SpaceX Files S-1 Seeking $1.75T Valuation Ahead of Historic June IPO

SpaceX提交S-1文件，在6月历史性IPO前寻求1.75万亿美元估值 SpaceX的招股文件披露了其与Anthropic达成的每年150亿美元巨额AI算力协议，与此同时，尽管面临广泛的宏观经济阻力，Nvidia的市场主导地位仍在向CPU领域持续扩张。 嘉宾小组分析了Anthropic的超高速增长、实现盈利的情况以及战略性引入Andrej Karpathy的举措，同时探讨了引发公众普遍反AI情绪的潜在经济与心理深层原因。他们深入拆解了SpaceX具有历史意义的S-1招股文件，剖析了其1.75万亿美元估值、Starlink的造血能力，以及xAI与Anthropic不断扩大的算力合作。此外，他们还分析了Nvidia单季高达816亿美元的爆发式业绩，审视了收益率飙升与顽固通胀等宏观经济预警信号，并对近期中美贸易磋商的战略成果进行了评估。 Andrej Karpathy 正式加入 Anthropic 领导预训练团队；同时 Anthropic 最近一个季度已实现净盈利，预计全行业大语言模型（LLM）年化经常性收入（ARR）到年底将达到 2000 亿至 4000 亿美元。 (0:00) 公众对 AI 的敌意日益加剧，主要源于经济收益不对称、潜在的外国信息战以及对人类智力霸权的心理威胁；而停止美国 AI 研发被视为极具博弈风险的举措，国际护栏与模型审查变得至关重要。 (12:42) 一项原计划发布的美国总统 AI 行政命令因包含强制联邦审查前沿模型的条款，在最后一刻被撤回。 (25:35) SpaceX 正式递交 S-1 招股书，拟募资 750 亿美元且目标估值达 1.75 万亿美元；受 Starlink 强劲盈利、火箭发射主导地位以及被称为“Elon Web Services”的 AI 算力基础设施推动，Anthropic 已与其签署每年 150 亿美元、为期 3 年的 Colossus GPU 集群租赁协议。 (45:18) Cursor 发布 Composer 2.5 模型，在利用 Colossus 算力结合专有代码 token 数据集进行强化学习后，实现了帕累托最优性能突破。 (49:39)

The All-In Podcast panel argues that rapid execution across frontier AI labs, SpaceX's emerging compute infrastructure, and Nvidia's hardware dominance continue to drive unprecedented commercial growth, even as macroeconomic debt pressures, rising bond yields, and public backlash threaten the broader environment. Frontier AI Scaling and Commercial Momentum (4:34) Frontier artificial intelligence development is accelerating both in capability and financial execution. Gavin Baker highlights that Anthropic recently achieved net-positive quarterly profitability, estimating that combined run-rate revenue for OpenAI and Anthropic is approaching $100 billion with roughly 80% gross margins on inference tokens, with total industry annual recurring revenue targeting $200 billion to $400 billion by year-end. Meanwhile, Andrej Karpathy's move to Anthropic to direct pre-training could compound capability improvements through recursive self-improvement. Model developer Cursor demonstrated this dynamic with its Composer 2.5 release, achieving Pareto dominance after three to four weeks of reinforcement learning on xAI's Colossus cluster using proprietary coding data. The competitive landscape is concentrating around four primary frontier model builders: OpenAI, Anthropic, Google (Gemini 3.1 Pro), and xAI (Grok 4.3). Beyond raw model scale, software runtime environments like Grok Build—which manage memory, state, and integrations across business applications—are becoming as critical as underlying foundation weights in driving enterprise utility. SpaceX's Compute Push and Valuation Underwriting (45:18) SpaceX has filed its S-1 registration statement targeting a $1.75 trillion valuation to raise $75 billion under the ticker SPCX. Starlink remains its financial core, delivering $11.4 billion in annual revenue at a 50% growth rate and $4.4 billion in operating profit across 10 million subscribers. However, capital expenditure reached $20 billion, with over 60% dedicated to computing infrastructure. Under what Baker calls 'Elon Web Services,' the company is industrializing data center deployment—compressing build times from 122 days down to 66 days—and securing massive off-take agreements, such as Anthropic’s three-year, $15 billion compute commitment on the Colossus clusters, subject to a 90-day mutual cancellation clause. Chamath Palihapitiya justifies an institutional case for a $2 trillion valuation by pointing to SpaceX's three interlocking moats: satellite communications infrastructure, unmatched launch cadence, and a rapid terrestrial hosting business projected to reach $100 billion to $200 billion in revenue by 2030–2032. Furthermore, Starship’s rapid orbital reusability and commercial off-the-shelf component architecture lay the foundation for viable space-based data centers, which Baker projects could become commercially operational between late 2028 and mid-2030. Semiconductor Economics and GPU Hardware Longevity (71:18) Nvidia delivered record Q1 results with $81.6 billion in revenue (up 85% year-over-year), $58 billion in net income, and $48 billion in free cash flow, while expanding its capital return program with an $80 billion buyback and a 25-fold dividend hike to $0.25 per share. Nvidia's data center growth continues to outpace custom ASIC alternatives from major hyperscalers, while its CPU division is tracking toward $20 billion in annual revenue. Baker notes an inefficiency in AI market valuations, observing that Nvidia trades at compressed multiples relative to peripheral cooling, power, and optical suppliers. The panel dismisses short-seller claims—such as those from Michael Burry—arguing that hyperscalers artificially inflate earnings by amortizing GPUs over four to six years instead of two. Baker explains that disaggregated inference architectures allow domain-specific accelerators (like Cerebras or Groq) to handle decode stages while legacy GPUs process core compute, extending older hardware lifespans to 10–15 years. This durability is reinforced by market behavior, where neo-cloud providers like CoreWeave secure 6% asset-backed debt and sign customers to non-cancelable six-year advance-pay contracts. Public Hostility, Labor Anxiety, and Regulatory Friction (9:06) Despite technological breakthroughs, artificial intelligence faces intensifying public skepticism. David Friedberg attributes this backlash to economic asymmetry—where immense wealth accrues to a handful of technologists while public benefits appear distant—alongside foreign information warfare and psychological distress over automated intellectual capability. Poor corporate messaging exacerbates this divide, highlighted by Cloudflare laying off 20% of its workforce while labeling departed staff 'measurers,' and Meta cutting 8,000 workers while deploying software to record desktop activity for model training. Policy responses remain divided. A planned federal executive order establishing oversight across frontier models was abruptly shelved by President Trump hours before announcement due to intrusive government review mandates. Meanwhile, Baker and Palihapitiya argue against protectionist measures like pausing autonomous vehicles or taxing warehouse automation, noting high turnover in manual logistics roles, that AV bans would open cities to wrongful death liability given human accident rates, and that civil tort law already imposes strict liability on developers for software harm. Macro Instability, Energy Advantages, and China Strategy (82:25) Broader economic indicators present severe headwinds: U.S. 10-year Treasury yields rose past 4.6%, forecasters project CPI inflation to reach 6.0% in Q2 amid prolonged Middle Eastern conflict, and global debt-to-GDP stands at 310%. Friedberg warns of a debt-servicing spiral forcing central banks into aggressive monetary expansion. However, Baker contends that domestic energy independence shields the United States, as domestic natural gas prices remain low while European and Asian liquefied natural gas costs surge 100% to 200%, inadvertently accelerating American manufacturing and data center onshoring. On geopolitics, the panel characterizes the recent U.S. diplomatic delegation to Beijing as transactional theater that secured minor trade purchases without altering structural friction. Baker argues that permitting Nvidia to export older, slightly throttled chips like the H100 and A200 to Chinese firms strengthens U.S. national security by keeping Chinese developers dependent on American architectures, disincentivizing the costly domestic development of indigenous semiconductor ecosystems. While the panel presents high conviction in the enduring capital moats of frontier model leaders, SpaceX, and Nvidia, unresolved vulnerabilities persist—namely whether runaway sovereign debt and spiking interest rates will spark a wider global credit contraction before AI productivity gains fully diffuse into the broader economy.

本期节目探讨了前沿人工智能的技术突破与日益加剧的公众抵触、SpaceX递交S-1招股书所揭示的庞大算力野心，以及英伟达历史性财报与全球宏观债务危机的激烈碰撞。讨论表明，尽管AI模型与算力基础设施展现出极高的商业增速，但行业正面临公众信任赤字、宏观流动性紧缩风险以及大国地缘博弈的多重制约。 前沿AI演进、公众反弹与叙事重塑 (0:00) Andrej Karpathy加盟Anthropic带领预训练团队，Chamath Palihapitiya将其比作早期的Google Fellows，预测聚焦于递归自我改进的模型开发或推动AI能力实现每年10倍的复合飞跃。Gavin Baker援引《华尔街日报》报道指出，Anthropic最近一个季度已实现净利润，预估OpenAI与Anthropic两家的年化经常性收入（ARR）合计接近1000亿美元且推理毛利率达80%，整个人工智能大模型市场年底ARR有望达到2000亿至4000亿美元。然而，单体模型参数膨胀带来的能耗压力促使David Friedberg主张转向分布式专业模型网络协同架构，以降低推理成本。 技术飞速演进的同时，AI行业遭遇了显著的公众抵触，例如Eric Schmidt在毕业典礼上面临学生嘘声。Friedberg分析了反AI情绪的三大根源：利益高度集中于少数投资者的经济不对称性、类似冷战时期的外国信息战放大民间对立，以及挑战人类智慧独占性的心理威胁。针对Cloudflare宣布裁减20%以上员工及Meta裁员8000人并监控桌面以训练AI所引发的社会恐慌，Chamath与Gavin呼吁科技领袖停止发布傲慢言论，转向突出实际民生价值，例如有病患家属利用大语言模型与蛋白质设计工具为患罕见神经疾病的女儿发现老药新用途，使其基线神经放电率从30–40%提升至80–90%。 SpaceX估值考量与算力基建扩张 (45:18) SpaceX已正式提交S-1文件，计划在6月中旬以股票代码SPCX上市，目标以1.75万亿美元估值筹集750亿美元，Polymarket数据显示其上市首日市值突破2万亿美元的概率达71%。招股书披露Starlink去年收入达114亿美元并实现44亿美元营业利润；火箭发射业务收入40亿美元，亏损6.5亿美元；AI基础设施业务营收32亿美元，亏损6.4亿美元，驱动了200亿美元资本开支中超过60%的算力投入。其中被称作“埃隆网络服务”（EWS）的算力租赁业务迅速崛起，Anthropic已与其签订为期3年、总额450亿美元的Colossus集群租赁协议，合同年化规模150亿美元并包含90天相互解约条款。 讨论强调，xAI的数据中心建设速度已从早期的122天压缩至66天，且正与英伟达联合开发直接直流（DC-to-DC）机架架构以减少转换热损耗。Cursor利用Colossus算力在专有代码数据上强化学习数周，推出的Composer 2.5模型实现了显著的性能跨越；xAI亦同步发布了集成桌面应用运行环境Grok Build。Friedberg指出，天基通信网络与数据中心为防范地面政府审查和专制提供了抗审查的人类文明备份；Gavin预计随着Starship推进日均多次飞行的快速完全可复用，轨道数据中心有望在2028年下半年至2030年上半年投入商业运行。 英伟达业绩韧性、算力寿命与芯片竞争 (71:18) 英伟达第一季度财报创历史新高，单季度营收816亿美元（同比增长85%，环比增长20%），净利润580亿美元，自由现金流480亿美元，毛利率达75%，并宣布追加800亿美元股票回购，将季度股息提高25倍至每股0.25美元。针对市场关于定制专用集成电路（如Google TPU、Amazon Trainium）正蚕食英伟达市场份额的论调，Gavin反驳称博通等厂商公布的AI芯片增长若剔除中国市场，在西方数据中心领域的实际增速不及英伟达，且竞争对手始终回避提交MLPerf标准化基准测试；此外，英伟达CPU业务年化营收正迈向200亿美元关口。 针对Michael Burry等投资者关于GPU实际有效寿命仅为2年、云服务商过度摊销折旧以虚增利润的质疑，Gavin阐明解耦推理技术改变了算力折旧规律。通过引入Groq或Cerebras等领域专用加速器负责解码阶段，老款GPU可置于后方承担核心计算，使其功能寿命延长至10到15年。Chamath引述CoreWeave首席执行官的陈述证实，客户不仅积极签署长达6年的不可撤销预付合约，该公司还能以约6%的低利率获得资产抵押贷款，表明具备真实资本敞口的产业界完全否定了做空者的短命论断。 宏观债务警报、地缘博弈与芯片出口政策 (82:25) 全球宏观经济指标发出严峻预警：中东战事延宕至第12周持续推升石油价格，Polymarket显示5月美国CPI达到或超过4.2%的概率为99%，市场共识预测二季度CPI或将攀升至6.0%，迫使降息预期转变为加息担忧。美债10年期收益率升破4.6%，日本30年期国债收益率创下5.1%的历史新高，英德长期国债收益率同样触及多年高点。Friedberg警告全球债务占GDP比重已达310%，政府过度赤字推高通胀将带来不可避免的债务下行危机；而Gavin指出，霍尔木兹海峡受阻导致海外LNG价格飙升100%至200%，但美国凭借粮食与能源自给自足以及本土低廉的天然气价格，反而在制造业回流与数据中心用能上获得了显著的不对称优势。 在科技监管与地缘博弈层面，特朗普在预定发布前数小时撤回了前沿AI行政命令，原因在于反对联邦政府对模型实施前置审查条款。近期中美北京会晤达成了有限的波音客机采购、农产品进口以及向百度出口降规英伟达H100/A200芯片的商业协议，但普京随即访华表明底层阵营对抗仍未化解。Gavin主张向中国出口降规旧款GPU符合美国安全利益，既能减缓中国构建完全自主芯片生态的迫切动力，又能维持美国底层计算标准的主导地位；而在潜在的台海危机中，控制海上能源通道依然是美国制约外部能源供给的核心战略杠杆。 本期讨论揭示了科技资本的几何级扩张与实体环境约束之间的深刻张力：一方面，前沿模型算力向吉瓦级乃至轨道空间激进拓展，核心硬件龙头展现出惊人的盈利能力；另一方面，全球债务高企与滞胀阴影、劳动力对自动化的恐慌抵制，以及主权国家在监管和能源安全上的地缘角力，正构成AI技术推进过程中必须直面的系统性阻力。

## 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。节目简介显示，本期特邀 Atreides Management 创始人 Gavin Baker 代班常驻主持 David Sacks，重点探讨了 Andrej Karpathy 加盟 Anthropic、美国社会对 AI 的舆论反噬、特朗普撤回 AI 行政令、SpaceX 递交 S-1 招股书冲击 2 万亿美元估值、英伟达最新财报与芯片做空潮，以及全球通胀、债市危机与中美地缘关系等重磅议题。 ## TL;DR 节目围绕前沿 AI 进展、资本市场超级巨头动态与宏观地缘风险展开了全方位拆解。四位投资人首先剖析了 Andrej Karpathy 加入 Anthropic 主攻模型递归自我改进的重大意义，并尖锐抨击了当下科技巨头以裁员引发公众反 AI 情绪的脱节公关。随后，众人详尽解构了 SpaceX 披露的 S-1 招股书，推演其由 Starlink、火箭运载与“Elon Web Services”算力底座支撑起的 2 万亿美元估值飞轮，并为英伟达财报大超预期却遭遇芯片做空进行了强力辩护。最后，讨论延伸至全球主权债券收益率飙升与通胀阴霾，以及美国科技高管随团访华背后的能源、地缘均势与芯片出口战略博弈。 ## Atmosphere & style 全场对话兼具硅谷顶级投资圈的硬核商业洞察与好友间毫无保留的插科打诨。主持人们在深入剖析工程架构、算力折旧与地缘算盘的同时，穿插着对彼此体型身材的毒舌嘲讽与对名企公关的犀利痛骂，现场笑料与火药味齐飞。听感上既有面对宏观债务危机与“末日过山车”俯冲时的冷峻严肃，又充斥着对太空算力、递归进化等未来图景的极致兴奋与行业狂想。 ## Bulletin - [45:18](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2718s) SpaceX 正式提交 S-1 招股书冲刺 1.75 万亿至 2 万亿美元估值，计划募资 750 亿美元，其中竞争对手 Anthropic 与其签下为期 3 年、总额高达 450 亿美元（每年 150 亿美元）的 Colossus 超算算力长协合同。 - [01:11:18](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=4278s) 英伟达第一季度实现营收 816 亿美元（同比增 85%，环比增 20%）和净利润 580 亿美元，今年 CPU 业务预计达到 200 亿美元，同时追加 800 亿美元股票回购并将季度股息提高 25 倍。 - [04:29](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=269s) 前 OpenAI 联合创始人 Andrej Karpathy 加盟 Anthropic 领衔预训练团队攻坚“递归自我改进”，而《华尔街日报》披露 Anthropic 上季度已实现 EBITDA 盈利，力破 AI 缺乏商业 ROI 的质疑。 - [01:22:25](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=4945s) 中东冲突推高油价引发通胀反扑，机构预测二季度美国 CPI 或达 6%，美债 10 年期收益率攀升至 4.6%，日本 30 年期国债收益率更突破 5.2% 创历史新高，加剧日元套利交易逆转的系统性风险。 ## Key takeaways - [02:08](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=128s) Andrej Karpathy 加盟 Anthropic 负责全新预训练团队，核心研发“递归自我改进”与持续学习；若 AI 能自主设计实验与迭代，模型能力有望开启每年数量级（10倍）暴增的新摩尔定律。 - [04:29](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=269s) 大模型行业正实质性兑现高毛利商业回报：Anthropic 达成正向 EBITDA 盈利，与 OpenAI 合计 ARR 逼近千亿美元且推理毛利率维持在 80% 左右，彻底粉碎了“左手倒右手无回报”的质疑。 - [20:10](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=1210s) 美国公众反 AI 情绪源于三重冲击：技术红利被少数寡头垄断引发的权力失衡感、外部地缘政治势力操控舆论挑起分裂，以及 AI 像“哥白尼日心说”一样击碎了人类智力优越感的哲学冲击。 - [23:01](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=1381s) 大国 AI 竞赛具有不可逆的曼哈顿计划属性：单方面放慢脚步只会打破战略平衡，中美当前模型差距不足 9 个月，双方应在生物武器等灭世风险上建立实名安全认证（KYC），而非人为窒息技术演进。 - [38:46](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2326s) 硅谷大厂裁员与监控引发反乌托邦公关危机：Meta 在利润创纪录之际裁员并监控员工屏幕以训练代码模型、Cloudflare 贬称被裁员工为“度量者”，展现出科技精英与社会沟通的严重傲慢与脱节。 - [45:18](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2718s) SpaceX 的 2 万亿美元估值由三大支柱支撑：年入 114 亿美元且盈利 44 亿美元的星链、营收 40 亿美元的火箭发射底座，以及在 Anthropic 每年 150 亿美元订单驱动下爆发的 AI 算力基础设施（EWS）。 - [55:26](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=3326s) 天基太空网络与轨道算力是人类文明的异地备份：面对地面主权政府日益收紧对信息、言论与数据中心的物理管控，不受单一政权干涉的卫星互联网和太空数据中心提供了独立延续商业与知识的终极自由底座。 - [01:05:07](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=3907s) 星舰追求的“极速可重复使用（Rapid Reusability）”形成代差垄断：唯有火箭降落后无需冗长拆解检修即可单日多次连续起飞，才能将每公斤入轨成本压至极限，改写深空基建与轨道算力的经济模型。 - [01:12:46](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=4366s) 二级市场在 AI 硬件链条上出现横向定价错位：市盈率仅 3-5 倍的存储芯片巨头及估值偏低的英伟达，与享有极高溢价的电力、液冷链条逻辑矛盾，算力与基础设施供应链终将迎来估值补涨或挤泡沫出清。 - [01:18:24](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=4704s) 解耦推理与生态绑定打破算力折旧神话：Grace CPU 年入 200 亿美元，且在专用芯片解码搭配旧款 GPU 预填充的混合异构架构下，老旧 GPU 经济寿命延长至 10-15 年，粉碎了空头唱衰新型算力云的论调。 - [01:23:57](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5037s) 全球债务见顶与日元套利交易反转濒临风暴：全球债务与 GDP 比例高达 310%，叠加日本 30 年期国债收益率异动至 5.2%，一旦流动性利差逆转，庞大的日元套利交易平仓将对国际金融体系带来毁灭性冲击。 - [01:30:38](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5438s) 能源自给与本土再工业化确立美国相对优势：美国发电主要依赖本土廉价天然气，在全球能源危机与霍尔木兹海峡受阻背景下，外部高油价压力反而成为倒逼美国制造业回流与再工业化的强力催化剂。 - [01:36:04](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5764s) 对华倾销上一代旧款 GPU 是维持美国 AI 统治力的战略筹码：全面断供只会倒逼中国全力突破自研独立生态与光互连，向其出口折旧芯片反而能将其研发深度锁定在基于 CUDA 的软硬件体系中。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=0s) 开场与重磅动态：Andrej Karpathy 加盟 Anthropic 节目在熟悉的轻松氛围中开场，迎来了第 274 期《All-In Podcast》。主持人 Jason Calacanis 宣布 David Sacks 本期缺席，但节目组请来了老朋友——阿特赖德斯管理公司（Atreides Management）的创始人兼首席投资官 Gavin Baker 代班。Jason 调侃道：“辛辣的观点必须奔涌流淌（The spicy takes must flow）”，热烈欢迎这位常客回归。本周科技圈大事接连不断，包括 SpaceX 与 OpenAI 的上市传闻、英伟达强劲的财报表现等，但几位主持人首先聚焦于震撼整个 AI 圈的重磅人事变动：前特斯拉 AI 负责人、OpenAI 联合创始人 Andrej Karpathy 正式加入 Anthropic。 年仅 39 岁的 Karpathy 已是科技界的传奇人物。Chamath Palihapitiya 透露，Karpathy 将在即将举办的峰会第二天上午发表主题演讲，而代班的 Gavin 不仅负责压轴当天的日程，还被当场“抓壮丁”担任最佳创投点子环节的评委。Jason 回顾了 Karpathy 近期的惊人影响力：他不仅创造了“氛围编程（vibe coding）”这一现象级词汇，前不久还利用周末随手写出了开源训练工具 AutoResearch——通过运行 5 分钟的微型实验帮助模型自我迭代，在 GitHub 上迅速斩获超过 8.2 万颗标星，掀起了普通开发者构建递归大语言模型的热潮；社区随后甚至针对 Claude Code 发布了基于其设计原则的“Karpathy Skills”扩展工具。如今，Karpathy 正式出任 Anthropic 全新预训练团队的领军人，核心研发方向正是“递归自我改进（recursive self-improvement）”，即让 Claude 学会自我提升与自我训练。 - **核心履历与成就**：OpenAI 创始团队成员、特斯拉 Autopilot/全自动驾驶（FSD）前负责人；提出“氛围编程”概念；开源 AutoResearch 项目在 GitHub 获得超 8.2 万标星。 - **Anthropic 新职责**：主管全新的预训练团队，全力攻坚“递归自我改进”，探索 AI 驱动 AI 自我进化的下一代前沿范式。 --- ### [02:08](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=128s) 查马斯评 Karpathy：从“苦涩的教训”到递归自我学习的新摩尔定律 Chamath 对这笔重磅人才引进给予了极高评价。他将 Karpathy 比作早期 Google 最顶尖的“谷歌研究员（Google Fellows）”——如 Amit Singhal、Sridhar Ramaswamy 以及至今仍在 Google 扛旗的 Jeff Dean。这些旷世奇才不仅是企业内的明星，更是始终站在历次技术巨浪最前端的弄潮儿。Chamath 认为，Karpathy 在 AI 领域的地位与之如出一辙，每一次技术范式转移都能看到他的身影。当年他在特斯拉主导 FSD 视觉研发时，是最早将强化学习先驱理查德·萨顿（Richard Sutton）著名论著《苦涩的教训》（The Bitter Lesson）付诸工业化落地的实践者——即彻底放弃繁琐的人工规则，坚定依靠海量算力与数据进行暴力破解。 Chamath 还分享了一段往事：在 2016 至 2017 年前后，贵为主管的 Karpathy 甚至亲自将约四分之一的工作时间花在给特斯拉车载视频做极其枯燥的手动数据标注上。从数据前线深潜，到联合创立 OpenAI，再到作为自由研究者随性开发，Karpathy 始终展现出纯粹的好奇心与顶尖的发明能力。Chamath 强调，递归自我学习技术就像是给基础模型同时挂上了“超频齿轮”和“自动驾驶仪”；一旦 AI 模型能够利用计算资源自动总结经验、自主迭代，大家很可能会见证模型能力出现一种新型摩尔定律——每年实现数量级（10 倍）的跨越式暴增，最终让模型质量曲线呈现出完全垂直的抛物线走势。 ```mermaid flowchart TD accTitle: AI递归自我改进与演进飞轮 accDescr: 展示从算力与数据驱动到AI模型自主反馈训练的闭环迭代流程。 subgraph Phase1["基础计算范式"] A["海量算力与数据"] --> B["基础模型前向传播"] end subgraph Phase2["Karpathy 递归闭环"] B --> C["自我实验与验证<br/>(如 AutoResearch 机制)"] C --> D["生成高质量合成数据与架构优化"] D -->|反馈注入训练循环| B end subgraph Phase3["系统质变"] D --> E["跨越'递归之谷'"] E --> F["模型能力呈现年度10倍抛物线增长"] end ``` --- ### [04:29](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=269s) 加文·贝克剖析 AI 商业回报：Anthropic 盈利与万亿美元 ARR 的远景 随后，Jason 转向代班的 Gavin Baker，询问他对 Anthropic 近期迅猛势头与激进招募的看法。Gavin 直言其商业表现“无可辩驳地非凡”。他指出，据《华尔街日报》最新披露，Anthropic 在最近一个季度已经实现了正向 EBITDA（息税折旧摊销前利润）。这一财务里程碑彻底打破了外界长期以来对 AI 行业“左手倒右手式循环融资（circular funding）”以及“缺乏实际投资回报率（ROI）”的巨大质疑。 Gavin 拆解了当下的商业逻辑：如果将 OpenAI 和 Anthropic 合计，两者的年度经常性收入（ARR）已向千亿美元大关迈进，且在模型推理端维持着约 80% 的极高毛利率。这证明真实回报已经实质性兑现。若再算上 Google Gemini、Cursor、xAI 以及开源阵营，到今年年底整个大模型行业在极高毛利下斩获 2000 亿至 4000 亿美元的 ARR 绝非难事。更何况，这甚至还没有计入 GPU 最赚钱的传统腹地——例如 Meta、Google 和亚马逊依靠 AI 算力在推荐系统、精准广告投放与效果衡量上已经攫取的巨额增量收益。 谈到技术内核，Gavin 认为 Karpathy 正在攻关的“递归自我改进”与“终身/持续学习（continual learning）”正是通用人工智能的最后两座堡垒： - **递归自我改进**：允许模型在前向传播训练期间直接向自身或外部监管模型输入优化反馈，推动模型能力加速跃迁。在这种机制下，Chamath 所说的“每年 10 倍提升”甚至可能显得过于保守。 - **持续学习（终身学习）**：让模型像人类一样在与现实交互的经历中实时积累经验、动态更新知识库，而无需从头进行昂贵的全量预训练，这被视为 AI 进化的终极圣杯。 --- ### [07:13](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=433s) 跨越“递归之谷”：小模型网络架构与本地化部署的未来 Jason 顺势抛出了一个更具哲学与工程意味的话题：我们何时才能跨越“递归之谷”——即 AI 在模型构建与迭代中所做的贡献，正式彻底超越人类工程师本身的投入？对此，David Friedberg 给出了冷静且务实的工程学视角。他认为，实现这一目标并不一定非要把整个模型硬塞进上下文窗口让其直接自我重写，技术路线上存在着大量不同的架构探索空间。 Friedberg 指出，业界一个巨大的机遇在于“模型重构”：制造更小尺寸的专用小模型（SLM），并将它们编织成协同运作的网络。相比于单体巨型模型（monolithic models），模型协同网络在完成复杂任务时能够大幅降低每个 Token 产出的能量损耗与综合成本。他前几周曾分享过来自麻省理工学院（MIT）的论文，证明该方向潜力巨大，哪怕仅仅取得微小的架构突破，也能瞬间将单 Token 成本砍去一半。 Jason 对此表示强烈赞同，提到专注于企业端垂直小模型的初创公司 Abacus.AI 正在这一赛道高歌猛进。他还提及近期被外界低调忽视的一则新闻：Google 在两周前悄悄将 4GB 大小的 Gemini Nano 模型内置进了全球用户的 Chrome 浏览器底层，用于在本地无感执行自动校对、语法纠错和智能自动补全。Jason 惊呼 Google 这是在“隐蔽地（covertly）”将模型硬塞进每一个人的操作系统；Chamath 随即打断并提醒道：“‘隐蔽’这个词太重了”，Jason 笑着退让：“好吧，至少是在完全没有提前告知用户的情况下”。众人一致认同，AI 正以迅雷不及掩耳之势深入端侧本地与各层基础设施。 ### [09:58](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=598s) 停止对模型的狂热追捧：聚焦真实落地场景与终端用户价值 Chamath 认为，科技界整天屏息凝神、大呼小叫地追捧模型每一次细微的迭代提升，纯粹是在浪费时间，根本没有任何投资回报率（ROI）。真正重要的是，人类目前正处于一条加速学习的轨道上，即将见证过去完全无法想象的终端用户突破，这才是应当关注的核心。例如，在人类的配合协助下，OpenAI 成功解决了一个悬而未决数十年的数学难题；在生物医药领域，原本束之高阁、不被看好的临床前候选药物（IND），如今借助 AI 重新获得了推进临床试验的希望。当前的技术已经深入到现实世界，正在切实为人们创造巨大价值。 然而，行业内很多人的沟通方式反而加剧了公众对 AI 的恐惧，将其塑造成一个非黑即白、敌我对抗的“可怕怪物（boogeyman）”，以至于只要一提起 AI，人们就如临大敌，仿佛在谈论一个肮脏的脏字。Chamath 认为，四位主持人以及整个行业有责任平衡叙事、摆出客观数据，否则全社会若陷入卢德主义（Luddite）开历史倒车的境地，对谁都没有好处。针对外界关于 Google 在后台运行大模型可能侵犯隐私的争议，Chamath 明确表示 Google 绝不是那种搞见不得人勾当（shady things）的公司，Jason 随即插话打趣讽刺：“你是说 Meta 吗？你在暗示你的前东家？”引发全场大笑。Chamath 回应自己不会指名道姓，但他确信 Google 的出发点纯粹是为了终端用户的实用性，行业必须把焦点转回给用户带来的实际效用上。 - **数学难题破局**：在人类引导配合下，OpenAI 攻克了困扰数学界数十年的未解难题。 - **医药研发提速**：多个原本被认为不可行而搁浅的候选药物分子与新药临床试验申请（IND），在 AI 辅助下重获新生、重回临床试验。 --- ### [12:27](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=747s) 科技行业的倡导责任与地缘政治竞争 讨论转向科技巨头的合规文化与宏观地缘格局。Gavin Baker 补充指出，业内通常有“极度依赖律师（attorney maxing）”的说法，而 Google 毫无疑问已经是“律师化到极致（attorney maxed）”的公司，绝不可能轻易违规越界。随后，Gavin 严肃地谈到，所有身处科技行业的美国人都肩负着一种内在责任，那就是成为 AI 带来的积极可能性的坚定倡导者，向公众展示乐观的未来图景。 Gavin 警告称，目前美国社会上弥漫的反 AI 与反数据中心声浪，其背后隐约让人感觉到有中国（CCP）资金在推波助澜的影子。从地缘政治和竞争逻辑来看，中国采取这种策略削弱对手是极其合理的，但这对于美国的发展却极其不利。因此，科技从业者不能任由负面叙事滋生，必须主动承担起发声的责任。 --- ### [13:21](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=801s) 各怀鬼胎的科技巨头：监管护城河与 Palantir CTO 的呼吁 Jason 随即抛出质问：究竟是谁在公关叙事上搞砸了？是 Anthropic 首席执行官 Dario Amodei 整天四处宣扬“所有人都会丢掉工作”，还是那些把自身裁员全部归咎于 AI 的企业 CEO？Chamath 犀利地反驳道，这些大厂掌舵人不过是在“为自己的账本交易（trading their own book）”。对于 Dario 而言，竭力营造危机感并推高监管门槛是一项极其精明的商业策略——因为 Anthropic 如今的体量已经足够庞大，只要监管高墙竖起，他就能躲在安全的大帐篷内“从里往外撒尿（inside of the tent pissing out）”，把后来的竞争对手全挡在门外。只要梳理 Anthropic 历轮融资规模与 Dario 言论音量的攀升曲线，两者完全同步吻合。 Chamath 指出，每个 CEO 在巨大的规模利益驱动下，都在执行一套精心设计的自我保护或反制策略，因此外界必须跳出这层迷雾、看清森林的全貌。他引用了 Palantir 首席技术官 Shyam Sankar 此前在福克斯新闻（Fox News）上的精彩呼吁：“不要再去屏息倾听那些大模型制造商的夸夸其谈了。去走到最前线的终端用户身边，去问问工厂里正在使用模型的工人怎么想，去问问医生和科学家的看法，去讲述那些真实的落地故事，这才是我们真正该讨论的东西。” ```mermaid flowchart TD accTitle: AI公关困境与利益驱动机制 accDescr: 展示头部模型厂商、社会舆论反噬与真正终端应用叙事之间的逻辑关系 subgraph BigTech["头部大模型厂商策略"] A["宣扬AI末日论/失业论"] --> B["推高AI安全监管门槛"] B --> C["构建自身监管护城河<br/>(将后来竞争者阻挡在门外)"] end subgraph Backlash["公众舆论与社会反噬"] C --> D["引发公众恐慌与卢德主义情绪"] D --> E["大学毕业典礼嘘声四起<br/>科技被视为权力失衡的工具"] end subgraph Solution["正确叙事转向"] F["深入一线终端用户"] --> G["工厂工人效率倍增"] F --> H["医疗与罕见病精准救治"] F --> I["打破垄断，普及生产力红利"] end E -.->|亟需纠偏| Solution ``` --- ### [15:08](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=908s) 拯救生命的真实案例与叙事重塑 Gavin Baker 赞同 Chamath 的观点，认为必须用普通人的真实故事来对抗虚无的恐慌。他分享了自己十天前在一场活动中听到的震撼经历：一位对冲基金经理的女儿出生时患有一种极其罕见的基因突变疾病，大脑中的神经元放电率仅有 30% 到 40%，几乎剥夺了孩子所有的快乐与生活质量，传统医学界对此束手无策，几乎等于宣判了“无解”。 这位父亲绝不接受认输的命运。他亲自利用大语言模型（LLM）进行了海量的文献挖掘与交叉研究，最终在市面上成功筛查出一种已获审批且安全、但此前从未被用于该病症的现有药物。药物使用后，女儿大脑的神经元放电率惊人地飙升到了 80% 至 90%，完全具备了过上正常人生活的能力。如今，这位父亲进一步利用 AI 推进蛋白质设计与分子定制，极有信心在几个月内研发出能够彻底治愈该罕见病的全新药物。Gavin 感慨，一个绝不向命运屈服的父亲凭借 AI 彻底改写了女儿乃至所有同病患者的人生，这才是科技应有的模样。相比之下，虽然马斯克描绘的“全民基本收入（UBI）与工作变成可选项”会让一部分人感到失去职业依托的恐慌，但黄仁勋等倡导者的努力和这些拯救生命的实例，才是整个行业应该倾力聚焦的光明面。 - **罕见病突破**：患儿大脑神经元放电率在 LLM 筛选药物干预下，从 30%~40% 跨越式提升至 80%~90%。 - **药物定制前沿**：结合前沿蛋白质生成设计，原本被判绝症的基因突变病症有望在数月内迎来彻底治愈的靶向药物。 --- ### [18:05](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=1085s) 年轻人为何狂嘘 AI？权力失衡、非对称回报与原子弹式恐惧 Jason 顺势将话题引向现实中正在爆发的“AI 公关危机”：在近期的多场大学毕业典礼上，只要演讲嘉宾一提到 AI 就会遭到学生们的激烈嘘声，前 Google 首席执行官 Eric Schmidt 在内的多位重量级人物甚至被当场起哄喝倒彩。Jason 困惑地发问：年轻人明明热爱智能手机、拥抱互联网，为什么偏偏对 AI 这一项技术产生如此强烈的厌恶与仇视？ David Friedberg 给出了深刻的社会学剖析。他指出，公众内心深处存在一种持久的恐惧：技术往往为极少数人赋予巨大的杠杆，进而造成难以弥合的权力失衡，而 AI 正是这种不平衡的终极象征。大众认为，极少数控制、受益并从 AI 中榨取暴利的人，正在攫取相对于全社会而言极不成比例的超额回报。尽管所有技术最终都会走向商品化和普惠扩散，但 AI 的技术扩散周期存在极大的时间滞后，导致在当前乃至未来相当长的一段时间内，社会各阶层所承受的回报与冲击呈现极端非对称性。Friedberg 将这种心理机制比作 20 世纪中叶原子弹诞生给人类思想带来的震荡：正是核武器的出现，让 20 世纪后半叶的公众对尖端科学和前沿工程产生了深刻的怀疑与戒心——人们越来越恐惧，那些掌握深奥知识并将其工程化落地的人，会为自己构筑不可动摇的超级优势，从而将其余所有人置于被动和脆弱的境地。 ### [20:10](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=1210s) 反 AI 浪潮的深层根源：贫富悬殊、地缘干预与哥白尼式的“非人类中心”冲击 David Friedberg 剖析了美国公众对人工智能产生强烈抵触情绪的三重核心原因。首先，最根本的痛点在于普通人无法从当前的技术浪潮中看到自身的切身利益：“这门技术究竟何时能让我受益？它到底怎么让我受益？”这些疑问在今天给不出确切答案。相反，大众每天接收到的叙事是巨额经济利益正疯狂向极少数人集中，仿佛演变成了一套少数人掠夺多数人的新型权力体系。普通人既不懂底层逻辑，也不清楚自身未来，却眼睁睁看着少数寡头狂赚数万亿美元，引发强烈反弹完全在意料之中。 其次，Friedberg 指出这种反科技情绪的背后存在外部势力的煽风点火。这不仅局限于中国通过非政府组织（NGO）的运作，更是一套可追溯至冷战时期克格勃（KGB）设计的舆论干预老剧本。长期以来，主权国家通过操控竞争对手国内的媒体和舆论场，刻意挑起分裂并放大抗拒技术的情绪，以此削弱竞争对手的发展动能与创新步伐。 第三重因素则触及深层的人类心理与哲学。Friedberg 将当前的 AI 冲击比作“哥白尼革命”。当年日心说的提出对教会与既有权力中枢造成了毁灭性颠覆，打破了“地球是宇宙中心、人类受上帝直接眷顾”的自大认知，让人类不得不接受自己只是宇宙微尘的现实。而当下的 AI 同样展现出一种极具颠覆性的“非人类中心（not human-centric）”甚至“反人文主义”特质，它毫不留情地击碎了人类的智力优越感与心理防线，成为许多人内心反感、排斥这项技术的潜意识催化剂。 ```mermaid flowchart TD accTitle: 驱动反 AI 情绪的三重核心力量 accDescr: 展示大众抵触 AI 技术的经济、地缘及心理哲学动因 subgraph Causes["反 AI 情绪的驱动维度"] A["经济分配失衡：利益归寡头，大众承担风险"] B["地缘信息战：外部势力放大抵触情绪以拖慢竞争对手"] C["哥白尼式哲学冲击：击碎人类中心主义与认知优越感"] end subgraph Effect["社会反应"] D["公众广泛的抗拒、焦虑与反技术浪潮"] end A --> D B --> D C --> D ``` --- ### [23:01](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=1381s) 曼哈顿计划的历史镜像：为何 AI 竞赛不可踩刹车 面对 Jason Calacanis 追问“我们究竟该不该放慢 AI 研发脚步”，Chamath Palihapitiya 斩钉截铁地表示“这根本不可能放慢”。Chamath 借用二战后曼哈顿计划的历史经验展开推演：当年美国建立国家实验室体系，初衷之一是防止研发核弹的顶尖科学家流失到苏联或德国。起初许多科学家基于道德良知反对制造核武器，甚至有人因担忧美国一家独大、全球缺乏制衡力量，而故意将核机密泄露给苏联。 然而，一旦技术扩散的齿轮开始转动，竞赛就彻底不可逆转。当苏联掌握机密并全力研制氢弹时，美国为了形成战略均势，别无选择只能开足马力领跑。核威慑的本质在于维持恐怖平衡，一旦失去平衡，拥有绝对非对称力量的一方将能在全球范围内肆意妄为。当前的 AI 格局正处于完全相同的历史关口：如果美国因行业限制、征税或监管而放慢脚步，中国或其他竞争对手绝不会止步。一旦中国在模型前沿能力与规模化部署上建立起压倒性优势，全球地缘政治将被彻底颠覆。因此，任何单方面的减速都会亲手打破这一关键的技术平衡。 --- ### [25:10](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=1510s) 机器人税与保护就业的假象：谁在为淘汰工种立牌坊？ Jason Calacanis 提出了当下华盛顿讨论的两种保护就业、缓冲技术冲击的方案： 1. **自动驾驶渐进式部署**：全美有超过 1,000 万依靠驾驶谋生的群体（出租车、Uber、公交与卡车司机），政策或要求分阶段落地，强制保留人类司机“在方向盘后”的岗位安全。 2. **拟人化机器人税（Robot Tax）**：针对进入亚马逊仓库的 Optimus 或刚刚完成整周分拣包裹测试的 Figure 机器人，按工作小时征收机器人税，税款专款用于受影响产业工人的转岗培训。 Chamath 对此进行了尖锐反驳。他指出，所有大谈“保护工人”的政策提案，从未真正做过针对卡车司机或仓库分拣工的真实民意调查： - **亚马逊仓库真实流失率**：亚马逊物流仓储部门的年化离职流失率（churn）高达 35% 至 40%。 - **工作本质认知偏差**：如果分拣包裹真是一份令人热爱、值得代代相传并以此体面养家的好工作，离职率应该只有 3% 到 4%。 - Chamath 直斥这纯粹是“场外围观党（peanut gallery）”为了抢占道德制高点、转嫁内疚感的伪善作秀，根本没人去问工人自己究竟愿不愿意一辈子做这种摧残身体的机械劳动。 --- ### [27:22](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=1642s) 特朗普叫停 AI 行政令：中美均势、勒内·吉拉尔与模型 KYC Chamath 与嘉宾们将话题转向当天突发的政坛地震：白宫原本计划在当天签署一项关于人工智能的重要总统行政令（EO）。前一天泄露的受邀名单包括各大前沿 AI 实验室与超大规模云厂商（Hyperscalers）的 CEO，甚至包括他们的好友、Palo Alto Networks 首席执行官 Nikesh Arora。然而，就在会议开始前一小时，总统突然紧急撤回了该行政令。根据透露出的信息，特朗普拒绝签署的核心原因是不满行政令中要求联邦政府对前沿 AI 模型实施强制性审查、监管和干预的条款。 针对监管与中美竞合，Chamath 阐述了务实的地缘战略思考： - **技术差距与战略均势**：中国在前沿 AI 上仅落后美国不到 9 个月，这反而是件好事，因为它促成了大国间的“战略缓和（détente）”，避免了单极霸权带来的误判。 - **勒内·吉拉尔（René Girard）的欲望模仿理论**：中美是两种制度正交（orthogonal）、组织架构截然不同的社会，这种异质性反而比两个一模一样的社会更能避开零和模仿冲突，从而更有可能摸索出和平与繁荣的共存路径。 - **AI 领域的实名安全认证（KYC）**：大国间真正需要协同的底线，是防止最极端的技术失控。中美应就模型分发达成“了解你的客户（KYC）”监管共识。既然中国在模型发布前已具备完备的训练审查机制，双方理应联手设立准入门槛，严格防止极端分子利用开源或前沿模型研发生物武器等灭世级威胁。除此之外，绝不应人为设限阻碍技术自身的发展。 ### [29:21](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=1761s) 前沿模型的监管难题：中美竞争与政府权力的“单向棘轮” 关于美国是否应该对前沿 AI 模型实施严格测试和监管框架，嘉宾 Gavin Baker 表达了审慎态度。面对 Jason 提出的二选一质问——是否应在泄露的行政命令（EO）所指的强大前沿模型发布前强制推行某种测试与监管机制——Gavin 坦言这一话题极其复杂，当前设立此类限制为时过早。他不赞成美国单方面推行自我约束而其他国家毫无限制的做法。在他看来，唯有在中美两国能够携手、彼此互信、目标一致且具备核查能力的理想情境下，设定武器放下（"weapons down"）的双边规则才具可行性。 Jason 随即换了一个角度：如果中美双方仅针对公认的高危领域（如生物武器、恐怖主义活动等极端威胁），设立一套类似 FDA 筛查食品药品毒素污染的简易准入测试，是否应当支持？Gavin 对此指出，美国体系本身具备自我调节机制以及完善的司法追责途径： - **普通法与侵权追责机制**：若 AI 公司行为不当造成危害，受害者完全可以通过民事诉讼寻求巨额赔偿（正如 OpenAI 目前正因未成年人与 AI 模型交流后自杀而面临法律诉讼）。 - **政府权力的“单向棘轮”效应（One-way ratchet）**：权力一旦赋予监管机构，几乎从未有过主动收回的先例，只会像棘轮一样单向膨胀。事前监管牌照很容易扼杀创新，而事后诉讼追责已经能倒逼模型开发商保持谨慎。 --- ### [31:56](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=1916s) 自动驾驶、就业替代与“过失致人死亡”的法律悖论 讨论转向 AI 与自动化对底层劳动力的冲击。针对波士顿、纽约等城市以及英国（如近期《60 Minutes》专题报道）部分群体主张禁止或严格限制无人驾驶以保住出租车司机饭碗的现象，Chamath 提出了尖锐反驳：如果人类司机撞死行人，而当地原本可以部署零致死率的无人驾驶系统，那些为了保护岗位而阻碍技术的决策者和市政当局很可能会面临“过失致人死亡”（wrongful death）的民事诉讼。无人驾驶关乎生命安全，这与在亚马逊仓库分拣包裹的性质完全不可同日而语；真想了解工人的想法，应该去问问仓库里的分拣员，他们究竟愿不愿意一辈子干这种苦力。 Gavin 完全赞同无人驾驶不可阻挡的趋势。他回忆起网约车刚普及的年代：当你降落在一个没有 Uber 的城市，那种低效会让人觉得极其不便和恼火，甚至“再也不想来这个地方”。同样地，未来如果某个城市不允许乘坐 Waymo 或特斯拉 Cybercab，只会让人觉得野蛮（barbaric）且缺乏安全保障。 - **交通事故伤亡数据**：美国每年因车祸死亡的人数约为 5 万人，全球每年高达 100 万人。在技术能大幅消灭事故的前提下，人为抵制自动驾驶在法律和伦理上难以维系。 - **地方竞争与“柯利效应”（Curley effect）**：美国各州与地方市镇自治形成了制度试验场。尽管政策碎片化可能带来短期摩擦，但城市之间的良性竞争最终会通过人们“用脚投票”淘汰落后政策。 --- ### [34:01](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2041s) 治安监控革命：“犯罪已成为一种选择” 科技与地方治理的交锋同样在公共安全领域上演。Jason 和 Gavin 以 AI 安防摄像头公司 Flock Safety 为例展开讨论。Gavin 抛出了一个深刻论点：“如今，犯罪已经变成了一种选择（Crime is now a choice）。”技术已经有能力彻底解决绝大多数街头犯罪，各地方政府完全可以自由选择做“支持犯罪”还是“打击犯罪”的城市。令人啼笑皆非的是，哈佛大学所在地的马萨诸塞州剑桥市议会（Cambridge City Council）日前竟然投票决定关闭枪声探测系统，理由竟然是担心系统会导致开枪的非法移民被捕。 与此形成鲜明对比的是拉斯维加斯的警务实践。Chamath 分享了他与妻子拜访风险投资人 Ben Horowitz 及其妻子 Felicia 的经历。Felicia 深度参与并资助了拉斯维加斯大都会警察局的技术升级： - **无人机应急网络**：警局楼顶配备了随时待命的无人机系统，一旦出现突发警情，数分钟内无人机便能飞抵现场并将高清画面传回指挥中心，对犯罪嫌疑人展开实时追踪。 - **极高的投入产出比**：相比犯罪给城市经济和居民安全带来的巨大损失，这类软硬件投入微不足道（de minimis）。每年只需投入 3,000 万至 4,000 万美元，就能让一座城市跃升为全美最安全的都会。 - **隐私保护与技术妥协**：对于公众担忧的隐私侵犯，Jason 指出已有成熟的工程化治理方案，例如限定车牌数据库滚动覆盖周期（如最多保存两到三年）、提供防篡改审计追踪记录（audit trail），以及暂不开启面部识别等。 ```mermaid flowchart TD accTitle: AI Security vs Traditional City Tech accDescr: Comparison between tech-forward law enforcement in Las Vegas and restrictive policies in Cambridge. subgraph TechForward["科技驱动型治安治理（以拉斯维加斯为例）"] A["街头传感器 / 枪声检测"] --> B["屋顶警用无人机数分钟内抵现场"] B --> C["指挥中心实时追踪嫌疑人"] C --> D["极低犯罪率与高安全感（年投入仅三四千万美元）"] end subgraph IdeologicalConstraint["意识形态驱动型治理（以剑桥市为例）"] E["出于特定政治诉求投票关停枪声检测"] --> F["放弃技术破案手段"] F --> G["治安恶化，居民被迫用脚投票迁移"] end ``` --- ### [37:16](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2236s) 大厂裁员不仅是“挤水分”：AI 替代正在真实发生 在进入下一议题前，Jason 总结了 AI 当前面临的严峻公关危机（PR Problem）。他严肃指出，硅谷科技大厂持续推进的裁员潮已经不能再简单归咎于“挤掉疫情期间的人员冗余”，也不是企业用来掩盖经营不善的借口（即所谓的“AI 洗白”/ AI washing），而是 AI 驱动的自动化正在真真切切地剔除白领岗位。 Jason 列举了两个令人不寒而栗的企业案例： 1. **Cloudflare 裁撤 20% 员工**：CEO Matthew Prince 两周前宣布裁减超过 20% 的员工。他在公告中坦言，公司此时正处于历史最佳状态——营收创纪录增长、自由现金流强劲、客户数量空前激增。裁员并非因为陷入困境，而是公司决定全面淘汰“度量者”（Measurers，即那些专门管理他人和统计分析数据的人），AI 已经使这部分管理层与数据中间层变得毫无必要。 2. **Meta 与扎克伯格的冷酷裁员**：马克·扎克伯格（Mark Zuckerberg）主导的新一轮裁员手法被舆论批评为缺乏人文关怀且极具“反乌托邦”（dystopian）色彩。在随后的公开讲话中，扎克伯格直言不讳地谈到了如何借由 AI 提升整座公司员工的平均智力门槛。 ### [38:46](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2326s) Meta 的 AI 悖论：监控员工屏幕、用工程师训练模型与大裁员恐慌 扎克伯格（Mark Zuckerberg）展示了一种冷酷的技术迭代逻辑：由于顶尖工程师解决编程问题的能力远超普通众包外包人员，Meta 开始利用内部成千上万名顶级软件工程师构建内部工具、解决复杂任务，以此来训练自家 AI 模型的编程能力。这种做法被认为能让 Meta 模型的代码编写水平以远超行业对手的速度飞速提升。然而，与这一宏大技术愿景同步发生的一幕却令人不寒而栗——扎克伯格在 Meta 创下历史最高利润的同时，宣布裁减 8,000 名极具才华的员工，其中许多人持有 H-1B 工作签证，裁员直接打碎了他们的个人生活和家庭未来。 更具讽刺意味和反乌托邦色彩的是，在这轮以及此前数万人的大裁员期间，公司要求在全体员工的电脑上安装行为记录监控软件，旨在全面记录和研究员工的一举一动以训练 AI 模型。不少员工在数月前的内部 AI 黑客马拉松上兢兢业业地开发各种自动化工具以提高工作效率，随后却在这一轮裁员中被直接解雇。Jason 指出，这在科技圈乃至整个社会引发了巨大的生存焦虑和信任危机：“你现在唯一能指望的，就是在岗位上再苟延残喘一段时间，亲自把替代你的 AI 训练出来，然后被扫地出门。”当一家科技巨头的掌舵人毫不掩饰地宣称“我们将全面研究你们以加速替代”，这彻底吓坏了所有人。 - Meta 内部利用高水平工程师的操作数据训练 AI 编程能力，效率远超第三方数据标注承包商。 - 伴随模型训练的是新一轮 8,000 人裁员潮（此前已有数万人裁员），重创持有 H-1B 签证的高技术员工。 - 员工电脑被安装监控记录软件，用于捕获日常工作行为作为模型训练数据，引发了“亲自训练自身替代品”的强烈恐慌。 --- ### [40:23](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2423s) 查马斯怒喷科技 CEO 与公关灾难：Cloudflare 备忘录与傲慢的精英脱节 面对这种社会层面的恐慌，Cloudflare 首席执行官马修·普林斯（Matthew Prince）近期发布的裁员备忘录彻底点燃了查马斯（Chamath Palihapitiya）的怒火。查马斯直言这份备忘录简直出自“弱智公关学校”，措辞之糟糕无出其右。在信中，管理层冷酷地将有血有肉的员工简化并贴上所谓的“度量者”（the measurers）标签，并宣称要将这批人全部清洗裁撤。查马斯愤怒地表示，这种做法无异于给员工背上刺上侮辱性的“红字”，让他们在未来求职时被外界戴上有色眼镜盘问，对解决实际问题毫无助益。 查马斯借此援引了 Palantir 首席技术官希亚姆·桑卡尔（Shyam Sankar）此前的真知灼见：外界在 AI 议题上过度倾听那些自诩天才的模型创造者，而真正值得倾听的是一线工人和护士——那些借助 AI 成功增加第三个生产班次、雇佣更多工人，或者得以省出时间陪伴病人、防止在换班时发生意外的普通劳动者。查马斯毫不客气地痛批：“闭上你们的臭嘴，滚回键盘后面老老实实干活！如果需要裁员管理就低调处理，别再写这种自我感动的公开信了，你们在这方面烂透了。”他更一针见血地指出，随着这些大厂在公布惊人裁员计划的同时，下一步监管文件必然是宣布巨额股票回购和分红提升，坐拥天量现金池却在公共沟通上极端脱节。 - Cloudflare CEO 马修·普林斯在公开备忘录中将员工分类并贬称为“度量者”（measurers）加以裁撤，遭查马斯痛斥为灾难性公关。 - 查马斯呼吁停止对 AI“发明家”们的盲目崇拜，转而关注借助 AI 提升效率、创造岗位的一线工人和医护人员。 - 科技巨头在裁员引发恐慌的同时，账面现金暴增并推行大规模回购与分红，其公关口径与社会情绪严重撕裂。 --- ### [43:09](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2589s) 节目间隙的插科打诨：假招聘风波与查马斯的“细狗腿”之争 在严肃的技术与商业探讨之余，主持人团队展开了一段招牌式的互损与恶搞。Jason 煞有介事地播报起“突发新闻”，戏称 Anthropic 近期又引进了重磅新员工：不仅 Sam Altman 个人宣布转投 Anthropic，保守派媒体人塔克·卡尔森（Tucker Carlson）也将加入 Anthropic 总部负责公关并开办播客，甚至连查马斯本人也“被加盟”。紧接着，屏幕上打出查马斯穿着卡其裤的照片，引发了演播室的一阵哄堂大笑。 Jason 随即抓住查马斯的下肢身材大肆嘲讽，取笑他胖了 5 磅，还穿着类似“护膝推高文胸”的裤子来掩盖自己瘦如细木条的“鸵鸟腿”，甚至质疑他是不是在股四头肌里塞了半圆平衡球（Bosu balls）来打肿脸充胖子，或是偷偷注射了 BPC-157 肽类药物来“大腿最大化”（leg maxing）。查马斯则强烈反驳，称自己身高 6 尺 2 寸（约 188 cm），肌肉线条在长骨骼上的分布自然无法与身高仅 5 尺的 Jason 相比。尽管场面一度陷入混乱的互怼，弗里德伯格（David Friedberg）最终还是打圆场，承认查马斯最近在腿部训练上的确下了不少苦功、体型有了显著改善。 --- ### [45:18](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2718s) SpaceX 提交 S-1 文件：冲击 2 万亿美元估值的世纪 IPO 随着气氛回归严肃，Jason 抛出了本期节目的重磅核心：SpaceX 已于周三正式提交 S-1 招股说明书，计划以 1.75 万亿美元的估值公开募资 750 亿美元，预计将于 6 月中旬（最可能为 6 月 12 日）挂牌上市，股票代码定为“SPCX”。如果顺利完成，这将成为人类历史上规模最大的 IPO，募资额将达到沙特阿美（Saudi Aramco）此前创下的 290 亿美元纪录的两倍以上。根据 Polymarket 的实时预测，SpaceX 在上市首日收盘市值突破 2 万亿美元的概率已高达 71%。 招股书详尽披露了 SpaceX 的三大核心业务板块： 1. **Starlink（星链业务）**：目前的现金牛与利润引擎。去年实现营收 114 亿美元，同比增速达 50%，带来 44 亿美元的营业利润。目前付费订阅用户已突破 1000 万，未来具备触达数亿订阅用户的广阔成长空间。 2. **Space（航天发射业务）**：传统核心业务，去年录得 40 亿美元营收，维持 17% 的稳健增长，但伴随 6.5 亿美元的营业亏损。 3. **AI 与算力业务（xAI / EWS）**：去年营收达到 32 亿美元，同比翻倍以上，但营业亏损高达 64 亿美元。去年公司整体资本支出（Capex）高达 200 亿美元，其中超过 60% 均投向了 AI 算力设施的建设。 ```mermaid flowchart TD accTitle: SpaceX 三大业务板块与财务全景图 accDescr: 展现 SpaceX 提交 S-1 招股书后披露的星链、航天发射与 AI 算力业务的财务数据及 Anthropic 算力长协订单。 subgraph SpaceX["SpaceX 拟 IPO 估值：$1.75T ~ $2.0T | 代码: SPCX"] Starlink["Starlink (星链) • 营收: $11.4B (+50%) • 营业利润: $4.4B • 订阅用户: >1000万"] Launch["航天发射业务 • 营收: $4.0B (+17%) • 营业亏损: $6.5亿"] EWS["AI 与算力 (Elon Web Services) • 营收: $3.2B (同比翻倍) • 营业亏损: $6.4B • 算力资本支出: 占 $20B Capex 的 60%+"] end Anthropic["客户: Anthropic"] -- "$1.25B/月 ($15B/年) 3年长协共 $45B" --> Colossus["Colossus 1 & 2 超级算力集群"] Colossus --> EWS ``` 最令资本市场震惊的爆发点在于被播客戏称为“EWS”（Elon Web Services，埃隆网络服务）的算力租赁业务。招股书透露，竞争对手 Anthropic 正在以每月 12.5 亿美元（折合每年 150 亿美元）的巨额代价向 SpaceX 租用 Colossus 1 及部分 Colossus 2 超级计算集群，签署了一笔为期 3 年、总价值高达 450 亿美元的天价长协合同。仅这一单算力协议，便凭空为 SpaceX 创造出了一个等同于 Starlink 全年营收体量的收入支柱。与此同时，SpaceX 对 AI 编程神器 Cursor 的收购案（预计将直接注入 20 至 30 亿美元且保持 100% 翻倍增速的营收）尚未计入本次 S-1 文件中，为后市估值留出了巨大想象空间。 - **拟募资与估值**：计划募资 750 亿美元，目标估值 1.75 万亿美元，预计 6 月 12 日挂牌，代码 SPCX。 - **Starlink 表现**：年营收 114 亿美元，同比增长 50%，运营利润 44 亿美元，全球订阅用户超 1,000 万。 - **航天发射业务**：年营收 40 亿美元，同比增长 17%，运营亏损 6.5 亿美元。 - **AI 算力业务亏损与 Capex**：年营收 32 亿美元，运营亏损 64 亿美元；200 亿美元 Capex 中逾 60% 砸向 AI 算力底座。 - **EWS 天价大单**：Anthropic 签订 3 年 450 亿美元长约（12.5 亿美元/月，150 亿美元/年），整包租用 Colossus 超算集群。 - **Cursor 增量预期**：收购 Cursor 将为集团再添 20~30 亿美元高增长营收（年增速超 100%）。 ### [47:49](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2869s) “Elon Web Services”的狂飙与 150 亿美元大单 关于 SpaceX 提交的 S-1 招股文件，Jason 引导嘉宾 Gavin Baker 将目光聚焦到被戏称为“Elon Web Services”（埃隆网络服务）的数据中心与算力托管业务上。Gavin 笑着指出，这个戏称虽然滑稽，但其展现的增长势头极其惊人：仅 Anthropic 一家公司就带来了高达 150 亿美元的算力订单，这意味着该项 AI 算力基础设施业务的体量几乎在一夜之间直接翻了两番（Quadrupled）。更令人震撼的是招股书中披露的惊人工程推进速度：马斯克团队建设首座数据中心耗时 122 天，第二座缩短到 91 天，而到了第三座数据中心，仅用了区区 66 天。以远低于行业标准的成本和前所未有的工程交付速度盖数据中心，马斯克团队正把这一套打法彻底流水线化。 Gavin 进一步分析道，既然目前已经有了明确的算力采购大户（Off-take partner），未来这种合作势必会演化为多客户生态，其盖机房、堆算力的速度没有任何放慢的理由。熟悉英伟达 CEO 黄仁勋（Jensen Huang）的人都知道，黄仁勋最看重的就是自己的 GPU 能否迅速点亮上电、投入使用；英伟达分配芯片的逻辑向来简单粗暴——谁能最快把芯片插上机架、接通电源、将电子源源不断地转化为 Token，谁就能优先拿到卡。正因如此，马斯克麾下算力业务的扩展速度远超所有人三个月前的想象。 Jason 随即补充了一个关键的交易条款：该笔 150 亿美元协议中规定，任何一方只需提前 90 天书面通知即可单方面解约。这意味着如果马斯克后续自身需要收回算力，或者 Anthropic 找到了更具性价比的替代方案，双方都拥有灵活的退出机制。Gavin 认同这是一个对彼此都非常合理的缓冲条款，但这一单的体量依然为算力基础设施赛道立下了全新的里程碑。 - **首座数据中心建设周期**：122 天 - **第二座数据中心建设周期**：91 天 - **第三座数据中心建设周期**：仅 66 天 - **Anthropic 算力采购协议**：150 亿美元规模，附带 90 天单方取消条款（90 days notice） --- ### [49:40](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2980s) Cursor 2.5 突围与“算力+专有代码数据”的飞轮 谈到算力与模型的配合，Gavin 将话题引申到本周 AI 领域的一项重磅进展——Cursor 发布的 Composer 2.5 模型。他特意让制作人切出帕累托前沿图（Pareto frontier），指出 Composer 2.5 在性能上展现出了绝对的帕累托主导（Pareto dominant）优势。其背后的实现路径极具启发性：这仅仅是在世界最大的单体算力集群 Colossus 2 上，结合 Cursor 自身的专有数据进行了 3 到 4 周的强化学习（RL）训练。据称，Cursor 内部拥有的私有编程交互数据量甚至超过了整个公开互联网上所有代码 Token 的总和（而且这还是至少一年前的统计数据），其在私有代码数据壁垒上与 Anthropic 并驾齐驱。 Composer 2.5 的底座依然是与 Composer 2 相同的 Kimmy K2.5 模型，这意味着仅仅通过几周深度的强化学习、尚未将其反哺进全新基座模型的预训练阶段，其综合表现就已经彻底甩开了行业曲线。接下来可以预见的技术路线更为恐怖：将 Cursor 的代码数据直接注入新一代基座模型的预训练中，再利用全球规模最大的紧密算力集群进行深度强化学习。此前，Cursor 因受限于算力瓶颈，在与 Google、Anthropic 和 OpenAI Codex 的竞争中一度显得难以为继，甚至濒临掉队；而马斯克向他们开放 Colossus 算力集群后，Cursor 瞬间重获生机、实现火箭式增长。 ```mermaid flowchart LR accTitle: Cursor 与 Colossus 协同提升模型路径 accDescr: 展示 Cursor 利用专有数据与 Colossus 算力通过强化学习跃升至帕累托前沿的过程 subgraph Data["数据壁垒"] D1["Cursor 海量私有代码数据\n(超过公网代码 Token 总和)"] end subgraph Compute["算力底座"] C1["Colossus 2 集群\n(全球最大紧密算力基础设施)"] end subgraph Training["训练演进"] M1["基座模型: Kimmy K2.5"] --> M2["3-4 周强化学习 (RL)"] M2 --> M3["Composer 2.5 发布\n(突破帕累托前沿)"] M3 -.-> M4["下一代预训练基座整合\n+ 超大规模集群 RL"] end D1 --> M2 C1 --> M2 ``` 当 Jason 发问未来一年内马斯克是否有可能把算力反向卖给 Google 或 OpenAI 时，Gavin 认为卖给 Google 的概率极高，甚至网络上已有风声传出。更重要的是，Google 绝不会想在“轨道算力”（Orbital compute）的太空布局中缺席。目前行业中只有完全脱离数据中心和太空产业的外行才会嘲笑轨道算力，而像 Google、Anthropic、Amazon 和英伟达等巨头，都坚信在太空中部署数据中心是必然趋势，而 SpaceX 无疑在这一领域处于压倒性的垄断位置。 --- ### [52:57](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=3177s) 四骑士格局与“模型外壳”的体系进化 Gavin 随后梳理了目前全球最顶尖大模型竞赛的残酷现实。当前真正稳稳站上“帕累托前沿”的仅剩“四骑士”：xAI（凭借拥有 5000 亿参数的 Grok 4.3 最新版本占据一席）、Google（Gemini 3.1 Pro）、OpenAI 以及 Anthropic。每一家在最顶级的前沿坐标轴上都有一到两个坐标点。 此前 xAI 的短板在于缺乏配套的应用外壳与工作流环境（Harness）——Anthropic 有 Claude Code，OpenAI 拥有深度绑定的 Codex。而随着 xAI 刚刚发布 Grok Build，这一短板被迅速补齐。Jason 与 Gavin 解释道，所谓外壳绝不仅仅是一个打通了 Slack、Gmail 和 Notion 的本地下载端，而是一个包含了运行环境（Runtime）、状态管理（State）与长效记忆的复杂体系。在前瞻性的 Agent 时代，业内顶尖专家一致认为“外壳与模型本身同等重要”，二者必须协同演进。 | 前沿阵营巨头 (The Four Horsemen) | 旗舰前沿模型代表 | 核心配套工作流 / 外壳 (Harness) | |:--- |:--- |:--- | | **xAI** | Grok 4.3 (约 500B 参数) | Grok Build (全新推出，提供状态/记忆环境) | | **Google** | Gemini 3.1 Pro | 自研生态环境与多模态原生接口 | | **OpenAI** | 旗舰 o 系列 / GPT 系列 | Codex 环境与开发者工具链 | | **Anthropic** | Claude 3.7 / 3.5 系列 | Claude Code 及 Agent 架构 | Gavin 强调，Composer 2.5 不仅在基准跑分上领先，还迅速成为 Cursor 上被调用最多的模型；在业界俗称的“使用氛围”（Vibes）层面，其实际交互体感也非常出众。Jason 补充说，马斯克此前对 xAI 研发进展缓慢曾公开表达过极度沮丧，但随着近期架构理顺、代码上线速度大幅提速，马斯克的焦虑感明显减轻，xAI 正在重新开足马力（Cooking with oil）。 --- ### [55:26](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=3326s) 抵抗陆地极权：太空网络作为人类文明的备份系统 针对 SpaceX 即将上市的传闻，以及外界盛传 Tesla 与 SpaceX 可能合并打造代号为“$ELON”、市值直奔全球第四大上市公司的构想，Jason 询问 Friedberg 的看法。Friedberg 没有单纯从市盈率倍数或财务模型切入，而是将思考升华到了全球地缘政治与人类文明演进的高度，呼应了本期播客早期关于反科技、反 AI 浪潮的讨论。 Friedberg 犀利地指出，全球范围内正出现一种危险的势头：各国政府正在系统性、有组织地收紧对信息流动、言论自由以及商业交易的控制，试图对数字社会施加更多干预。互联网诞生之初被赞誉为超越实体商业的自由乌托邦，但人们忽视了一个物理事实——互联网最终必须物理性地落脚在地球的某片土地上。近期美国本土对数据中心扩张的围剿和政治阻力，恰恰暴露出这种集中在陆地上的脆弱性。 在 Friedberg 看来，SpaceX 的真正战略意义在于从物理底层构建一套“替代性互联网”（Alternative internet）。一旦拥有不受任何单一主权政府干涉、管制或物理破坏的低轨卫星通信网与太空轨道数据中心，并实现天对地的无线宽带直连，就等于为人类文明乃至科技进步打上了一份永不熄灭的“云端异地备份”（A backup for civilization）。哪怕地球某些区域陷入极权压迫、封锁言论或阻断信息流动，这套天基基础设施依然能保障人类知识与商业文明的延续。抛开商业估值不谈，这种提供独立于主权束缚之外的自由信息传输与存储能力，才是马斯克航天体系最深远的价值所在。正当 Jason 想要接着补充马斯克当年创立 SpaceX、满世界寻找俄罗斯廉价火箭运载火箭的早期轶事时，全场的讨论已然触及了商业背后的底层哲学。 ### [58:30](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=3510s) 备份地球生物圈：太空数据中心的构想与争论 Jason Calacanis 回忆起 Elon Musk 当初创立 SpaceX 的原始愿景——为了“给生物圈做备份”（back up the biosphere）。当年 Elon 从俄罗斯考察买火箭未果回国后，私下对 Jason 坦言，既然真正的瓶颈在运载工具上，还不如自己造火箭；他甚至构想在太空建造测地线穹顶（geodesic domes），把地球上的植物、野生动物和各类生灵送入太空。如今，这种宏大愿景延伸到了计算领域。Jason 认为，把数据中心送入太空并非天方夜谭，当年无数人质疑 Starlink（星链）行不通，现在轨道上已有上万颗 Starlink 卫星在运转；从概念上讲，数据中心卫星与通信卫星并没有不可逾越的鸿沟。 Chamath Palihapitiya 则立即提出了技术异议，指出两者在物理结构上完全不同，数据中心的散热翼片、机翼结构和体积都要庞大得多。不过 Jason 辩称，随着运载能力提升 10 倍的全新星舰（Starship）登场，物理体量的限制将被打破。Chamath 承认这在技术上并非不可能，且 Elon 极有可能是全球第一个攻克这一难题的人，但看待估值依然需要回到更切合实际的商业数字上。 ### [01:00:01](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=3601s) 2 万亿美元估值推演：飞轮效应与超级计算基础设施 针对市场如何给 SpaceX 支撑起 2 万亿美元估值的问题，Chamath 给出了自己的投资测算逻辑。SpaceX 去年的营收约为 180 亿至 190 亿美元，今年预计达到 250 亿至 300 亿美元。虽然以目前营收来看估值倍数昂贵，但投资者实际上是在为三大核心业务买单： ```mermaid flowchart TD accTitle: SpaceX 三大业务支柱与估值逻辑 accDescr: 展示 SpaceX 如何通过 Starlink、火箭发射平台以及地面巨型计算中心形成资本与技术飞轮，支撑 2 万亿美元估值。 subgraph CoreAssets ["三大业务支柱"] A["Starlink 星链基础设施<br/>互联网诞生以来最重要项目<br/>服务数亿用户，成本极速下降"] B["天地运载发射网络<br/>GDP+10% 至 15% 稳健复合增速<br/>承载一切未来业务的底座平台"] C["AI 算力基础设施 (Colossus)<br/>吉瓦级（Gigawatt）地面数据中心<br/>顶层应用 + 底层极致算力输出"] end subgraph MoatFlywheel ["护城河飞轮加速"] D["资本护城河 (Capital Moat)"] --> E["技术护城河 (Tech Moat)"] E --> F["执行与工程迭代护城河 (Execution Moat)"] F --> D end CoreAssets --> MoatFlywheel MoatFlywheel --> G["2030-2032 远景：地面算力达 1000-2000 亿美元营收<br/>2 万亿美元估值对应仅 10-20 倍市销率"] ``` Chamath 指出，不能仅用传统的当期市盈率或现金流倍数来衡量这家公司，因为高速增长的营收赋予了 Elon 极高的高经营杠杆，去跨界投资并巩固护城河。这种飞轮先由庞大的资本护城河驱动，进而加速技术护城河，再进一步转化为极致的工程迭代与学习护城河。此外，Elon 旗下各个资产未来极可能进一步整合，将物理移动能力（特斯拉）、全网连通性（星链）与底层算力融为一体。放眼全球科技界，Elon 是唯一一位能在台上带来如同当年乔布斯 "One more thing" 那般、产生文明级颠覆创新的领军者。 - **地面数据中心蓝图**：根本不需要押注虚无缥缈的太空数据中心，单凭地面建设的 Colossus 3、Colossus 4 等吉瓦级数据中心，到 2030 至 2032 年就能产生 1000 亿至 2000 亿美元营收。 - **直击电源架构痛点**：目前机房通行的“直流-交流-直流”（DC-AC-DC）多次转换导致了严重的电能损耗与高昂制冷成本。Nvidia 的黄仁勋正在全力推行全流程“直流对直流”（DC to DC）革新，而 Elon 是全球唯一兼具电力系统设计能力与巨型机房场景、能与 Nvidia 联手落地的顶级工程合作伙伴。 ### [01:05:07](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=3907s) 星舰的体量革命与“快速可复用”的降维打击 Jason 展示了两张震撼的对比图表：第一张是运载火箭体型图，星舰的有效载荷空间堪比一架波音 747 客机，内部甚至能容纳 300 人，让现役的主力猎鹰重型（Falcon Heavy）相形见绌；第二张是 1957 年至今人类历史累计发射入轨质量统计图，自 2012 年以来，SpaceX 凭借一己之力，正在迅速超越人类有史以来所有国家与机构发射载荷的总和。 Gavin Baker 随后深入剖析了航天发射经济学的分水岭。他指出，目前图表展现出的惊人数据，甚至全部是在星舰尚未正式商业运营前、仅靠 Falcon 猎鹰系列达成的。更关键的技术代差在于**“可复用”（Reusable）**与**“快速可复用”（Rapidly Reusable）**的本质鸿沟： - **普通可复用（Falcon / 竞争对手路径）**：火箭回收后需要耗费大量时间对发动机、整流罩进行深度翻修与检测，通常需要 30 到 60 天才能再次执飞。即便蓝色起源（Blue Origin）或某些国家在未来攻克了回收，其技术水平也仅仅是 SpaceX 十年前的状态。 - **快速可复用（Starship 目标路径）**：同一枚火箭在降落后无需冗长拆解检修，能在单日内完成多次起降并连续发射。这种能力将把每公斤入轨成本压低到前所未有的地步，彻底改写全球航天与空间基建的经济模型。 ### [01:07:42](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=4062s) 星舰的极速复用挑战与轨道算力的商业曙光 星舰（Starship）之所以被视为航天工程的终极试炼，关键在于它所追求的并非普通的可重复使用，而是“极速可重复使用”（Rapid Reusability）。如果目标仅仅是造出一枚能重复利用的火箭，工程难度虽然不低，但远比目前要容易得多。然而，埃隆·马斯克（Elon Musk）的野心是建立月球基地、火星殖民地以及在月球表面部署质量加速器（Mass Drivers），这就必须实现极速周转与发射。极速复用一旦彻底攻克，即使其他航天机构全部掌握火箭回收技术，向轨道运送载荷的份额也将迅速收敛，其他竞争者能拿到的蛋糕将微乎其微。关于实现这一目标的时间表，行业普遍共识是在一到两年内，甚至可能更快。无论当天的试飞是完美着陆还是化作一团巨大的火球，SpaceX 的核心文化就是“从失败中快速迭代学习”，没有失败就谈不上真正的技术跃进。 与此同时，轨道算力（Orbital Compute）已经不再是遥远的科幻构想。太空中目前已经有一块正在运行的英伟达 H100 GPU，此前安德雷·卡帕斯（Andrej Karpathy）甚至已经在轨道上对其完成了模型训练与推理任务。不过，将数据中心芯片直接送入太空面临着截然不同的工程约束：数据中心使用的庞大散热模块和结构配重在太空中毫无必要，但芯片本身必须进行结构加固，以承受发射时惊人的重力加速度（G力）与剧烈颠簸，防止物理结构震碎解体。 ```mermaid flowchart TD accTitle: 轨道计算芯片工程适配与商业化路径 accDescr: 展示英伟达数据中心GPU经航天工程改造并结合SpaceX低成本上行方案后部署于轨道的演进路径 subgraph Ground["地面数据中心级设计"] A["标准英伟达 H100 GPU"] --> B["超重被动散热片与高抗压机架架构"] A --> C["未考虑极高震动与高 G 力物理冲击"] end subgraph Space["太空级工程改造"] D["重新设计太空热交换与轻量化架构"] E["抗震结构加固与特殊封装"] end subgraph SpaceX["SpaceX 差异化降本策略"] F["消费级/非抗辐射芯片选型"] G["火箭舱段力学与系统级防辐射工程"] H["通过太空耐受测试即投入 Starlink 部署"] end subgraph Orbit["轨道部署与应用 (2028-2030)"] I["Andrej Karpathy 在轨训练与推理验证"] J["Starlink 卫星群分布式轨道算力组网"] end Ground --> Space Space --> Orbit SpaceX --> Orbit ``` 值得注意的是，SpaceX 最被低估的核心能力之一，在于其运载系统和载荷舱出色的工程兼容性：他们无需使用价格高昂的航天特种抗辐射芯片（Radiation-Hardened Chips），就能让廉价的商用半导体在太空中正常运行。例如 Seraphim Space 投资的 Excite Labs，其供应给每颗 Starlink 卫星的芯片原本完全不是为太空环境设计的，但在接受抗辐射测试时表现优异，直接被 SpaceX 大规模采用。对于轨道算力真正形成规模化商业闭环的时间节点，业界的基准预测普遍落在 2028 年下半年至 2030 年上半年之间。 - 轨道计算验证现状：已有英伟达 H100 部署于太空并成功运行，完成过端到端模型微调训练及推理。 - 芯片适配工程差异：剥除数据中心冗余散热与结构配重，针对发射阶段高 G 力颠簸做加固。 - 商业化落地时间点：预计在 2028 年下半年到 2030 年上半年全面铺开。 --- ### [01:11:18](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=4278s) 英伟达爆炸级一季报与 5.3 万亿美元市值的增长奇迹 英伟达（Nvidia）公布的最新第一季度财务报表再次震撼了华尔街，其财务扩张速度被评为资本市场上“前所未见的异类”。公司单季度实现营收 816 亿美元，同比暴增 85%，环比增幅亦高达 20%。在传统美股市场中，20% 的年化增长率即可被定性为高成长股，而英伟达却以 5.3 万亿美元的庞大体量，实现了 20% 的“环比”飞跃。更为惊人的是其盈利能力：单季度净利润录得 580 亿美元，自由现金流达到 480 亿美元，综合毛利率稳定在 75% 的极高水平。 尽管基本面业绩呈现断层式领先，英伟达股价年内涨幅却相对温和地收于 16% 左右。面对海量现金沉淀，管理层公布了一份力度空前的股东回馈计划：在 2023 年初启动的 1000 亿美元回购计划基础上，追加 800 亿美元股票回购额度，两者合计锁定了公司流通股本的约 4%；同时将季度股息提高 25 倍，由每股 0.01 美元跳升至 0.25 美元。首席财务官更公开承诺，未来将把公司自由现金流的 50% 直接通过分红与回购返还给二级市场股东。 - 单季度核心业绩指标：营收 816 亿美元（同比 +85%，环比 +20%），净利润 580 亿美元，自由现金流 480 亿美元，毛利率达 75%。 - 股东回报矩阵：新增 800 亿美元股票回购授权，季度股息扩增 25 倍至每股 0.25 美元，承诺回馈 50% 自由现金流。 --- ### [01:12:46](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=4366s) AI 估值结构错位与英伟达对决博通的 ASIC 算力之争 面对英伟达亮眼的业绩，市场深层却暗流涌动。知名 AI 投资人利奥波德·阿申布伦纳（Leopold Aschenbrenner）的基金规模短时间内从零飙升至 50 亿美元，其 13F 持仓披露中赫然出现了对芯片半导体板块的巨额看跌期权（Puts）。不过，这一持仓实际上记录的是第一季度末地缘政治担忧达到顶峰时期的防御性对冲动作，随着风险落地，并不必然意味着他对芯片长周期赛道全面看空。 当前的二级市场定价存在极其明显的“横向逻辑割裂”（Cross-Sectional Inefficiency）。从估值倍数来看，存储芯片（DRAM）制造巨头的市盈率（P/E）被压制在 3 到 5 倍的极低水平，英伟达本身的市盈率相对其成长性而言也处于偏低区间，部分加速器芯片商的估值同样中规中矩；但与此形成鲜明对比的是，电力设备、精密温控、光模块等 AI 基础设施周边链条却被赋予了极高的溢价倍数。这两种估值逻辑绝不可能同时成立：如果电力和冷却环节的高估值是合理的，意味着下游对算力吞吐的需求极端迫切，英伟达与存储芯片的市值严重低估，后市必将大幅补涨；反之，若芯片端当前的低市盈率真实反映了终端需求的降温，那么电力与温控供应链的估值泡沫将面临严厉出清。 | 板块类别 | 典型代表领域 / 标的 | 当前估值倍数特征 | 市场隐含预期与定价悖论 | |:--- |:--- |:--- |:--- | | **底层核心算力与存储** | 英伟达 (Nvidia)、DRAM 存储巨头 | 市盈率极低（DRAM 仅 3-5x P/E，英伟达倍数偏低） | 隐含硬件周期见顶或资本开支缩减假设 | | **基础设施支持链条** | 变压器与电力设备、液冷/温控系统、光模块 | 估值溢价高企，享有多倍高倍率估值 | 隐含算力中心建设永续扩张、负荷无限增长假设 | 与“市场割裂”并行的，是关于英伟达与博通（Broadcom）定制芯片（ASIC / 谷歌 TPU）的市场份额争夺战。市场上近期盛传英伟达正在被 ASIC 夺走市场份额，博通更给出了其 AI 半导体收入同比激增 143% 的超预期业绩指引。但英伟达在近期调整了财报统计口径，将业务拆解为数据中心、边缘计算，并在 AI 领域细分为超大规模云厂商（Hyperscalers）、新型 AI 专业云（AI Clouds）以及工业与企业客户。若剔除受政策限制的中国区市场，仅以欧美主流市场构建可比口径，将超大规模厂商与以 CoreWeave、xAI 为代表的新型 GPU 云采购合并计算，英伟达的核心 AI 算力业务增速实质上依然稳稳跑赢博通，打破了其在主干数据中心建设中落后于 ASIC 的市场叙事。 - 估值悖论：3-5 倍市盈率的存储巨头与高估值的电力温控产业链无法长期共存，当前 AI 市场定价呈现横向无效状态。 - ASIC 竞争实质：剔除中国市场影响后，英伟达在由 Hyperscalers、CoreWeave 及 xAI 驱动的西方算力中心建设中，实际增长率仍然超越博通的 AI 半导体业务线。 ### [01:16:58](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=4618s) 英伟达的估值困境与 ASIC 芯片的“阴影之战” 加文·贝克（Gavin Baker）与查马斯（Chamath Palihapitiya）深入探讨了英伟达（Nvidia）管理层面临的巨大挫败感。英伟达交出了破纪录的财务业绩，但在二级市场上却几乎“得不到任何奖赏”。市场上甚至弥漫着一种“英伟达正在丢失市场份额”的看空叙事，这在数据面前显得极其荒谬——哪怕不做任何额外调整，英伟达的实际营收增速依然远超各大超大规模云服务商（Hyperscalers）的资本支出（CapEx）增速。 令黄仁勋格外愤怒的另一点，是各类自研 ASIC 芯片拒绝参与客观的基准测试。无论在计算机视觉分析、InferenceMAX 还是公认的 MLPerf 榜单中，都见不到这些竞争芯片的身影。加文直言，它们之所以不敢提交测试数据，核心原因就是“拿去测一定会输”。在没有公开发布干净透明的基准数据之前——无论是英伟达 GB300 对决谷歌 TPU v7，还是叫板亚马逊的 Trainium 或 Inferentia——市场都处于一种在“与阴影作战”（fighting shadows）的状态。 加文进一步指出，当一家企业成长为全球市值最大的超级巨头时，其交易逻辑往往呈现“阶梯式”（stairstep）形态： ```mermaid flowchart LR accTitle: 巨头估值阶梯式演进模型 accDescr: 展示超大市值公司在市盈率压缩后重新估值的阶梯演进过程 subgraph S1["估值承压期"] A["业绩创新高"] --> B["市场质疑体量天花板"] B --> C["市盈率倍数持续压缩"] end subgraph S2["重估与新台阶"] C --> D["估值重新定价 (Re-rating)"] D --> E["在更高基数确立新底部"] end ``` - **超级巨头的阶梯模式**：市场因对超级公司的庞大体量心存疑虑，会不断压缩其估值倍数（multiple compression），直至业绩彻底打破天花板完成重新估值（re-rating），确立新的股价底部。 - **ASIC 真实战力存疑**：尽管市场上看好 Trainium 等专用芯片的声音不断，但缺乏透明客观的公测打榜数据让其实际竞争力成谜。 --- ### [01:18:24](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=4704s) 200亿美元CPU业务、DSA演进与“解耦推理”架构 本季度英伟达财报中另一个震撼市场的关键数据是：公司预计其 CPU 业务今年的营收规模将达到惊人的 200 亿美元。这意味着英伟达几乎是在一夜之间跃升为全球最大的 CPU 制造商之一。加文强调，这体现了英伟达独一无二的产业生态位——它是业内唯一一家能与全球每一家前沿 AI 实验室展开深度合作的公司。这种全方位的触角让英伟达能够与大模型团队深度协同设计（co-design），精准预判下一代架构演进。 查马斯借此反思了行业关于“领域专用架构”（Domain-Specific Architecture, DSA）的认知偏差。此前业内普遍认为 Groq、Cerebras 等专用架构会取代通用 GPU，但现实却是“DSA 市场的演化直接发生在英伟达体系内部”。例如英伟达在 Colossus 3 集群中与埃隆·马斯克展开高压直流（DC-to-DC）供电合作，直接将技术标准推高数个台阶。 加文随即提出了关于算力资产寿命的核心洞见：在“解耦推理”（disaggregated inference）的新范式下，GPU 依然是系统架构的核心锚点。老旧的 GPU 可以与 Groq 或 Cerebras 的专用解码（decode）加速芯片混合编组，由专用芯片负责高频 token 输出，老旧 GPU 则负责预填充计算。这一架构让早先的 GPU 算力资产获得了长达 10 到 15 年的有效经济寿命。 - **超强 CPU 变现**：Grace 等 CPU 产品线在 AI 集群协同效应下，一年内斩获 200 亿美元体量。 - **混合异构架构**：前端部署 Groq/Cerebras 芯片进行快速 decode，后端串联旧款 GPU，彻底打破了算力换代即报废的旧逻辑。 --- ### [01:19:44](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=4784s) 算力折旧神话破灭：英伟达如何“拯救”新型算力云 加文解释称，CoreWeave 等新兴云厂商（Neo Clouds）乃至主流超大规模云厂商此前一直饱受空头质疑。著名做空者迈克尔·伯里（Michael Burry）曾公开唱空，称这些公司将 GPU 和 CPU 按照 4 至 6 年进行折旧完全是会计造假，坚称 GPU 的真实物理与经济寿命只有 2 年，借此抨击整个行业的账面利润严重虚高。查马斯等人对此戏谑道，听伯里评价前沿 AI 硬件架构，“就像找过气说唱歌手讨论现代音乐一样不靠谱”。 英伟达本季度的产品路线与算力生命周期重塑，从根本上为新型算力云扫清了致命隐患。由于异构加速架构让老旧芯片在推理场景中保持了长效经济价值，CoreWeave 等企业能够以极低的资本成本为其 GPU 获得资产抵押贷款（Asset-backed financing），其借贷利率甚至已低至 6%。 杰森·卡拉卡尼斯（Jason Calacanis）分享了他采访 CoreWeave 首席执行官迈克尔·英特拉托（Michael Intrator）的细节。英特拉托表示，客户不仅毫无心理负担地签署长达 6 年的算力租赁合同并提前预付款，而且在合同期满后还在抢购进入第 7、第 8 乃至第 9 个年头的延役算力设备。真实的商业订单已经用真金白银彻底击碎了空头的短视逻辑。 --- ### [01:22:25](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=4945s) 宏观红灯频闪：通胀回潮与全球债市崩盘边缘 节目话题转向剧烈恶化的全球宏观经济环境。中东伊以局势持续紧绷，冲突已拖入第 12 周，远超原先预估的 4 至 6 周速决战。“战争从来不会迅速结束”，高企的原油价格正在全面推升通胀： - **通胀预期急剧恶化**：Polymarket 预测 5 月通胀率达到 4.2% 及以上的概率高达 99%；专业预测机构甚至普遍预期第二季度 CPI 将飙升至 6%，“降息梦”迅速演变为“美联储加息”恐慌。 - **主权债券收益率全面暴动**：美国 10 年期国债收益率攀升至 4.6%，大幅背离财政规划目标；英国国债收益率创全球金融危机以来新高；德国国债创 2011 年以来新高；日本 30 年期国债收益率更是触及历史纪录的 5.1%～5.2%。 - **亚洲投机情绪过热**：在韩国，散户投资者正通过大举借贷融资狂赌 AI 芯片概念股，杠杆规模打破历史纪录，呈现出类似加密货币牛市见顶时的危险过热信号。 --- ### [01:23:57](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5037s) “末日博士”弗里德伯格的过山车隐喻与套利交易危机 面对杰森调侃其“末日博士”（Dr. Doom）人格是否会再度上线，戴维·弗里德伯格（David Friedberg）给出了极为冷峻的宏观判断。他打比方说，当过山车已经被拉到了最高点并即将俯冲时，所谓的担忧已经毫无意义——重力是无法抗拒的，唯一的反应只能是举起双手高喊“Whee!”迎面承受冲撞。 弗里德伯格指出，当前全球债务与 GDP 的比例已达到触目惊心的 310%。即便把美国的全球储备货币地位撇开不谈，全球各级政府为了维系虚假的经济增长和偿付既有杠杆而疯狂借贷支出的问题，必然会引发连锁反应，系统最终必定崩裂。一旦防线被撕破，本币价值暴跌将带来剧烈的恶性通胀，政府将被迫通过疯狂印钞推高资产名义价格，从而陷入永无止境的恶性螺旋。 针对日本 30 年期国债收益率跃升至 5.2% 的极端异动，弗里德伯格向市场参与者发出严厉警示：这极有可能是点燃新一轮全球信贷危机的导火索。国际金融体系中沉淀着规模天文数字的日元套利交易（Carry Trade），一旦底层流动性与利差逻辑逆转，平仓风暴将对全球金融市场带来不可估量的毁灭性冲击。 ### [01:25:37](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5137s) 宏观闪烁红灯：查马斯极度集中的“五只股”投资哲学 当市场上各种宏观风险信号频频闪烁、被戏称为“末日博士（Dr. Doom）”的悲观论调再次出现时，Jason 调侃这究竟是“每月一次的恐慌发作”，还是第 17 次成功预测了接下来的 6 次衰退。面对这些闪烁的警报，Chamath 承认宏观红灯确实在亮起，但他认为市场中依然存在能够承保（underwrite）、真正代表未来的优质资产。如果投资者能找到几家这样的企业，能够安心持有 10 年，那就坚定买入并持有；除此以外的所有东西，无论牛市还是熊市，都应该尽量远离、拒绝投机。Chamath 直言，随着年龄增长，市场的剧烈波动在上涨时带给他的“糖分冲动（sugar high）”远不如从前，而在下跌时却会让人痛不欲生，这完全不值得。因此，他选择将资产高度集中在极少数几家自己真正深信不疑的公司上，埋头做好自己的本业（stick to my knitting）。 在谈及普通人能承受的持仓数量时，Chamath 强调自己的大脑无法同时跟踪 30 只股票，因为“没有那么多时间，也没有那么聪明，信息量实在太庞大了”。对他而言，能在夜里安稳睡着且维持长期持有的美股公开市场标的，极限就是“5 只甚至更少（five or less）”，并且其最大仓位和前两大重仓在个人净资产中占据了极高的比重。他认为与其在数十个标的间疲于奔命，不如将认知收缩在三四个核心标的上。 --- ### [01:27:39](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5259s) 加文·贝克的三重现实：利率、AI超高速增长与霍尔木兹海峡危机 与个人投资者的极端集中策略不同，对冲基金经理 Gavin Baker 经营着一个超过 30 人的专业投研团队，管理着上百个投资头寸。在他看来，当前的宏观与科技图景中，有“三件事同时成立（three things can be true）”： ```mermaid flowchart TD accTitle: 宏观与AI市场的三重现实 accDescr: 展示利率攀升、AI基本面爆发式增长以及地缘政治危机对美国相对优势的三重影响。 subgraph S1["三重现实并发"] A["1. 宏观逆风<br/>长端国债利率走高"] B["2. AI基本面爆发<br/>Anthropic超速增长且盈利<br/>英伟达估值仅十几倍PE"] C["3. 地缘地带收紧<br/>霍尔木兹海峡受阻<br/>全球能源危机"] end subgraph S2["经济与资产结果"] D["全球高通胀与债务困境"] E["生产力革命与企业强劲现金流"] F["美国相对受益<br/>能源独立与再工业化倒逼"] end A --> D B --> E C --> F ``` 首先，无风险利率与长端收益率（10年期与30年期国债）上升确实令人担忧，但回顾历史，互联网泡沫破裂时的基准利率远高于今天，且当年泡沫核心思科（Cisco）的前瞻市盈率高达 100 倍，而今天的算力中枢英伟达（Nvidia）在买方共识的真实盈利预期下，实际估值仅处于十多倍的“低至中十位数（low to mid teens）”。其次，AI 正在经历人类历史上前所未有的超大规模扩张，Anthropic 等头部公司的体量与复合增长速度甚至超越了上百个国家，更令人震惊的是它们已迅速实现了季度盈利。最后，地缘政治冲突导致霍尔木兹海峡航道受阻，虽然这对全世界都堪称灾难，但相对而言美国受到的冲击最小，甚至相对获益最多。 --- ### [01:30:38](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5438s) 能源优势、AI季节性与美国相对优势 Gavin Baker 进一步拆解指出，美国在当下动荡的国际局势中拥有极强的相对优势：粮食与能源完全自给自足，并且已经成为全球最大的石油与天然气生产国及出口国。在债务过高、危机四伏的全球“烂街区（bad neighborhood）”里，即便美元面临外部压力，由于德国联邦银行和德国马克时代早已过去、全球缺乏基本面更健康的替代货币，美国依然是“最好的一栋房子（best house）”。 - **工业电价与天然气脱钩**：电力是工业与制造业几乎所有环节的核心基础要素。美国发电主要依赖本土极其廉价且年内持续走低的天然气（NG1），而欧亚等全球其他地区严重依赖液化天然气（LNG），其成本已暴涨 100% 至 200%。 - **霍尔木兹海峡封锁与再工业化**：霍尔木兹海峡长期关闭对极度依赖中东原油的日本、中国、菲律宾、印度以及欧洲都是沉重打击，但这一外部压力恰恰成了类似疫情时期的强力倒逼机制（forcing function），有力推动了特朗普政府追求的美国本土再工业化目标。 - **AI 产品的季节性特征**：科技投资需要注意季节周期性（如“Sell in May and go away”）。以往由于大学放暑假以及夏季天气良好导致人们使用减少，ChatGPT 和 Claude 等 AI 工具在第二、第三季度往往会出现活跃度下滑；随着生成式 AI 深入企业级场景，这一季节性波动是否会重演仍有待观察。 --- ### [01:32:44](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5564s) 科技巨头随总统访华：表面握手与缺失的“重大协议” 节目尾声转向了刚刚结束的重磅外交行程：美国总统率领一众顶级科技公司 CEO 展开了为期 48 小时的访华之旅，并与中国最高领导人会晤。Jason 指出，虽然现场充满了击掌、握手和友好的外交姿态，但此前外界期待的能扭转关税政策、全面开放高端芯片出口（除少量 H100、A200 获准向百度等巨头供货外）、或中美联手应对霍尔木兹海峡危机等全局性战略突破（grand deal）完全没有出现，实质成果仅停留在采购大豆、部分农产品以及少量波音飞机等常规业务拓展（BD）层面。 David Friedberg 剖析认为，市场原本憧憬本次访问能达成一项为双边长期合作定调、从根本上缓和地缘紧张的大协议，但这一预期显然落空了。访问后发布的声明仅限于维持合作意愿与继续探索贸易协议的框架性表态。Friedberg 特别提醒注意访问结束后的国际政治表演：就在美方团队离开的同时，普京随即与中方最高领导人举行会晤，这直观地表明大国间的战略对抗与地缘博弈远未通过一次商务外交得到解决。 ### [01:34:29](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5669s) 中美关系、高层访问与台面下的地缘政治协议 嘉宾们深入探讨了中美之间复杂且漫长的博弈。随着中俄之间关系与信任纽带的持续加深，这场崛起中的大国对美国霸权构成的挑战绝不会迎来“彩虹般的童话结局”（rainbow-colored chapter three），而是一场长期的戏剧性历史大戏。谈及近期的高层访华行程，主持人和嘉宾们指出，虽然表面上看并没有达成某种轰动性的“全面大交易”，仅仅是卖了一些波音飞机、大豆，甚至可能涉及少量 Nvidia H100 芯片，但双方能够坐下来沟通本身就是巨大的积极信号。 表面上的新闻通稿往往掩盖了闭门会议中的真实博弈。Chamath 认为这次行程相当成功且极富成效，双方核心的成果在于对地缘政治“井字棋”下一阶段的落子走向达成了某种共识与分歧管控。这种大国博弈的本质在于划分地缘势力版图——“找到一种既能满足他们、也能满足我们利益的棋盘划分方式”，而这种真实的利益交换与默契，是绝不可能写在公开发布的新闻通稿里的。 - 表面经贸交易包括大豆、飞机以及少量芯片出口，但核心价值在于闭门会谈中的地缘政治对表。 - 中美两强相争是长期历史趋势，双方高层沟通是避免直接冲突失控的关键机制。 --- ### [01:36:04](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5764s) 芯片出口争议：倾销旧款 GPU 是保持美国 AI 统治力的战略工具？ 关于是否应允许 Nvidia 向中国销售芯片，Gavin Baker 提出了深刻且反直觉的行业战略思考。他主张美国应当向中国出售已折旧、非顶尖的旧款 Nvidia GPU。他分析指出，如果完全封锁所有芯片，中国将被迫全力自研替代性的计算生态，那将加速其采用光互连技术搭建高能耗的本土集群；相反，若向其倾销上一代旧款 GPU，中国构建自研独立生态系统的动力和几率都会大幅下降。 Gavin 强调，将旧架构芯片持续卖入中国，使其锁定在基于 CUDA 的软硬件生态中，恰恰是全球地缘稳定的基石，也是“确保美国在 AI 领域持续领先并掌控 AI 主导权概率最高的一条道路”。Chamath 随即插话澄清，自己在这个问题上完全支持 Gavin 的逻辑，甚至认为“把所有能卖的芯片都卖给他们”才是上策。 ```mermaid flowchart TD accTitle: Nvidia GPU Export Strategy Logic accDescr: 比较全面封锁与倾销旧款GPU对中国自研AI生态的影响路径 subgraph Block["全面封锁策略 (全面断供)"] A["全面切断先进与旧款GPU"] --> B["倒逼中国全力突破自主生态"] B --> C["加速光学互连与替代架构演进"] C --> D["彻底摆脱对美依赖并形成抗衡"] end subgraph Export["选择性倾销策略 (Gavin Baker 观点)"] E["出售已折旧旧款GPU (如H100)"] --> F["降低中国自研替代生态的迫切性"] F --> G["深度锁定在CUDA软硬件生态中"] G --> H["保持美国对AI底层技术的绝对控制"] end ``` --- ### [01:37:39](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5859s) 特朗普的外交风格、修昔底德陷阱与能源封锁战略 讨论随后转向中美如何规避“修昔底德陷阱”（Thucydides's Trap）。嘉宾们指出，中国对这一历史宿命论不仅极度清醒，而且经常公开主动提及，而高层持续对话正是打破这一历史诅咒的基石。在谈到特朗普的外交特色时，嘉宾们打趣称，特朗普最大的“超级超能力”就是他能够与各路强人、君主、海湾王室成员建立极佳的私人关系并产生“化学反应”。特朗普不仅毫不介意亲自登门拜访，甚至会邀请他们去现场观看 UFC 格斗赛——如果未来中国领导人访美并与达纳·怀特（Dana White）并肩坐在八角笼边观赛，那就证明两国关系彻底稳住了。 在中美战略筹码的深层推演中，嘉宾们指出能源才是美国遏制冲突最致命的底牌。现代战争是吞噬海量石油的无底洞，而中国主要从伊朗、委内瑞拉和俄罗斯三地进口石油。如果台海或印太局势失控，美国及其盟友可在棋盘上迅速剥离这些供应： - 委内瑞拉石油、美国石油及巴西石油将立刻切断供应。 - 通过海湾合作委员会（GCC）盟友重新封锁霍尔木兹海峡，中东及伊朗的石油供给将被彻底清零。 - 仅凭俄罗斯一国的管道与运力，根本无法填补战争状态下巨大的能源缺口。 - 面对美、日、韩、澳等盟友在西太平洋形成的联合战线，仅靠俄国石油打赢长期全面战争在物理上是不可能的。 嘉宾们总结认为，随着中东及伊朗局势在未来的逐步演变，全球地缘格局有望走向一种基于实力制衡的全新稳定态。 --- ### [01:40:00](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=6000s) 现场趣谈与节目收尾 在节目的尾声，主持人团队对因故缺席的“雨人”David Sacks 表达了思念，并热烈感谢了特邀嘉宾 Gavin Baker 的深度加盟。随后，主持人揭秘了一段幕后爆料：Gavin 的岳父 Jeff Painter 是《All-In Podcast》的忠实铁杆粉丝。原本 Gavin 当天因为日程冲突一度推辞录制，主持人随即使出了“杀手锏”，直接对 Gavin 施压：“告诉你岳父，只要你今天出席，我们就送他接下来两届峰会（Summit 和 Liquidity）的后台 VIP 门票；要是你不来，我就私底下向他告状！” 这招软硬兼施的“场外施压”立竿见影，Gavin 当即调整日程赶来录制，并笑称自己“为了岳父甘愿赴汤蹈火”。最终，节目在轻松调侃、经典的片尾混音与听众二创歌曲声中圆满落幕。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=0s) The spicy takes must flow — Jason Calacanis > [13:21](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=801s) inside of the tent pissing out — Chamath Palihapitiya > [13:21](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=801s) 不要再去屏息倾听那些大模型制造商的夸夸其谈了。去走到最前线的终端用户身边，去问问工厂里正在使用模型的工人怎么想，去问问医生和科学家的看法，去讲述那些真实的落地故事，这才是我们真正该讨论的东西。 — Shyam Sankar > [34:01](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2041s) 如今，犯罪已经变成了一种选择（Crime is now a choice）。 — Gavin Baker > [38:46](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2326s) 你现在唯一能指望的，就是在岗位上再苟延残喘一段时间，亲自把替代你的 AI 训练出来，然后被扫地出门。 — Jason Calacanis > [40:23](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2423s) 闭上你们的臭嘴，滚回键盘后面老老实实干活！如果需要裁员管理就低调处理，别再写这种自我感动的公开信了，你们在这方面烂透了。 — Chamath Palihapitiya > [58:30](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=3510s) 给生物圈做备份（back up the biosphere） — Elon Musk > [01:16:58](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=4618s) 拿去测一定会输 — Gavin Baker > [01:19:44](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=4784s) 像找过气说唱歌手讨论现代音乐一样不靠谱 — Chamath Palihapitiya > [01:25:37](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5137s) stick to my knitting — Chamath Palihapitiya > [01:27:39](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5259s) three things can be true — Gavin Baker > [01:36:04](https://www.youtube.com/watch?v=HGbA6ze0_3M&t=5764s) 确保美国在 AI 领域持续领先并掌控 AI 主导权概率最高的一条道路 — Gavin Baker ## Mentioned Andrej Karpathy, Anthropic, Claude, Claude Code, AutoResearch, Karpathy Skills, Gavin Baker, Atreides Management, Jason Calacanis, Chamath Palihapitiya, David Friedberg, David Sacks, OpenAI, Tesla, Autopilot, FSD, Cybercab, Richard Sutton, The Bitter Lesson, Jeff Dean, Amit Singhal, Sridhar Ramaswamy, Google, Gemini, Gemini Nano, Chrome, Meta, Mark Zuckerberg, Amazon, Cursor, xAI, Grok, Grok Build, Composer 2.5, Kimmy K2.5, Abacus.AI, Dario Amodei, Palantir, Shyam Sankar, Eric Schmidt, Nikesh Arora, Palo Alto Networks, Donald Trump, Optimus, Figure, Flock Safety, Ben Horowitz, Felicia Horowitz, Cloudflare, Matthew Prince, Tucker Carlson, Sam Altman, SpaceX, Starlink, Starship, Falcon Heavy, Colossus, Colossus 2, Colossus 3, Colossus 4, Saudi Aramco, Polymarket, Nvidia, Jensen Huang, H100, GB300, Broadcom, TPU, Trainium, Inferentia, Leopold Aschenbrenner, Groq, Cerebras, CoreWeave, Michael Intrator, Michael Burry, Excite Labs, Seraphim Space, Blue Origin, Baidu, Boeing, Vladimir Putin, Dana White, Jeff Painter ## Video description (0:00) Gavin Baker joins the show! (0:30) Andrej Karpathy joins Anthropic; hypergrowth and profitability (12:42) Why Americans have turned on AI, anti-human perception (27:22) Trump pulls AI EO, US-China AI relationship, dystopian AI layoffs (45:19) SpaceX S-1 tear down! Breaking down the three major businesses and the case for a $2T valuation (1:11:22) Nvidia smashes earnings but stock falls, why people are shorting chips (1:22:25) Market update: Flashing red signals, oil, inflation, yields up (1:32:45) China trip flops, or was progress made behind the scenes? Follow Gavin Baker: https://x.com/GavinSBaker 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 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://www.cnbc.com/2026/05/19/anthropic-hires-openai-cofounder-andrej-karpathy-former-tesla-ai-lead.html https://github.com/karpathy/autoresearch https://github.com/multica-ai/andrej-karpathy-skills/stargazers https://techcrunch.com/2026/05/19/openai-co-founder-andrej-karpathy-joins-anthropics-pre-training-team https://www.wsj.com/tech/ai/mind-blowing-growth-is-about-to-propel-anthropic-into-its-first-profitable-quarter-7edbf2f4 https://x.com/i/broadcasts/1dxYljYVREYJX https://apnews.com/article/trump-ai-executive-order-ee318f35acc8a2c43e47f3ebf26cb459 https://x.com/wallstengine/status/2057378437485216031 https://x.com/MorePerfectUS/status/2056842597117636890 https://x.com/lulumeservey/status/2057239284487201043 https://polymarket.com/event/spacex-ipo-closing-market-cap-above https://x.com/elonmusk/status/2057228707606196434 https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm https://s201.q4cdn.com/141608511/files/doc_financials/2027/Q127/NVDA-F1Q27-Quarterly-Presentation-FINAL.pdf https://www.ibtimes.co.uk/leopold-aschenbrenner-investment-shift-agi-over-ai-chips-1797606 https://polymarket.com/event/may-inflation-us-annual https://www.cnbc.com/2026/05/15/inflation-rate-projected-to-hit-6percent-in-the-second-quarter-top-economic-forecasters-say.html https://polymarket.com/event/fed-rate-hike-in-2026 https://www.cnbc.com/2026/05/18/treasury-yields-inflation-bond-rout-oil.html https://www.cnbc.com/quotes/US10Y #allin #tech #news [Source](https://www.youtube.com/watch?v=HGbA6ze0_3M)

[0:00] Jason Calacanis: The spicy takes must flow [13:21] Chamath Palihapitiya: inside of the tent pissing out [13:21] Shyam Sankar: 不要再去屏息倾听那些大模型制造商的夸夸其谈了。去走到最前线的终端用户身边，去问问工厂里正在使用模型的工人怎么想，去问问医生和科学家的看法，去讲述那些真实的落地故事，这才是我们真正该讨论的东西。 [34:01] Gavin Baker: 如今，犯罪已经变成了一种选择（Crime is now a choice）。 [38:46] Jason Calacanis: 你现在唯一能指望的，就是在岗位上再苟延残喘一段时间，亲自把替代你的 AI 训练出来，然后被扫地出门。 [40:23] Chamath Palihapitiya: 闭上你们的臭嘴，滚回键盘后面老老实实干活！如果需要裁员管理就低调处理，别再写这种自我感动的公开信了，你们在这方面烂透了。 [58:30] Elon Musk: 给生物圈做备份（back up the biosphere） [76:58] Gavin Baker: 拿去测一定会输 [79:44] Chamath Palihapitiya: 像找过气说唱歌手讨论现代音乐一样不靠谱 [85:37] Chamath Palihapitiya: stick to my knitting [87:39] Gavin Baker: three things can be true [96:04] Gavin Baker: 确保美国在 AI 领域持续领先并掌控 AI 主导权概率最高的一条道路

[16:34] Gavin Baker: That's just one person—one dad who was unwilling to accept defeat for his daughter, and who changed her life and the life of everyone else with that disease. [25:35] Chamath Palihapitiya: If it was such a great job, I suspect the churn would be 3% or 4%. So, what exactly is it that we want to protect? And have you asked them? [37:16] Shyam Sankar: We're listening too much to the inventors of AI. [...] We need to be listening to the frontline factory workers who are using AI... or the ICU nurse who says, 'I have more time to spend with my patients.' [45:18] Jason Calacanis: This would be the largest IPO ever by more than double Saudi's Aramco $29 billion IPO a couple years ago. [71:18] Gavin Baker: The AI market is cross-sectionally inefficient right now... If the multiples on the power, cooling, optical names are correct, Nvidia and memory are going up a lot. If the multiples on Nvidia and memory are correct, everything else is probably going to underperform. [83:57] David Friedberg: What is the point of being concerned when you have ridden the roller coaster to the top and it is beginning its descent?... The force of gravity is inevitable.

[0:00] All right, everybody. Welcome back to [0:01] the number one podcast in the world. [0:03] It's the All-In podcast episode 274. [0:07] Sacks is out today, but we're very lucky [0:10] to have Gavin Baker from a Treaties [0:13] Management joining us. The spicy takes [0:15] must flow. [0:17] Welcome back to the program, bestie [0:19] Gavin. [0:19] >> Thanks for having me. Always love it. [0:21] >> It's been a huge week in tech. We can [0:24] start with the SpaceX and OpenAI IPOs. [0:26] We've got Andrej Karpathy joining [0:28] Anthropic, Nvidia crushing it. So many [0:30] different to go, but I think we'll start [0:32] with Andrej Karpathy joining Anthropic. [0:35] Karpathy is only 39 years old. He's [0:38] already a legend in the tech industry, [0:39] if you don't know him. I believe he's [0:41] also coming to Liquidity, Chamath. [0:42] >> He's going to keynote on Monday morning. [0:44] >> Oh, fantastic. [0:45] >> I did not know. Tuesday, Tuesday, [0:46] Tuesday, day two. I think he's [0:47] keynoting. Okay. [0:49] >> As is Gavin. Gavin will be there. [0:51] Founding member [0:52] >> Gavin anchoring day two. [0:54] >> Excellent. Yeah, and this is Gavin's [0:55] second appearance at the event. [0:56] >> I look at those two bookmarks, Andrej [0:58] Karpathy and Gavin Baker. [0:59] >> Yeah, you know. [1:01] >> Liquidity pulls in the stars. [1:03] >> Obviously, Andrej was a founding member [1:06] of OpenAI. He led the self-driving team. [1:08] >> Also, hold on. Gavin is going to help us [1:10] judge the best ideas section as well. [1:13] >> Excellent. I don't know if you know [1:14] that, Gavin, but you're a judge. You're [1:15] going to be a judge. [1:16] >> I'm a pretty thing, man. I'm easy. [1:18] >> Yes. Karpathy also coined the term vibe [1:20] coding. He recently built Auto Research. [1:22] I think we talked about that here a bit. [1:24] That's an open-source training tool. It [1:25] helps AI models improve themselves by [1:28] running 5-minute experiments. That got [1:29] over 82,000 stars on GitHub. He did that [1:32] like as a weekend experiment, and all [1:34] these civilians started building their [1:36] own uh recursive LLMs. Really inspiring. [1:40] And the Andrej Karpathy skills is a tool [1:43] based on his set of principles for [1:45] Claude code, and somebody just released [1:48] that. And so, it's just pretty crazy [1:51] when you think about it. He's going to [1:52] be in charge of a new pre-training team [1:54] at Anthropic. The focus obviously being [1:56] recursive self-improvement. In other [1:58] words, [1:59] they're going to have Claude improve [2:01] itself and they've already talked a [2:02] little bit about AI improving AI over at [2:06] Anthropic. Chamath, what's your take on [2:08] this? Is [2:09] this super important in 2026? Obviously, [2:12] Karpathy is super well respected. He's [2:15] obviously, you know, one of the true [2:17] talents in the space, but hey, we're in [2:19] a we're in a different inning than we [2:20] were say 10 years ago when he was at [2:22] Tesla or 5 years ago when he co-founded [2:25] OpenAI. [2:25] >> You know what's interesting? If you go [2:26] back to like Google, [2:28] the culture of Google which they got [2:30] right was [2:31] the singular technical talents there. [2:35] They were singled out and they were [2:36] called Google fellows. I don't know if [2:37] you guys remember this like Amit [2:39] Singhal, [2:40] Shreedhar Ramaswamy, Jeff Dean. [2:44] These guys are stars. And what's [2:46] interesting is if you track what folks, [2:48] particularly Jeff Dean I guess now [2:50] because the other two aren't there [2:51] anymore, but [2:53] what they did inside of Google, it's [2:55] like wave upon wave, they were at the [2:56] foot of those waves. What's interesting [2:58] about Andrej, he's been at the wave upon [3:00] wave of AI. He was probably the first [3:03] person that really commercialized the [3:06] Richard Sutton bitter lesson essay [3:10] when he was leading FSD at Tesla, [3:13] which was really about the brute force [3:15] computation. And I remember him telling [3:17] me the story. I don't know if he said [3:18] this publicly or not, but where he spent [3:21] a portion of his time, I want to say a [3:22] quarter of his time labeling data. Could [3:24] you imagine like 2016, 17 like hand [3:27] labeling video data from Teslas? [3:31] So he did that, then he's co-founder of [3:32] OpenAI. He's a star. [3:34] And he's an exceptional human being and [3:36] he's super curious. And [3:38] then what he's done as a kind of a free [3:40] agent [3:41] is also quite impressive. So I think [3:44] that this is a really important deal. I [3:46] think he's one of these [3:48] really curious people that can be sent [3:50] off and they'll just go and [3:52] invent new things. And I think this idea [3:55] of recursive self-learning puts these [3:57] models on [3:59] a combination of overdrive and [4:02] autopilot. And so if you put those two [4:04] things together, I think that you start [4:06] to [4:08] you can potentially live out this idea [4:10] that there's an order of magnitude [4:12] improvement on a yearly basis. So like [4:14] this new form of Moore's law. [4:17] Well, so then the model quality just [4:18] goes absolutely parabolic. Just like [4:20] this. Straight up. [4:21] >> And I think that a bunch of compute at [4:23] the problem and these things learn [4:25] really quick. [4:26] I think is the high-order bit there. [4:29] Gavin, what do you What's your take on [4:31] Anthropic's recent success and their [4:33] massive hiring binge? [4:34] >> The success is extraordinary. It's [4:37] undeniable. I think the fact that they [4:40] are now they were even positive per the [4:42] Wall Street Journal in the most recent [4:44] quarter [4:45] is a really important fact for kind of [4:48] the whole AI narrative. Because now [4:50] there's [4:52] you know, you could talk about circular [4:53] funding, you could talk about ROI, and [4:56] we could go look at the ROIC of the [4:57] hyperscalers. [4:59] But if OpenAI and Anthropic are at call [5:03] it $100 billion of ARR now with 80%ish [5:08] gross margins on inference, like the [5:10] returns are there. And then if we add in [5:12] and they're growing really fast, if we [5:14] add in Gemini, we add in Cursor, we add [5:16] in xAI, we add in open source, [5:19] you know, it's it's not hard to see 200, [5:23] 300, 400 billion dollars of ARR at the [5:26] end of this year at high [clears throat] [5:28] market [5:28] >> Across all of the Across all of the [5:30] language models. And you're talking [5:31] specifically about the private language [5:33] model companies, maybe not Google, which [5:35] is obviously public. You're including [5:37] Google. Okay. [5:38] >> but I was excluding, you know, a lot of [5:40] the returns to this GPU spend have come [5:42] from you know, better recommender [5:44] systems at Facebook and Google, [5:48] Amazon, better ad targeting, better ad [5:50] measurement. [5:51] >> Sure. [5:52] >> So, excluding that and just narrowing it [5:54] to LLMs, which I think [5:56] is the [5:57] possible definition, [5:59] and it seems like there's going to be a [6:01] really strong ROI this year, even [6:04] excluding what are still some of the [6:06] most economically important and [6:07] profitable use cases for GPUs and AI [6:10] infrastructure. [6:12] I do think what Karpathy is working on, [6:13] recursive [6:15] self-improvement, is really important [6:18] and unlocking that and continual [6:20] learning, [6:21] you know, maybe the two final frontiers [6:24] for for AI. [6:26] And just the idea of recursive [6:28] self-improvement that the model, [6:30] while it is training, you know, during a [6:33] forward pass, [6:36] has input into its training, or another [6:39] model has input into the training, [6:42] I think that could be really powerful [6:44] and I think [6:45] Chamath's statistics of [6:48] of, you know, 10Xing every year, [6:52] you know, might seem conservative if [6:53] that comes to pass. And then, of course, [6:55] continual learning is the holy grail, [6:58] where the model learns from experiences [7:01] the way humans do. Yeah. And that's [7:03] something we haven't unlocked yet. [7:06] And those those those two combined, I [7:08] think would um [7:10] they might pull the future forward in a [7:12] very real way. [7:13] >> Yeah, and we have right now Anthropic [7:14] has a decent lead on everybody else, [7:17] whether it's 3 months or 6 months. [7:19] Obviously, they're probably 6-12 months [7:20] ahead of open source. Maybe they're [7:21] 3-6-9 months ahead of their [7:23] contemporaries, but they have a lead. [7:25] You put Karpathy in there, Friedberg, [7:27] now you have Karpathy, he does recursive [7:31] and at some point, and it may have even [7:33] occurred at Anthropic, the AI is going [7:36] to be improving [7:38] the language model more [7:41] than [7:43] the humans in the loop are doing it. [7:45] Obviously, they're orchestrating it, [7:46] Friedberg, but at what point do we think [7:49] this, let's call it, super recursiveness [7:52] occurs? When will we cross the recursive [7:55] valley [7:56] and AI is doing more to build a language [7:58] model than humans are? [8:02] >> I'm not sure when this idea that you [8:05] feed the whole model into a context [8:06] window to train itself and build a new [8:09] model is going to happen, but I think [8:11] there's probably a lot of different [8:12] architectural paths that could be walked [8:14] here. [8:15] One of which is this idea that you could [8:17] make much smaller models and then create [8:19] networks of smaller models that work [8:21] together where you ultimately have [8:24] less energy or less cost per token [8:25] produced out of a [8:27] aggregation of models than you did with [8:29] one single large model. I've said this [8:32] probably three or four times now. [8:33] There's a lot of work and a lot of [8:35] opportunity ahead [8:37] in kind of re-architecting models and [8:39] re-architecting how models work together [8:42] to solve problems. My guess is a lot of [8:44] leadership that that he can bring [8:47] to exploring those paths. And all it [8:49] takes is a minor breakthrough and your [8:51] cost per token drops in half. That's a [8:52] tremendous efficiency gain [8:54] that is seems very much on the horizon [8:56] because some of the early papers, I [8:58] think I shared one from MIT a few weeks [8:59] ago, indicate that there's a lot of room [9:02] to run here. [9:03] >> In terms of re-architecting models and [9:05] deployment of models. [9:06] >> These very small models, [9:10] uh, small language models and then [9:11] verticalized ones are the future. We've [9:14] got a company Abacus that's doing it for [9:16] corporations. Crushing it. Everybody's [9:18] got an interest in doing this. And I [9:20] don't know if you saw the news this [9:21] week, uh, Chamath, it happened about 2 [9:23] weeks ago very quietly. Chrome included [9:27] Gemini or Google included in their [9:29] Chrome browser the Gemini Nano model [9:32] without telling anybody. 4 GB on your [9:34] computer. And that's the one that does [9:37] like proof reading, spelling, auto [9:38] complete, all that. So, now we have [9:41] Google [9:42] covertly installing this on everybody's [9:45] operating system. [9:46] >> on. Hold on. Covert is a strong word, so [9:48] let's not use that word. [9:49] >> Without telling people, without giving [9:51] people a heads-up. Let's say it in a way [9:53] that we can both agree. [9:55] >> We're in the phase now where [9:58] I think breathlessly talking about every [10:00] model improvement is a waste of time. [10:02] There's no ROI in it. [10:04] We are on a path of accelerated [10:06] learning, and we're going to start to [10:08] see end user achievements that were [10:12] heretofore impossible. That should be [10:14] the focus. So, for example, [10:17] we were able to solve, I'm just [10:18] collectively saying we, in this case it [10:20] was specifically [10:22] OpenAI. [10:24] By the use of a human, and this is [10:26] important, [10:27] you know, a math problem that stood [10:28] outstanding, not been solved for decades [10:31] and decades. [10:33] I can tell you in a different example, [10:35] there are drug candidates that [10:37] are about to enter clinical trials and [10:38] INDs that were sitting on the shelf and [10:41] people didn't think were very viable at [10:43] all. [10:45] So, we're at the phase now where these [10:46] things are front and center, they're [10:49] useful to people, they're increasingly [10:50] valuable. [10:52] I think what we should do now is focus [10:54] on these end user [10:56] use cases [10:57] because the way that you say it, in my [11:00] opinion, is part of the problem because [11:02] it starts to create this boogeyman us [11:04] versus them thing, and I'm not saying [11:06] you're doing it on purpose, but I'm [11:07] saying this is exactly why I think [11:10] so many people are becoming sort of like [11:13] it's a four-letter word now when you [11:14] mention AI because it's presented as [11:17] this thing, and I think we have to [11:18] present the other side of it, at least [11:19] so that people have the data. So, I [11:21] don't think Google's in the business of [11:23] doing shady or shady things. I don't [11:26] They're not that company. There are [11:27] other companies that would. [11:29] >> Meta. [11:30] >> [laughter] [11:30] >> You're referring to your alma mater? [11:32] >> I'm not going to say which ones, okay? [11:34] But Google is not that company. So, I I [11:37] think that the reason they did it was [11:38] probably because there's user utility. [11:39] And my point is we should focus on the [11:41] user utility because I think that's the [11:43] story we're telling from now on because [11:44] I think we, collectively, the four of [11:46] us, [11:47] can responsibly tell both sides of the [11:49] story in a well-balanced way because I [11:51] think nobody wins if we become Luddites [11:54] and go back in time. [11:55] >> Yeah. [11:55] >> And I think that if we don't if we're [11:57] not careful with our words, that's what [11:59] will happen. [11:59] >> Yeah, and by the way, [12:02] obviously not a Luddite, but this is [12:04] what [12:05] you know, has been reported by a lot of [12:06] folks that people were [12:10] surprised, shocked when they saw the [12:12] size of the model being done in the [12:14] background, and it has triggered some [12:17] people looking at it around privacy, and [12:19] I do agree that Google is not a bad [12:23] actor in the space. So, probably a speed [12:25] hour more than anything, I would say. [12:27] >> May I just add two things? [12:30] Um [12:30] there's attorney maxing, and then [12:32] there's attorney maxed. [12:34] And Google is probably [12:35] >> [laughter] [12:36] >> attorney maxed. [12:37] >> Yes. [12:38] >> Yeah. [12:38] >> It happened for a long time. And the [12:40] second second thing I would say is I do [12:42] think it's incumbent on all of us as [12:44] Americans who are involved in the [12:46] technology industry in one way or [12:47] another [12:48] to be advocates for a the positive, [12:51] optimistic possibilities [12:54] that AI introduces to to everyone in [12:56] this world [12:58] because it it is starting to feel or [13:00] seem [13:02] like there may be a CCP-funded [13:05] campaign against AI and data centers in [13:09] America. [13:10] And that's very logical for China, but [13:13] it is not good for America. [13:15] And so, I just I think it's [13:19] it we all have responsibility, is what I [13:21] would say. [13:21] >> Yeah. Who do you think's doing [13:24] a poor job at that and responsible for [13:25] this? Is it Dario with his constant hey, [13:29] everybody's going to lose their job? [13:31] Who's responsible for this? Is it the [13:33] CEOs blaming AI for their layoffs? [13:36] What's your take on this, Gavin? Look at [13:38] Hold on a second. Everybody is [13:40] trading their own book. [13:42] It makes enormous sense for Dario to try [13:45] to create the boundary conditions for a [13:47] regulatory moat because he will be [13:50] inside of the tent pissing out. [13:52] He's big enough now. [13:54] And if you notice that [13:56] a lot of the breathlessness has ramped [13:58] up. [13:59] And Jason, we've talked about this. You [14:01] can annotate successive rounds of [14:03] fundraising and successive scale with [14:06] the volume. [14:07] So, I think that it's a reasonable [14:09] business strategy and I think that he's [14:11] quite clever. And I think that look, if [14:13] you actually and I and I do this, if you [14:15] actually just have an ash bot, an ash [14:17] agent inside of Quad and you ask it what [14:19] it would do, it would come up with this [14:20] strategy. And meanwhile, there are other [14:22] versions of other counter strategies and [14:25] counter exploitative strategies. The [14:27] point is that each CEO has a clear [14:29] incentive. They're operating at such a [14:31] level of scale [14:33] that they're they're just reading their [14:35] own book. So, it's it's up to us to take [14:38] a step back and actually see the forest [14:40] from the trees. I think Nick, can you [14:42] find this? There was a clip of Shyam [14:44] Sankar, friend of the pod, fabulous guy, [14:46] the CTO of Palantir and he was I think [14:48] he was on Fox News [14:50] and he said, "Stop breathlessly asking [14:53] these model makers what they think. [14:56] Go to the end user and ask the person in [14:58] the factory that's using the model and [15:00] ask him what he or she thinks. Ask what [15:02] the doctor thinks. Ask what the [15:04] scientist think and start to tell those [15:06] stories. That's what we should be [15:07] talking about." [15:08] >> Yeah, and Gavin, you were [15:10] I was sort of asking you your opinion on [15:12] what's what's who's causing this and [15:15] then what's the solution? Like do you do [15:17] you have folks you think in the industry [15:20] who are representing it particularly [15:21] well we can point out, "Hey, Elon has [15:24] said we're going to move to a world of [15:26] incredible abundance and working will be [15:28] optional." I think that's on the margin [15:30] a little scary for people to hear [15:31] because they hear no job. [15:33] But, he does say, "Hey, universal basic [15:35] income is probably going to have to come [15:36] into, you know, place." And he said that [15:39] multiple times. You have Chamath, [15:41] according to Chamath, talking his own [15:43] book, scaring the bejesus out of people [15:45] in order to get regulatory capture. [15:48] What do you think is going on here? And [15:50] how can we do better as an industry? [15:52] >> I think Chamath outlined like a very [15:56] viable and positive path forward. We're [15:58] just, you know, real people who are not, [16:01] you know, at the tip of the spear. [16:04] These are the positive impacts AI's had [16:06] on my life. I was at a [16:08] I was at I was at an event maybe 10 days [16:10] ago [16:11] and someone who runs a hedge fund, his [16:14] daughter was born with a very rare [16:16] genetic mutation [16:18] that effectively would have [16:20] normally condemned her to a life devoid [16:22] of joy, meaning everything. [16:26] The neurons in her brain were not [16:27] firing. So, she wouldn't [16:30] you know, who knows what her life [16:31] expectancy would have been or what the [16:33] quality of her life would would have [16:34] been. And it's it's a tragic disease. [16:38] He said he didn't accept that as an [16:40] answer. He found He did an enormous [16:43] amount of research with LLMs and found [16:45] an existing safe drug on the market that [16:48] they thought would have a meaningful [16:50] impact on his daughter's condition. [16:53] And it did. [16:55] It took I think the percentage of times [16:56] the neurons were firing was 30 or 40%. [16:59] And it took it up to 80 or 90%. And that [17:02] means that she can live a normal life. [17:05] She may not be as smart as she would [17:06] have been, but she will live a normal [17:08] life. [17:09] And he's now figured out how to use AI, [17:13] how to to further tailor that drug. And [17:15] you know, there've been all sorts of [17:17] advances in [17:18] in protein design, etc., etc. [17:21] And he's reasonably confident he's going [17:23] to have a drug in months [17:25] that is a complete cure. [17:27] And that's just one person [17:30] one dad who was unwilling to accept [17:32] defeat for his daughter [17:34] and who changed her life and the life of [17:36] everyone else with that disease. And we [17:38] tell those stories. So, I think Elon's [17:40] doing a good job. A future where work is [17:42] optional, I think that sounds great to [17:44] some people and [17:45] you know, scary to others. You know, [17:47] four-day workweek, you know, I think [17:49] it's is probably something that sounds [17:50] sounds good to a lot of people. I think [17:52] Jensen is doing a good job of being an [17:54] effective advocate. And I do think [17:57] anyone who's trying to drive reg- [18:01] I just we need to stay focused on the [18:03] positives. That's what I'd say. [18:04] >> Yeah. [18:05] Freeberg, [18:06] what's your take here on the AI [18:10] PR crisis, if we'll call it that. Uh we [18:13] had three different [18:15] commencement speeches that were booed. [18:18] Eric Schmidt being one of them, two [18:20] other ones by maybe less notable folks. [18:23] When you hear young people booing AI [18:26] vociferously, why are they doing that, [18:28] Freeberg, and what's your take on the [18:30] overall PR problem and how to turn it [18:32] around? [18:35] >> Uh that's a [18:38] There's a long answer to that question. [18:39] Um [18:42] It relates in some ways to your concerns [18:44] about socialism and polarization. [18:46] >> answer? [18:47] >> Yeah. What's the long answer? [18:48] >> I mean, that's like [18:50] like why do people hate technology? [18:54] The greatest technological [18:56] >> Why do they hate this technology? They [18:58] love their phones. They love the [18:59] internet. This technology they hate. [19:02] >> Mhm. [19:04] I think that there's like an underlying [19:07] view that technology creates leverage [19:09] for a small group of people [19:11] which creates power imbalances and [19:15] nothing represents that more than AI. [19:18] That a small number of people that [19:21] control and profit from and benefit from [19:25] AI [19:26] are going to end up getting outsized [19:28] returns [19:29] relative to the broader population. That [19:32] the time to diffusion of the technology [19:35] cuz ultimately all technologies like [19:37] commoditize and diffuse. [19:39] But the time to diffusion here is such [19:41] that it's going to be like extremely [19:44] asymmetric for society. And I think that [19:47] there is something fundamental about [19:48] that. [19:49] It's like, you know, nuclear bombs I [19:51] think really created this this moment in [19:53] people's minds [19:54] in the mid-20th century that by the back [19:56] half of the 20th century gave everyone a [19:58] high degree of skepticism about [19:59] technology and science generally. [20:02] That those who have the knowledge and [20:04] those who engineer solutions with the [20:06] knowledge can create outsized advantages [20:09] for themselves and it puts the rest of [20:10] us at risk, the rest of the world, the [20:12] rest of the population at risk. [20:14] And because those questions about when [20:16] does this benefit me, how does it [20:18] benefit me can't be answered today [20:21] the economic benefit that's accruing to [20:23] the few today becomes the narrative, it [20:26] becomes the story and it becomes this [20:27] like power system [20:29] that a few people take from the many. [20:32] And so there's something deeply [20:34] disturbing for the average person about [20:36] that. They don't understand how it [20:38] works, why it works, what it'll do for [20:39] them, when it will do it, and all that [20:41] they're being told is that some people [20:42] are making trillions of dollars. [20:44] So, I think that it's pretty obvious why [20:46] this has got such a backlash. Secondly, [20:49] I think that there's a deep amount of [20:51] external energy that's fueling this [20:53] anti-technology sentiment in the United [20:54] States and has been for decades. [20:56] I think to Gavin's point, I don't think [20:58] it's just China with NGOs today. I think [21:00] that there is a long history [21:03] of [21:04] state actors intervening in media [21:07] activities in foreign nations to try and [21:10] create the sentiment [21:12] and fuel a sentiment that reduces [21:15] progress in that competitive state. [21:18] I think this goes all the way back to [21:19] KGB design [21:21] during the Cold War and it's been [21:24] refined and honed and improved over [21:26] time. This is not just some conspiracy [21:27] theory. There are plenty of great books [21:29] about this. The techniques of what's [21:30] going on specifically today, I don't [21:33] know enough. I don't have any great [21:34] details on that. But I don't think that [21:36] there's no foreign interest in seeing [21:39] technology advancement slow [21:42] in competitive nations. The United [21:43] States probably does similar things to [21:45] other nations and I think that that's [21:47] probably a key part of this. [21:50] And then I think this this like third [21:51] piece is like when [21:54] the Copernican revolution happened, it [21:56] was a mind, you know, like [21:58] heliocentricity was a totally new way of [22:00] thinking for humans and it was uh deeply [22:03] disruptive to the church. [22:05] And it was deeply disruptive to the [22:06] power centers, which were [22:08] the centers that could tell people Earth [22:11] is at the center of the universe, we're [22:13] in control, we're the direct channel to [22:14] God. [22:15] And the idea that the sun is at the [22:17] center of the solar system and we spin [22:18] around it and we're a tiny speck in the [22:20] universe was very hard for people to [22:22] grasp. There's something about AI that's [22:24] very like not human-centric and it kind [22:27] of shifts [22:29] and with the ego of the human. It [22:31] it's it's almost anti-humanist. [22:33] And I think that that's like a deep [22:35] psychological current for a lot of [22:37] people and their disdain for this [22:38] technology. It fuels it. It's not the [22:41] cause, but I think it fuels it. So I [22:43] think there's a lot of complicated [22:44] aspects to this, J Cal. You know, I [22:45] don't think there's like a simple put [22:46] put Cham on on a podcast and he'll solve [22:49] the the problems of AI right now. I [22:50] think that there's a real set of shifts [22:52] happening and there's a real set of [22:54] global competition underway [22:57] where, you know, various state and [22:59] actors and interests are competing with [23:00] each other. [23:01] >> Yeah. Cham, do you think that we should [23:03] slow down? [23:06] >> I don't think you can. [23:07] >> No, no, no. Do you think we should slow [23:09] down? [23:10] >> No, I I think I was just talking to some [23:11] people on a Zoom right before this, but [23:13] I think after the Manhattan Project, the [23:15] research labs were stood up to maintain [23:17] our scientists that worked on the the [23:19] Manhattan Project from uh effectively [23:21] leaching back or leaking back to [23:24] Russia and Germany and and and other [23:26] places that that were adversary to the [23:27] United States. [23:30] And they all were against the nuclear [23:32] bomb. They worked on it because it was [23:34] necessary for the United States [23:35] security, but then when Russia got a [23:37] hold of the secrets, they were leaked [23:39] because people were worried that if the [23:40] US had all the power, [23:42] there would be no counterbalance to the [23:44] power. [23:45] And so the the nuclear secrets were [23:47] leaked to Russia for that purpose. Then [23:49] when Russia had the nuclear secrets and [23:51] they began developing hydrogen [23:53] fusion bombs, and it it was clear that [23:56] they were going to race ahead. The [23:57] United States raced ahead with [23:58] developing nuclear bombs as a [24:00] counterbalance to Russia. When the [24:02] proliferation began, there was no [24:04] stopping it. It began and you had to [24:06] have this balance in the world, [24:07] otherwise you have effectively an [24:09] asymmetric power that can do whatever it [24:12] wants globally. [24:14] I think there's that moment in the world [24:15] right now where [24:17] if the United States does not [24:19] advance its AI technology, the [24:22] availability of it, TBD, industry, [24:24] taxation, all these all these things [24:25] that we're talking about doing, there [24:27] will be someone else that will. [24:29] And if someone else does, we can go [24:32] through what would happen. There's a [24:33] There's a complicated game theory on [24:35] this, but what would happen if China had [24:37] sufficiently advanced models and [24:39] sufficiently advanced scaled deployment [24:42] of those models relative to the United [24:43] States, as you do that analysis, you [24:45] realize, wait a second, that's probably [24:47] not a healthy place for the world to be. [24:49] It's also probably not a healthy place [24:50] for the United States to be the only one [24:52] with AI. [24:53] And so I I think what we end up seeing [24:57] is if we do try and slow down AI, [25:00] we kind of lose this moment of balance [25:02] that's necessary when you have a [25:03] technology proliferation like we saw [25:06] with the arms race after World War II [25:08] that, you know, we're going to see again [25:10] here. [25:10] >> Chamath, you [25:12] I think bring up a good point. You know, [25:14] should we slow it down or could we slow [25:16] it down? There actually have been some [25:18] discussions about ways to do this. One [25:21] of them would be, hey, with [25:22] self-driving, people are scared that all [25:24] these cab drivers are going to lose [25:26] their jobs, Uber drivers, cab drivers, [25:28] bus drivers, truck drivers. This is, you [25:31] know, over 10 million people in the [25:32] United States driving things for a [25:33] living. [25:35] Would you be in favor of some of the [25:37] announcements that [25:39] that will be [25:40] a paced rollout, it won't happen all at [25:42] once. In other words, those people will [25:44] be given some amount of job security to [25:46] stay behind the wheel with it. Another [25:48] example that's been given is if you put [25:50] Optimus into Amazon factories or the [25:53] Figure robot just did like a week of [25:55] just sorting packages. I'm sure [25:56] everybody saw that video, we'll insert [25:58] it here. [26:00] That Figure robot sorting these, hey, if [26:01] Amazon deploys those, there'll be a tax [26:03] on those per hour and we'll tax humanoid [26:06] robots in some ways and then use that [26:08] for, say, retraining people. Those are [26:10] two very specific conditions and [26:13] approaches that people have been [26:14] promoting. Do either of those resonate [26:16] with you in any way? [26:19] >> I think it's [26:20] interesting that in all of those [26:22] discussions [26:23] I've yet to see an actual survey of only [26:26] the truck drivers and only the package [26:28] sorters. [26:29] >> [clears throat] [26:30] >> The question that I would have is do the [26:31] people that do these jobs want these [26:33] jobs? [26:35] And if they do, then there's a [26:36] reasonable claim to make to keep those [26:37] jobs the way that they are. [26:39] If you're saying this is the job that I [26:40] do, I love it, I'm able to provide for [26:42] my family, great. That's a very [26:44] different argument than well, you know [26:46] what? Amazon has 35 or 40% churn inside [26:50] of their warehouses [26:52] and we should probably ask the question, [26:53] why is that? Because if it was such a [26:55] great job, I suspect the churn would be [26:57] 3% or 4%. So, what exactly is it that we [27:01] want to protect? And have you asked [27:04] them? [27:05] And I think that this is just a again, a [27:07] bunch of [27:08] people in the peanut gallery who want to [27:10] take a moral high ground and try to make [27:12] some other group of people feel guilty [27:15] or feel bad. At no point are we actually [27:17] asking the conversation that [27:20] that we should be having, which is [27:22] it's interesting to me that the there [27:24] was supposed to be an EO, an [27:26] presidential executive order that was [27:27] announced today. [27:28] And then it was pulled. It was scrubbed [27:30] at the last minute. [27:31] >> Mhm. [27:31] >> Did you guys notice that? [27:32] >> Yeah. [27:32] >> And yesterday, what was leaked was [27:34] everybody that was attending. It was all [27:36] the big neo labs CEOs, and it was all [27:38] the big hyperscaler CEOs. [27:40] Including friend of the pod, Nikesh [27:42] Arora. Shout out to Nikesh. [27:44] And then it was scrubbed an hour ago. [27:47] Why was it scrubbed? [27:49] And the president said that there were [27:50] aspects of the bill that he didn't agree [27:52] with. And as far as we can tell, the [27:54] aspects would have required some amount [27:57] of [27:58] supervision, insight, review from the [28:00] federal government. [28:01] >> Of language models specifically. Of [28:03] these frontier models is what I read. [28:05] >> Not language models, cuz I think just of [28:07] AI, because there's going to be many [28:08] different kinds. They're not always [28:09] going to be language models, but of AI. [28:11] So, look, I think Freeburg is is right. [28:14] We are in a proliferation [28:16] with China. I think it's actually good [28:18] that China is less than 9 months behind [28:20] us. [28:21] I think it allows us to find a detente [28:24] where we have a certain magnitude of [28:26] capability that they also have. [28:29] And that allows all of us to then seek [28:31] peace and abundance. And the fact that [28:33] we are orthogonal societies, we are [28:35] organized differently, increases the [28:37] probability of finding peace using the [28:39] René Girard kind of framework of mimetic [28:42] theory, than if if it was like us and [28:44] another country that was exactly similar [28:46] to us. [28:48] So, I think what we need to do [28:51] we probably need KYC. I think that that [28:53] should be something that us and China [28:55] get together and say, "You don't want it [28:58] to get into the hands of people you [28:59] can't control." You probably already KYC [29:01] those models anyway inside of China. You [29:03] already review those training runs [29:04] before you allow these models to get [29:06] released. We already know that that's [29:07] happening. [29:08] >> Yes. [29:08] >> So, we should probably do some sort of [29:10] KYC so some crazy person doesn't create [29:12] some biological weapon. I think that [29:14] those are like some reasonable ground [29:15] rules, but otherwise Freeburg is right. [29:17] You have to take a a little bit of a [29:19] deterministic view here, which is that [29:21] we are in this existential race [29:24] and we need to get to the place where [29:25] each of us, meaning us and China, can [29:27] look each other in the eye and say, "All [29:28] right, weapons down." So to speak. [29:31] >> Kevin, I'm going to hold you to answer [29:33] two questions. One, [29:35] should we [29:37] run frontier models cuz that's [29:39] specifically what was mentioned in the [29:40] leak about the EO frontier models, the [29:42] powerful ones. Should they be run [29:43] through some sort of testing before [29:45] they're released and should there be [29:46] some regulatory framework for that? [29:48] That's my first question to you. Yes or [29:50] no question and then you can explain [29:51] your answer. [29:55] >> Jeez, like I just think it's such a [29:57] complicated topic. It feels [30:01] we're a little early for that. [30:04] I don't love the idea of the United [30:06] States [30:08] um doing it and no one else doing it. [30:11] I [30:13] like I think in a world where we hold [30:15] hands with China [30:17] like I think that's that's much more [30:19] palatable and we are aligned and we [30:21] trust each other and have kind of [30:23] verification [30:25] uh capabilities. I do think [30:27] >> Right. Yeah. Let Let me then rephrase [30:28] it. Should China and the US come up with [30:31] a simple battery of things that have to [30:33] be [30:34] tested before these go out including [30:36] bioweapons, terrorism, and in genre of [30:41] and that vertical of just really known [30:43] dangerous things just like the FDA might [30:45] test for poisons or contaminants in a [30:47] food or a drug. Would you be in favor of [30:49] that? I'm curious. [30:52] >> So, a few things like the one one thing [30:54] that's great about America is there is [30:57] one thing that's [30:59] just [31:00] we have other forms of regulation. [31:03] >> Self-regulation, sure. [31:04] >> We have self one we have self-regulation [31:06] also we have the courts. [31:08] And if an AI model company [31:11] behaves responsibly [31:13] they know that there are [31:15] ways that people who have been harmed [31:17] can seek recourse. [31:19] And so we already have a system that [31:22] encourages responsible behavior on the [31:25] part of the model makers. [31:28] >> That's a great point because OpenAI is [31:29] being sued right now by a kid who who [31:31] killed themselves after talking to [31:32] OpenAI's model. So, you're you're [31:34] correct in that. Yes, after the fact. [31:35] Yeah. [31:36] >> And we will see we'll see we'll see more [31:37] of that. I just [31:40] you know, to me once you give something [31:42] give a power to the government [31:45] it's almost never taken back and it [31:47] seems to grow. [31:48] And it's kind of a one-way a one-way [31:50] path. [31:51] >> One-way ratchet. Yeah. And then the [31:53] one-way second yeah, second question [31:55] then. [31:56] Chamath is saying hey nobody listens to [31:58] the you know, these cab drivers or maybe [31:59] the [32:00] people sorting the packages do they want [32:02] their jobs or not. Actually the UK [32:04] there was just a 60-minute special in UK [32:07] and also Boston and New York are pretty [32:09] adamant that they want humans to stay [32:12] and they want to ban self-driving in [32:13] those locations or severely limit it or [32:15] maybe limit it in some way to let those [32:18] people keep their jobs. How do you feel [32:20] about that possibility? Is that [32:22] something you think society should be [32:24] open to some gradual licensing [32:26] >> They're going to get sued for wrongful [32:28] death. When somebody runs over somebody [32:30] else and you could have implemented a [32:32] solution that has a zero death rate, [32:33] that's very different from package [32:35] sorting. [32:36] >> Okay. [32:36] >> Go talk to the package sorters is what I [32:38] say. Go talk to the people inside the [32:39] Amazon warehouse. Ask them what they [32:41] would rather do at Amazon. Ask them. [32:44] >> Yeah, sure. But Gavin, what are your [32:45] what are your thoughts here on either [32:46] one of those examples here? [32:48] >> I think going to a city where you can't [32:49] get in a Waymo or a cyber cab is going [32:53] to feel barbaric and unsafe until you [32:56] >> Agree. I strongly agree. [32:57] >> I don't know if you remember like but [32:58] the early days of Uber, sometimes you go [33:00] to a city where there was no Uber. [33:02] >> Yeah, that'd be incredibly frustrating. [33:04] >> What I'm not going to come back until [33:05] they have Uber. It's so inconvenient. [33:08] And I think so whatever individual [33:11] municipalities decide, I do think one, [33:14] Chamath's point is really powerful. [33:16] There's 50,000 automotive deaths per [33:18] year in the United States if I recall [33:20] correctly and a million globally. [33:24] And you know, that's not tolerable and [33:26] there will be for sure be wrongful death [33:27] lawsuits and then just from a [33:29] convenience and quality of life [33:31] perspective, I just don't think it's [33:32] going to persist. [33:34] And that's another great thing about [33:35] America is you know, you have this [33:37] patchwork of different states and [33:38] municipalities and [33:41] each one doing things in a different way [33:42] and I'm not suggesting that's good for [33:44] AI, [33:46] but it does tend to, you know, [33:49] as historically the Curley effect aside, [33:53] led to, you know, I think more positive [33:55] outcomes where cities and states [33:57] compete. You know, the Curley effect [33:58] being that [33:59] >> Yeah, that yeah. That this is a really [34:01] important point you're making, Gavin. [34:03] With Flock Safety as but one example, we [34:05] had a an AI there's an AI tool called [34:08] Flock Safety. It's cameras that use AI, [34:10] monitor people who are committing [34:12] crimes. There's a privacy issue around [34:13] it. It is bottom-up. You just do it by [34:15] town. It's not top-down and states can [34:18] regulate it. Same thing will probably [34:19] happen with self-driving and states will [34:22] probably have some say in how AI is [34:24] deployed even if maybe some centralized [34:26] governments don't want to do that. [34:29] I really think this It comes down to [34:31] >> The Flock [34:31] >> thing I think it's so good, Jason. Crime [34:33] is now a choice. [34:34] >> Yeah. [34:35] >> You know, I think the the Cambridge City [34:36] Council voted [34:38] to turn off gunshot detectors 2 days [34:41] ago. [34:42] And [34:43] >> which city did that? Which municipality? [34:44] >> Cambridge. Cambridge, Cambridge, Mass. [34:46] >> That's the place where Harvard is. [34:48] >> That's the place where Harvard is. [34:49] >> So, the geniuses coming out of Harvard [34:51] in that town decided [34:53] gunshot detection [34:56] shouldn't occur. [34:57] >> We don't want gunshot detection. Just so [34:59] we're clear. [35:00] >> It's wild because, you know, there's [35:01] there's a theory that it disadvantages, [35:03] you know, that it might lead to an [35:05] illegal migrant [35:06] who's shooting a gun being apprehended, [35:09] and we don't want that. [35:10] >> Got it. [35:10] >> And A16Z had a great essay on Flock. [35:14] We can really, really solve crime, [35:18] and it's just a choice. And different [35:20] states and municipalities will make [35:22] different choices to be pro-crime or [35:24] anti-crime. And I'm sure they don't cast [35:26] it as pro-crime. There's [35:29] an, you know, some sort of, you know, [35:31] moral or ethical reason they may make [35:33] that choice, but people will vote with [35:35] their feet over time, and then voters [35:37] will vote with their votes, and we'll [35:39] see what works. [35:40] >> Have you guys been to Vegas recently? My [35:42] wife and I went to visit Vegas, and we [35:44] spent the [35:45] the afternoon with Ben Horowitz and his [35:47] wife, Felicia. She has done this [35:49] incredible job with the Las Vegas Police [35:51] Department. [35:52] It is one of the most impressive things [35:54] I've ever seen, and to your point, [35:56] crime is an option, and they've said no. [35:58] So, what happens there is they have [36:00] gunshot they have drones that get [36:02] deployed off the roof of the police [36:04] building. We were sitting inside of [36:06] mission control where you see it [36:07] happening, Jason. If something happens, [36:10] they have eyes on site within minutes. [36:12] They can track offenders and bad guys [36:15] all the way to wherever they're hiding. [36:17] And you walk out of it, and you feel [36:19] incredibly safe, like they're really on [36:21] top of it. And when you understand the [36:23] level of investment, it's not it doesn't [36:25] take billions of dollars compared to the [36:28] cost of the crimes. It's de minimis. [36:30] It's de minimis. Especially when [36:30] compared to the cost of the crime [36:32] occurring. [36:33] >> Exactly. If you gave the Las Vegas [36:34] Police Department [36:36] 30, 40 million dollars a year, it would [36:39] be the safest city in America and that's [36:40] all it would take. [36:41] >> Yeah, exactly. And the for the privacy [36:44] concerns Freiberg, there are very simple [36:45] solutions to this. I I am a privacy [36:47] advocate myself, of course. We all want [36:49] some level of privacy. [36:51] I had the Flock CEO on This Week in [36:52] Startups twice in the past 10 years. [36:54] He's very considered. And the way they [36:56] do it with Flock is they allow you to [36:59] have a rolling database and I think [37:01] there's a maximum you can save the [37:03] license plates for and they don't do [37:05] facial recognition. I I don't see why [37:07] not, but let's put that aside. You can [37:09] only keep it for two or three years. And [37:10] then they insist on having an audit [37:12] trail on it. So there are all little [37:13] things you can do on the back end to [37:14] protect privacy with audit trails, etc. [37:16] We got a lot more to get to in the [37:17] docket. I just want to just give my [37:19] final thoughts on [37:21] what we're talking about here in terms [37:22] of the AI problem [37:24] and the PR problem. I think we have to [37:27] recognize that the layoffs that are [37:29] occurring in Big Tech and in a lot of [37:31] these places are not just the bloating [37:35] issue anymore. And I'm just going to [37:36] point to two factors that I think are [37:38] scaring the bejeezus out of people and [37:40] we just have to admit that this is [37:42] occurring as opposed to we've been [37:43] debating it here. Is it occurring? Is [37:45] this just cover? And are we AI washing? [37:47] The first one I want to give you an [37:48] example of is Matthew Prince [37:51] who's the CEO of Cloudflare, incredible [37:53] company, public company. Two weeks ago I [37:55] laid off more than 20% of my workforce. [37:56] I didn't do it because Cloudflare is [37:58] struggling. We posted record revenue [38:00] growth, have strong free cash flow, and [38:02] are adding an unprecedented number of [38:04] customers, yada yada yada. [38:06] And he says basically he's getting rid [38:08] of measurers. Measurers are the people [38:11] who manage people and who measure data. [38:13] And he just says we're getting rid of [38:15] all those people. They're unnecessary [38:17] because of AI and we'll be adding to [38:19] people in other positions. At the same [38:21] time Zuckerberg did another round of [38:23] layoffs, and they were done in a way [38:25] that people felt was not considered [38:28] and a bit [38:31] um [38:32] What's the word? [38:34] Dystopian. Dystopian, thank you, sir. Uh [38:37] he did him in a in a pretty dystopian [38:38] way. [38:40] Here's Zuckerberg for 30 seconds. [38:42] >> In general, the average intelligence of [38:44] the people who are at this company is [38:46] significantly higher than the average [38:48] set of people that you can get to do [38:50] tasks if you're working through the [38:52] contract um [38:53] through through these contractors. So, [38:55] if we're trying to teach the models [38:57] coding, for example, then having people [39:00] internally uh build tools that or or [39:03] solve tasks that um that help teach the [39:06] model how to code, we think it's going [39:08] to dramatically increase our models' [39:09] coding ability faster than what others [39:11] in the industry have the capability to [39:13] do who don't have thousands and [39:15] thousands of extremely strong engineers [39:17] at their company. [39:18] >> [clears throat] [39:18] >> Okay, so what Zuckerberg did at the same [39:21] time, concurrently, he told everybody, [39:23] "We're laying off these 8,000 people." A [39:24] lot of those people are incredibly [39:25] talented. [39:27] Some of them are on H-1B visas, creates [39:29] all kinds of chaos for them in their [39:30] personal lives, and obviously they're [39:31] having record profits there as well. [39:34] At the same time, he was laying off [39:36] those 8,000 people. This is after tens [39:38] of thousands of layoffs before, which [39:40] were obviously because of bloating. He [39:41] said, "We're putting recording software [39:42] on every single person in the company's [39:44] computer to study and train our model." [39:47] And people were like, "Oh." And previous [39:49] people said, "I built during the AI [39:52] hackathons they had months ago. I [39:55] built all these AI tools to make my job [39:57] uh more efficient." And then Zuckerberg [39:59] laid me off. So, the now per- the [40:02] perception people have now, and it's [40:04] quite correct, I think, is the most you [40:06] can hope for here is you keep this job [40:09] for some amount of time and train your [40:10] way out of it, and hopefully there's [40:12] some more work for you, but they're [40:13] studying you, and Zuckerberg just said [40:15] it plainly there, "Hey, we're going to [40:16] study everybody here, and that's going [40:18] to lead to more replacements. This is [40:19] scaring the bejesus out of people and we [40:22] need to have an answer for it. Yeah. [40:23] >> I thought the Matthew Prince note was [40:26] horrible. [40:27] >> Okay, explain. [40:28] >> This was like from the [40:30] PR school of retards. [40:32] >> Okay, here we go. [40:34] >> You could not have written a worse memo. [40:36] It's like you reduce humans to a label [40:40] called the measurer [40:42] and then you're like I'm going to lay [40:43] off all the measurers. [40:45] I mean [40:47] I just think that part of this again, [40:49] I'll go back to the [40:51] maybe the Sham Sankar quote that I'm [40:53] thinking about should be extended beyond [40:55] the model makers. Can you just play this [40:57] for 1 second and I'll tell you why I [40:59] think this is just so [41:00] >> And then we'll go on to the S1 from [41:01] SpaceX. [41:02] >> We're listening too much to the [41:04] inventors of AI. I know that's [41:05] appealing. They're geniuses, they're [41:07] smart. We need to be listening to the [41:08] frontline factory workers who are using [41:10] AI saying, "Wow, I was able to add a [41:12] third shift. I was able to hire more [41:13] workers." Or the ICU nurse who says, "I [41:15] have more time to spend with my [41:17] patients. I'm able to ensure they don't [41:18] code during a shift change down." [41:20] >> so look, my my point is like the first [41:22] part of what he said applies here, which [41:24] is who cares what Matthew Prince thinks? [41:26] Because the reality is that if this is [41:28] the way that you're going to message [41:30] something as critical as this, I think [41:31] you did a horrible job. And now you [41:33] label these people and you put a scarlet [41:35] letter on their back. So now when they [41:36] try to get a different job, they're [41:37] like, "Oh, you're one of the Cloudflare [41:39] measurers?" How does that help anybody? [41:41] It didn't needed to be done this way. [41:43] There's enough of these tech [41:45] CEOs that are now public. You can hear [41:48] them, you can understand them. [41:50] And I think what we're learning is men, [41:52] they're really good at one thing and [41:53] they're not necessarily as good at all [41:55] the other things. [41:55] >> Yeah. Uh okay. [41:57] >> And so I would say shut the up. [42:01] Get behind the keyboard, just do your [42:03] job. And if you need to manage [42:05] something, just manage it, but don't [42:06] write these missives. You're terrible at [42:08] it. All of you. You're all terrible. You [42:10] suck at this. [42:13] >> All right. [42:14] >> End of my TED Talk. Thank you for coming [42:16] to my TED Talk. [42:16] >> Thanks for coming to Chamath's 18-minute [42:18] TED Talk. And we'll we'll we'll uh keep [42:21] moving on. [42:21] >> Sorry. And sorry. When everybody gets [42:23] upset, this will be why. [42:25] >> Yes. I mean I I do think the Zuckerberg [42:28] this and Jack saying, you know, "Hey, [42:29] we're going to have half as many people [42:30] everybody reports directly to me." This [42:32] is all building this fear in society and [42:34] I think people are right rightfully [42:36] scared. If the people building it tell [42:38] you be scared, your job's going away. [42:40] >> Wait till the next [42:42] regulatory filing comes out from these [42:44] companies and they authorize a massive [42:46] share buyback and an increase in their [42:48] dividend. [42:49] >> Yeah, and their cash pile grows. [42:51] >> All I'm saying is there's a right way to [42:53] do this, make these decisions, and then [42:54] there's a wrong way to do it, which is [42:56] to message it in the way that they're [42:57] doing it. So, whoever is running PR and [43:00] comms and approves and reviews these [43:02] things are really at their job. [43:05] >> Mhm. [43:06] >> They don't understand the moment. [43:08] >> Oh, some breaking news here. It looks [43:10] like Anthropic has hired three more [43:13] people. Here we go. [43:14] Oh, here we go. Personal job news from [43:16] Sam Altman. He'll be joining Anthropic. [43:18] I say that's pretty good. Okay. Who else [43:20] is joining Anthropic? Let's see if we [43:21] checked everybody's socials. Oh, Tucker. [43:25] Tucker Carlson is also joining [43:26] Anthropic. He'll be doing their PR and [43:28] podcast from Anthropic headquarters. And [43:31] who else? Oh, Chamath. Looking good. [43:34] [laughter] Well, you look like you put 5 [43:35] lb on. [43:36] And using the pants [43:38] >> First of all, this is not This is not [43:40] what I look like. And [43:41] >> Did you gain 5 lb? [43:43] Looks like the same [43:45] guy. Are you not wearing Are you not [43:46] wearing underwear? [43:47] He's not wearing underwear and he's [43:48] wearing his khakis, but I think it's [43:50] just he's trying to not show off those [43:51] scrawny those scrawny splints those [43:54] those little slats he calls legs. He's [43:56] covering them up now. [43:57] >> Uh hold on. I'm going to send Nick [43:59] I I I knew that this was going to come [44:01] up. I'm going to send Nick an updated [44:02] picture of my legs, and you you can deal [44:04] with this. [44:05] >> You can Photoshop your legs all you [44:07] want. [44:07] >> I didn't Photoshop [44:08] >> look like ostrich, brother. [44:10] >> He's better than you today. I'M WORKING [44:12] ON [44:13] IT. [44:14] >> [laughter] [44:14] >> HE MAY NOT BE PHOTOSHOPPING. He may have [44:16] been leg maxing. [44:17] >> He could be leg maxing. [44:19] >> Are you leg maxing? Are you BP 157 your [44:21] legs? What are you doing here? [44:22] >> of all, first of all, I'm 6'2, you [44:24] goons. Okay, so [44:26] >> Ostrich legs. [44:27] >> All you little people, you know, Jason, [44:29] you're 5' tall, so you know, your [44:31] ability to have legs [44:32] >> I'm on a good day with my platforms. [44:33] >> to have legs is is different cuz like my [44:36] muscle mass won't show on my legs the [44:39] way it shows on your leg. [44:40] >> Oh, I don't think you're help but I look [44:42] at You know what he's doing there? [44:43] Freeberg, you see it, right? Look at how [44:45] skinny the calves are and then look at [44:47] how he's wearing a [44:48] >> Oh my god, Jason. [44:49] >> bra for your quads. [44:50] >> Oh my god, stop. [44:51] >> equivalent of a push-up bra. He has [44:53] those Bosu balls. He put Bosu balls [44:55] under his hammies to You did it. You're [44:58] using filets to pump those up. [45:00] >> my legs and I flat Okay, you know what? [45:01] Move on. [45:02] >> You know, I was just saying I think we [45:03] should give credit where credit is due. [45:06] >> Yeah. Jamath's legs [45:08] >> better [45:08] >> Jamath has been doing a lot of work on [45:09] his legs. [45:10] >> Thanks, man. [45:10] >> It's better. [45:11] >> Thank you. [45:12] >> It's better. [45:12] >> Thank you. I'll I'll give him better, [45:14] but I do think that he's pumping them up [45:16] here. Okay, let's keep going here. All [45:18] right, topic two is SpaceX just filed [45:20] their S-1 on Wednesday. They are aiming [45:23] to raise 75 billy at a 1.75 trillion [45:27] with a T valuation. This would be the [45:28] largest IPO ever by more than double [45:31] Saudi's Aramco $29 billion IPO a couple [45:34] years ago. [45:36] Listing is expected mid-June, likely [45:38] June 12th. Ticker will be SPCX. [45:41] We got a lot of interesting information [45:43] in the S-1 teardown. Obviously, SpaceX [45:46] has three main business units. [45:48] Starlink is the money printer right now, [45:52] but there's a second one that's [45:53] emerging. Starlink did 11.4 billion in [45:55] revenue last year on 50% growth with 4.4 [46:00] billion in operating income. [46:02] Over 10 million people are now [46:03] subscribing to Starlink. That business [46:05] could easily be hundreds of millions of [46:08] paying subscribers. So, that that's a [46:11] lot of growth potentially there. The [46:13] space business is but 4 billion in [46:16] revenue. It's growing 17% growth, which [46:17] would still be strong growth. [46:19] Uh but had 650 million in operating [46:21] losses. AI did 3.2 billion in revenue. [46:24] That's more than double year-over-year [46:25] growth, but it had 6.4 billion in [46:27] operating losses. [46:29] SpaceX had 20 billion in capex spend [46:32] last year. Over 60% was for the AI [46:34] compute buildout. [46:36] And obviously they were trailing [46:38] Anthropic and OpenAI and Gemini in terms [46:40] of XAI, uh playing catch-up. And he did [46:43] a big reboot of that as we saw on the [46:45] Twitter. [46:46] But here's the big one. EWS, Elon Web [46:49] Services, as we call it here on the [46:51] All-In Pod, has exploded. Anthropic is [46:55] paying SpaceX, wait for it, 1.25 billion [46:59] a month to rent out Colossus 1 and parts [47:01] of Colossus 2. It's a $45 billion deal [47:05] over 3 years, 15 billion a year. In [47:08] other words, they added a Starlink in [47:09] terms of revenue to the party. Plus, if [47:12] they buy Cursor, that's going to add [47:14] another 2 or 3 billion. Not [47:15] >> Not if. I already told you they already [47:17] bought it. [47:17] >> Okay. I'm just trying to, you know, dot [47:19] the i's and t's here. But when they buy [47:22] Cursor, that adds 2 or 3 billion. That's [47:24] not in the S1. [47:26] >> That's also growing and doubling. [47:28] >> Yeah, that's probably growing 2x [47:30] year-over-year. And who knows how much [47:32] faster it'll grow. Polymarket, 71% [47:35] chance SpaceX closes its first day of [47:37] trading with a market cap above 2 [47:39] trillion. Thank you to our partner, [47:41] Polymarket. [47:43] I'll stop here. Gavin, you've been [47:45] involved in the company for a long time. [47:46] Chamath, you I think were a big investor [47:49] in the satellite company that became [47:51] part of Starlink, which is the revenue [47:54] driver there. So, you both have a lot to [47:56] say about this. Gavin, your take on the [47:58] S1 and I think specifically Elon web [48:01] services. [48:02] >> Well, I think what's important about [48:04] Elon web services does make me laugh. [48:07] But 15 billion, that means the AI [48:09] business right there is going to [48:11] quadruple. It has already effectively [48:13] quadrupled. [48:15] I think what's important about that is [48:17] there's a stat in it that for I think [48:19] the the first data center was 122 days. [48:22] With the second one, it took them 91 [48:24] days. [48:25] The third one was I think 66 days. [48:28] They build data centers dramatically [48:31] faster than anyone else at a lower cost. [48:36] And now that you have a [48:38] clear off-take partner and I would [48:41] expect partner to become partners, [48:44] there is no reason they can't start [48:46] stamping these data centers out really [48:49] fast. [48:51] And having watched Jensen for a long [48:53] time, it is important to Jensen that his [48:55] GPUs be used. [48:57] And so GPUs will be allocated to who can [49:00] plug them in, turn them on, and start [49:02] converting electrons into tokens. [49:05] And so I think this business can grow [49:09] dramatically faster than I think [49:12] you know, maybe what anyone could have [49:14] contemplated three, you know, three [49:16] three months ago. But 15 billion dollars [49:19] from Anthropic [49:21] is is extraordinary. [49:22] >> Important note, it can be canceled by [49:24] either party with 90 days notice. Just [49:26] want to make sure we also have that in [49:27] there. So that means Elon might want his [49:29] compute back or Anthropic may find [49:32] another solution. So they do both have [49:33] an out. [49:34] >> And I think that's that's a [49:36] you know, I think that's that's that's [49:37] probably an important provision [49:39] >> Yeah. [49:39] >> for everyone. But I think the other [49:41] thing that came out this week, which was [49:43] not in the S1. Nick, can you throw up [49:45] the parade of frontier and maybe don't, [49:47] you know, include the email and the [49:48] names and everything, but the composer [49:50] 2.5 stat. I think this is really [49:53] extraordinary. [49:54] So, Cursor's Composer 2.5 model came out [49:57] this week. [49:59] And I mean, this is Pareto dominant. [50:02] And this is just [50:04] you know, three, four weeks of doing [50:06] reinforcement learning [50:08] on Colossus 2 [50:10] with Cursor's data. [50:13] And Cursor has We will never know, but [50:16] but Cursor allegedly has more tokens of [50:20] coding data than exist on the public [50:22] internet. [50:24] And that is a stat from I think more [50:25] than a year ago, so I'd imagine it's [50:27] grown significantly. [50:29] And I think Cursor and Anthropic [50:32] probably have the most proprietary [50:35] tokens of coding data. And what this [50:38] this jump from Composer 2 to Composer [50:40] 2.5 showed us [50:42] is that when you do an appropriate [50:44] amount of reinforcement learning using [50:47] that data, let alone injecting it into [50:50] the pre-training of a new base model. [50:52] Cuz Composer 2.5 is the same base model [50:54] as Composer 2, which is Kimmy K K.25. [50:59] Like this is amazing. This is three or [51:01] four weeks, and it is Pareto dominant. [51:03] The Pareto frontier, if you draw a curve [51:07] of the blue dots, [51:09] you can see [51:10] Composer 2.5 is literally [51:14] well outside the Pareto frontier. And [51:17] that's after three weeks. [51:18] And what's going to happen next is [51:20] you're going to have a [51:21] new base model [51:23] with a Cursor model in it. [51:25] >> Yep. [51:25] >> Then the Cursor model [51:27] RL'd [51:29] using the biggest coherent [51:31] compute cluster in the world. [51:33] And I think [51:35] this is [51:36] I I think this may [51:37] >> It's significant, yeah. [51:38] >> It's extremely significant for XAI and [51:41] Cursor. [51:42] >> And Cursor was dead in the water in [51:45] terms of access to compute and they were [51:48] falling very far behind Kodaks, [51:51] Google, Anthropic, and then Elon let [51:54] them on Colossus and boom, instantly [51:58] their models are [52:00] growing faster and this could be we [52:02] could be sitting here a year from now [52:04] and they're the dominant player. [52:06] And [52:07] could we be sitting here Gavin in a year [52:09] and Elon is selling compute to Google [52:13] and Open AI? Is that a possibility or [52:16] not? [52:18] >> Well, I think it's much easier to see [52:20] him selling compute to Google and I [52:22] think that's going to happen. [52:23] There've already been posts about that. [52:25] >> Yeah. [52:26] >> And for sure Google is going to want to [52:28] be part of orbital compute. [52:31] You know, it's it's a very funny. The [52:33] only people who are skeptical of orbital [52:35] compute are those people who are not [52:37] involved [52:39] in data centers or space. Google, [52:41] Anthropic, Amazon, Nvidia. [52:45] They are all very convinced that orbital [52:46] compute is going to be reality and [52:48] obviously SpaceX is extraordinarily well [52:49] positioned for that. [52:51] But I do think that composer 2.5 data [52:54] point is really powerful. [52:55] >> Keep an eye on it. Yeah. [52:57] >> And then the other thing that's come out [52:58] is Grok builds. So, what Grok lacked [53:01] that a lot of other models had and I do [53:03] think it's important to [53:05] remember that the newest version of Grok [53:07] 4.3 is on the Pareto frontier for all [53:09] frontier models. [53:11] And you're either on the frontier or [53:12] you're not. [53:13] And the the companies on the frontier [53:16] are xAI [53:18] with one build of Grok 4.3, which is a [53:20] 500 billion parameter model, [53:22] Google 3.1 Pro, and then Open AI and [53:25] Anthropic. And that's it. Those are the [53:26] companies on the frontier. [53:28] >> And then the four horsemen, yeah. [53:29] >> Google today each have one dot on the [53:32] Pareto frontier. [53:34] And obviously you want as many dots as [53:35] possible, but Grok lacked a harness. [53:40] So, Claude had Claude code, OpenAI had [53:43] Codex, and now with Grok build, there is [53:46] a harness that is available to to Grok, [53:49] and you know, as I'm as I'm sure [53:51] >> a downloadable app to translate into [53:54] English that has integrations to all [53:56] your favorite stuff, whether it's [53:57] Notion, Gmail, Slack, etc. And if you [54:00] don't have that, it's just like using a [54:03] basic chatbot from a year ago. So now [54:04] they have their downloadable, it's in [54:06] market, and they are cooking with oil on [54:08] it, and they're playing catch-up, but [54:09] they're moving fast. [54:10] >> And and it's it's more than just an app. [54:12] It's it's a runtime, it's an [54:13] environment. It manages state, it [54:15] manages memory. [54:17] It makes these models dramatically [54:21] more useful to the extent that I think [54:23] the people at the frontier all agree [54:25] that the harness is essentially as [54:27] important as the model, especially in an [54:29] energetic world, and the harness and the [54:31] model need to be developed together. So [54:34] the release of Grok build and the pace [54:36] at which they're iterating [54:37] is I think also really encouraging. So [54:40] now you have cursor, you have the cursor [54:42] data, you have a clear existence proof [54:44] that the cursor data is really [54:45] important. [54:46] Composer 2.5 is now Pareto dominant and [54:49] the most selected model on cursor, and [54:52] that's also important because, you know, [54:54] these evals don't capture everything. [54:57] You know, this is why people on X talk [54:59] about the vibes, and the vibes on cursor [55:02] 2.5 are also really good. [55:04] >> They're immaculate, yeah. [55:05] >> That together [55:06] >> Yeah. [55:06] >> with Grok build, I think these are [55:09] really important developments. [55:10] >> Yeah, I think they're There was Elon was [55:13] incredibly frustrated by the state of [55:15] affairs at xAI. He was very public about [55:17] that, and he's less frustrated now, and [55:19] he's shipping a lot faster. And so I [55:22] think that says something, and he has [55:24] been very focused on it. [55:26] Freeberg, your thoughts on the SpaceX [55:29] IPO and what this collection of [55:32] companies might look like a year or two [55:35] from now, especially if, like many [55:38] people believe [55:39] Tesla and [55:41] SpaceX merge, what do you think of [55:43] dollar sign E L O [55:45] E L O N as an entity and what impact it [55:49] might have as if those two were put [55:51] together the market cap would put them [55:52] in the fourth largest company in the [55:53] world. We can revisit our earlier [55:55] conversation about an anti-tech, [55:57] anti-AI, anti- [56:00] progress [56:01] world and society ahead. And if there is [56:04] an effort a concerted effort and [56:06] organized effort by governments to stop [56:08] or block access to information, [56:11] restrict freedom of speech, restrict [56:14] freedom of purchasing or buying things [56:17] to [56:18] control more things. And I think there's [56:20] a trend line in this direction right now [56:22] globally. The internet has always been [56:24] lauded as this kind of [56:26] system that provides an open alternative [56:29] to physical commerce that you could [56:31] create digital commerce, digital [56:33] uh [56:34] information, digital media that you [56:36] could share. And um it's almost this [56:38] digital representation of society, but [56:40] the internet has to sit physically [56:42] somewhere. [56:43] And the assault on data center builds [56:45] out in the United States right now, [56:48] I think [56:49] may indicate the importance of having an [56:52] alternative internet from the ground [56:54] layer up. If you have a communication [56:56] network that isn't restricted and [56:59] controlled [57:00] by a government [57:03] on Earth, [57:05] it's almost like a backup [57:07] for civilization, but it's a backup for [57:10] progress. [57:11] And I don't own any SpaceX shares, and [57:13] I'm not trying to sell the book of [57:15] SpaceX, [57:16] but I think that there's like an [57:17] important aspect of [57:19] can you create a system that's not under [57:21] the control of governments [57:23] as a way to ensure humanity's progress, [57:26] to ensure civilizational continuity if [57:29] things go south, if things aren't good, [57:31] if things are restricted, and if you [57:32] know, fundamental forms of tyranny start [57:34] to restrict speech, restrict commerce, [57:37] restrict information flow and whatnot. [57:40] And I think having like a space-based [57:42] communication network, space-based data [57:45] centers, [57:46] and space-based communication back down [57:48] to Earth wireless, [57:51] I think it's generally a good thing. [57:53] It's good to have a backup. [57:54] >> Yeah. [57:55] >> So, put all the economics aside and the [57:57] multiples and the valuations and whatnot [57:59] and whether it's SpaceX or not, [58:01] I think the idea that you could have [58:02] data centers, store information, [58:04] transmit information, route information, [58:06] and access information through [58:08] space-based systems that can't be [58:10] controlled, manipulated, or destroyed by [58:12] governments is is important. And I I [58:15] just I like that. [58:17] Yeah, if you [58:18] >> Most people don't remember this, but [58:19] when Elon was starting SpaceX, the [58:22] original idea [58:24] when he was running around with the DAO [58:25] and they were [58:27] looking at some rockets and getting [58:28] carriage from Russian rockets was to [58:30] back up the biosphere. And he came back [58:33] from that trip, and I remember talking [58:34] to him about it, and he said, "I think I [58:35] just have to make my own rockets because [58:37] that's actually where the problem is, [58:38] and it would be easier just to make my [58:39] own rocket to back up the biosphere." [58:41] And he wanted to put geodomes, [58:44] like geodesic domes in space [58:47] with all the plants and wildlife and and [58:49] creatures. [58:50] Uh and what incredible vision, and then [58:53] it you know, there was the necessity of [58:54] actually getting that up into space, and [58:56] that's that's the the unknown origin [58:58] story. I will say this, Chamath, the [59:01] idea of putting [59:03] uh data centers in space seems [59:05] completely doable, even though there are [59:07] a bunch of people who are saying it's [59:08] not, [59:10] when you compare it to what happened [59:11] with SpaceX um with Starlink, which [59:14] people said also wouldn't work. And now [59:16] he's got 10,000 Starlinks up there. The [59:18] difference between a Starlink satellite [59:20] and a data center satellite is really [59:22] not that different. And [59:25] >> No, they're they're pretty different. [59:27] >> Well, conceptually, yeah, of course [59:29] they're physically different, but [59:31] conceptually Elon put 10,000 Starlinks [59:34] up. [59:35] Is he capable of putting 10,000 [59:37] >> No, look, the size is much the size the [59:39] size is much bigger, Jason. The foils [59:41] are much bigger, the wings are much [59:42] bigger. [59:43] >> but it's it's yeah. My point is it's not [59:45] different if he has the new Starship. [59:48] Cuz that's 10 times bigger, yeah. [59:50] >> You can't just scale like this. That [59:51] being said, it's technically possible. I [59:53] think he will be the first one to figure [59:56] it out. [59:58] But I'll just take a much more [59:59] pedestrian take, which is okay, you're [60:01] sitting here [60:02] and if I'm asking myself, Chamath, how [60:04] do I underwrite SpaceX at $2 trillion? [60:07] Here's the basic math that I would do. [60:09] Well, last year it did 18 19 billion [60:11] dollars. [60:12] It'll probably do 25 to 30 this year. [60:14] Okay. So, I'm buying this thing [60:17] at a fairly [60:20] costly premium, right? [60:22] So, what am I buying? [60:25] Well, I'm buying probably the most [60:28] important internet infrastructure [60:30] project that's happened since the [60:31] internet itself. [60:32] That's going to scale to hundreds of [60:34] millions of users. And the reason that's [60:35] going to scale to hundreds of millions [60:36] of users is [60:37] it's just very useful and it's just [60:38] going to become cheaper and cheaper and [60:40] cheaper. [60:41] So, that's number one. [60:43] I'm buying a delivery infrastructure. [60:45] But I think over time [60:48] GDP plus 10, GDP plus 15 kind of a [60:51] grower. So, good business, [60:53] valuable business, but it's the [60:54] underlying platform that allows [60:55] everything else to happen. [60:58] And then I'm buying an AI business which [60:59] will be at the top level the apps, but [61:01] at the bottom layer all the compute [61:03] capability. [61:05] And I think when you scale that out, [61:09] like why is Colossus so valuable to [61:12] Anthropic? Maybe that's like a good [61:13] question to ask. [61:15] It's because if you look at who's [61:17] actually capable of delivering a [61:19] gigawatt data center, [61:21] these guys are the closest. [61:23] Like an actual gigawatt. And And the [61:25] reason is is that this stuff is very [61:28] complicated and very, very hard. I think [61:30] you've probably heard this famous story [61:31] where Jensen was like, "Yeah, he was the [61:33] one that figured out this one thing that [61:35] we that nobody else could figure out so [61:36] that you could strip a bunch of racks [61:38] and drive a bunch of east-west traffic [61:40] and make the whole thing work together." [61:43] So, I suspect what happens is next year [61:46] it's probably 40, 45 billion. And then [61:50] the year after that it probably doubles [61:51] again. So, now I'm buying it at 20 times [61:53] revenue. And you would say, "Well, why [61:55] can you buy a company like this on [61:56] revenue versus earnings and cash flow?" [62:00] And I think the reason is because what [62:01] the revenue does is it gives him the [62:04] operating leverage to go and invest in [62:06] all of these other businesses [62:08] that ultimately consolidate [62:11] his differentiation and his competitive [62:13] moat. Because what he creates is a [62:15] capital moat that then accelerates a [62:17] technology moat that then accelerates an [62:19] execution and a learning moat. [62:22] And that's flywheel when it starts to [62:23] spin very quickly. And you would say, [62:26] "Hey, hold on a second. It's probably [62:27] spinning quickly now." [62:28] I would say we're at the beginning of [62:30] the beginning. [62:32] Because he's Again, he still has all [62:34] these disparate assets. I still don't [62:35] like the fact that Tesla's over here. [62:38] And as I've told you, that will get [62:39] merged in. And now you have this [62:42] incredible corpus of [62:44] physical capability, [62:46] movement of all kinds, X, Y, and Z, [62:49] right? [62:51] You have learning capability. You have [62:53] infrastructure. You have all the [62:54] connectivity. [62:56] That thing will look very cheap, I [62:58] think, in a few years. [63:00] And he has this one thing that nobody [63:02] else, if you look at the big CEOs, [63:05] who steps on stage where you're always [63:08] curious, "Okay, what does he got up his [63:10] sleeve?" You know, the Steve Jobs, "Oh, [63:12] and one more thing." [63:14] This is the only guy at the scale of [63:17] civilizational [63:19] out of left fields. [63:22] He's he's the guy. Whether you like him [63:24] or you hate him, he's the guy. And [63:26] there's a premium that is well-deserved [63:28] that comes with that. [63:29] So, if you had to pick an underwriting [63:31] case, Jason, I would flex the revenue [63:33] and realize that terrestrial data [63:34] centers alone [63:36] are 100 or 200 billion dollars of [63:38] revenue by 2030, 2032. Just And that [63:41] means just building it. So, already [63:44] you're buying it at 20 times revenue [63:45] just for that business. Everything else [63:47] is gravy. [63:48] >> on the ground earth-based. [63:50] >> it'll be Colossus three. No, no, forget [63:52] space for a second. It's a Colossus [63:54] three, Colossus four. [63:55] >> It pencils out with that, yes. [63:57] >> Getting a nameplate 1 gigawatt Look, it [63:59] is freaking hard, man. Getting a [64:01] gigawatt nameplate working [64:05] is almost And then, by the way, there's [64:07] all the stuff that he can do on land [64:08] that he's the best positioned to do. [64:10] I'll give you one example. There's a [64:11] great push that Jensen's making, which [64:13] he needs a partner, and I think Elon [64:15] becomes a natural partner, to do DC to [64:16] DC. Forget all this DC to AC to DC [64:19] nonsense that goes inside of a data [64:21] center. All the lossiness, all the [64:22] lossiness. [64:22] >> that is in English, yeah, for everybody. [64:24] >> Just like, look, you go through a bunch [64:25] of power transformations to to actually [64:28] deliver the electrons into the rack so [64:30] that, as Gavin said, you can generate [64:32] the token on the other end. [64:34] Today, it's it's very inefficient, it's [64:36] very costly, it requires a lot more [64:37] power, it requires a lot of cooling, it [64:40] requires complexity. And what people [64:42] have said is, "Wow, if we could just do [64:43] DC to DC like it comes in as DC, direct [64:46] current, it goes right to the rack as [64:47] DC." But it requires a fundamental [64:50] rearchitecture. [64:51] Jensen needs a design partner and a [64:53] thought partner to get that done. [64:56] He's probably the only one. So, I just [64:58] think there's a lot of reasons where you [64:59] can underwrite this to a multiple of [65:02] revenue plus the X factor, which is just [65:04] the creativity and the the one more [65:06] thing. [65:07] >> Love it. And then here's two charts, and [65:10] I'll have you comment on these, Gavin, [65:11] after it. [65:13] Here's the rocket sizes. [65:16] Just in terms of scale, and I most [65:18] people have not actually seen a Starship [65:20] in person. When you see this thing in [65:22] person, and I I've been inside that [65:24] rocket. [65:25] I think you were we were together, [65:27] Gavin, when we were in the first build. [65:29] And like inside of that, you can fit 300 [65:32] people. It's basically like [65:34] a giant [65:35] If you thought of a commercial air [65:37] aircraft, that's what it feels like when [65:40] you're inside, right? Like a 747 in [65:42] terms of the amount of space in it. [65:44] Especially when you compare the Falcon [65:45] Heavy, which is their workhorse, [65:47] correct, Gavin? [65:48] >> Yeah, and Starship's going to get [65:49] bigger. Based on their road map, it's [65:51] going to get a lot bigger. [65:52] >> A lot bigger. And then this one is the [65:55] most interesting that this started [65:56] trending this week. This is cumulative [65:58] payloads launched 1957 to today. SpaceX [66:01] is basically about to in just that and [66:05] and this is really what exponential [66:06] growth is about, and this is what [66:07] disruptive technologies are about. [66:10] Just from 2012 [66:12] to today, [66:13] SpaceX is about to dwarf [66:16] the rest of the world's cumulative [66:18] payloads into space. So, Gavin, maybe [66:23] take the other side of it. [66:25] When do these data centers in space [66:27] happen? What has to happen for those to [66:30] be a reality? When does that hit [66:32] SpaceX's bottom line? We've heard from [66:34] Chamath, "Hey, here's all the things [66:35] that hit the bottom line in the short [66:37] term and mid term." But I think data [66:39] centers in space would be a mid term to [66:41] long term play. 3 years is what I'm [66:43] hearing. So, [66:45] tell us about that business in relation [66:47] to the two charts I just shared. [66:49] >> Well, the one thing I would just say, [66:50] well, first, all those charts about [66:53] launch are before Starship was [66:54] operational, and most most of that mass [66:57] to orbit was done by Falcon. [66:58] >> Yes. [66:59] >> And Starship [67:01] the Falcon is reusable. Starship is [67:04] designed to be rapidly reusable. And [67:07] this is a critical difference. Like, [67:09] let's say Blue Origin successfully [67:11] solves reusability. [67:13] They're where SpaceX was 10 years ago. [67:16] Let's say China solves it 10 years ago. [67:19] Rapid re- The reusability means that you [67:21] extensively refurbish the rocket, [67:24] you know, the engines, everything, the [67:26] fairing. It takes a lot of time. [67:29] You know, maybe you can fly that rocket [67:31] again in [67:33] 30 days, 60 days. [67:35] Rapid reusability means that you can fly [67:38] the same flying land the same rocket [67:40] multiple times per day. [67:42] So, if SpaceX and and it's really hard [67:46] to do rapid reusability. [67:49] I think it would have been It would have [67:50] been not trivial, but much easier to [67:53] have Starships working if it was just [67:55] designed to be reusable. [67:57] >> Mhm. [clears throat] [67:57] >> That's not enough [67:59] for what Elon wants to achieve of, you [68:02] know, a a moon base, a colony on the [68:05] moon, a colony on Mars, mass drivers on [68:07] the moon. You need rapid reusability, [68:10] and that is why Starship is such a an [68:13] engineering challenge and will be such [68:15] an impressive achievement when they have [68:17] rapid reusability. But, I do think that [68:19] mass to orbit [68:21] rapid reusability of Starship means [68:24] if they get [68:25] >> you When do you predict they'll have [68:26] that, rapid reusability to space, you [68:28] think? [68:28] >> I mean, we're going to find out We're [68:29] going to find out, you know, we find out [68:31] I'm I'm I'm I'm going to be at Starbase [68:34] today for the whole launch. [68:35] >> Yeah. [68:36] >> So, you know, we we we turn over cards [68:38] and, you know, it's important for [68:39] everyone to remember, [68:41] like, let's just [68:42] let's just say it's a fireball. SpaceX [68:45] will still learn [68:47] from this. [68:48] >> Yes. [68:48] >> They They They learn from failure. If [68:50] you don't fail, you're not learning. [68:52] Same way, if you're not wrong, you [68:53] didn't learn anything in that day. [68:55] And this is a brand new rocket, a brand [68:57] new booster, a lot of new technology. [68:59] There's a lot of instrumentation on it. [69:02] So, whatever happens today, SpaceX is [69:05] going to learn and rapidly iterate. [69:07] I don't know when. I don't want to make [69:09] a prediction. [69:10] I would guess a year or two. [69:13] Maybe sooner. [69:14] >> I think that's most consensus. A year or [69:15] two is I think perfect consensus, yeah. [69:17] >> We'll see. So, like the Even if [69:19] everybody else solves reusability, [69:21] master orbit from everyone else will [69:22] quickly [69:24] asymptote to a very small number. [69:29] As far as when will orbital compute be a [69:31] reality, I would say, "Well, it is [69:33] important to realize there is a working [69:36] H100 and NVIDIA H100 GPU in space today. [69:40] >> Yeah. [69:40] >> Andrej Karpathy both trained a model on [69:43] and used for inference. So, this is [69:46] Yeah. [69:47] You know, it's it There's a working GPU [69:50] in space today. [69:51] >> And NVIDIA's making a space-designed [69:55] version of this, which will be different [69:57] because the heat sink has to be [69:58] different. There's a bunch of weight [69:59] that you put on it when it's in a data [70:01] center that you don't need in space. And [70:03] you also have to reinforce it for the [70:05] journey to space because these things [70:06] are going to shake and break apart. The [70:09] data center ones are not made to have [70:12] that many Gs put on them. So, you're [70:13] going to need an an an an industrial an [70:16] industrial-strength one that gets to [70:18] space that has a different profile. [70:19] Yeah, Gavin? [70:20] >> Well, one of the things that's been so [70:21] magical about SpaceX is they're very [70:23] good at engineering the rocket and the [70:25] payload so that you can use [70:28] semiconductors that are not designed [70:31] >> Hm. [70:32] >> be in space or satellites in space, and [70:34] those semiconductors are a lot cheaper. [70:36] We have a couple My my firm, Seraphim [70:38] Space, is an investor in a company [70:39] called Excite Labs that [70:41] it's a matter of public record is going [70:43] to be in essentially every Starlink. [70:46] And the chips were not designed to go to [70:49] space. They're not radiation-hardened. [70:52] You know, everyone, [70:53] you know, SpaceX really liked a lot of [70:55] the the specifications on the chips. and [70:58] then it's like, well, we'll see how they [70:59] do with rad testing and they just so [71:01] happen to pass. And so that is one of [71:04] like one of something that's very under [71:07] appreciated I think about space out. [71:08] >> One of their specialties. [71:09] >> One of their specialties. [71:11] >> Yeah. [71:11] >> But I think [71:12] second half of 28 to first half of 2030 [71:16] would be my point prediction. [71:17] >> All right, let's do Nvidia and then the [71:19] the market recap since we have you here [71:20] Gavin and since Freeberg you wanted to [71:21] get in on that. Nvidia blew out its [71:23] earnings again. [71:25] Q1 performance is just mind-boggling. [71:27] 81.6 billion in revenue up 85% [71:30] year-over-year, 20% [71:32] quarter-over-quarter. High growth in the [71:33] stock market for those people who don't [71:35] participate, 20% would be a high growth [71:37] company. [71:38] Year-over-year. They did that [71:40] quarter-over-quarter. 58 billion of net [71:42] income and 48 billion in free cash flow. [71:46] They're doing all this at 75% gross [71:48] margins. [71:50] They're growing massively and [71:53] they're obviously the most valuable [71:54] company in the world at a 5.3 [71:56] trillion-dollar market cap. Stock's up [71:59] but 16% year this year with all that [72:02] growth. That's a magnitude of that 16%. [72:05] And [72:07] they've announced another 80 billion in [72:10] additional buybacks on top of the 100 [72:11] billion in buybacks they did at the [72:13] start of 2023. So they're buying back [72:15] about 4% of the company. [72:17] They raised the quarterly dividend 25x [72:19] from 1 cent a share to 25% cents per [72:22] share. [72:23] And their CFO said they're going to [72:25] return 50% of the free cash flow to [72:27] shareholders. [72:29] Never been a company like this, huh [72:30] Freeberg? The the scale of this is just [72:33] extraordinary. [72:35] >> Mhm. [72:36] >> Yep. [72:37] >> Don't don't seem to [72:39] >> They have first there's your market [72:40] report from Freeberg. Don't seem so [72:42] enthused. [72:43] >> Yeah, it's a mhm mhm. He's got potatoes [72:45] in the oven. [72:46] >> I have a question for Gavin. He did a [72:48] really interesting talk with Patrick [72:50] O'Shaughnessy and there was this one [72:52] thing that I wanted to ask you about [72:54] cuz I thought it was so interesting. You [72:56] said when you look at the revenue [72:58] multiples of the chip companies and you [72:59] look at the revenue multiples of the [73:01] DRAM companies, both cannot be true. In [73:03] the context of Nvidia's earnings, can [73:04] you just explain maybe in plain language [73:07] for folks? I just thought it was so [73:09] fascinating cuz it explains it explains, [73:12] I think, just to set it up where is [73:14] value over the next 5 years? Like, I [73:16] think if you looked at Leo Ashan [73:17] Brenner, his fund has gone from like 0 [73:20] to 5 billion dollars overnight and it [73:22] looks like he's just got massive puts on [73:23] the chip sector and he's kind of [73:25] rotated. So, just give us context, [73:28] Gavin. Where Where's the puck going? [73:30] >> Well, so maybe take the question in the [73:32] reverse orders. For Leopold, who's [73:34] clearly a brilliant man, I think he was [73:36] a road scholar like 19 and I think my [73:39] understanding is he's putting up pretty [73:41] extraordinary numbers. I I've yet to [73:42] meet him. [73:44] He actually shares an office in San [73:45] Francisco with a friend of mine, so I [73:47] think I'll probably meet him sometime [73:48] soon. But, it's got to for that 13F that [73:51] he filed was at the end of the first [73:54] quarter [73:55] when, you know, I would say we were in [73:57] the, you know, the the thick of [73:59] geopolitical fears. [74:01] And I think you saw a lot of puts on a [74:03] lot of 13Fs and I don't know that those [74:06] puts are still there. [74:07] >> Okay. [74:08] >> You know, I think a lot of people wanted [74:09] to be hedged for on [74:11] and, you know, now I think it's a little [74:13] more clear. So, I wouldn't read I [74:15] wouldn't read Leopold's 13F as being [74:18] super negative on on semi. [74:21] >> Chips, okay. [74:22] >> On chips. [74:24] Second thing, I think cross actually, if [74:25] you look at the valuations for all these [74:26] AI names, they just they can't all be [74:30] accurate. You have memory makers that, [74:33] you know, [74:34] three to five times PE, you have Nvidia [74:37] at a really low PE. [74:40] You actually have, um, you know, some [74:42] other accelerator companies at [74:43] reasonable multiples. [74:45] And then you have everything else. [74:46] Everything in power, everything in [74:48] cooling. And when I say power, I don't [74:49] mean utilities. So IPPs are actually [74:51] quite [74:52] reasonably valued. But power, cooling, [74:55] even probably some of some of the [74:58] optical names. [75:00] These are discounting very different [75:03] things. If the multiples on the power, [75:05] cooling, optical names are correct, [75:09] Nvidia, memory, they're going up a lot. [75:12] If the multiples on Nvidia and memory [75:14] are are correct, [75:16] everything else is probably going to [75:17] underperform. The AI market is [75:19] cross-sectionally inefficient right now, [75:22] which is what I was trying to say. As [75:24] far as the Nvidia quarter, I do they [75:26] went to a new reporting structure, data [75:28] center and AI, and then with no data [75:30] center and edge, and then within within [75:33] AI they have hyperscalers, [75:35] and then I think they call it AI clouds, [75:37] >> AI clouds, [75:38] >> industrial and enterprise. What I [75:40] believe is if we were to make a true [75:42] apples-to-apples comparison, [75:45] and Broadcom, you know, there's a [75:46] narrative that Nvidia is losing share to [75:49] the TPU. [75:50] And Broadcom guided for 143% [75:54] year-over-year growth in their AI [75:56] semiconductor revenue in the quarter [75:58] that they will report. That's comparable [76:00] to the one Nvidia just reported. [76:02] I think [76:04] if you were to and I just [76:06] >> [laughter] [76:07] >> I so wish they had reported slightly [76:09] differently. I wish they'd done [76:10] hyperscalers, AI clouds, and then [76:13] industrial and enterprise. Because I [76:15] think the segment that is comparable is [76:18] the sum of hyperscalers [76:20] plus AI clouds, [76:23] stripping out China, because Broadcom [76:25] just did not have the China business [76:26] that Nvidia did. [76:28] And I think on that basis, in other [76:29] words, in within the western AI world, [76:32] within data centers that are being [76:34] built, whether they're being built by [76:35] CoreWeave, XAI, Amazon, Google, [76:39] >> Mhm. [76:39] >> Nvidia's AI business is growing faster [76:42] than Broadcom's. [76:44] And faster than a lot of other companies [76:47] that are, you know, seen as part of this [76:49] ASIC share gain story. [76:51] And, you know, I think Jensen has become [76:55] and you could you can hear it [76:58] increasingly frustrated, and rightfully [77:00] so, at two things. [77:02] One is to say, what is the performance [77:04] of stock? Uh [77:05] two two [77:06] >> vocal about that. Like, what is going on [77:08] here? You're putting up record numbers, [77:10] and we're getting no [77:11] like, uh credit. [77:12] >> Yeah, I get it. It It just [77:15] how can there be a share loss narrative [77:18] if I'm gaining share? And it is [77:20] indisputably true that he's growing [77:21] faster than hyperscaler CapEx, even [77:24] without these adjustments. [77:26] >> Yeah. [77:26] >> And I think the other thing that's so [77:27] frustrating to him is these other ASICs [77:30] >> Mhm. [77:30] >> are not being submitted for benchmarks. [77:33] They're not in the CV analysis in for [77:34] inference max. They're not in MLPerf. [77:38] And I think the reason they're not being [77:39] submitted is they will lose. [77:40] >> Mhm. [77:41] >> And you can't fight shadows. And until [77:45] we see a clean benchmark [77:48] of whether it's GB300s versus TPU [77:51] V7s or, you know, versus [77:54] >> versus inferentia, yeah. Yeah, yeah. [77:56] >> versus yeah, versus Tranium. [77:57] >> Yeah. [77:58] >> If we're we're not going to know. And [77:59] that's why the a lot of these other [78:01] chips [78:02] >> I think Tranium's in a great spot. [78:03] Aren't being submitted. [78:05] But nonetheless, if Nvidia's doing well, [78:07] once you become the largest company the [78:08] world, you kind of you tend to trade by [78:11] observation would be at stairstep [78:12] patterns. [78:14] Where you kind of the multiple [78:15] compresses, compresses, compresses, cuz [78:17] people are skeptical of the size [78:19] >> Then you have a re-rating. Yeah. [78:20] >> And then you re-rate. [78:21] >> New floor is established at a higher [78:23] rate. Yeah. [78:24] >> I think there was one other really [78:26] important thing in the Nvidia quarter. [78:27] It's they said that they thought their [78:29] CPU business was going to be $20 billion [78:31] this year. [78:31] >> Yeah, yeah, yeah. [78:32] >> Put it there. It means overnight you're [78:35] one of the world's largest CPU [78:37] manufacturers. And I think that is a [78:39] testament to Nvidia has a unique [78:41] position. They're the only company that [78:44] works with every lab. And so that puts [78:48] them in the best position to architect [78:51] their chips, they call co-designed, for [78:53] where the models are going. [78:55] And I think that $20 CPU figure is [78:58] pretty extraordinary. [79:00] >> This is the thing like at the end of [79:01] this Grok transaction last year, my kind [79:03] of prevailing thought on this is we're [79:05] going to move to these domain-specific [79:07] architectures. I thought that was like a [79:09] a fait accompli, we're just now waiting [79:11] for which models, but the reality is [79:13] that that DSA [79:15] market evolution is actually happening [79:17] inside of Nvidia. That's what's so [79:19] insane to me. That was my takeaway from [79:21] the quarter as well, which is like holy [79:23] These guys actually have domain-specific [79:25] architectures because they're doing [79:26] these design programs with every This is [79:28] why back to sort of the, you know, when [79:29] he does DC to DC with Elon and Colossus [79:32] 3 or whatever, it's just this it's [79:33] another game-changer for everybody. [79:35] >> He doesn't get a one-year flip. He makes [79:37] nine. And then I don't think the cost at [79:40] which you can finance these chips and [79:41] these useful lives is really important. [79:44] >> You had an incredible insight, which is [79:45] the amortization schedule for Coreweave [79:47] and all these guys, they got saved. [79:49] You may want to just explain what that [79:51] is and what why you said that. I thought [79:52] that was a great insight. [79:53] >> No, thank thank you Chamath, I [79:54] appreciate it. So, when Coreweave and [79:56] all these neo clouds came public, and by [79:57] the way, this goes for the hyperscalers [79:58] too, there was a big bear case that hey, [80:01] the these guys are amortizing their GPUs [80:03] and CPUs over four or five six years, [80:06] and that's way too short of a lifespan. [80:08] The true lifespan of a GPU is more like [80:10] two years, and therefore, you know, the [80:12] profits of all these businesses are [80:13] overstated. [80:15] The reality is [80:15] >> This was Michael Bury who put this out [80:17] there. [80:17] >> Yes, to be clear, yeah. Yeah, and you [80:20] know, thank you Michael Bury, we need [80:22] bears. Thank you. [80:23] >> Yeah, that's like that's like asking [80:25] Gerardo about modern music. [80:28] >> Well, I don't I don't want to cast [80:30] aspersions on Michael Bury, he's a he's [80:31] a brilliant man, but we need Barry. [80:33] >> Let me Hey, somebody call Vanilla Ice [80:35] and ask him what he thinks. [80:37] >> [laughter] [80:37] >> Oh my god. Milli Vanilli, check your [80:39] mail. [80:40] >> waste. What a waste of time. What a [80:42] waste of time. [80:43] >> Oh, poor Michael Come on the program [80:44] anytime, Michael Berry. [80:45] >> [laughter] [80:45] >> Go ahead, Gavin. [80:46] >> Yeah, I know. Keep it on track. [80:48] >> I'm happy to chat. But now that we've [80:49] disaggregated inference, we have these [80:51] different domain-specific accelerators. [80:53] You can mix and match them. And I think [80:55] the GPU stays in a lot of ways at the [80:57] center of this constellation for a [80:59] while. And you can put whether it's a [81:02] Groq accelerator, whether, you know, [81:05] it's a Cerebras accelerator in front of [81:07] old GPUs, use Groq or Cerebras for [81:10] decode. And then those older GPUs, they [81:12] have a useful life for 10 or 15 years. [81:14] And this means that you can finance GPUs [81:17] I think Corey's lowest financing, I [81:18] can't forget if it's 6 or 7%. [81:20] >> 6%. 6%. [81:22] >> Yeah. [81:22] >> And if you can get an asset-backed loan, [81:24] an asset-backed financing [81:26] for GPUs at a lower rate than other [81:28] chips. [81:29] >> No, that That's a profound advantage. [81:31] >> That quarter single-handedly saved the [81:33] Neo Neo's clouds. This quarter Nvidia [81:35] [laughter] [81:36] saved I mean, they single-handedly saved [81:38] them all. I think they they should all [81:40] They should all say an incredible thanks [81:42] to Jensen because [81:44] I interviewed the CEO of Core Weave, [81:45] Michael in Trader Michael in Trader, [81:48] yeah. And he was saying, "Hey, listen, [81:49] people have no problem [81:51] buying and financing these over a 6-year [81:54] period." And people are asking for [81:56] things that are coming off and that he [81:58] thinks they're going to have year 7 8 9, [82:00] they'll have some useful life. You know, [82:03] uh in addition to that. So, [82:04] Yeah, so he's like, "I I I don't know [82:06] what anybody's talking about here." [82:08] Like, this is just not informed analysis [82:10] was his point. Like, I the game on the [82:12] field and people are betting with their [82:14] dollars with him. He has them pay in [82:17] advance and sign 6-year contracts. If [82:19] they didn't think it had 6-year [82:20] lifespan, they wouldn't be signing a [82:22] 6-year contract. Pretty straightforward. [82:24] And they can't get enough of them. Okay, [82:25] let's end on this market update, macro [82:27] picture not great. Oil remains elevated, [82:30] although there might be a settlement [82:32] every week we have there is a settlement [82:33] coming. Maybe this time 16 time it's a [82:36] charm in the Iran war is going to wrap [82:38] up, but we're in week 12 of it. And this [82:40] was supposed to be 4 to 6 weeks. So, [82:41] wars never get resolved quickly. That's [82:43] one thing we've learned in our [82:44] lifetimes. Oil is driving inflation [82:47] massively higher. Polymarket says 99% [82:49] chance May inflation comes in at 4.2% or [82:52] higher. Survey of professional [82:54] forecasters projecting CPI hits 6%. You [82:58] heard that right, folks. We weren't just [82:59] talking about a 3% handle, which we just [83:01] said. Now people are saying 4, 5, and 6% [83:04] in Q2. And obviously, that's a huge [83:07] revision. And the narrative was, "Hey, [83:10] more Fed rate cuts coming." Now we're [83:12] talking about Fed [83:15] rate increases. Inflation is causing [83:18] obviously bond yields to rise. 10-year [83:19] hit 4.6%. You remember we've had Bessen [83:22] on the [83:23] pod multiple times, and his goal was to [83:26] get that under 4%. Now it's [83:28] significantly above that number. And [83:31] also, if we go around internationally, [83:32] Japan's 30-year is at a high of 5.1%. [83:36] Highest ever recorded. UK yields highest [83:39] since the great financial crisis. [83:40] Germany highest since 2011. And in [83:42] Korea, retail investors are borrowing [83:45] borrowing record amounts of money to [83:47] trade in AI chip stocks. They also had [83:50] an incredible run in Korea betting on [83:52] crypto at the peak. So, that's some sort [83:55] of an interesting signal. [83:57] Friedberg, is your alt personality going [84:00] to come out right now? Are you [84:01] concerned? [84:03] Is Dr. Doom making an appearance here, [84:05] or do you think this is manageable? How [84:06] concerned are you? [84:07] >> What is the point of being concerned [84:09] when you have ridden the roller coaster [84:11] to the top and it is beginning [84:13] [laughter] its descent? I I don't know [84:14] what there is to be concerned about. The [84:17] the the force of gravity is inevitable. [84:20] The roller coaster will roll down. We [84:22] will throw our hands in the air and we [84:24] will scream, "Whee!" as we go for the [84:26] ride. Global debt to GDP is 310%. [84:32] Reserve currency status [84:34] to the side, [84:36] the spending problem at the federal, [84:38] state, local level, the spending problem [84:41] at every country to basically keep [84:43] economies growing to support existing [84:45] leverage, ultimately creates a cascading [84:48] effect. It ultimately breaks. And as it [84:51] starts to break, you have massive [84:53] inflation because the value of your [84:56] underlying [84:57] currency collapses and then you have [84:59] money printing and all this other sort [85:01] of stuff, which inflates the value of [85:02] assets, which allows you to keep [85:03] servicing your debt. [85:05] And the spiral takes off. [85:07] And so we will just enjoy the ride. This [85:10] is the moment, you know, 30-year 5.2% [85:14] This Japanese yield [85:16] Some argue might, you know, you should [85:18] talk to more active market participants [85:20] than me, but [85:21] and probably some economists who trade [85:23] the market, but I would think that this [85:25] is one of those things that could be a [85:26] catalyst for a for a credit crisis. [85:30] Because there's a lot of people that are [85:31] in this carry trade. [85:33] And we'll see, you know, this is [85:36] >> Okay. [85:37] >> This is water leaking out of the bucket. [85:38] There it is. [85:39] >> he is. Dr. Doom is here. Chamath, your [85:41] thoughts on [85:43] Dr. Doom's panic attack of the month. Is [85:46] this Is this time real? Is this the 17th [85:49] prediction of the next six recessions? [85:52] What do we got here, Chamath? Are you [85:53] concerned? How concerned are you about [85:55] these signals that are flashing? [85:58] >> I think I think that's exactly what that [85:59] is, Jason. There are signals that are [86:00] flashing. I think there's pockets of the [86:02] market that still make sense that you [86:03] can underwrite if you [86:06] want to buy businesses that represent [86:07] the future. [86:09] And if you can find a few of those and [86:11] you can get comfortable with that and [86:13] you can own it for 10 years, [86:15] I think you buy those companies [86:17] and generally everything else I think [86:20] you should not speculate and you should [86:22] generally avoid. Not just because it's [86:24] an up market, but in every market. I've [86:26] learned this the hard way. We've all [86:28] kind of gone through this. As I get [86:29] older, I it's just not worth it. The [86:31] vicissitudes of the market [86:34] um don't give me [86:36] anywhere near the sugar high it used to [86:37] on the way up and it makes me feel [86:40] horrible on the way down. So, how I [86:41] manage myself is I have a few companies [86:44] that I really believe in. I have [86:45] extremely concentrated large holdings in [86:47] those. Large for me doesn't mean large [86:49] for everybody else. [86:51] And then otherwise, I just kind of stick [86:53] to my knitting and keep my head down. [86:54] It's a much more rational way to behave. [86:56] So, [86:56] >> How many public stocks can you keep in [86:59] your brain and still sleep at night [87:01] holding for the long term, Chamath? [87:03] What's the Is there a number for you? Is [87:04] it five? Is it 10? Is it [87:05] >> Seven. It's five It's five or less. [87:08] >> Five or less? [87:09] >> Five or less. [87:09] >> And so, what your largest holding right [87:11] now is what percentage of your net [87:12] worth, which is a [87:15] >> I don't know. [87:15] >> Or like top two, maybe. [87:16] >> Yeah. [87:18] Oh, top two? [87:19] >> Yeah, like one and two. One is 20, two [87:21] is 15, or one is 40, two is 20. [87:24] >> I don't know. Again, it depends on the [87:26] day. But I don't know. But it's it's [87:27] >> Just curious. [87:28] But I think that's I think that's really [87:29] important. [87:30] >> My point is there's no 30 things that [87:31] I'm tracking. I don't I don't have the [87:33] time. I'm not smart enough. [87:35] There's too much information. There's [87:37] like four things that I stay on top of. [87:39] >> Gavin, you you do this for a living. How [87:41] many positions do you manage and what's [87:43] your take on some of the flashing signs [87:45] that are saying, "Hey, slow down." or [87:47] maybe there might be a wreck around the [87:49] corner here, you know, when they do the [87:50] checkered flag in the F1 or whatever the [87:52] metaphor you want to use is. [87:54] >> Well, so what I I manage [87:57] more than 100 [87:58] positions at my firm. Um and I do that [88:01] with a team. We're over 30 people now. [88:04] So, it's not just me and I work with [88:05] some some great people. [88:06] I then [88:07] three things can be true. Rates going up [88:10] is very concerning. [88:12] What is happening with AI right now with [88:14] Anthropic growing faster than [88:17] any country any company in history at [88:19] massive scale. [88:20] >> And certainly any country. [88:22] >> Absolutely unprecedented. Yes. They're [88:24] actually Yeah, they are they are now the [88:25] size of you know, yeah yeah they're we [88:29] pull up where they [88:29] >> They're bigger than a hundred different [88:31] countries. [88:31] >> For sure. [88:32] >> Exactly. [88:32] >> And compounding really fast. [88:34] >> And now profitable. [88:36] >> Yeah. [88:36] >> Which I think really changes the moment. [88:37] >> Which is bizarre. [88:38] >> Yeah. [88:39] >> Yeah, how did that happen? [88:40] >> So, those things can both be true and I [88:43] think we should all remember the tech [88:44] bubble happened with you know, the [88:45] 10-year and the 30-year much higher than [88:47] they are today. And you know, [88:50] the the Nvidia of of of the tech bubble [88:52] was Cisco and it traded a hundred times [88:54] forward earnings. And you know, I think [88:56] Nvidia is probably at a low teens [88:58] multiple of kind of low to mid teens [89:00] multiple of real earnings, you know, if [89:03] you that that'd be a buy side consensus [89:05] not the trade number. And then the third [89:07] thing that I think is true is a straight [89:08] up formula is being closed. [89:10] While it's terrible for everyone, it is [89:12] relatively the best for America because [89:15] we are self-sufficient in energy, we are [89:17] self-sufficient in food, we've become a [89:20] massive exporter of oil, we're now the [89:22] world's largest not only oil and gas [89:24] producer, but oil and gas exporter. [89:28] And you know, those three things can be [89:29] true. What do they mean? I think it's [89:31] probably hard for me to see [89:33] with America being the most advantaged [89:36] by what is going on. [89:38] >> We're still the best currency. [89:40] >> Yeah. [89:40] >> the best economy. We still have the best [89:42] public companies and we have the best [89:44] private market companies. [89:46] >> Yeah. [89:46] >> And we are [89:48] one of the greatest producers of oil in [89:50] the world. So, we're in good shape [89:52] despite this international chaos. [89:55] >> think a dollar crisis is around the [89:56] world. Now listen, like if you know, the [89:58] Bundesbank [90:00] in was still in charge and you had the [90:02] Deutsche Mark and there was a currency [90:04] with better fundamentals, we would be at [90:06] very high risk, probably. But just cuz [90:09] we're the best house in what is globally [90:12] a, you know, bad neighborhood of high [90:14] debt levels, and we have AI in our [90:17] corner, and we have energy [90:19] self-sufficiency, and every day the [90:21] Strait of Hormuz is closed, I think it's [90:23] relatively good for the [90:24] reindustrialization of America. [90:27] Like I think, you know, you have to you [90:29] have to all [90:30] >> You're saying it's a forcing function. [90:32] It makes us, just like COVID did, be [90:35] more resilient and, yeah, more [90:37] self-reliant. [90:38] >> Yes, and [90:41] electricity is a base input to every [90:44] manufacturing or industrial process, [90:46] essentially all of them. [90:48] >> Mhm. [90:48] >> And what we make electricity with in [90:50] America overwhelmingly is natural gas, [90:52] and you can look it up. NG1, it is down [90:55] this year. [90:56] >> Mhm. [90:56] >> The input cost for electricity in the [90:58] rest of the world, you know, lots of [91:00] different things, but LNG is a very [91:02] important one, and it's up 100, 200%. [91:06] And so, the Strait of Hormuz is [91:07] absolutely [91:09] bad for everyone, but relatively good [91:11] for America, [91:13] and relatively good for Trump's policy [91:15] goals. [91:16] >> Mhm. [91:16] >> And that's why I think he's in no hurry. [91:19] Every day the Strait is closed, and for [91:21] whatever he does seem like a relative [91:23] thinker. Every day the Strait of Hormuz [91:25] is closed is relatively good for [91:28] America. It's terrible for Europe. It is [91:31] terrible for Asia. [91:33] >> Japan and China need that oil. [91:34] Philippines needs it. India needs it. [91:37] Yeah. [91:37] >> Yeah, it's so all these things can be [91:39] true, but the one thing I do just want [91:40] to say is rates going up [91:42] and inflation going up is never good. [91:46] But we have to hold what's happening [91:47] with AI, where the fundamentals are [91:49] getting a lot stronger in our mind. And [91:52] the the one thing I would just add is [91:55] AI has been seasonal, the market's [91:57] seasonal. [91:58] You know, it often, you know, sell in [92:00] May, go away. [92:02] And AI fundamentals also appear a little [92:06] seasonal. [92:08] In the past that's been cuz you know, [92:09] cuz college students they they use a lot [92:11] less chat GPT [92:13] and Claude in the summer. And generally [92:15] people maybe work a little less hard [92:17] when the weather's nice. Now with the [92:18] generative AI, will the fundamentals [92:20] still be seasonal? We will see. [92:23] >> Oh, that's a really interesting point, [92:24] right? We see that e-commerce, apps, [92:26] subscriptions as investors in a lot of [92:28] these companies, we would always have [92:30] these board meetings Chamath. Oh, uh Q3, [92:33] uh yeah, people are out gallivanting and [92:36] they're not playing Candy Crush or [92:38] buying calm or whatever, but oh hey, [92:40] Uber and DoorDash went up, people are [92:42] traveling, etc. Okay, final story of the [92:44] week. We had a 48-hour jaunt [92:47] by a bunch of tech CEOs and the [92:50] president [92:51] to hang out with Xi. A lot of high [92:53] fives, a lot of handshakes, [92:55] a lot of great vibes, but coming out of [92:58] it we haven't seen anything [93:00] >> definitive Friedberg [93:02] in terms of policy. This was supposed to [93:04] be some big breakthrough. It would have [93:07] downstream effects on tariffs, on us [93:10] selling chips to China. We did see a [93:12] little movement there uh and the Strait [93:14] of Hormuz, we would become, you know, [93:16] Wonder Twins with Xi and Trump reopening [93:19] it, but nothing really definitive other [93:21] than some soybeans being sold and some [93:25] H100s, A200s getting sold to Baidu and [93:28] some of the top folks. So what do and [93:30] maybe some planes got sold, too. So [93:32] other than a little BD, a little [93:33] business development is [93:35] uh Friedberg, what's the outcome here or [93:37] was it just a bit performative in your [93:40] mind? [93:40] >> There was a a question [93:44] would the administration [93:46] leave China with a grand deal that made [93:48] everyone feel like there's a long-range [93:51] view on a partnership, but I don't think [93:52] that that's [93:53] what happened. There were a few [93:55] announcements obviously around an [93:56] intention to continue to work together [93:58] in a cooperative way and find a path to [94:00] partnership, an intention to [94:03] establish additional trade deals, and [94:05] and, you know, there were some purchases [94:07] of aircraft and [94:08] and some agricultural product [94:10] commitments, but fundamentally, the [94:11] grand deal, the big deal that I would [94:14] say reduces or de-escalates tension [94:17] probably didn't manifest as some had [94:19] hoped it would. And I don't think it's [94:21] any surprise that Putin is with Xi [94:24] today. [94:25] And this is also performative, that [94:27] following the US visit, there's now a [94:29] relationship bonding moment happening [94:32] between China and Russia. [94:34] So, the story continues. You know, there [94:37] is no happy ending, and there is no [94:39] rainbow-colored chapter three in this [94:41] book. It's going to continue to be a [94:44] dramatic arc [94:45] as this rising power continues to [94:47] challenge the United States. [94:50] And I think the story continues. [94:51] >> Were you expecting a happy ending? [94:54] >> CAN'T ANSWER THAT QUESTION. [94:56] >> [laughter] [94:58] >> I MEAN, TO THE STORY OF THE CHINA VISIT. [95:01] Not not I'm not talking about any other [95:03] things going on in your life, but [95:05] it seemed like some planes and soybeans [95:07] got sold, some H-100s, perhaps, but [95:10] >> the [95:11] >> it wasn't like there was some grand deal [95:13] that occurred, but it's nice to see them [95:15] together, right? I mean, that is nice. [95:16] >> I think it was successful. I think [95:18] there's what you see on the surface, and [95:19] then there's what happens behind closed [95:21] doors, and [95:22] without speculating too much, I think [95:23] that it was a useful and productive [95:25] trip. I think the the biggest thing that [95:27] they probably got alignment on is just [95:29] geopolitically, the tic-tac-toe of what [95:32] has to happen next. And and I think that [95:34] there's some amount of agreement there. [95:36] I'm just guessing. [95:37] >> Yeah, and so that guess, if I was going [95:38] to unpack it, hey, we got we have [95:42] Venezuela, we have Iran, and Taiwan. [95:45] You have lots of things to [95:47] >> here's [95:48] geopolitical chessboard, and [95:51] here's what I would say, just very [95:52] generally. Like it's just [95:54] I think that there's a that there's a um [95:57] a way to divide up the game board in a [96:00] way that helps them and helps us. [96:03] >> Mhm. [96:03] >> [clears throat] [96:04] >> Gavin, any thoughts on specifically [96:06] Nvidia being able to sell more chips [96:08] into China? Material for the company? [96:12] Good for America? I mean, it's obviously [96:15] a pretty debatable issue. [96:16] >> I think I disagree with Chamath on this. [96:18] I think selling [96:21] depreciated Nvidia GPUs [96:24] to China [96:26] lowers the odds of them developing their [96:28] own alternative ecosystem, which would [96:31] be a lot power hungrier. Um cuz you do [96:33] use optical you bring in optical a lot [96:36] earlier for scale-up fabrics. I think [96:37] there's sound arguments that this [96:41] is stabilizing for the world [96:44] and is the best highest probability path [96:49] for keeping America ahead in AI and kind [96:52] of keeping control of AI. [96:54] And that's almost a shame we've had to [96:56] have this debate because now people like [96:58] me have said this many times and try to [97:00] If they didn't understand it, they [97:01] probably do really understand it. And by [97:02] the way, that's not to say we shouldn't [97:03] have had the debate. [97:05] But um [97:07] that is what I believe, you know, [97:09] reasonable minds can disagree. [97:11] >> No, wait. What Where do you think we [97:12] disagree? I'm not sure I agree with you. [97:14] >> Oh, good. I'm glad we [97:15] >> I was going to I was going to jump in [97:16] No, no, you start you start with I'm in [97:18] agreement on [97:18] >> No, no, no. I'm like sell [97:20] >> I'm like sell everything to them. No, [97:21] what I was just saying is that there was [97:23] what you see on the surface of what they [97:25] can speak to the press, but I think the [97:26] most important thing was the negotiation [97:29] of hey, listen, like [97:31] we're going to do these things, you do [97:32] these other things. [97:33] And that's never going to get put out in [97:35] a multi, you know, memoed press release. [97:37] That's my point. That's all I'm saying. [97:39] >> Yeah. And I I would just say, listen, it [97:41] like [97:42] America and China talking is only good. [97:44] >> Only good. [97:45] >> the these trap that is being discussed. [97:48] And China talks about a lot. They're [97:50] very aware of it. And they brought it [97:52] up. [97:53] >> [laughter] [97:54] >> She brought it up by name. Yeah. Talking [97:56] is a [97:57] integral step of avoiding this [98:00] situation. [98:01] Having a nice resolution. From my [98:03] perspective, like the greatest [98:04] superpower Trump has is his ability to [98:07] bond with dictators, monarchs, royal [98:12] families, Gulf monarchies. He's just [98:15] great at it. They see eye to eye. They [98:17] vibe. He has no problem going to see [98:19] them. He has no problem inviting them to [98:21] UFC fights. Like this is like if if she [98:24] comes to the United States and he's [98:25] sitting courtside with Dana White, like [98:27] that's when we know things are going to [98:28] be okay. I do think he's probably giving [98:31] them the green light on like, "Hey, [98:33] Taiwan's yours. Just let's not have it [98:35] during my administration. Maybe like we [98:38] do a 30-year deal or a 20-year handoff [98:40] deal." I wouldn't be surprised if [98:42] something like that happens. Or a [98:43] 100-year deal or a 200-year deal. But [98:45] the one thing I would just say that I'm [98:47] sure was communicated is, "Hey, [98:50] wars consume a vast amount of oil. [98:53] You buy your oil from Iran, [98:55] Venezuela, and Russia. Russia alone can [98:58] supply a fraction of what you need. [99:01] And now it should be clear to the world [99:03] Two of the three are off the chessboard. [99:05] Yeah, if you Two of three are off the [99:06] chessboard, if you do something we don't [99:08] like, [99:09] Venezuelan oil, gone for you. American [99:12] oil, gone. Brazilian oil, gone. [99:16] And we'll say to all of our good friends [99:18] in the GCC, "We're so sorry, but we have [99:20] to close the Strait of Hormuz again." [99:22] So, Iran, all of Middle Eastern oil, [99:26] gone for China. Now you just have [99:28] Russia, and good luck fighting a war [99:31] with just Russian oil against us, [99:35] Japan, [99:36] South Korea, Australia, UK, France, [99:39] Germany, the world. I mean, who knows [99:42] about Europe? But for sure Japan would [99:44] be there. For sure Japan would be there. [99:47] Oh, and Australia for sure, Korea for [99:49] sure. Yeah. [99:49] >> And so I think it's going to be a more [99:51] stable world on the other side of Iran. [99:56] However it resolves, and I think that's [99:58] nothing but good. [99:59] >> Yeah, I think that's a good insight. I [100:00] think it's a great insight. All right, [100:02] listen. [100:03] We missed you Sachs. Come back soon. [100:06] And Gavin. [100:07] >> Big shout out to bestie Gavin. Thank [100:09] you. [100:09] >> Yeah, you're so great. Thanks for [100:11] coming. We appreciate it. Hey, and your [100:13] father-in-law, what's his name again? [100:15] >> Jeff Painter. [100:16] >> Jeff Painter. [100:18] We love you. Thank you so much for all [100:21] the kind words. We'd love to invite you [100:22] to come to Liquidity or the Summit. I [100:24] know you're a big fan of the show. Want [100:26] to give you a shout out here on the [100:27] show. [100:29] Thanks. I used your fandom of the show [100:33] to leverage Gavin who was like, "I can't [100:34] make it today." And I was like, "Tell [100:35] your father-in-law [100:37] we're going to get him backstage VIP [100:40] tickets to the next two events [100:42] if you show up today. And if you don't, [100:44] I'm going to back channel it to him." [100:45] And all of the sudden Gavin made it to [100:47] the show. [100:48] Love you boys. A little bit of pressure. [100:50] That's my way of the That's my hormone [100:52] straight. [100:53] >> Yeah, so I'm always there for my [100:55] father-in-law. [100:55] >> Absolutely. All right, everybody. We'll [100:57] see you next time. Bye-bye. [101:00] >> [music] [101:01] >> You let your winners ride. [101:04] Rain man David Sacks. [101:09] We [music] open sourced it to the fans [101:11] and they've just gone crazy with [101:15] >> [music] [101:21] >> Besties [music] are gone. [101:31] >> We should all [music] just get a room [101:33] and just have a one big huge orgy cuz [101:34] they're all just useless. It's like this [101:36] like sexual tension that they just need [101:38] to release somehow. [music] [101:40] >> Wet your bed [101:41] Wet your bed Wet your bed [101:44] Wet [101:44] >> [laughter] [101:45] >> We need to get merch. [101:46] >> Merches are back. [101:46] >> I'm going all in. [101:48] >> [music] [101:54] >> I'm going all in.