listen, I'm out. We're going to be [4:07] putting our efforts into [4:10] Florida." And this is the same thing [4:12] that happened to Chicago, and he [4:13] basically said like I really felt [4:14] offended, and I was, you know, nervous [4:17] about this [4:18] because of my personal safety. And he [4:20] called him out. Mondami came out with [4:22] like [4:23] a mealy-mouthed response that didn't [4:25] even apologize for what he did, [4:27] just doubled down on it essentially. All [4:29] right, let's get to the I don't know if [4:30] you guys saw it or not, but who cares? [4:33] And New York is becoming a flyover city. [4:36] >> It's an interesting way to put it. Don't [4:37] disagree. [4:38] >> All right, first story. Elon just leased [4:40] all of Colossus 1. [4:41] >> New data center? [4:42] >> He did? [4:44] >> Nostro What? What? [4:46] >> Yes, shocking to Dario and Anthropic [4:49] Chamath on this week's pod. Go ahead and [4:51] give yourself a pat on the back. You [4:53] said Elon and Dario should do a deal [4:54] tomorrow. It didn't happen. The next day [4:56] it happened 5 days later, so you came [4:58] close, Chamath, but no cigar because of [5:01] Anthropic's obvious compute constraints. [5:04] Anthropic just added over 220,000 Nvidia [5:07] GPUs, over 300 megawatts of energy. [5:10] The deal is already having an impact as [5:13] we've discussed here. Claude users have [5:15] been experiencing rate limits. Well, [5:17] Claude has now doubled the Claude code [5:19] rate limits, removed peak usage caps for [5:21] paid users, and increased API volumes [5:23] for Opus models. XAI is now trading [5:26] their models at Colossus 2. So, they [5:29] have more than enough compute. Elon made [5:31] a great bet [5:32] on compute and built up those data [5:34] centers really fast, and that is now [5:37] paying off. We had the cursor deal we [5:38] talked about last week. Let's talk about [5:41] the emergence of Elon Web Services, EWS, [5:45] Chamath. He is now in the hyperscaler [5:48] competing against Google Cloud, Amazon [5:50] Web Services, and Azure. And uh I don't [5:52] know if you had inside information or [5:54] just a brilliant uh epiphany, but uh [5:56] take us behind the call and what do you [5:58] think about the deal itself? [6:00] >> I think the deal is fantastic. I'll say [6:03] maybe three quick things. [6:06] The first is, as I mentioned a couple [6:08] weeks ago, [6:10] Anthropic [6:11] and OpenAI's revenue performance has [6:14] nothing to do with demand. Zero. [6:18] It is entirely to do with the supply [6:21] constraints that exist in data centers [6:24] and specifically in power. [6:27] If they had infinite power, I think that [6:29] their revenues would probably be even [6:31] more parabolic. [6:33] And so all the breathlessness about [6:35] either exceeding or underperforming a [6:37] forecast, in my opinion, mean nothing. [6:40] I think the five-year view [6:42] for those two companies [6:44] is quite robust. [6:46] The thing that they really need is more [6:49] compute and more power. That's the first [6:52] thing. [6:53] The second thing is, while they need [6:55] that, we have a very big problem, [6:58] which is we unfortunately have very poor [7:00] leadership [7:02] at the head [7:03] of most of these AI firms. [7:07] I think they are coming off as [7:09] untrustworthy or too self-interested. [7:13] The political reaction now [7:15] is starting to turn negative. [7:18] The community reaction is negative. [7:21] You have about 9 gigawatts that are [7:24] supposed to come online this year. [7:26] Almost 50% of it now is being protested. [7:29] More than likely, if if history holds, [7:33] most of that will get turned off. So [7:35] they will get even more supply [7:36] constrained. So that's the setup. [7:39] So what's the opportunity? I think for [7:41] Elon, if you look inside of how people [7:44] try to nitpick the SpaceX valuation [7:46] case, or let's not even let's give them [7:48] sorry, let's be more generous. [7:50] When people try to paint the bear case, [7:52] or they try to red team the valuation, [7:54] the biggest element is the on the come [7:57] value around the orbital data centers. [8:01] And by actually landing a bunch of [8:03] terrestrial capacity, [8:05] I think you start to blunt that. Because [8:07] you can now start to say that even if [8:09] the orbital data centers get delayed by [8:11] a few months or a few quarters, even if [8:13] the technological [8:15] de-risking of it takes longer, [8:17] he now has a structural core business [8:19] that will effectively subsidize [8:22] his ability to train Grok, which I think [8:25] is a really important and underreported [8:27] theme. [8:28] So, you have all this infrastructure, [8:30] he somehow saw the tea leaves [8:32] before most people, he built to a level [8:35] of scale and secured power [8:37] before most people. It is now become the [8:39] critical asset. Now, he's [8:41] kind of king-making. [8:43] And I think that that's a really [8:44] interesting valuation [8:47] reinforcement as SpaceX goes through [8:49] testing the waters in the in the road [8:51] show. Brad, your take. [8:53] >> Yeah, no, I think it's well said. I [8:54] mean, first we we know that there's [8:57] nobody better on planet Earth than Elon [8:58] at converting electrons to tokens. It's [9:01] a critically important evolution to the [9:04] story. You know, I think our friend [9:05] Shawn Maguire, he he he sent out a tweet [9:08] that summed it up well, and he said [9:09] SpaceX has this five-layer cake, launch, [9:12] connectivity, compute hyperscaler, space [9:15] data centers, and then applications and [9:18] models, and then other bets, right? The [9:20] question on the road show has been, "But [9:23] x.ai doesn't on the revenue trajectory [9:27] of OpenAI and and Anthropic, and yet [9:29] there are huge commitments. And now we [9:32] see the ace card that Elon's playing." [9:35] He said he was building AWS all along or [9:38] EWS all along. And so I estimate that [9:41] this is going to generate in this year [9:43] an incremental 45 billion dollars of [9:46] revenue on top of what I have seen [9:48] analyst estimates in the mid-20s. That's [9:50] a material amount of incremental revenue [9:52] to offset the cost of the investments [9:54] that he's made here. And that will [9:56] subsidize to Chamath's point all that [9:59] he's investing to build the next [10:00] generation of growth. Remember too that [10:04] he has three facilities. Colossus, macro [10:07] hard and macro harder. 1.2 gigawatts in [10:10] macro hard and macro harder in [10:11] Blackwell. So he's given the one that's [10:14] kind of less connected H100s great for [10:16] inference to Anthropic. He's monetizing [10:19] it in a big way. It's terrific for [10:21] Anthropic. And it solves what I think [10:24] was the biggest question in the [10:26] valuation story, which is what if he [10:29] spends ahead of x.ai's revenue. It takes [10:33] the pressure Chamath off x.ai delivering [10:36] immediate revenue. Now he becomes [10:38] an immediate competitor in the [10:40] hyperscale. I don't think this is the [10:41] last announcement. I think he's going to [10:44] make a lot more, [10:46] you know, moves in this direction. I [10:48] think it will be a material part of [10:49] their story and their revenue [10:50] projections [10:52] as they come together. And I would just [10:54] say finally, you know, again, everybody [10:56] has talked about how we don't have [10:58] enough power, how we don't have enough [10:59] compute, how the revenues would not show [11:01] up in this year. You know, but the chaos [11:03] that is American capitalism somehow [11:06] finds a way, okay? And there's [11:08] tremendous demand for Anthropic and we [11:11] find a way. I was so happy to see kind [11:14] of the detente and the kind exchange [11:17] between the team at Anthropic and Elon [11:20] because we need all of this in order to [11:23] produce American frontier models to stay [11:25] at the frontier. [11:27] And then finally I just say, you know, [11:28] Chamath, you referenced these activists [11:30] that are protesting delaying these data [11:32] centers [11:33] in in these localities. One thing I want [11:35] to dispel this myth. These This is not [11:37] like organic hyper local protests by [11:40] people in a community that aren't being [11:42] spurred on. [11:43] This is highly organized activists that [11:46] are moving across the country to stir up [11:48] trouble in the exact same way they did [11:51] to stop all fission reactors being built [11:53] 30 years ago in America. Now we have no [11:56] nuclear reactors being built. China's [11:58] got a hundred of them. Who was funding [12:01] those activists? I think we need to [12:03] really look into who's funding the [12:04] activists now. I'm not saying that there [12:06] aren't any concerns, but the [12:07] misinformation about water, the [12:09] misinformation about electricity bills [12:12] electricity bills are going up in the [12:14] places that are not building data [12:16] centers. New York and California because [12:18] they haven't built any supply on the [12:21] grid. In Texas, where you're building [12:22] the most data centers in the country, [12:24] electricity costs are going down. So, um [12:27] I'm I I I think that's a boogeyman that [12:28] we got to take on. [12:31] >> Sacks, sure that's [12:33] >> Well, look, the deal is highly [12:35] complimentary as Chamath and Brad [12:37] pointed out. SpaceX has a profitable, I [12:40] think very profitable, [12:42] space and telecommunications Starlink [12:44] business, the satellite business, but [12:47] the xAI business had huge losses. The [12:49] reason's pretty straightforward. You [12:51] need these super large training [12:52] clusters, but they cost a lot of money. [12:54] And until you have a model that's [12:56] capable of competing at the frontier, [12:57] you're not making any revenue. And that [12:59] problem was compounded by the fact that [13:01] right now all the revenue is in [13:02] enterprise, which is to say coding. We [13:04] know that xAI just did that deal with [13:07] Cursor to try and catch up, but they [13:09] don't have a coding product yet. So, [13:11] they're not participating in the [13:12] revenue, but they're participating in [13:13] all of the cost. So, this deal fixes [13:15] that problem. [13:17] Elon's now able to have a frontier model [13:20] company, but he's able to now not have [13:23] these massive unpaid for CapEx [13:26] commitments, right? Because he's able to [13:27] kind of lease that capacity. So, I think [13:30] it solves a major problem for them and [13:32] their balance sheet. And then you have [13:34] to say that for Anthropic, this is a [13:36] really great thing because they were [13:38] compute constrained. [13:39] And just to build on that point, I mean, [13:42] I guess let me be the first to [13:43] congratulate Dario on winning the AI [13:45] race. [13:46] >> And you've been Let's be honest, Sachs. [13:48] You have been [13:50] on this podcast, you've been moderately [13:52] critical of that company and Dario [13:54] himself for being um you know, a little [13:58] P-doomer 110. And on your X account, [14:00] you've been even a little spicier. So, [14:03] now that there's peace in the Middle [14:04] East of [14:06] of the AI business, what's your take [14:08] here? [14:09] >> My take is, look, let's just honestly [14:11] and accurately assess where the state of [14:14] this AI market is at right now and [14:16] Anthropic's place within it. So, for the [14:18] last 3 years, Anthropic has been growing [14:21] at a rate of 10x a year. [14:23] I think going into this year, [14:25] probably the conventional wisdom was [14:27] that there'd be no way to sustain that [14:29] kind of rate of growth at this level of [14:31] scale. And what happened in the first 4 [14:33] months of the year? First, we find out [14:35] that from January 1st to March 31st, [14:37] they grew from roughly 10 billion of ARR [14:40] to 30 billion. So, it tripled. And then [14:42] in April, if anything, the rate of [14:44] increase seemed to accelerate. They went [14:46] from 30 to 44 billion of ARR. Nobody in [14:50] Silicon Valley has ever seen anything [14:52] like it. Forget about the rest of the [14:53] country. I mean, all we do in Silicon [14:55] Valley is deal with exponentials, and [14:57] still people have never seen that kind [14:59] of growth at that level of scale. [15:01] The only thing holding them back in the [15:03] future was compute. Now they've made [15:05] this deal. They've made other deals as [15:07] well to get that compute. I think it's [15:09] pretty much a foregone conclusion that [15:10] they will hit that forecast of 10x this [15:13] year, exiting the year at call it [15:14] roughly 100 billion of ARR. And now the [15:17] only question is whether they hit a [15:19] trillion [15:20] in 2027. [15:22] And we can debate whether that's [15:24] >> on board. [15:25] >> No, we can't we can debate whether [15:26] that's true or not. But look, if they do [15:28] that, [15:29] I think they'll easily be the most [15:31] valuable tech company in history. In [15:33] fact, they might even be more valuable [15:35] than the rest of the Mag 7 put together. [15:37] Just to give people some basis for [15:39] comparison here. You know, the biggest [15:41] tech companies Apple, [15:42] >> Nvidia, [15:43] >> Nvidia, Google. I think they they kind [15:45] of do around 4 to 500 billion a year [15:47] right now [15:48] of of revenue. I mean, I guess Nvidia's [15:49] a little bit of a different category, [15:51] but you look at you look at Google. [15:53] >> the three [15:53] >> The hyperscalers, yeah. I mean, Google [15:55] is doing what? Like 120 billion a [15:58] quarter? Something like that. 100 [15:59] billion a quarter. [16:00] >> Correct. [16:01] >> But growing at what? 20% year over year? [16:03] Not 100% and certainly not 1,000%. [16:06] So, the fact that Anthropic could be on [16:09] track, in fact, let me correct that. [16:11] >> them going to the Mag 8? It'll be a Mag [16:13] 8. [16:14] >> I'm saying something else, which is that [16:17] unless something about their current [16:19] trajectory [16:21] changes, [16:22] Anthropic will be the most powerful [16:24] monopoly ever created in human history. [16:27] >> Go. [16:28] >> Again, it will be, you know, a trillion [16:30] dollars of ARR growing at some [16:32] exponential [16:33] >> Interesting. [16:34] >> Dario calls it AGI. I call it the [16:36] biggest monopoly in human history. [16:38] >> Interesting to hear that word monopoly, [16:40] Sacks. Very interesting placement. [16:42] Chamath, go ahead and then I'll move on. [16:44] >> Apple in 2025 was 420 billion. [16:48] Microsoft was 300 billion. Alphabet was [16:51] 390 billion. Amazon 700 billion. Nvidia [16:55] 190 billion. Meta 185 billion. Tesla 110 [16:59] billion. Total [17:00] about 2.3 to 2.35 trillion. So, if Sacks [17:03] is right [17:04] and Anthropic, [17:06] you know, can tack on a trillion, it [17:09] won't be the Mag 7, it'll be the Mag 1. [17:12] >> Just to put it in perspective though, [17:13] Dario and Dwarkesh said he thought the [17:16] combined AI revenue [17:18] of the market leaders would be about a [17:19] trillion in 2029. I love what you're [17:21] saying, Zach. I think there is unlimited [17:23] TAM. We may be over our skis a little [17:26] bit in terms of, you know, the forecast. [17:28] If you back your way from compute, [17:30] right? They're They expect to have 5 [17:32] gigs by the end of this year, 10 gigs by [17:33] the end of next year. It's kind of hard [17:35] to get to those numbers for a single [17:36] company, but I do believe that you know, [17:39] the trajectory that they're on, I [17:41] totally agree with you, is on an [17:42] exponential that not many people [17:45] believed in 4 months ago. [17:46] >> Right. So then the question is, okay, I [17:48] think we all agree they're on an [17:49] exponential curve and that the TAM is [17:51] big enough to support that. Just one [17:53] data point on TAM. My understanding of [17:56] the total market size just on coding [18:00] is 1 trillion, meaning that a trillion [18:01] dollars a year roughly is spent on [18:04] software developers and all things [18:06] related to the creation of software. [18:08] Now, I'm not saying that they eat that [18:09] entire market, but I can easily see the [18:12] market for software doubling. [18:14] >> Well, hold on. It's doubling from a [18:16] trillion to 2 trillion given that [18:19] coding tokens basically 10 X's or 100 [18:21] X's the value of that market and the [18:25] ability to generate code. So, I think we [18:27] all agree that the TAM here is large [18:29] enough to support a trillion dollars of [18:31] revenue. Brad, I think you bring up a [18:33] couple of really important constraints. [18:34] First, there may not be enough compute [18:36] and there's not enough energy. I'd say [18:37] the second big one is, what's the [18:39] competitive reaction going to be? [18:41] Because I would say at the beginning of [18:43] this year all these frontier labs were [18:45] playing around with a lot of different [18:46] things. I mean, [18:48] Anthropic was the porcupine. They [18:50] believed in one thing. All these other [18:52] companies were kind of acting like the [18:53] fox who thinks they're good at a lot of [18:54] different things. They were doing nano [18:56] banana, they were doing Sora, they were [18:59] doing, you know, they were doing image [19:00] generation, they were doing fantasy [19:02] character chatbots. In hindsight, they [19:05] were doing a lot of things that appeared [19:06] to be kind of a waste of time. The whole [19:09] market appears now to be coding and the [19:11] things that we built on coding tokens [19:13] like co-work, like agents. And so there [19:16] is going to be a competitive response [19:17] here where all the other guys realize, [19:19] "Oh, wait a second. We were misfocused." [19:21] They're going to get focused. I just [19:23] don't know how much share they're going [19:25] to be able to take. It does look like [19:27] OpenAI has already made the pivot. We [19:29] hear very good things about Codex now [19:31] based on GPT 5.5. 5.5 is based on a new [19:35] base model called Spud. I think they're [19:36] very optimistic about continuing [19:38] improvements. Their rate of growth [19:40] appears to be accelerating now because [19:43] of 5.5. [19:45] So, look, there's reason to believe that [19:47] OpenAI can take some share here. I'm [19:49] sure that Google won't be asleep at the [19:52] wheel. They're very, very good at [19:53] coding. They've got a really good team. [19:54] And Elon just tied up with Cursor. So, [19:56] there is going to be more competition. [19:58] But still, what you have to say, [20:00] and I think all of us know this from [20:02] Silicon Valley, is you always want to be [20:04] the company in the lead that's on that [20:06] trajectory where all you have to do is [20:08] maintain inertia, whereas the other [20:11] people have to change something in order [20:13] to put themselves back in the race. So, [20:15] this is when when I say somewhat sort of [20:18] facetiously, "Congratulations, Dario, on [20:20] winning the AI race." I don't mean that [20:22] he's won it, but he is winning it right [20:24] now. [20:25] >> Well, here's the the brilliance of what [20:26] Elon's doing. If you look at the [20:28] existing business, which is Starlink and [20:30] basically the launch services at SpaceX, [20:33] incredible business, obviously. [20:36] 20 billion this year, I think is the [20:37] estimate. But if you look at the [20:38] footprint of Amazon Web Services, Azure, [20:41] and GCP, you're looking at, you know, [20:44] 300 billion dollars in revenue [20:48] and a market cap of combined 5 trillion, [20:53] 4 trillion if these were independent [20:54] companies. And if you look at what is [20:57] Elon's core competency at Tesla, it's [21:00] building factories. And if you look at [21:02] the footprint of these factories, [21:04] they're huge. What are data centers? [21:05] They're basically big giant factories. [21:07] And then if you look at energy, what [21:10] else is Elon extremely good at? This is [21:12] the battery deployment and he's also got [21:14] solar deployment from the SolarCity [21:16] often criticized acquisition he did [21:18] years ago. So you put this all together, [21:21] if this is 5 billion as I think you [21:22] referenced Brad, if it's $5 billion in [21:25] incremental [21:27] Elon web services business and he's a [21:29] neo cloud, [21:30] what could he build on planet Earth? [21:33] What could he build inside of Tesla's in [21:35] terms of extra compute? What could he [21:37] build inside the power wall? What if the [21:39] power walls had his new fabs in them and [21:41] you built a distributed system from home [21:43] to home? The power wall has compute in [21:46] it, the cars have compute in it and of [21:48] course [21:49] the ultimate manifestation of this where [21:51] nobody can complain is you go right out [21:54] into space. And that's what he's going [21:56] to do and the the sneaky small part of [21:58] this announcement from Elon and from [22:01] Anthropic and Dario was they're also [22:04] interested in space. So look for the [22:06] race to go from factories and data [22:09] centers to homes, the power wall with [22:12] compute in it. It's already online, [22:14] right? And Starlink also gives him the [22:16] ability to do distributed compute to [22:18] people's homes. Again, you could be [22:20] paying people to put power walls with [22:22] computer in it. That's going to be the [22:23] next shoe to drop, I believe. [22:25] >> Did you guys see the deal that was [22:26] announced [22:28] yesterday [22:30] between Pulte Homes, which is a huge [22:31] builder, [22:32] >> Yes. They're doing it as well. [22:33] >> and Span? [22:34] >> Yes. [22:35] >> Nick, just throw this up here. It's [22:36] super cool. What's happening is that [22:37] these guys [22:39] are putting [22:40] many data centers with Nvidia GPU [22:42] clusters [22:44] beside every home and then allowing [22:46] people to actually run those things and [22:47] that's just incredible. I thought that [22:49] was so cool. [22:50] >> It's a great pivot. What this company [22:51] did originally Chamath was they did the [22:53] power panel. They made smart power [22:55] panels. So you know when you flip your [22:56] breakers, [22:57] all those breakers are in an app. I [22:59] looked at it for my house, but I guess [23:00] they pivoted to add this and I think [23:02] Base Power Brad urine industrial unit [23:04] they're going to do the same thing. [23:05] >> Yep, Zach Dell's doing that. One of the [23:07] things I just would say in response [23:08] Jason to what you just said about Elon, [23:10] right? This is why the SpaceX IPO is [23:13] going to trade at 40 to 50 times [23:16] revenue. Okay, so next year if they do [23:18] 40 to 50 billion, and this thing goes [23:21] out at 2 trillion, right? That they're [23:23] going to trade at a really high revenue [23:25] multiple compared to the Mag 5 that are [23:28] trading at like 25 times earnings. And [23:30] there's only one person on the planet [23:33] who has a future pipeline of innovation [23:36] and the largest TAM in the world because [23:38] he's playing in all these different [23:39] spaces that can command that multiple, [23:41] and it's Elon, and it's deserved, and [23:44] it's great for the country. [23:45] >> same Tesla's has that same Elon [23:48] variable in it as well, which is people [23:51] value his companies at I would say two [23:53] times market, three times market, four [23:55] times market because of the future [23:56] pipeline. You guys And they devalue [23:58] Apple because they don't have somebody [24:00] like Elon or Steve Jobs there who is [24:02] giving them the future. [24:03] >> I don't think it's devalued. I think [24:04] >> Or properly valued if you don't have an [24:06] Elon and you have somebody like I think [24:08] that's exactly what it is. We talked [24:10] about this last week, but explain why [24:11] you think it's different. [24:12] >> all of these companies are actually very [24:14] fairly valued, and then [24:16] Elon world gets a premium. [24:18] >> Totally. [24:19] >> And that premium is because of what you [24:21] guys said. That I agree with. The big [24:23] message that I take away from this, [24:25] which the markets [24:27] and retail are telling you is [24:29] you guys have stopped innovating. [24:32] There's a lot of incrementalism, [24:34] and we as a society aren't benefiting [24:37] broadly the way that you told us we [24:39] would be. [24:40] And so maybe this is the best way for [24:42] them to get this message, which is to [24:44] whack their valuation. And by the way, [24:46] I'll just say it again, when Tesla and [24:48] SpaceX merge and we have all things Elon [24:51] in Elon Corp, okay, which will happen [24:54] probably by at end of the year, maybe [24:55] it'll happen the middle of next year. [24:59] It's going to then break everybody's [25:00] brains again because you'll have this [25:02] one asset, as you guys said, that will [25:05] trade at a valuation premium. And some [25:07] people will say it's unexplainable. And [25:09] I think it's logically explainable, [25:11] which is everybody else has stopped [25:14] innovating. People know how to draw more [25:16] blood from the stone, how to target [25:19] better ads. [25:21] That does nothing for society anymore. [25:23] >> Yeah. That's it. Literally. [25:24] >> In fact, it does the opposite. It There [25:26] is no good left. [25:27] >> That was literally the exact point I was [25:29] making when you cut me off. If you look, [25:30] Tim Cook's greatest innovation [25:33] Tim Cook's greatest innovation, before [25:35] you cut me off, was Apple TV. Not even [25:37] the hardware product. It was just [25:39] spending money and making a Netflix [25:40] knockoff. There's been no other product [25:42] >> Hold on. Let me finish again before you [25:45] interrupt me making my point. [25:46] >> that? You don't like being interrupted? [25:47] >> Oh, no. Oh, yeah. Well, okay. Okay. [25:49] >> Interesting. [25:50] >> Pot meet kettle. [25:52] If you look at their track record, and I [25:54] think this is why we had a change there [25:57] is they have not done anything [25:59] innovative. And in fact, the things they [26:00] were doing that were innovating in AI or [26:02] self-driving cars, they shut down. They [26:05] won't take any swing for the bat. So, [26:06] they are getting penalized in their [26:08] valuation. [26:09] >> They're just not getting a premium. [26:11] They're not getting penalized. [26:12] >> I think they're getting penalized. [26:13] >> metric, they're trading at incredible [26:15] valuations. Just look at them. [26:17] >> Oh, no. I don't I I mean, if you compare [26:18] the two valuations, I think they're [26:20] being penalized. Anyway, let's Anybody [26:22] else want to get in on this before we [26:23] move on to the next one? Yeah. [26:25] >> There is no world in which Google and [26:27] Meta and Apple and Amazon could be [26:32] viewed as being penalized in valuation. [26:35] There is very clearly a world where Elon [26:37] gets a massive premium because he's [26:39] innovating. [26:41] >> You're saying the same things. You're [26:42] saying the same things. [26:43] >> we're saying the same thing. It's not [26:44] the same thing. [26:45] >> We're I think we're we're debating [26:46] semantics here. I'm not letting you off [26:48] the hook, Saxby poo. [26:51] When Sax is very deliberate in how he [26:53] speaks, they said he's the captain of [26:54] the debate club in his 20,000 word [26:56] article this week and that he's a master [26:58] debater. He's a masturbator. And [27:01] you slipped in you slipped it in. [27:04] Are you saying that the FTC or whoever [27:07] should be going in and looking at [27:09] Anthropic? Oh, Brad's book is getting [27:12] attacked headwinds. You said they're a [27:13] monopoly or they're heading to monopoly [27:15] tactics Sacks is are that what you're [27:17] saying? [27:19] >> Well, look, I mean we know that tech [27:21] markets have a history of consolidating [27:25] down and turning into either monopolies [27:27] or duopolies. And if you just look at [27:29] the revenue right now, there's only two [27:31] companies making substantial revenue on [27:35] AI. It's Anthropic and Open AI. We know [27:38] that Open AI is growing at three to four [27:39] X, which is incredible at the level of [27:41] scale they're at. Anthropic though we [27:43] said is growing at an exponential 10 X a [27:46] year. And if they just do that for 18 [27:49] more months, they'll be by far the most [27:51] valuable company in human history and [27:53] they'll have unprecedented control over [27:55] the most important technology of our [27:57] time. So, I don't know what you call [27:59] that, [28:00] but it is something to think about. And [28:02] I guess I do have a thought experiment [28:04] for you guys, [28:05] which is [28:07] I just want you to think for a second [28:09] about the case of of John D. [28:11] Rockefeller, who I think is known as [28:14] probably [28:15] the most successful greatest most [28:17] ruthless monopolist in in American [28:20] history. But he wasn't very good at PR. [28:22] He was terrible at PR. Everyone sort of [28:24] recognized how ruthless he is. We see [28:26] movies like There Will Be Blood, which [28:28] is basically about him. In any event, [28:30] imagine if John D. Rockefeller was way [28:32] better at public relations. And instead [28:35] of calling his company Standard Oil, he [28:38] called it Safe Oil. Okay, let's just [28:40] let's just play this thought experiment. [28:42] >> Clean beautiful coal. Yes. [28:43] >> Safe oil. He called it safe oil because [28:46] as we know kerosene is dangerous. Their [28:48] first big product was kerosene. And [28:51] kerosene can light your house or it can [28:52] burn it down. [28:54] And in the wrong hands, it can torch a [28:56] city or you can use it to make a bomb. [28:59] So, John D [29:01] let's say should have called for the [29:03] creation of a new government agency to [29:05] regulate the safety of his product. And [29:07] they could have done rigorous testing, [29:09] licensing, common sense regulation. [29:11] There would have been a very intense [29:12] debate over safety standards. [29:15] You know, what should the proper wick [29:16] thickness be? And should we allow all [29:18] those [29:20] dangerous independent refiners, right? [29:23] And I think people would have gotten so [29:24] wrapped up in this debate over what [29:26] constituted safe oil or safe kerosene [29:30] that they would have missed what was [29:31] really going on, which is that [29:33] Rockefeller was building the richest, [29:34] most powerful monopoly of all time. In [29:36] fact, people might even have called [29:38] Rockefeller an effective altruist [29:41] because of course he was so concerned [29:43] about the safety [29:45] of his product. [29:46] >> it. Shout out to David Sacks and [29:48] writers. Great, great, great [laughter] [29:50] writers. [29:50] >> Newman? Newman wrote this? [29:52] >> No, I wrote it. [29:53] >> Well, Emmy award for best writing in a [29:55] dramatic monologue goes to Newman. Wow, [29:58] Sacks. [29:58] >> No, that's my writing. [29:59] >> You landed it. Very good, Sacks. [30:01] >> I thought after the Elon Anthropic [30:03] detente where Elon said, you know, [30:05] complimented Anthropic that and David [30:08] started off with a bit of a compliment, [30:09] I thought we maybe were past this. [30:11] First, it's ridiculous to think of this [30:12] as a monopoly. You know, we're talking [30:15] about annual run rate revenues, David, [30:17] but on a GAAP basis, they're doing about [30:19] the same revenue as OpenAI in the month [30:21] of March. Okay? So, we're way ahead of [30:23] ourselves. By the way, 5 months ago, [30:25] everybody thought OpenAI was going to [30:27] run away with this. Google's revenues [30:29] are very substantial in AI. And by the [30:31] way, Google, Amazon, etc., these [30:34] companies are producing $100 of free [30:36] cash flow to justify their incremental [30:38] investment. At the same time, you have [30:40] these two startups that are still [30:42] fledgling, that are still fragile in the [30:44] scheme of things, [30:45] you of all people should know we've got [30:47] the best competition in AI on the [30:49] planet, which is why we're at the [30:51] frontier and kicking the tail of [30:53] everybody else on the planet. So, I just [30:55] want to see these companies compete. I [30:57] want to see DC stay out of the way. The [30:59] last thing I want to be doing is is is, [31:02] you know, seeing people talk about this [31:04] and throwing roadblocks into the way of [31:05] the competition. [31:06] >> Um I think Anthropic is [31:08] >> Well, let me let me hold on. Let me [31:10] >> Translate Brad for you. Don't with [31:12] my paper is what he's saying. He's [31:13] [laughter] got bets on these. So, Sax, [31:16] Washington, D.C., don't with Brad's [31:18] paper. Sax, uh you want to get into the [31:21] uh regulation stuff right now as a segue [31:23] or [31:24] >> Let me respond to Brad and also [31:25] translate what I'm saying satirically, [31:27] okay? [31:28] >> Satirically. [31:29] >> First of all, nobody wants to see these [31:32] companies compete vigorously more than [31:35] me. That was the whole premise [31:37] of the action plan that we worked on [31:39] last year is we want to bring out the [31:41] best in everyone. This is how America's [31:43] going to win the AI race. We have five [31:45] major labs vigorously competing. [31:48] And as long as that competition is [31:49] taking place, that I think that's a good [31:51] thing. Doesn't mean we can't have [31:52] guardrails and the rest of it, but [31:53] basically, competition should be our [31:55] North Star. [31:57] All of that being said, okay, what I am [31:59] pointing out and I I think it's [32:01] historically true [32:02] that people in Washington have woken up [32:05] to monopolies [32:07] on the late side, not early, right? [32:09] Because I mean, once a company has won [32:12] 80% of the market, that's when they wake [32:14] up and say, "Oh, we have a monopoly [32:15] here." [32:16] And I'm not saying that they have a [32:18] monopoly yet, but if the trajectory [32:20] continues for just 18 more months, [32:23] then I think it will be in this [32:25] unpresently powerful position. [32:28] >> I mean I mean [32:28] >> And and hold on. And I don't think [32:31] people should be distracted from that [32:32] fact [32:34] by this rhetoric around safety because [32:36] someone like Rockefeller could have used [32:38] it, too. [32:39] And I do think let me just like one one [32:42] last point on this. I do think that if [32:43] you actually look [32:45] at what a lot of the the safetyist [32:48] policies are calling for, [32:50] they're basically calling for a form of [32:51] regulatory capture and they're calling [32:53] for things that would create a stronger [32:55] moat around this monopoly or duopoly [32:58] that's in the process of being created [33:00] and it would get in the way of [33:02] competition. So again, I think that [33:05] people [33:06] might not have such a terrible view [33:08] of all of this safety rhetoric if they [33:10] understood [33:12] that what was being created here is the [33:13] biggest monopoly in human history. [33:16] And I think we should just be a little [33:17] bit more skeptical about some of these [33:19] altruistic claims. I can't believe that [33:22] David is like you know, talking [33:25] monopolies when we haven't even left the [33:27] starting gate [33:29] of AI. I I I I think this is a uh [33:34] to me [33:34] >> There's only two companies with revenue. [33:36] >> thing I want is DC trying to [33:38] preemptively preemptively, which would [33:41] be like a disastrous consequence, get in [33:44] the game of picking winners and losers [33:46] at the starting line of AI. That would [33:48] be a disaster. [33:49] >> Brad, did you just put another soapbox [33:51] on top of the soapbox you were standing [33:52] on? [33:53] >> [laughter] [33:54] >> Look Brad, like I said, my North Star is [33:57] competition. As long as there's [33:58] competition going on, I support it. [34:01] Hold on. Hold on. We know that [34:03] monopolists want to stop competition and [34:06] they use regulatory capture to do it. [34:08] And furthermore, they do things like ban [34:10] their competitors from using their [34:11] product. What conceivable reason [34:14] did Anthropic have for banning open claw [34:17] using its models? That is [34:18] anti-competitive. Is it not? [34:21] >> I I I would double click on it. I would [34:23] double click on it. I might not, you [34:24] know, file, but I would double click. [34:25] Okay, listen. Chamath, the girls are [34:28] fighting. Let's keep moving through the [34:29] docket. We're going to be here all day [34:31] with these two and one thing that you're [34:34] going to need to act on very quickly is [34:36] the All-In Summit. It's selling out [34:37] fast. Don't miss it. Speakers are top [34:39] tier again. Freeberg busy working on [34:42] some amazing speakers. Sachs will be [34:43] there. He's flying in and out every day [34:45] for 4 hours and then [34:47] >> [laughter] [34:47] >> we're [34:48] we're going to have a lot of networking [34:49] stuff going down. We're building some [34:50] networking software. So, when you come [34:52] to our events, you get to meet people. [34:53] That's what we always say. My playbook [34:55] for events, if you learn something from [34:58] the speakers every day, one or two [34:59] things, if you meet somebody new and you [35:02] eat some great food and have some fun, [35:04] you get two or three of those things, [35:05] man, even if you get one, you're going [35:06] to come back to the event. You're going [35:07] to get all three all day long. [35:10] allin.com/events Los Angeles September [35:12] 13th, 14th, and 15th. Apologies to [35:15] everybody asking, but liquidity is sold [35:17] out and we've shut down the waitlist. [35:19] There's just no more room. All right, [35:20] the White House allegedly possibly is [35:24] considering, according to reports, [35:26] an FDA for AI. That would vet You heard [35:30] that correct, folks. That would vet new [35:32] models for safety. The thing we've been [35:35] talking about not doing here. The thing [35:37] David Sacks has spent the last year on. [35:39] The White House is considering. New York [35:41] Times reported Trump is considering an [35:43] executive order to create an quote AI [35:45] working group. This group would include [35:47] tech execs and government officials who [35:49] would quote examine potential oversight [35:52] procedures [35:54] including quote a review process for new [35:57] AI models. Oy. According to the report, [36:00] the catalyst was wait for it, [36:02] Anthropic's model, [36:05] which reportedly scared, spooked, made [36:07] people really nervous at the White [36:09] House. Quote, the White House wants to [36:11] avoid any political repercussions if a [36:14] devastating AI-enabled cyberattack were [36:17] to occur. They want to see why A, [36:19] according to the New York Times. Kevin [36:21] Hassett, that guy, the director of the [36:24] National Economic Council confirmed the [36:26] report on Fox Business. Here's your [36:27] 15-second clip. [36:29] >> We're studying possibly an executive [36:31] order to give a clear roadmap to [36:33] everybody about how this is going to go [36:35] and how future AIs that also potentially [36:37] create vulnerabilities should go through [36:40] a process so that, you know, they're [36:42] released into the wild after they've [36:44] been proven safe, just like an FDA drug. [36:46] >> Additionally, [36:48] friend of the pod, [36:49] Scott Bessent, had something to say. [36:51] >> had in the past month was a step change [36:55] in the power of one large language [36:57] model, but we're going to see it from [36:58] the other [36:59] AI companies. What we are determined to [37:02] do is work with our AI companies to [37:05] allow them to continue innovate, but our [37:08] charge of the US government is [37:09] maintaining safety, and there there is a [37:12] very important calculus here between [37:15] innovation and safety, and at the the US [37:18] government, we're going to make sure [37:19] that things stay safe. [37:20] >> There you go. Kevin Hassett and [37:23] Bessent. [37:24] Slightly different positions here, Brad. [37:26] What do you think? [37:27] >> Actually, I don't think they're slightly [37:28] different positions, but I I I would [37:30] agree that Kevin bringing up the FDA [37:32] kind of muddied the waters. I talked to [37:33] Kevin last night after that clip ran. [37:37] You know, and I asked him. I I just [37:38] said, "Do you think FDA is the right [37:40] analog here?" And he said, "You know, I [37:42] was I was only bringing it up to say [37:44] that they we want them to show us the [37:45] models so that we can coordinate them. [37:47] Obviously, our job is to make sure that [37:50] the government is prepared, that we [37:52] harden our systems, that our [37:54] intelligence agencies are up to speed, [37:56] but he does not think and I can't find [37:57] anybody [37:59] on the right [38:00] you know, that believes that we're going [38:02] to move to an approval regime, right? [38:04] The approval regime, this idea that [38:06] you're going to have to share every [38:08] model with an FDA in Washington, and [38:10] they're going to have to pre-approve the [38:12] model is a disaster. Sachs has been [38:14] effectively fighting against this [38:16] correctly over the course of the last [38:18] year. It would just it would lead to [38:20] three bad things. Number one, we do not [38:22] want to put the Washington in in the [38:24] position of picking winners and losers [38:25] when it comes to these models. We're [38:27] winning. We're on the winning horse in [38:29] America. We're out in front of the rest [38:31] of the world. There's no reason to [38:32] change horses and regimes at this point. [38:35] And we don't want to burden this with [38:36] more democracy. But at the same time, [38:39] obviously, I call these pre-AGI or AGI [38:41] models, Mistral, Spud, etc. I see a lot [38:44] of coordination going on between the [38:46] industry and government. I think we can [38:48] do an even better job of a evolving that [38:50] framework so that everybody in [38:52] government is on the same page. We need [38:54] to build more capacity in government to [38:56] quickly be able to do the cyber review [38:58] on these models. Right now, it takes too [39:01] long when the coordination does occur. [39:03] So, we need to have a finite amount of [39:04] time that they give government feedback, [39:06] etc. [39:08] But the last thing that we want is an [39:09] FDA of models sitting in Washington. [39:12] Kevin understands that. Scott Bessent [39:14] understands that. So, I expect that we [39:16] will continue down the path that we've [39:18] been on. Chamath, obviously, I think we [39:21] all agree we don't need an FDA for AI, [39:23] but [39:25] there are things that a reasonably [39:27] people would want to have guardrails [39:30] around AI. I'm sure you would agree. It [39:32] shouldn't be a total free-for-all. So, [39:34] what's your take on this? Is it just [39:37] somebody gave a bad analogy here? Or [39:39] maybe some people were weaseling their [39:41] way into the White House to try to shift [39:42] things when Sachs was back at home or [39:44] something? What what's going on here? [39:46] Give us the uh Cuz that's what I That's [39:48] what people say. They say the last [39:49] person to talk to Trump kind of has his [39:52] ear and that things can bend a certain [39:53] way. [39:55] >> I don't think it's that. I think that [39:56] there's a pretty profound vibe shift [39:59] with respect to tech tech oligarchs, [40:03] Silicon Valley, and particularly the AI. [40:05] That vibe shift has already happened on [40:08] Main Street. [40:10] And I think that that's starting to seep [40:12] into Washington. I think that [40:13] regulations are coming. [40:16] I think they'll be worse under a [40:18] democratic regime, [40:19] but I think that some form of oversight [40:21] is going to exist under a Republican [40:23] regime. [40:25] The question that I think is worth [40:26] asking is why. [40:29] And if you listen to everybody's tone, [40:33] it's all around the negatives of AI. [40:36] So, I think we suffer from two things. [40:38] Number one is we have horrible [40:39] messaging. [40:40] Nobody spends the time [40:42] and the money to articulate the positive [40:46] upside case [40:47] so that there's broad-based support. And [40:50] two, [40:51] the idea that there's going to be a [40:52] Sachs said earlier, [40:54] a few winners and many, many, many [40:57] potential losers, [40:59] I think is really disconcerting to [41:01] everybody. [41:02] And the response from the tech [41:04] community, [41:05] again, should be the leadership of the [41:07] tech world [41:08] coming together [41:10] and actually reinvesting in America writ [41:13] large. They're not doing that in enough [41:15] of a scale that blunts this. So, what [41:18] you're seeing is the build-up of [41:19] antibodies. [41:20] Is it avoidable? Yes. [41:22] Are we doing a good job of avoiding it? [41:25] Absolutely not. We're doing a horrible [41:26] job. I'd give the community, the tech [41:28] leaders, a D- minus [41:30] trending to an F. [41:32] The response is what we're seeing. So, I [41:34] think the question, Jason, isn't [41:37] regulation, no regulation, it's why did [41:39] we get here? [41:41] And I think we got here because the [41:43] other version, the glass half full [41:45] version, the demonstrated investment, [41:48] the broad-based uplifting of American [41:50] society hasn't happened. And if it has, [41:53] it's been very poorly communicated. And [41:55] so, the response is, [41:56] "Hey, hold on. We're going to give three [41:58] guys trillion-dollar net worths, and [42:01] we're going to allow them to control the [42:02] keys?" [42:03] That's why this is happening. [42:04] >> Exactly correct. And it's very easy, [42:07] Sachs, to imagine all the bad things [42:09] that That happen. Our minds are [42:11] constructed to do that. We're vigilant. [42:13] We look out for the tiger or, you know, [42:15] the tornado to keep ourselves safe. [42:17] Humans have a bias towards safety and [42:21] they're going to think about, you know, [42:22] deep fakes. They're going to think about [42:24] robotics. They're going to think about [42:25] self-driving cars taking people's jobs. [42:27] They're going to think about, you know, [42:28] all the dark things that could happen, [42:30] bio weapons, etc. And we don't have [42:32] anybody out there really talking about [42:34] all the positives that could happen. [42:37] What's your take on [42:39] the palace intrigue we all have here? [42:41] What's going on in the palace, in the [42:44] 47th administration, around this debate? [42:47] Who's leading Trump down the path of [42:50] regulation and creating this AI FDA? We [42:54] know you're part of the camp that wants [42:56] to keep this train moving and not [42:58] overregulated, not have regulatory [43:00] capture. Who are the people trying to [43:01] slow this down? [43:02] >> Well, look, I I think there's several [43:04] things going on here. The first one is [43:06] there's a lot of fake news. This whole [43:08] idea of an FDA for AI, I don't think any [43:11] senior official supports it, just like [43:13] Brad was saying. [43:14] I spoke to Hassabis well. That's not [43:16] where his head is at. So, I don't think [43:18] anybody in the administration is saying [43:20] they want an FDA for AI. Certainly, I [43:23] don't think that's the way the president [43:25] thinks about these issues. He's the most [43:27] pro-innovation president we've ever had. [43:29] And the White House Chief of Staff, [43:30] Susie Wiles, just put out a statement [43:31] last night that I think pretty much [43:33] shoots this down. So, I think there's a [43:35] big fake news component. Remember, it [43:36] was not really the White House who was [43:38] saying it was the New York Times. [43:40] >> Mhm. [43:40] >> And really, I think actually Andrew Ross [43:42] Sorkin, [43:43] now I'm not criticizing him, but he's a [43:45] commentator and he's the one who said [43:47] this first and then somehow that spin or [43:50] that gloss [43:52] somehow took on a life of its own. And I [43:54] think [43:55] Silicon Valley reacted accordingly. [43:57] There's a very visceral negative [43:59] reaction here because we know how [44:00] damaging that would be to innovation. [44:02] But, look, I think the good news is that [44:03] that was fake news. Second, [44:06] I think that there's another thing going [44:08] on, which is a straw manning of what [44:12] the Trump administration did on AI in [44:14] its first year. [44:16] And in the same way that they want to [44:18] spin this FDA for AI, they're also [44:20] trying to spin what we did as this [44:22] completely laissez-faire attitude where [44:25] there'd be no regulations whatsoever, no [44:27] guardrails. It's a way of criticizing [44:29] what we did. They're trying to portray [44:31] it as unsafe. In fact, if you look on [44:34] March 20th, White House released [44:36] a national AI regulatory framework that [44:39] I worked on, [44:40] in which we put out a four-page bulleted [44:42] list of legislation that we would [44:44] support if Congress wants to pass it. [44:48] So, we have not been against every [44:51] conceivable regulation or every [44:53] conceivable law. We just believe that [44:55] there should be specific solutions to [44:57] specific problems, as opposed to a giant [45:00] power grab by Washington that would [45:02] squash innovation. So, I think that's [45:04] point number two. [45:06] Point number three is, there is a [45:08] legitimate thing happening here with [45:10] let's call it Mythos or cyber. Okay, we [45:13] know that it's not just Mythos, [45:16] OpenAI now has a model that's just as [45:18] cyber capable as Mythos, and within [45:21] three to six months all the major [45:23] frontier labs, and including Chinese [45:25] models, will have cyber capabilities. [45:28] In response to that, we do need there to [45:31] be a hardening of systems, and we do [45:34] need there to be a scanning of codebases [45:37] to find these vulnerabilities and patch [45:39] them before the hackers do it, because [45:42] the hackers will have these capabilities [45:44] in a matter of months. That's a [45:46] certainty. Because the same capabilities [45:48] you use for cyber defense can also be [45:50] used for cyber offense. It's the same [45:52] tool set. And the open source models [45:54] will have these capabilities anyway. [45:56] >> have it to a certain extent, let's be [45:57] honest. They have 80% of it, 90% of it. [46:00] >> it's simply the case that AI will be [46:02] good at cyber. And so, we do need a [46:04] response to that. Now, my view on what [46:07] should that response be should be first [46:10] of all, we should want the government [46:12] and the private sector to work [46:14] cooperatively, and I think they are. We [46:16] have a giant cybersecurity industry in [46:19] the United States whose sole job it is [46:22] to protect systems and protect against [46:25] breaches. [46:26] >> We have the best companies in the world [46:27] at doing that. We have CrowdStrike, we [46:30] have Palo Alto Networks, we talked about [46:31] that before. We have the best defense. [46:34] >> Right, exactly. And so, what we should [46:36] be doing, I think, is getting these [46:38] tools, Methos and then the OpenAI model [46:40] and and others like it in the hands of [46:43] our cybersecurity industry. And by the [46:45] way, not just the public companies like [46:47] Palo Alto Networks and CrowdStrike, all [46:49] those certainly. They're two of the most [46:50] noteworthy, but there's also some [46:52] incredibly strong startups on the way [46:54] out. [46:55] >> a long tail of hundreds of companies [46:56] that are doing Yes. [46:58] >> AI-powered pen testing and all the rest [47:00] of it. We need to get these tools into [47:02] their hands as quickly as possible [47:03] because they're a force multiplier. For [47:05] all the companies out there that aren't [47:07] that good at cybersecurity or maybe [47:10] they've got IT departments, they can use [47:12] these companies as vendors. So, I think [47:14] that there is a role for us to play. [47:16] >> Yeah. Do you think that the models [47:18] should [47:20] have a KYC wrapper going forward? [47:22] >> KYC for the audience is know your [47:24] customer. [47:25] >> Yeah. [47:25] >> Yeah, so what really what it would mean [47:26] is that before you can use [47:29] Methos, you have to identify yourself so [47:31] that we can try to know that you're not [47:33] a [47:34] state-sponsored actor or, you know, a [47:37] bad guy. [47:38] >> I think that's the type of thing that we [47:39] should be thinking about. So, first of [47:41] all, I want to say that both Anthropic [47:43] and OpenAI acted responsibly here. No [47:46] one was trying to release these super [47:48] powerful models. So, in a way, all the [47:50] people who are saying that we need [47:52] pre-release approvals for models, [47:54] they're trying to solve a problem that [47:55] didn't exist. Anthropic wasn't [47:57] >> which [47:58] is the ideal situation. [47:59] >> trying to release this. They all [48:00] understood the power and they were all [48:02] acting responsibly. [48:03] >> They understood the ramifications, [48:04] because they would have been sued. So, [48:06] there is a self-policing going on here, [48:08] which is the ultimate way to do this. [48:10] >> Yes. And but to your point, Chamath, [48:11] yeah, look, I think that before giving [48:15] your API for a super powerful model, you [48:17] should not give that to a company or an [48:19] actor you don't know who they are. So, [48:21] yeah, some basic KYC makes sense. They [48:23] should know who they're giving these [48:24] tools to. And I guess my view on the [48:26] Mythos preview and whatever the [48:28] equivalent is of what Open AI is doing [48:31] is that we very rapidly need to get [48:33] these tools into the hands of more good [48:35] guys. You need to know who those good [48:37] guys are. You need to know who they are. [48:39] So, yeah, KYC is like a predicate for [48:43] that, right? You got to know who they [48:44] are. [48:44] >> Just to be clear, we'd all agree that if [48:46] you did have identity for those frontier [48:49] models, which they're probably doing [48:50] anyway right now, and you logged what [48:52] people were doing with them to look for [48:53] security breaches, that wouldn't [48:54] necessarily happen when you released it [48:56] to the public because of privacy issues. [48:58] Here's your Polymarket [49:00] for Trump ordering a federal review of [49:02] AI models by May 31st, 21% chance. I [49:06] think [49:07] to our partner at Polymarket, man, I I [49:09] got to get in here. I don't Do I have [49:11] inside information here being [49:13] the world's greatest moderator on this [49:14] podcast or I can Can I collect this [49:15] money, Chamath? What am I going to do [49:16] here? [49:17] >> I would not Do not Do not place a bet, J [49:19] Cal. [49:20] >> Don't place a bet. [49:21] >> Don't place a bet. But [49:23] to Chamath's point, I mean, look, I [49:24] think we're kind of work shopping this [49:25] in real time. [49:26] >> We are. We should say [49:27] >> I think that I think that for the [49:28] preview period, we should definitely [49:29] have KYC. [49:31] Maybe [49:31] >> logging? What about logging? [49:33] >> Well, look, once you're past the preview [49:34] period and it's in general release, [49:37] I'm not sure if the KYC matters as much [49:39] because so many people are going to have [49:40] it. But during the preview period, there [49:42] should be KYC. [49:43] >> Let me Let me just say one thing. All [49:45] the labs are already tracking API use, [49:48] okay? And anything suspicious, because [49:50] they they're they're there are major [49:52] anti-distillation efforts going on by [49:54] all the labs. There's a ton of [49:56] coordination going on with the [49:57] government. There's way more happening, [49:59] I think, in terms of our API and API [50:02] use. Um and anything suspicious is being [50:05] flagged [50:06] and being shared with the government. [50:07] So, the idea that we have no idea who's [50:09] doing it, I think, is not the case. And [50:12] in fact, in some cases, we may want to [50:15] allow people to use it so that we can [50:17] see exactly the types of things that [50:18] they are extracting. So, I would just I [50:21] would say we're already down that path, [50:23] but better coordination [50:25] may may may, in fact, be called for. [50:27] >> Yeah. And just one last point in this [50:28] whole thing is I just want to build on [50:30] my point that pre-release approvals are [50:32] solving a problem that didn't really [50:33] exist because, again, Anthropic and [50:35] OpenAI weren't trying to release these [50:36] models yet is that there is a [50:38] substantial faction of, let's say, AI [50:42] ideologues or doomers who [50:45] are basically employing the classic [50:47] never let a crisis go to waste strategy, [50:49] right? That, yes, we do have this cyber [50:52] issue that is real. You know, everyone [50:55] needs to harden their systems now over [50:57] the next three to six months. That is a [50:58] real issue. [51:00] But that is a problem that we will solve [51:02] over the next six months. We have to. [51:03] But what they're trying to do is use [51:05] that issue to try and create a permanent [51:09] new infrastructure in Washington. Again, [51:11] I don't think the That's not the [51:12] administration's intention. That's not [51:14] the administration's agenda. But you saw [51:16] a lot of people [51:18] on social media, a lot of the think [51:19] tanks, and even Bernie Sanders weighed [51:21] in. And he said, "For the first time, I [51:23] like something that [51:25] >> This is great. [51:26] The administration understands the 1% of [51:29] the 1% tax, and everybody understand [51:31] that this is out of control. The AI is [51:33] going to take the jobs. They're going to [51:34] take my summer home. It's going to be [51:36] terrible." [51:37] >> So, there are people who have this [51:38] agenda. Look, Bernie Sanders just wants [51:40] to stop the progress. I mean, he's [51:41] >> Of course, I do. [51:43] >> He wants to ban data centers. He's put [51:44] out a bunch of things. [51:46] >> He He basically has bought into the [51:47] whole doomer narrative. So, look, that's [51:50] why he likes the FDA [51:52] idea is because it would put the kibosh [51:55] on innovation. [51:56] >> already. Let's give Let's have a go back [51:57] to paper and pen. It was a better [51:59] society, Sachs. [52:00] >> Jason, what do you think? [52:01] >> I think there's two really interesting [52:03] things I want to build on here. The [52:04] first is your point, Chamath, [52:06] around how do we turn around the sort of [52:10] bad vibes around AI? I think we have to [52:13] have two strategies here. One is giving [52:16] what you've been working on, Brad, with [52:17] your project. We should see more people [52:19] giving. There's no reason why Nvidia, [52:22] SpaceX when they go public, [52:24] Anthropic when they go public, OpenAI if [52:27] and when they go public or if they [52:29] become stay as nonprofit, there's no [52:31] reason those folks in an IPO couldn't [52:33] give a portion of the IPO to every [52:35] American citizen. So, IPOK, IPO for [52:38] kids, they all take, you know, whatever [52:40] it is, 5%, 1%, whatever they choose, and [52:43] they put it into the Invest America [52:45] accounts. And we should see some major [52:47] giving from the people who are becoming [52:50] trillionaires, hundred billionaires, [52:52] whatever it happens to be. There's no [52:53] reason not to. But those people haven't [52:55] been doing that. We had this giving [52:56] pledge, which was a little bit of virtue [52:58] signaling, and it wasn't real. It was [53:00] just, you know, at the end of your life [53:01] you promised to give away half your [53:02] money. So, let's have something real. [53:03] Let's have something where, you know, [53:05] people say I'm going to give away 1% of [53:06] my stock over the next 20 years of my [53:09] life. Every year 1% will go into Invest [53:11] America, whatever it is. It won't cost [53:13] anybody anything. You can't spend this [53:14] money, whether it's Bezos or whatever. [53:17] Second, [53:18] and that same thing in terms of giving [53:20] back, we have not talked about how [53:21] massive this could be for health and [53:24] extending people's life and reducing [53:25] suffering. We need to work on that. [53:27] That's where contributions to basic [53:29] science could come in, and obviously [53:31] education and lowering the cost of [53:32] education. And if you look at what [53:35] Americans on the bottom have, you were [53:36] talking about the, you know, cup half [53:38] empty. There's really two or three [53:40] things they really feel anxiety about. [53:43] One of it is income, and the second is [53:46] health care. [53:47] And on the margins, housing and their [53:49] kids, you know, their kids' education [53:51] and the cost of those things. [53:52] We should really take a look deeply at, [53:55] and I know this is very unpopular [53:56] amongst capitalists, including myself. [53:58] We should really look at the minimum [53:59] wage and study what happened in New [54:01] Zealand, Sweden, Switzerland, [54:04] Australia when they raised it. What [54:06] actually happened when they raised it [54:07] and there was a lot of hand-wringing [54:08] about it. But when they slowly raised it [54:09] what they found was those consumers [54:11] don't save money, they spend it. They're [54:13] always behind the eight ball in terms of [54:15] their spending. We should opt in to [54:17] trying to raise the minimum wage company [54:20] by company by company and just give [54:22] people who are at the end of the [54:24] spectrum that understanding that, "Hey, [54:26] year over year, whether it's Amazon or [54:29] Target, etc., restaurants, we're all [54:32] collectively going to add a little bit [54:34] to that minimum wage and try to lift the [54:36] bottom third of society." That's the [54:38] stuff we're not talking about. We don't [54:40] talk about it here on this podcast, we [54:41] don't talk about universal health care, [54:43] we don't talk about the minimum wage, [54:44] but that's what capitalists should be [54:45] talking about. And if we did that, if we [54:48] increased the minimum wage, and I'm not [54:50] a socialist, I'm a capitalist who thinks [54:51] this is good for capitalism. If we [54:53] increased the minimum wage just modestly [54:55] each year [54:56] and we opted into doing that and we [54:58] figured out a way to give universal [54:59] health care, companies wouldn't have to [55:00] deal with universal health care and we [55:02] would have customers, and we're a [55:03] customer driven economy. Like 60, 70, [55:06] 80% of what happens in this country is [55:08] driven by the consumer. We need consumer [55:10] spending. It's great for companies if we [55:13] had more people being able to buy [55:15] Netflix or order on Amazon. Anyway, [55:17] that's my that's my TED Talk. Thanks for [55:19] coming. How do we get from AI to the [55:21] minimum wage? I'm still a little bit [55:23] confused. I don't know. No, the black [55:25] eye we have in this country with [55:27] polarization of wealth and people scared [55:29] of losing their jobs. We We should look [55:30] at why are they scared, David? And I've [55:33] talked to you privately and you said to [55:35] to privately, you can strike this if you [55:36] want, But, you said to me privately, you [55:38] wouldn't be against necessarily figuring [55:41] out a way to you do universal health [55:42] care if there was a way to do it. You [55:44] want to see every human have health [55:46] care, yes? [55:48] >> Sure. I mean, the issue is not the [55:50] desirability of it, it's the cost. I [55:52] mean [55:53] >> So, you're the real entrepreneur of our [55:55] time. How would you do it? Have you [55:56] given any cycle to it? [55:58] >> I haven't studied that issue, so I don't [56:00] know. I just know that [56:01] countries [56:03] I remember what P.J. O'Rourke once said, [56:05] which is if you think health care is [56:06] expensive now, just wait until you make [56:08] it free. [56:10] >> [laughter] [56:10] >> And [56:11] uh so, you take away all the incentives, [56:13] and you have a even bigger problem. [56:14] >> Well, what do you think of minimum wage? [56:15] Yeah, go ahead. [56:16] >> Well, let me just Can we just get back [56:17] to AI? Listen, you guys are right about [56:20] the unpopularity of AI. We've all seen [56:21] those polls, but I want to just put up [56:23] this additional poll that came out about [56:27] the salience of this issue, which is how [56:29] important do people think it is. [56:31] And AI ranked 29 out of 39. So, although [56:34] AI is not very popular, it is certainly [56:37] not top of mind for voters. It's not in [56:39] the top 10 issues. It's not in the top [56:41] 20 issues. What is top of mind for [56:43] voters? Number one, cost of living. [56:44] Number two, the economy. And we know [56:47] that AI is deflationary. It helps with [56:50] the cost of living, and it's creating an [56:53] economic boom right now, okay? It's 75% [56:56] of GDP growth in Q1. By the way, that [56:58] that economic growth is not just limited [57:00] to startups in Silicon Valley. We're [57:02] seeing a construction boom. We're seeing [57:03] a blue collar boom. We're seeing 25 to [57:06] 30% wage increases for uh construction [57:09] workers. [57:10] >> And so on down the line. So, [57:12] >> And Brad, if you if you look at that [57:13] chart, there's health care in there, [57:14] too. [57:15] >> So, look, my point is that AI may not be [57:17] popular, but the effects of it actually [57:20] are popular if the media would honestly [57:22] report what was happening, which is AI [57:25] is creating an economic boom right now. [57:27] >> It couldn't be better said, David. You [57:29] know, Bernie Sanders calling for a [57:31] moratorium, shutting down all data data [57:33] centers. We'd have negative GDP growth [57:35] this year. The stock market would be [57:37] down 15 to 20%. Unemployment would be on [57:40] the rise. You know, there is a [57:41] consequence to the government [57:43] controlling through command and control [57:45] the economy. In 1929, we had 4% [57:48] unemployment. 3 years later, it was [57:50] unemployment was 23% [57:52] because government got involved in [57:54] regulating everything and shutting down, [57:56] you know, what was working. That is the [57:58] greatest threat we have here. AI is [58:00] delivering huge net benefits today in [58:02] terms of unemployment rate, in terms of [58:04] economic growth and productivity growth. [58:06] We need to tell the story. But, to [58:07] Chamath's earlier point, we also need to [58:09] deliver net benefits. Jason, thanks for [58:12] the shout-out. Yes, every American [58:14] having an investment account that [58:15] compounds with the upside of AI, we're [58:17] going to do that. Going to deliver that, [58:19] and that's going to be massive. But, I [58:20] also think ideas like, if we're going to [58:23] put a data center in Abilene, Texas, [58:25] let's make electricity in Abilene free [58:27] for the households in Abilene, Texas, [58:29] right? There are ideas that can deliver [58:31] net benefits. We got to deliver those. I [58:33] think we're Optimism will be on the [58:35] march. I think we're in the trough right [58:37] now. [58:37] >> What's your take on minimum wage and, [58:39] you know, how do you think about it as a [58:40] capitalist, as an innovator, and in the [58:41] face of AI, which could have a dramatic [58:42] impact on these issues? You know, like [58:42] Zach's, I I you know, to me I haven't [58:43] spent a ton of time thinking about those [58:44] except that as a society, we're $38 [58:45] trillion in debt. We haven't been able [58:45] to afford to deliver those things. I [58:46] generally think the market works out [58:47] those [58:48] You know, like Zach's, I I you know, to [58:50] me I haven't spent a ton of time [58:51] thinking about those except that as a [58:53] society, we're $38 trillion in debt. We [58:55] haven't been able to afford to deliver [58:57] those things. I generally think the [58:59] market works out those [59:01] issues better than the government [59:02] top-down trying to, you know, the [59:04] government gets more and more involved [59:05] in healthcare, and the only thing that [59:06] happens it gets more expensive. So, [59:08] we've all seen the charts of the most [59:09] expensive categories where we've had [59:11] inflation, education, healthcare, etc. [59:13] It's where the government's involved. I [59:15] actually think if you just let the [59:16] markets work, we're entering into an age [59:18] of abundance. A lot of these problems [59:20] are going to be solved. People are going [59:21] to have a lot of AI coaches in [59:24] healthcare and education, etc. Let the [59:26] market work. Government stay at bay, [59:28] keep things safe. We're on a good march. [59:31] >> I think this [59:32] proves my point perfectly. If you talk [59:34] to any founder, they're not thinking [59:35] about housing, they're not thinking [59:37] about higher education costs, they're [59:39] not thinking about minimum wage, you [59:41] know, and they're not thinking about [59:42] healthcare all that often. Some Some do [59:44] though, there's some innovation there. [59:45] And it's because it's so regulated that [59:48] entrepreneurs and VCs are just like [59:49] that's Kryptonite, the government has [59:51] poisoned the well, we can't participate [59:53] in that. And that's the roadblock and [59:54] that's where Americans are suffering and [59:56] that's where it would be great if [59:58] founders actually put their minds to it [59:59] and the government's got to get rid of [60:01] all that regulation and let us cook in [60:03] those specific verticals. All right, the [60:05] market is in hyperdrive. Hyperscaler [60:08] revenue has made the markets move up. We [60:12] hit on this briefly, but we didn't have [60:13] you here. Fifth best EB G cloud [60:16] computing on a tear. I referenced it [60:19] earlier, but AWS is now on a $150 [60:21] billion run rate. Azure 108 billion, GCP [60:25] Google Cloud 80 billion. There's a [60:27] little bit of fun with numbers there, [60:28] Chris. Azure and Microsoft include some [60:30] of their software products in there and [60:31] Google Cloud includes things like Google [60:33] Office or Google Suite in there. [60:35] But the growth numbers are tremendous. [60:38] AWS, which is the more pure play of the [60:40] three, 28% growth on a very big number. [60:43] Azure 39%, Google Cloud stunning [60:46] everybody with 63% [60:48] growth. It is incredible what the ARR [60:52] numbers are. Google Cloud added 10, AWS [60:55] 10, Azure 9.5. So basically 30 billion [60:57] collectively. Jamin Ball, who works for [60:59] you, I think put out some data on the [61:02] Twitter. [61:04] Brad, markets at all times high all time [61:06] highs. [61:08] Mag 7 cooking. [61:11] Uber blowing out growth, Disney blowing [61:14] out growth. The consumer seems absurdly [61:17] strong based on those two bellwethers. [61:20] Tech seems extremely strong based on the [61:22] cloud computing. What's your take on the [61:24] overall market and overall economy? [61:26] Obviously, inflation up a bit, people [61:27] hand-wringing about the never-ending war [61:30] and the cost of oil. [61:32] >> Let's just telescope way out. You know, [61:35] the level of criticism directed to this [61:37] administration, right? Tariffs were [61:39] going to cause hyperinflation, we're [61:41] going to destroy GDP, [61:43] uh conflicts in Venezuela and Iran were [61:45] going to do the same. You know, we've [61:47] heard all of the negative stories, but [61:49] what's happening? Accelerating GDP. [61:52] A 10-year that's sitting at 4.3, [61:54] inflation totally under control. AI, AI, [61:58] AI, compute, compute, compute. We're [62:00] leading the world. It's contributing [62:02] massively, right, to GDP growth in the [62:05] country. We see the S&P only up 8% this [62:08] year, right? So, we're not in the bubble [62:10] territory here. Meta's trading at 17 [62:13] times fully taxed GAAP earnings, Nvidia [62:15] at 19 times, Microsoft at 20 times, [62:18] Google at 24 times. And then the memory [62:21] stocks that everybody's excited about, [62:23] we have 25% of our portfolio in SK [62:26] Hynix, five times fully taxed GAAP [62:29] earnings, Samsung six times, Micron [62:31] seven times, right? This is not the [62:34] stuff that bubbles are made of. You [62:36] know, David referenced it earlier, we [62:38] started the year Open AI and Anthropic [62:40] were doing combined about 30 billion in [62:43] revenue, now combined four months later, [62:46] 80 billion in revenue. The the policies [62:49] of this administration on the economy [62:52] are working, they're working in in [62:54] spades. Our gap on the rest of the world [62:56] in AI is growing. And so from from my [62:58] perspective, you know, uh you know, [63:01] we've been all in on the market. I [63:02] talked about it earlier in the year, [63:04] we're heavily tilted [63:05] >> you make that switch to go all in on the [63:06] market? Cuz you were bearish. [63:09] >> Uh I'd say toward the toward the end of [63:10] last year, the market had run up a lot. [63:12] We had a lot of these questions. Listen, [63:14] entering this year, there was a huge [63:16] question hanging over the market. Would [63:19] the AI revenue show up? [63:22] If the Anthropic revenues hadn't shown [63:23] up and we didn't see this [63:25] re-acceleration out of the hyperscalers, [63:27] the market would be down 10 to 15% [63:29] because people would say there's no ROI [63:32] on all of this investment in [63:34] infrastructure. Exactly. When I saw the [63:36] numbers start showing up in December and [63:38] into January, we went from medium to [63:41] large in terms of our exposures and 80% [63:44] of our exposures or more have been in [63:46] compute, AI, memory, etc. [63:49] >> And this is why it's great to operate in [63:50] the private market and the public market [63:52] because you can see things in the [63:53] private markets that inform the public [63:55] markets. But, the question remains, [63:57] Brad, how much better would the economy [63:59] have been doing [64:01] you know, as much credit as you're [64:02] giving to the administration if they [64:04] didn't start a $100 billion war that we [64:06] did not need to go into, according to [64:08] all reports, and if we didn't do a bunch [64:11] of tariffs that wound up being [64:13] unconstitutional by the Supreme Court, [64:15] which Trump himself put in. We would [64:18] have been further ahead. That's uh my [64:20] take on it. We would be ripping even [64:22] more if we didn't have those uh silly [64:25] diversions. [64:25] >> It's It's hard to imagine, okay? Just to [64:28] set up again here. [64:29] It's hard to imagine a more Goldilocks [64:32] situation for the United States of [64:34] America. We have reset the table [64:35] geopolitically. The discount rate [64:37] globally is actually coming down, not [64:40] going up, evidenced by markets at [64:42] all-time high and the bond market in [64:43] control. And then, look at the private [64:45] markets. We have multiple [64:47] trillion-dollar companies that have been [64:49] created in the private markets that are [64:51] now coming public. SpaceX coming public [64:54] is going to be a multi-trillion dollar [64:56] you know OpenAI, Anthropic, like at some [64:59] point you just have to acknowledge USA [65:01] is winning. Of course, there are always [65:04] things that we could be doing better, [65:05] but there's not a country on the world [65:07] that wouldn't trade all of its fortune [65:10] for the United States fortune today. [65:12] >> 100% in agreement. American [65:13] exceptionalism as embodied by the great [65:16] companies in America, SpaceX, Google, [65:18] etc. All the ones we've been talking [65:20] here. That is the story and I give [65:23] infinite credit to this administration [65:24] for being business-friendly. I do think [65:26] they've made two critical mistakes. I [65:27] think the tariffs were poorly executed [65:29] and I think we shouldn't have gone to [65:30] this war [65:32] um and we should find a quick resolution [65:34] to it which the administration seems to [65:35] be desperately doing. Sachs, your take [65:37] on the economy. [65:39] >> Well, look, we have an AI boom going on [65:40] right now and I think that's thanks to [65:41] President Trump's policies. Remember, [65:43] the first week he was in office, he [65:44] rescinded the Biden policies on chips [65:47] and models. And what were those [65:49] policies? It was the approval regime [65:51] that we're talking about. Models would [65:54] have to go to Washington to get approved [65:55] if they were trained with some number of [65:57] flops and then every sale of a GPU [66:00] worldwide would have to be licensed from [66:02] Washington unless it fit into some [66:04] narrow exemptions. So, the whole [66:06] approach of the Biden administration [66:08] >> Exactly. [66:08] >> that President Trump inherited was [66:10] everything approved in Washington. He [66:13] rescinded that. He declared that we had [66:14] to win the AI race and he unleashed our [66:17] companies to do that. Now, one other [66:19] really important thing is energy. [66:21] Remember, it was this president, going [66:23] back a decade, who said, "Drill, baby, [66:24] drill." He said, "We have to unleash [66:26] American energy. That's the basis for [66:28] the American economy. It's also the [66:29] basis for AI." [66:31] He also has said that he wanted to allow [66:35] our AI companies become energy companies [66:37] so they could bring their own power to [66:39] these data centers. So, they're not [66:40] drawing off the grid. They're not [66:42] competing with consumers for [66:43] electricity. They're generating their [66:44] own power. And it's thanks to this [66:47] president that we have seen this [66:49] blue-collar construction boom right now [66:52] powering all of this infrastructure. [66:54] What would the alternative been? [66:56] We know. I mean, Bernie Sanders has said [66:58] it. Would have been ban on data centers. [67:00] So, [67:01] >> Yeah. [67:02] >> Right now, Biden has banned data [67:03] centers. It's That would be a much [67:05] a much worse choice. Chamath, I'll give [67:07] you the last word here as we wrap on the [67:10] economy generally. [67:11] >> I think the markets are going to keep [67:13] going up for a while. [67:15] And then at some point they're going to [67:16] go down. [67:18] >> Okay, I wrote it down. Chamath, you said [67:21] markets are up and then eventually [67:23] they're going to come down. [67:25] Those are the two things. Up and let me [67:27] put a U here and then a D here. [67:29] >> It's got the I think you're doing a an [67:31] impression of the taking notes emoji. Is [67:33] that what you're doing? [67:34] >> Yes. Yeah, I just Okay, [67:36] I'll shut down. [67:37] Wow. Thank you for tuning in to All-In [67:39] where you can get your great calls and [67:41] market action advice. It's going to go [67:43] up and then down folks. Act accordingly. [67:45] Good. Chamath, in all seriousness. But [67:47] what what what makes you bullish, you [67:48] know, let's say in the next 6 months, 12 [67:51] months catalysts and then what do you [67:52] think the headwinds are as well? Let's [67:53] take the, you know, the short to [67:55] mid-term 6 months to 2 years. [67:58] >> I think that in the short term [68:00] the people that [68:02] makes the new thing [68:06] needs [68:07] to get valued and needs to demonstrate [68:10] value. So, who are the people making the [68:12] new thing? [68:13] It's the Nvidias, it's the memory [68:15] makers, it's the Anthropic's, it's the [68:17] SpaceX's and uh it's the Open AIs. But [68:20] eventually it all comes home to roost. [68:23] And [68:25] you can't just make things for a market [68:28] who then doesn't have a measurable [68:29] benefit themselves. [68:32] To be very clear and blunt, there is [68:34] literally not a scintilla of evidence [68:37] that AI has helped lift the operating [68:39] margins of the S&P 500. [68:42] There's all kinds of bluster. There's [68:43] going to be a an important fork in the [68:45] road. It's probably 2 or 3 years from [68:47] now. One path will be [68:50] OpEx shrinks [68:51] hence margins increase [68:53] and the other path is revenues grow and [68:56] margins expand and OpEx stays flat but [68:58] it or maybe it even goes up. [69:00] Those two things are very important [69:02] differences, because in the former [69:04] you're talking about shrinking workforce [69:06] and shrinking off access a percentage of [69:08] operating margin and revenue. In the [69:10] latter, you're actually growing through [69:12] it. [69:13] The answer to that question I think is [69:14] critical about how the markets will [69:15] respond. [69:17] And how society will respond. So, I [69:19] think we have kind of call it 500 days [69:22] where [69:24] you just got to be net long. [69:26] But, I think it's literally in the [69:28] hundreds of days from now [69:30] 500. [69:33] You're going to have to have an [69:34] important reckoning moment. The people [69:37] that are paying for all these tokens [69:38] need to see it an actual benefit. [69:41] >> Yeah, that that's reasonable. Yeah. [69:42] >> That's a really interesting point. Let [69:43] me connect a couple of dots here between [69:45] something Chamath said and what Brad [69:46] said, which is [69:47] Brad said at the beginning of the year, [69:48] we went into this year with this massive [69:50] CapEx, this massive investment [69:52] infrastructure, but people weren't sure [69:54] that the ROI was going to be there in [69:57] terms of model revenue. And that was [69:59] true, and then the model revenue has [70:01] proven out. And now what Chamath is [70:02] saying is that we're going to be at [70:03] another fork in the road soon in terms [70:05] of whether there's going to be ROI [70:08] on all those tokens that are being sold [70:10] and generating the revenue for those [70:13] model companies. And I agree with you [70:15] that that is not proven out yet, but I'm [70:17] optimistic that it is going to be proven [70:20] out, and otherwise these you wouldn't be [70:22] seeing [70:23] >> Look, I [70:23] >> enterprises continue to to buy. Hold on, [70:25] let me just finish my point. [70:27] You wouldn't be seeing enterprise [70:28] continue their month-over-month spend [70:31] on coding tokens if they didn't feel [70:33] like the ROI was going to be there. But, [70:35] you make a good point, which is what is [70:37] the impact on the economy going to be [70:39] when all of this new software, this [70:41] bespoke software that's being created [70:44] through again all these coding tokens [70:45] that are being bought, is going to power [70:47] a wave of productivity like I think [70:49] we've never seen before. So, I think [70:51] what you're seeing is the ROI is getting [70:54] is sort of trickling down from [70:55] infrastructure to model to application [70:58] to end user. And I think it's going to [71:00] create an economic boom. [71:02] >> I got I got it. I'm with you, Sachs. [71:03] It's This is déjà vu all over again. We [71:05] watched this happen with the PC [71:06] revolution, the internet revolution, [71:08] cloud revolution, mobile revolution. We [71:10] had all this hand-wringing. Will this [71:12] ever pay off? Should I build an app? [71:13] Should I build a website? Should I not? [71:15] Should I move to the cloud? Should I [71:16] keep it on prem? All of these questions [71:17] over and over and over again. And then [71:19] they went from question marks to [71:20] exclamation points. I can tell you [71:21] inside of my firm, we have started We [71:23] were using agents then we started [71:25] building code. And I've got three people [71:26] on the team who are making all the [71:28] interfaces and products that a 22-person [71:31] investment firm should not be making [71:33] internally. They should be using SaaS [71:35] software. And they are shipping product [71:38] day in and day out. The ROI is fait [71:41] accompli, Brad. It is fait accompli. I [71:43] think this has been decided. I think [71:45] it's been decided. [71:46] >> It has not been decided at all. It has [71:48] not been decided at all. [71:49] >> You have 80/90. You're working with the [71:50] big enterprises. I invest in 100 [71:52] startups a year. I work with the small [71:53] ones. It is fait accompli with startups. [71:56] They are building software. They're [71:57] shipping. They are getting massive value [71:59] from these tokens. And they're getting [72:00] so much value that they don't have to [72:02] add, you know, but half the number of [72:05] employees that they would with the same [72:06] amount of capital. They're getting [72:08] further with less money. It is working [72:11] in startup land. I don't know what's [72:12] happening at 80/90. You would have a [72:14] better picture, obviously, of the [72:15] enterprise. Tell us what you see there. [72:16] >> I mean, our business is doing well, but [72:20] what I'm trying to get across to you [72:21] guys is that [72:24] you can't will [72:25] profits to go up. [72:28] Okay? So, ultimately what happens is I'm [72:30] just going to take a company randomly. [72:33] Anheuser-Busch. [72:35] They have to eventually sell more beer. [72:39] Okay? Take Nike. [72:42] They ultimately have to sell more shoes. [72:44] Take [72:45] a medical devices company. They have to [72:47] sell more artificial hips and knees. [72:49] So, the The I'm trying to get across is [72:51] right now there's an enormous amount of [72:54] very constructive and creative [72:56] experimentation. [72:58] But I think it's what is also true is a [73:01] lot of that [73:03] has not yet proven value. I don't think [73:05] that means it's going to stop. All I'm [73:07] just trying to say is [73:09] until a company can trace very directly, [73:12] I spent X and I made Y where Y is now [73:14] greater than X and it's lifted my [73:16] margins, [73:18] that is the thing that causes the [73:20] flywheel to spin faster. [73:22] And right now we've started the first [73:24] part of that equation. We've spent the [73:26] X. [73:28] And we have not seen the Y. [73:30] You would see it in global GDP. You [73:33] haven't. You would see it in global [73:35] productivity. You haven't. You would see [73:37] it in the global profit margins of the [73:38] S&P 500. We haven't. It doesn't mean [73:40] it's not coming. [73:41] >> Brad, you want to you want to pick up on [73:43] this because I yeah, I'm definitely [73:44] taking the other side of it cuz I'm [73:45] seeing with a lot of these companies [73:48] massive lowering of costs. Their ads are [73:50] getting more effective. At the same time [73:52] they've stopped hiring. They're not [73:53] adding positions in a lot of cases. And [73:56] things like just pick the Nike example, [73:59] a lot of the photo shoots [74:00] they used to do for their app, excuse [74:02] me. A lot of the imagery they used to [74:04] make, now they're able to make more of [74:05] it without having to hire photographers [74:08] and do that stuff. I know this example [74:09] cuz we have a startup that does this [74:10] specifically for brands like Nike. [74:12] They've seen a massive drop. We have one [74:14] that helped DoorDash with their food [74:16] pictures. All those pictures used to [74:17] have to be taken by photographers. Now [74:18] it's all done by AI. Massive reduction [74:20] in cost and they're using ads and ad [74:23] creative now at that is, you know, you [74:26] know, double-digit percentage more [74:27] effective while costing half as much. [74:30] So, I definitely think we're seeing it [74:32] on the earnings, but is is that true? [74:34] Are you seeing it in the earnings of [74:35] these companies yet? [74:36] >> Yeah, so so two data points. Number one, [74:39] we just saw Azure grow 39% in the [74:41] quarter. We saw Google Cloud grow 63% in [74:44] the a [74:45] Headcount growth for those companies the [74:47] last 3 years mag 5 combined is about 3%. [74:50] So their operating margins are all [74:52] expanding. If you look at the S&P 500 [74:55] writ large in Q1 of 24 operating margins [74:58] were about 11.8% [75:00] that was up from 11% in 23. This year [75:03] they're 13%. So we've had a 200 basis [75:06] point improvement in the operating [75:09] margins of the S&P 500 which is massive. [75:12] Do you think that's AI across those [75:14] businesses? [75:16] >> I think that that's the question where [75:18] where it could dovetail with what you're [75:20] saying. It's dollars to donuts it's not [75:21] AI. [75:21] >> Yeah, so [75:22] >> So but any amount of money is not AI. We [75:24] are the same financial engineering that [75:26] got these earnings to rise in the last [75:28] decade. [75:29] >> Yeah, so I I I think that's the [75:30] question. Are is this margin expansion [75:33] durable? The forecast the consensus [75:35] forecast and estimate is that margins [75:38] are going to continue to expand over the [75:39] course of the next 2 years. You and I [75:42] both know back in 22 23 we went from the [75:45] age of excess to the age of fitness, [75:47] right? A lot of these companies were [75:48] able to shed people, you know, with the [75:50] excuse of AI just because they had [75:53] become you know, [75:55] too excessive during the period of [75:57] COVID. So I I think it's a legitimate [75:59] question whether or not that's all from [76:01] AI, but I will tell you anecdotally it [76:03] maps for me. I'm hearing like Jason and [76:05] David a lot of these companies that are [76:08] you know, really growing their top lines [76:10] at an accelerating rate without [76:11] expanding head count nearly at the same [76:13] pace. [76:14] >> Okay. [76:15] Sacks, I'll give you the final word [76:16] while we're at it. [76:17] >> Brad was talking about how we got all [76:19] these operating efficiency improvements. [76:21] The unemployment rate stayed at [76:23] historical lows during that time. I mean [76:25] the economists consider full employment [76:27] to be 4 to 5% and we've stayed at you [76:30] know, the low low 4% 4.2% roughly during [76:34] this time. So you're able to get these [76:36] efficiency improvements while [76:38] unemployment is still extremely low. [76:40] Moreover, there was just a big article [76:42] saying that the unemployment rate for [76:45] young college graduates has dropped. So, [76:48] you know, there was this whole narrative [76:50] recently that recent college graduates [76:52] were going to have the hardest time [76:53] finding jobs because you know, there's [76:54] going to be no no work left for [76:56] entry-level jobs because of AI. And in [76:58] fact, it has gotten easier for recent [77:00] college graduates to find work recently. [77:02] Maybe that's because they're AI natives. [77:04] Maybe that's because they know how to [77:06] use AI better. [77:07] So, in any event, I mean, we're just not [77:09] seeing any evidence yet of these [77:11] theoretical downsides of AI around job [77:14] loss and unemployment. And we are [77:16] starting to see big productivity gains. [77:17] >> Yeah. I mean, this is going to be a [77:19] circular discussion, but yeah, there's a [77:21] lot of conflicting evidence. The last [77:23] piece of conflicting evidence, [77:24] obviously, is the labor participation [77:26] rate because if you are not even opting [77:28] into participate, then you [77:31] you know, don't get counted as [77:33] unemployed. And that's been, I think, a [77:35] big challenge. 61.9% in March labor [77:38] participation rate. [77:40] Back in before COVID, we're at 63.3. [77:43] >> Yeah, and college graduates are hearing [77:44] different stories. Certain degrees [77:46] getting jobs, other ones not getting [77:48] jobs. [77:49] It's too early to tell, I think, is [77:51] probably what we all agree in. And it's [77:52] a it's a mixed bag. [77:54] >> No, no, I don't agree with that. [77:55] >> No. Look, whenever whenever I have data [77:58] to refute one of your narratives, you [77:59] always say it's too soon to tell. [78:01] >> No, no, no. What I do [78:02] >> Headline, Wall Street Journal. Nick, put [78:04] it on the screen. College graduates are [78:05] finally catching a break in this job [78:07] market. [78:08] >> Yeah. [78:08] >> J Kyle, you should be happy about this. [78:10] >> I'm Listen, I'm happy anybody gets a [78:12] job. But what you do is then you say we [78:14] don't trust the numbers and we should [78:15] get rid of the Fed and we should get rid [78:16] of the numbers. [78:17] >> We should get rid of the Fed? [78:18] >> No, that was Schmalz that you you chose [78:20] Schmalz. Let's get rid of the Fed cuz we [78:21] don't like the numbers. You listen, it's [78:23] all great. Welcome to the debate club. [78:25] >> I say? What did I say? [78:25] >> You said abolish the Fed. [78:28] Abolish the Fed. [78:30] What is the Fed here for? All right, [78:31] listen, enough. We're getting into Trump [78:33] Derangement Syndrome or Trump Ben the [78:35] Knee Syndrome. It's the end of the show. [78:36] We had a great show, everybody. We had [78:38] some laughs. We all learned. We [78:39] workshopped some stuff. Let's leave it [78:41] where it is. Great job, President Trump. [78:44] >> to I want to congratulate [78:46] >> Oh, here we go. [78:47] >> I want to congratulate all of our [78:48] innovators, and I want to congratulate [78:50] Elon and um Dario D. Rockefeller on [78:53] their recent deal. [78:54] >> Oh, SHOTS [laughter] FIRED. THIS IS [78:56] GETTING STRAIGHT. Come on the program [78:58] anytime, Dario. Hey, you know Dario [79:00] well, Brad. Get him on the program. Next [79:02] week, I want him on the program. Have [79:03] him come on. [79:04] >> I mean, I I I'm going to ask [79:05] >> Ask him for me. [79:06] >> I [79:06] um [79:07] >> Will you ask him for me? [79:09] >> Well, I I sure I'll ask him. And you you [79:10] you know, the the the the fact of the [79:12] matter is [79:13] I thank our lucky stars that we have [79:15] Elon, that we have Anthropic, that we [79:18] have OpenAI, that we have Google, that [79:20] we have Amazon all innovating in this [79:23] country. And you know, I know we like to [79:25] you know, kind of poke poke fun on the [79:27] edges of these things, but the fact of [79:29] the matter is [79:30] you know, I see them all [79:33] showing up, sharing their models, [79:35] driving as hard as they can to innovate. [79:37] We have the best competitive framework [79:38] in the country. David's right, it's been [79:41] transformed over the course of the last [79:43] 14 months. We need to stay the course. [79:45] We're on the winning horse. We just had [79:47] the Derby last week. [79:49] >> There it is. [79:49] >> the winning horse. Stay on the horse. [79:51] >> America for the win. [79:53] >> America for the win. [79:54] >> There it is, Senator Brad Gerstner. I [79:56] think if you're going to run, you got to [79:57] get rid of the red glasses. We got to [79:58] get rid of the corduroy [laughter] [79:59] shorts in there, but I think you got to [80:01] you got a serious shot, Senator. Senator [80:04] >> I I like I like [80:05] I like secretary better. Secretary. [80:07] >> Secretary. [80:07] >> Uh [80:08] >> Who's Yeah, Secretary of the Treasury, [80:10] Brad Gerstner. [80:12] >> Secretary of [80:13] >> Secretary of State, David Sacks. [80:15] >> Secretary [80:16] >> of cashmere and wine. [80:18] >> Chamath Palihapitiya. How are you doing [80:19] with your uh [80:19] >> There's so much fake news out there [80:21] because I mean, look, I totally agree [80:22] with everything Brad said. Look, I poke [80:24] fun at at some of these companies for [80:25] some of the things they do, but I am [80:27] happy that they are American companies [80:28] and that they're innovating here. So, [80:31] congratulations on your Monopoly diet. [80:33] Absolutely. Yes, and look, there's so [80:35] much fake news out there. I mean, we [80:37] just covered on this podcast how [80:39] beneficial some of these economic trends [80:41] are. You never get it from the media. [80:43] And they are trying to derail us from [80:46] from, you know, the the policies that [80:47] have been so successful. [80:49] >> Yes. I but but they did some great [80:51] inspiring coverage of micro looting, so [80:53] get your micro looting on. [80:54] >> [laughter] [80:55] >> Congratulations New York Times. We'll [80:57] see you next time everybody. Bye-bye. [80:59] >> Love you boys. Love you best. [81:02] >> [music] [81:03] >> Let your winners ride. [81:06] >> Rain Man David Sacks. [81:10] >> And [music] I said, we open sourced it [81:12] to the fans and they've just gone crazy [81:13] with it. [81:14] >> Love you Sacks. [81:15] >> Ice Queen of Quinoa. [81:19] >> [music] [81:19] >> Let your winners [81:20] ride. [81:23] >> Besties are back. [81:25] >> [music] [81:26] >> That is my dog taking a in your [81:27] driveway, Sacks. [81:31] >> Oh man. [81:32] >> My hot Natasha will meet me at the [81:33] Sacks. We should all [music] just get a [81:34] room and just have a one big huge orgy [81:36] cuz they're all just useless. It's like [81:38] this like sexual tension that they just [81:39] need to release somehow. [81:40] >> [music] [81:41] >> What you talking about, B? [81:43] >> Let your queer feet. [81:45] >> What? [81:46] >> [laughter] [81:47] >> We need to get merch. [81:48] >> Besties are back. [81:50] >> [music] [81:57] [music]