SpaceX IPO Pop Precedes Shift to $50 Billion Gigawatt Data Centers

SpaceX IPO大涨，随后转向500亿美元吉瓦级数据中心 Gavin Baker认为，在廉价的Starship发射将轨道算力成本减半之前，SpaceX将通过地面数据中心的快速变现来推动近期估值。 Gavin Baker分析了SpaceX IPO的顺利推进及其近期市场驱动因素，重点强调了其快速将吉瓦级数据中心算力变现以及扩展Cursor等企业工具的能力。他指出，在全球算力短缺的背景下，未来5至10年仅作为纯粹的“token工厂”运营便已足够，同时实体基础设施资产（EV / Net PP&E）将稳步升值。讨论还涉及利用太空算力绕过地面瓶颈成本的兴起、主权AI战略对抗美国前沿实验室的现实局限，以及SpaceX成为有史以来最重要公司的潜力。 SpaceX 顺利完成 IPO 并在高盛与摩根士丹利的主导下首日上涨约 20%；其近期核心估值驱动力在于快速落地的地面算力变现（如与 Google 达成的 500 亿美元/GW 协议）及 Cursor 的企业渗透，而非长远的深空项目。 (0:00) 鉴于全球算力持续处于赤字状态，云基础设施提供商在未来 5 至 10 年内仅作为纯粹的“Token 工厂”运营便已足够；同时，衡量实体电力与数据中心资产的 EV / Net PP&E 指标日益受到市场重视。 (4:46) 散户投资者在 2024 与 2025 年的表现大幅超越了绝大多数专业基金经理；与此同时，针对小众硬件供应链的“瓶颈交易”投机正走向尾声，资本开始重新青睐具备持久特许经营价值的企业。 (12:41) 相比地面数据中心单吉瓦约 600 亿美元的建设成本，依托 Starship 完全可复用发射的太空轨道数据中心单吉瓦成本预计仅为 300 亿美元，被动散热与持续太阳能供电展现出极具吸引力的经济可行性。 (19:03)

Investor Gavin Baker argues that while SpaceX holds unmatched long-term potential as a multi-planetary infrastructure provider, its public market valuation will primarily be driven in the near term by rapid power monetization and operating as an AI 'token factory.' Near-Term Token Monetization Anchors SpaceX's Valuation (0:00) Following SpaceX's public listing managed by Goldman Sachs and Morgan Stanley—which saw a clean ~20% pop—Baker contends that public markets will evaluate the company on concrete near-term execution rather than distant speculative projects like Mars colonization or lunar mass drivers. The primary near-term growth drivers over the next year are terrestrial compute scaling and rapid enterprise adoption. SpaceX reportedly ordered 20% of NVIDIA's Rubin chips and is monetizing gigawatts at an estimated $50 billion per gigawatt on contracts like its Google partnership, roughly double to triple standard Neocloud pricing. Baker addresses the operational tension between reserving compute for internal frontier model development and selling cloud capacity to third parties like Google and Anthropic. Rather than needing to build out a comprehensive enterprise cloud with storage and database services mimicking Amazon Web Services, Baker argues that serving purely as an infrastructure 'token factory' will be more than sufficient for the next 5 to 10 years given the persistent global shortage of high-density compute capacity. Orbital Compute and Infrastructure Cost Disruption (19:03) Baker argues that the investment strategy focused on hunting physical land and supply-chain bottlenecks is nearing exhaustion, as key suppliers have declined to raise prices indefinitely. Instead, the ultimate disruption to land constraints lies in space. Baker calculates that building a 1-gigawatt terrestrial data center costs roughly $60 billion—comprising $25 billion for power and cooling infrastructure alongside $35 billion for IT hardware. By contrast, orbital data centers could halve total capital expenditure to approximately $30 billion per gigawatt. In orbit, continuous solar power and passive thermal cooling eliminate the $25 billion terrestrial facilities cost, replaced by an estimated $5 billion in launch expenses via reusable Starships. Meanwhile, edge compute platforms like Content Delivery Networks (CDNs) capture high-margin premiums for latency-critical voice applications, but Baker estimates CDNs currently deliver under 1%—and likely less than 0.1%—of total global AI tokens. Divergence in Frontier and Sovereign AI Strategies (21:31) Baker forecasts that frontier foundation model development will remain confined to the United States and China due to the intense capital, engineering talent, and live human feedback loops required for frontier breakthroughs. Consequently, other nations pursuing sovereign AI will largely abandon building frontier models from scratch. Instead, their standard playbook will involve customizing open-source base models using reinforcement learning, supervised fine-tuning, and local system prompts, hosted in domestic data centers for national security and government operations. Simultaneously, Baker observes that Chinese AI labs are slipping further behind the global frontier. He attributes this gap to Beijing's restrictions preventing domestic firms from importing advanced chips like NVIDIA's H200s or B200s, combined with frontier American labs locking down access to reasoning traces and APIs, which effectively cuts off the industrial-scale model distillation Chinese developers previously relied upon. Generational Scale and Public Market Realities (25:18) Drawing on his early investments in Tesla and NVIDIA, Baker describes SpaceX as a uniquely generational enterprise that could ultimately evolve into the commercial backbone of the solar system. He defends large-scale wealth creation by citing Elon Musk's role in accelerating clean energy adoption, Starlink's connectivity programs, and the internal liquidity created for more than 10,000 SpaceX employees who participated in the IPO after a decade of structured secondary tenders. Contrasting SpaceX's institutional scale with broader public market trends, Baker notes that retail investor indices significantly outperformed professional money managers across 2024 and 2025. However, he warns that public micro-cap equities under $2 billion have turned into an unregulated speculative arena where coordinated social media campaigns inflate micro-cap valuations, echoing dynamics from the late-1990s dot-com bubble. While Baker outlines an expansive thesis for SpaceX spanning space-based compute and interplanetary exploration, the company's valuation remains anchored to near-term execution in physical data center deployment, GPU supply allocation, and high-margin compute contracts.

投资人 Gavin Baker 认为，SpaceX 近期核心价值主要由地面算力中心扩建与高效商业化变现驱动，而非遥远的天体探索；未来若实现 Starship 的完全可复用，太空轨道计算有望在成本上击败地面基础设施，使 SpaceX 成为人类历史上最重要的商业实体。 近端算力变现与“代币工厂”模式 (0:00) Baker 驳斥了公开市场无法承受重资产资本支出周期的观点，指出 SpaceX 的 IPO 在高盛与摩根士丹利的承销下稳健上涨约 20%。他认为公司股价在未来一年的核心驱动力在于地面算力的快速扩张：SpaceX 在将吉瓦级电力变现为数据中心方面的速度领先行业，其与 Google 的合作达到每吉瓦 500 亿美元的变现水平，是普通新兴云服务商定价的两到三倍。硬件方面，SpaceX 订购了 NVIDIA 约 20% 的 Rubin 芯片并拟整合 Groq LPU，同时其赋能的 Cursor 已渗透至约 50% 的财富 500 强企业。 面对算力分配中的内外博弈，Baker 将 SpaceX 对外输出算力的行为解读为对其持续并网吉瓦级电力能力的绝对自信。他指出，鉴于全球算力长期处于赤字状态，SpaceX 乃至 Meta 等巨头在未来 5 到 10 年内无需急于构建涵盖冷存储、缓存等全套工具链的传统云架构，专注于充当高产出、低复杂度的“纯代币工厂”（Token Factory）便足以攫取巨大市场价值。在此背景下，“企业价值倍数除以固定资产净值”（EV / Net PP&E）正在成为衡量实体数据中心与电力稀缺性的关键估值指标。 边缘网络现状与太空算力经济学 (19:03) 在探讨算力交付路径时，Baker 指出尽管部分内容分发网络（CDN）签署了数十亿美元的高额合同且终端用户愿为极低延迟支付溢价，但 CDN 目前交付的全球 AI 代币总量可能不足 0.1%，绝大多数算力仍集中在大型核心数据中心内部运行。更重大的行业转折在于过去一年追逐单一供应链短板的“瓶颈交易”正在失效，类似味之素（Ajinomoto）等半导体封装材料供应商并未无限制提价，市场正从寻找偶发瓶颈转向评估长期商业壁垒。 关于土地与电力瓶颈，Baker 提出了轨道算力替代地面算力的成本模型：在地球上建设 1 吉瓦级数据中心需耗资约 600 亿美元（250 亿美元用于电力与冷却，350 亿美元用于 IT 硬件）；而在太空中，由于具备持续太阳能供电与被动辐射散热，可直接免去 250 亿美元的能源与冷却成本。一旦 Starship 实现完全可复用，单吉瓦有效载荷的发射成本预计可压缩至 50 亿美元，使轨道吉瓦级数据中心的总体成本降至约 300 亿美元，仅为地球建设成本的一半。 主权 AI 的分化与前沿模型护城河 (21:31) Baker 预测，除美国和中国外，其他国家自主训练达到前沿水平（Frontier）的主权模型极难成功，原因在于顶尖模型研发不仅依赖资金与芯片，更需要顶尖团队和海量连续使用反馈形成的数据飞轮。其他国家的主流可行路径将是基于优质开源底座，利用本国语言、文化和法律价值观进行监督微调（SFT）和强化学习（RL），并将模型部署于本国受控的物理数据中心以满足国防、政务及公共安全需求。 就大国竞争而言，Baker 认为中国实验室因难以获得顶级芯片（如 H200 或 B200）而在算力硬件上进一步承压。过去通过设立大量 API 代理端点采集数以万计推理过程（如提取 o1 或 o3 轨迹以构建 DeepSeek）的工业级知识蒸馏模式，正在随着 Anthropic 等前沿机构对思维链推理轨迹、合成数据及访问接口的严格封锁而难以为继。 资本结构异化与市场投机周期 (12:41) 在资本市场流动性方面，Baker 修正了对散户投资者的偏见，指出散户在 2024 至 2025 年的投资表现显著优于大部分专业基金经理。SpaceX 在 IPO 之前长达十年中每半年提供一次内部流动性回购，使得早期投资者与老员工能够提前变现，加上本次 IPO 中超万名员工直接申购股票，形成了不同于传统新股上市的健康筹码分布。 与此同时，Baker 警告了当前公开小盘股市场（低于 20 亿美元市值）的投机乱象。与拥有严格合规监管的卖方研报不同，金融社交媒体目前处于缺乏监管的野蛮状态，匿名账号可通过海外工具操纵数千万美元市值的微盘股暴涨十倍。他将此现象与 1990 年代末的互联网泡沫类比，指出狂热投机最终会引来监管整肃，但这一过程中投入的资本客观上也为少数具备时代级影响力的基础设施奠定了基础。 Baker 将 SpaceX 比作未来太阳系中的“东印度公司”，但其商业叙事的完全兑现仍建立在关键假设之上：短期内取决于地面超大规模数据中心的如期并网与前沿代币定价权的维持，长期则完全取决于 Starship 能否如期实现常态化、低成本的轨道复用回收。

## Background 该内容由科技商业频道 TBPN 于 2026 年 6 月 23 日发布在 YouTube 平台上。页面简介显示，本期节目是主持人与知名投资人 Gavin Baker 的完整对话，围绕 SpaceX 上市后的市场反响展开，深入探讨了其为何有望成为史上最伟大的企业、AI 基础设施与“Token 工厂”对资本市场的重塑、主权 AI 的现实局限，以及数据中心、轨道计算与 AI 决战终局的未来演变。 ## TL;DR 加文·贝克深入拆解了 SpaceX IPO 后的资本结构与长线逻辑，指出其核心驱动力在于地面算力的极速交付变现与 Cursor 等 AI 应用的闭环落地。他分析了超大规模云计算巨头的资产定价与竞争博弈，指出上游盲目的“瓶颈交易”正走向终结，并算清了未来依托星舰复用大幅降低成本的“轨道计算”经济账。贝克还剖析了中美 AI 模型代差扩大的原因与工业化蒸馏的瓶颈，并直言绝大多数主权 AI 仅能立足于安全防御。最后，他警示了当前社交媒体上利用匿名马甲操纵 AI 微盘股的投机乱象，强调在合规真空期投资者需独立审慎调研。 ## Atmosphere & style 整场对谈兼具华尔街顶级对冲基金合伙人会议式的严密算账，以及硅谷硬核科技信徒的前瞻狂想。贝克语气笃定、反应机敏，既有早期重仓英伟达与特斯拉的战绩底气，又充斥着对航天与算力工业的真挚敬畏；主持人与嘉宾之间默契十足、抛接迅速，在拆解复杂金融数据与回击投机泡沫之余爆梗不断。听众如同置身于一场高密度拆解未来商业命脉的思想风暴中，既能感受到星舰呼啸升空的史诗震撼，又能嗅到资本市场清醒冷峻的周期机警。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=0s) SpaceX IPO 表现稳健首日涨幅达 20%，不仅已锁定英伟达下代 Rubin 芯片 20% 的产能，在与谷歌的交易中地面算力变现效率更高达每吉瓦 500 亿美元。 - [20:00](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1200s) 轨道计算展现出显著的成本优势：一旦星舰实现完全复用，部署 1 GW 太空算力仅需约 300 亿美元，较地面数据中心的 600 亿美元直接减半。 - [18:27](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1107s) 资本市场对上游物理瓶颈的炒作（如味之素 ABF 薄膜）已逼近强弩之末，企业拒绝盲目涨价显现出金融炒作与产业实操的巨大脱节。 - [23:43](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1423s) 海外前沿实验室加速收紧思考链（CoT）日志与推理输出，国内单纯依靠抓取 API 推理轨迹的“工业化蒸馏”套利窗口正在迅速关闭。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=0s) 公开市场对于具备宏大愿景和巨额资本开支的企业具有远超 VC 预期的长期耐心，正如早期的特斯拉和亚马逊建设 AWS 时一样。 - [01:58](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=118s) SpaceX 短期商业价值的核心支柱在于地面算力的极速交付与通电变现，而非遥远的火星或小行星开采计划。 - [04:46](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=286s) AI 基础设施在未来 5 到 10 年的核心价值在于成为高吞吐的“Token 工厂”，并不需要在短期内照搬传统 AWS 式的全栈数据库与存储堆栈。 - [07:19](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=439s) 投资估值正在向“企业价值 / 净固定资产”（EV / Net PP&E）倾斜，已通电部署的物理硬件资产具有高价值与低陈旧风险（HALO 交易逻辑）。 - [09:53](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=593s) 科技巨头在 AI 时代的算力商业化策略具备极高灵活性，股价与股权激励是留住顶尖“万倍工程师”的生死线。 - [12:41](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=761s) SpaceX 员工在 IPO 中自掏腰包大量加仓，且过去十年频繁的要约回购已充分释放套现需求，解禁抛压被外界严重高估。 - [16:14](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=974s) CDN 服务商虽占据低延迟端侧优势，但实际处理的全球 Token 吞吐量不足 1%，复杂的深度推理仍将在中心化数据中心闭环完成。 - [18:27](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1107s) 追逐供应链上游所谓垄断瓶颈的“瓶颈交易”正在衰退，单纯的材料或土地卡位无法保证持久的特许经营价值与长期定价权。 - [20:00](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1200s) 星舰完全可重复使用将颠覆算力经济学，太空轨道计算凭借自然冷却与太阳能优势可规避地面面临的电冷基建通胀压力。 - [21:31](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1291s) 除中美两强外，多数国家的主权 AI 终局主要定位于开源模型加本土文化微调的防守型应用，而非追赶前沿基础大模型。 - [23:43](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1423s) 依靠数万个 API 节点逆向抓取思考链的工业化模型蒸馏模式，随着前沿实验室封锁数据通道而即将走到尽头。 - [28:00](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1680s) 若能在地外天体建立持久城市，SpaceX 的终局将演变为贯穿整个太阳系商业生态的“不列颠东印度公司”。 - [30:22](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1822s) 公开二级市场正充斥着利用匿名社交账号拉抬微盘 AI 概念股的投机操纵行为，相比 2000 年互联网泡沫期面临更严峻的合规监管真空。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=0s) SpaceX 成功 IPO 与公开市场的长期耐心 主持人与 Gavin Baker 就 SpaceX 备受瞩目的首次公开募股（IPO）展开讨论。主持人感慨整个上市流程执行得极其顺畅，股价涨幅约 20%，几乎完全符合投行的预期，甚至平稳到没有任何剧烈起伏的戏剧性场面。Gavin 对承销商高盛（Goldman Sachs）和摩根士丹利（Morgan Stanley）的出色运作表达了高度赞赏。随后主持人抛出核心疑问：在经历了上市前的多年长线叙事后，公开市场是否会像对待普通企业一样，开始对 SpaceX 进行严苛的逐季（quarter by quarter）业绩考核？ Gavin 借用 X（前 Twitter）上长跑爱好者的争论打了个生动的比方：有人说“马拉松就是 26 次 1 英里连着跑”，虽然跑单英里与跑完全程本质不同，但每一个季度显然都至关重要。不过，回顾特斯拉在过去五六年的表现，以及亚马逊当年为了建设 AWS 和零售仓储物流网络而经历重度资本开支周期的历程，公开市场对大规模前瞻性投资的容忍度和时间跨度，远比风险投资（VC）圈所认为的要宽容得多。 - 主持人称赞投行操盘精准，首日股价涨幅达到教科书般的 20%（"20% pop almost to a tee"）。 - 承销投行：高盛（Goldman Sachs）与摩根士丹利（Morgan Stanley）。 - 长期视角参照物：特斯拉（Tesla）在产能扩张期的资本容忍度，以及亚马逊（Amazon）在重投入建设 AWS 和线下基础设施时的市场反馈。 --- ### [01:58](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=118s) 近期核心变量：地面算力部署与 Cursor 商业飞轮 Gavin 介绍了同在现场的 Atreides 分析师 Foxy，指出外界对 SpaceX 价值周期的判断存在误区——其商业驱动力其实远比大家想象的更切近现实。Gavin 明确提出未来一年决定公司走向的两个关键变量：第一是其地面算力（terrestrial compute）的通电与交付速度；第二则是类似 Cursor 这样的顶尖 AI 应用带来的实际商业落地。 关于算力货币化能力，根据 Altimeter 的估算数据，SpaceX/xAI 在地面算力上的变现效率极其惊人，每吉瓦（GW）的变现收入可达平均 Neocloud 价格的 2 到 3 倍；在与谷歌（Google）达成的大宗交易中，甚至达到了每吉瓦 500 亿美元的规模。英伟达 CEO 黄仁勋（Jensen Huang）也公开证实，SpaceX/xAI 团队搭建与通电数据中心的速度超过业内任何竞争对手。与此同时，在软件侧，Cursor 的 Composer 2.5 仅经过三周的强化学习（RL）和监督微调（SFT），在 Colossus 2 集群上就展现出了帕累托最优（Pareto dominant）的强悍表现，并且已经渗透进半数财富 500 强企业。火星殖民、小行星采矿或月球质量投射器固然激动人心，但地面上的算力工厂与 AI 工具才是眼下实打实的营收支柱。 ```mermaid flowchart LR accTitle: SpaceX Ground Compute Value Chain accDescr: 算力基础设施与应用生态的商业闭环 subgraph Infra ["基础设施层"] A["土地与电力资源"] --> B["极速通电与交付<br/>(Colossus 2/3/4...)"] end subgraph Monetization ["算力变现层"] B --> C["高溢价算力租赁<br/>(500亿美元/GW)"] B --> D["自研模型训练/推理"] end subgraph Application ["应用落地层"] D --> E["企业级工具 (Cursor)<br/>500强企业渗透"] end ``` - 地面算力变现估值：比普通 Neocloud 高出 2 至 3 倍，谷歌交易对应每吉瓦高达 500 亿美元。 - 算力建设能力：黄仁勋亲口背书其数据中心落地速度冠绝行业。 - 集群演进序列：Colossus 系列从 2、3 一路规划到 10，内部亦有幽默别称 "macro hard, macro harder, macro hardest"。 - 软件与应用：Composer 2.5 经 3 周强化学习和监督微调即达帕累托优势，Cursor 已打入近半数 Fortune 500 公司。 --- ### [04:46](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=286s) 算力分配张力：自研模型与云基础设施的博弈 主持人探讨了 SpaceX 算力业务发展中出现的权衡与张力：一方面是自身的模型研发与应用程序，另一方面则是对外提供算力云服务。外界此前对 Anthropic 获得算力支持的消息感到诧异，因为任何顶尖研究实验室的负责人都渴望独占所有算力配额。此外，所有 AI 实验室目前都面临着将算力划拨给“预训练”还是“推理”的棘手难题。 Gavin 认为这正体现了管理层对自身极速交付吉瓦级电力的充分自信。Altimeter 数据显示 SpaceX 预订了英伟达下一代 Rubin 芯片多达 20% 的产能。不同于上一代 Blackwell 在系统工程和上线过程中的诸多挑战，Rubin 作为相对平滑的替换升级（drop-in replacement）性能极其强悍；同时，未来 6 到 9 个月内团队还将整合 Groq 的 LPU。针对“马斯克网络服务（Elon Web Services）”是否需要像 AWS 那样提供涵盖数据库、冷热存储等全套云组件的疑问，Gavin 断言，在未来 5 到 10 年里，成为一个极致高效的“Token 工厂”（token factory）就已足够且极度稀缺，未来如果确实需要其他传统云套件，团队也完全有能力随时补齐。 - 芯片与架构配置：锁定了英伟达下代 Rubin 芯片约 20% 的订单份额；预计在 6 至 9 个月内完成 Groq LPU 的集成。 - 业务模式定位：无需在短期内照搬 AWS 的全栈数据库与存储堆栈，聚焦成为高吞吐的“Token 工厂”在未来 5 至 10 年内即具备充沛的市场需求与护城河。 --- ### [07:19](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=439s) 超大规模厂商对比、EV/Net PP&E 与 AI 终局前哨 话题转向其他超大规模云计算巨头（Hyperscalers），尤其是坐拥海量现金流与自建数据中心经验的 Meta。主持人指出，将内部算力转化为商业化的 Token 工厂是一种完全不同的业务机制。Gavin 提到 Meta 也有轻巧的路径，例如直接给 Fireworks 这样的团队划分 10 万张 GPU 进行收入分成。 为了衡量这套逻辑，Gavin 提出他正越来越看重“企业价值与净固定资产比率”（EV / Net PP&E）这一估值指标。其背后的核心投资逻辑在于，地球上已经通电部署的物理硬件资产（"installed atoms on earth"）未来将会持续升值，这呼应了高盛提出的“光环交易”（HALO trade：高资产价值、低淘汰率）。Meta 目前较低的 EV / Net PP&E 乘数反映出公开市场对其将庞大资本资产高效变现的能力抱有巨大疑虑。扎克伯格（Mark Zuckerberg）提出的“个人超级智能”（Personal Superintelligence）概念目前仍停留在口头，Meta AI 既未霸榜应用商店，Instagram 内也尚未出现颠覆性的杀手级变现功能。最后，Gavin 观察到各家科技巨头的动作表明，全行业正提前进入决战终局（the endgame），OpenAI 近期大幅削减 Codex 价格，正是因为高质量的代码 Token 稀缺且极其昂贵，一旦缺乏足够的代码数据，就将丧失进入下一阶段递归自我改进（Recursive Self-Improvement）的入场券。 - 估值新锚点：EV / Net PP&E（企业价值 / 净物业、厂房及设备），押注实体物理硬件资产的稀缺性与升值潜力。 - 华尔街策略概念：HALO Trade（High Asset value, Low Obsolescence，高资产价值、低陈旧过时风险）。 - Meta 面临的困境：缺乏类似于爆款消费级应用或企业级 API 的清晰变现闭环，宏大愿景尚未体现在直接业绩上。 - 行业进入终局阶段（Endgame）：OpenAI 激进下调 Codex 价格以争夺关键的编程数据流（Coding tokens），以此驱动递归自我改进的迭代周期。 ### [09:53](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=593s) Meta的战略灵活性、算力商业化与工程师留存 在讨论科技巨头的战略演进时，嘉宾们提到了电影《塔拉迪加之夜》（Talladega Nights）的经典台词：“所以你的意思是，我们还有机会？”（"So you're saying there's a chance."）以及那句更残酷的“如果你不是第一，你就是最后一名”（"If you're not first, you're last."）。对话迅速聚焦到扎克伯格（Mark Zuckerberg）身上。Gavin Baker 坚信，扎克伯格历来证明了自己是一位极具调整能力的出色创业者，随时愿意根据现实情况迅速转向。尽管目前 Meta 一再向投资者强调“绝不会向外部商业化出租自己的 GPU 算力”，但在 Gavin 看来，科技巨头的承诺往往瞬息万变。 这与 2022 年底 Meta 的大裁员如出一辙。当时在 Altimeter 的 Brad Gerstner 给扎克伯格写了一封公开信（直言“你是国王，你想做什么都可以，但我谨代表投资者提出建议”）之后不久，Meta 还在声称绝不削减投资，然而仅仅几天之后，战略就急转弯变成了极度专注于控制营运支出（OPEX）。Gavin 调侃道，如果今天 Meta 告诉你他们不打算对外出租算力，你不妨“过几个小时再来看看”。商业世界里每个人都有自己的筹码与价格，就像马斯克与 Anthropic 之间的关系，数月前看似毫无瓜葛，但随着机遇出现，双方迅速展开了合作。 归根结底，驱动这些超大市值巨头行为的除了生存危机感，还有至关重要的股价。如果股价持续下跌，员工手中的股权激励缩水，就会导致工程师心生怨怼。在如今的 AI 研发竞争中，人们早已不再讨论“十倍工程师”（10x engineer），而是在争夺“万倍工程师”（10,000x engineer）。如果无法通过强劲的股价留住这批顶尖人才，任何宏大的 AI 愿景都无法落地。 ### [12:41](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=761s) 散户力量的崛起与SpaceX IPO的筹码结构 在谈到零售投资者（散户）在当前资本市场中的角色时，主持人指出许多人嘲笑 SpaceX 的 IPO 散户认购仅超额 2 倍，而一些规模仅 30 亿美元的 IPO 却能获得 10 倍超额认购。但这种对比完全忽略了两者的数量级差异。随着超级云厂商（Hyperscalers）在激进资本开支下逐渐转为自由现金流为负，加之中东地缘冲突引发的重建资金需求，散户资本正扮演着前所未有的关键托底角色。 Gavin 借用《阿甘正传》的名言“傻人有傻福”（"Stupid is as stupid does"）来反驳机构投资者对散户的鄙视态度。数据显示，追踪散户青睐标的的指数在 2024 年和 2025 年均表现极为抢眼，散户的实际收益甚至跑赢了绝大多数由私募股权、风险投资和公募基金打理的专业理财经理。过去三年半里，散户对市场定价的影响力超过了他职业生涯中的任何时期，甚至强于 2000 年的互联网泡沫（他强调当前的 AI 热潮与当年的纯粹泡沫完全不可同日而语）。 ```mermaid flowchart TD accTitle: SpaceX IPO筹码结构与锁定分析 accDescr: 展示SpaceX长期流动性释放与IPO员工买入对二级市场抛压的吸收机制 subgraph 内部筹码稳定性 A["过去10年每6个月一次的要约回购<br/>(想套现的早期股东/员工早已退出)"] B["超过10,000名SpaceX员工<br/>选择在IPO时主动真金白银买入"] end subgraph 市场预期偏差 C["华尔街机械分析:<br/>SPV层层穿透、解禁抛压担忧"] D["实际供求关系:<br/>长线筹码沉淀，浮动筹码极度稀缺"] end A --> D B --> D C -.->|未能反映真实供求| D ``` 针对 SpaceX 的 IPO，Gavin 指出华尔街分析师犯了一个严重的认知偏差。外界都在大谈特谈解禁期（Lockup period）分析，担忧层层套嵌的特殊目的载体（SPV）会带来天量抛压，却完全忽视了以下两点： - **员工高度看好并加仓**：有超过 10,000 名 SpaceX 员工在这次 IPO 中自掏腰包认购了公司股票。 - **流动性早已充分释放**：在过去 10 年里，无论是员工还是在股权架构表（Cap table）上的投资机构，几乎每 6 个月就有一次要约出售套现的机会（例如去年下半年的股份回购）。那些真正想套现离场的人早已卖出，因此解禁后的实际抛压并不会如外界预期的那样严峻。 ### [16:14](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=974s) 何谓“Token路径”：边缘网络、低延迟溢价与算力分布 讨论随后深入到了当下极为火热的概念——“身处 Token 路径之中”（in the token path）。主持人以 Cloudflare 为例发问：作为在网络边缘向终端用户传输数据包的服务商，随着海量 AI Agent 开始与各类网页进行实时交互，Cloudflare 是否算真正意义上处于 Token 路径之中？ Gavin 阐述了内容分发网络（CDN）在 AI 时代的真实定位与局限： - **低延迟的高额溢价**：CDN 厂商（如 Cloudflare 与 Anthropic 达成的 18 亿美元合作）依托遍布全球的网络接入点（PoP），为高价值用户提供极低延迟的 Token 输出。从芯片公司 Cerebras 的市场表现可以看出，用户愿意为了极端速度支付巨大的溢价。 - **实际 Token 处理体量极低**：从全球实际消耗的 Token 总量来看，全球所有 CDN 分发处理的 Token 占比可能不足 1%，甚至不到千分之一（10 个基点）。 - **计算逻辑的发生地**：绝大部分复杂的推理与计算都是在数据中心内部闭环完成的，最终只有提炼后的简要结果才经由边缘送达用户。因此，CDN 只能说是“拥有一条通往 Token 路径的通道”，并正竭力把自己往核心路径上靠拢。 对于不同的模型，延迟要求的差异导致了算力部署的分化：需要实时交互的语音模型必须紧贴边缘，甚至需要将英伟达服务器直接与 CDN 节点共址部署；而编写代码等深层思考模型，即便让它在弗吉尼亚州的数据中心“烹饪”（cook）20 分钟甚至一个小时再返回最终代码，用户也完全能够接受。 ### [18:27](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1107s) 算力泛在化与“瓶颈交易”的终局 随着算力下沉，端侧边缘设备的形态正在走向极端。英伟达甚至投资了一家在住宅外墙安装 4 卡 GPU 服务器的初创公司。这种将计算单元直接装在房屋外墙甚至未来的各类日常硬件中的设想，与特斯拉 Powerwall 在电网之外进行本地离网储能的逻辑如出一辙，也让人联想到黄仁勋在早期加密货币挖矿热潮时描绘的“全民挖矿”图景。 ```mermaid flowchart LR accTitle: AI供应链瓶颈挖掘链条 accDescr: 展示资本市场沿着半导体供应链逆流而上寻找瓶颈标的的传导逻辑 subgraph 供应链溯源挖掘 A["先进制程芯片 / GPU"] --> B["半导体基础材料与封装"] B --> C["上游特种材料<br/>(如味之素ABF薄膜)"] C --> D["极度泛化的物理资产<br/>(土地、电力、甚至砂石)"] end subgraph 交易周期演变 D --> E["瓶颈交易狂热<br/>(Bottleneck Bros)"] E --> F["瓶颈交易走向终结<br/>(厂商拒绝提价/产能释放)"] end ``` 然而，资本市场对于供应链上游物理瓶颈的盲目炒作正面临反噬。投资者一路向上游溯源，试图在每一个环节寻找垄断瓶颈——从抢购数据中心土地，到挖掘为半导体供应链提供特种材料的卫浴陶瓷厂商，再到追溯至将沙子变为硅的源头。Gavin 明确指出，华尔街所谓的“瓶颈老哥”（bottleneck bros）热衷的“瓶颈交易”（bottleneck trade）正在迅速走向尾声。 以《华尔街日报》报道的日本味之素（Ajinomoto）公司为例，该公司生产的 ABF 薄膜（Ajinomoto Build-up Film）是半导体制造封装中不可或缺的绝缘材料。大量投机资金涌入这只股票，押注其会利用垄断地位大幅提价。然而味之素官方却表态绝不涨价，认为趁机提价对公司的长期客户关系与企业价值是错误的。这种产业实操与金融炒作之间的巨大脱节，预示着一味追逐物理瓶颈的投资逻辑已逼近强弩之末。 - **核心观点**：依靠炒作特定原材料或物理资产稀缺性的上游“瓶颈交易”正在衰退，单纯的供应链上游卡位并不能保证定价权与长久超额收益。 ### [20:00](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1200s) 从寻找算力瓶颈到轨道计算的经济学账本 在过去一年里，整个资本市场都在为“下一个供应链瓶颈是什么”而狂热追逐，投资人不断地在各种硬件短缺之间来回切换。然而加文·贝克（Gavin Baker）指出，寻找瓶颈只是上一阶段的游戏，下一阶段的核心命题是“谁拥有能够穿越周期的持久特许经营价值（enduring franchise value）”。在探讨土地与电力是否会成为终极瓶颈时，贝克提出了一套关于“轨道计算（Orbital Compute）”的颠覆性算账逻辑：一旦 SpaceX 的星舰（Starship）实现完全可重复使用，把算力搬上太空的经济账将变得极其诱人。 贝克详细拆解了地面与太空部署 1 GW（吉瓦）算力的成本对比：在地球上建设 1 GW 算力中心大约需要 600 亿美元，其中约 250 亿美元用于电力基础设施和散热冷却系统，剩下的 350 亿美元则是 GPU、CPU、网络交换机、内存与存储等纯 IT 设备；如果转移到太空，环境自带的散热特性和太空太阳能意味着你可以基本省去那 250 亿美元的电冷基建支出。取而代之的变量是火箭发射成本——在星舰成熟复用的前提下，将相应设备送入轨道的发射成本预计仅需约 50 亿美元。这意味着在太空搭建 1 GW 算力总共只需约 300 亿美元，仅为地面 600 亿美元成本的一半，而且地面电力与冷却基础设施的通胀压力还远高于太空发射成本。因此，普通的商业用地甚至很难挤进 AI 财富创造的“代币链条（token path）”中；真正能从这轮 AI 财富浪潮中持续升值的实物资产，或许只剩下稀缺的海滨豪宅、私人飞机和顶级豪车。 ```mermaid flowchart TD accTitle: 地面数据中心与太空轨道计算成本对比 accDescr: 对比1吉瓦算力在地面600亿美元与太空300亿美元的资本支出构成 subgraph Ground["地面数据中心 (1 GW): 总计 600 亿美元"] G1["IT 算力硬件 (GPU/CPU/交换机/存储)<br/>350 亿美元"] G2["电力与冷却系统设施<br/>250 亿美元 (面临通胀压力)"] end subgraph Space["太空轨道计算 (1 GW): 总计 300 亿美元"] S1["IT 算力硬件 (同等规模)<br/>约 250 - 350 亿美元"] S2["星舰完全复用发射成本<br/>约 50 亿美元"] S3["空间自然冷却与太阳能<br/>0 亿美元电冷基建"] end Ground -.->|"星舰完全复用后成本降低 50%"| Space ``` - **地面 1 GW 算力总投资**：约 600 亿美元（350 亿硬件 + 250 亿电力冷却配套）。 - **轨道 1 GW 算力总投资**：约 300 亿美元（350 亿硬件折算 + 50 亿星舰发射运力成本）。 - **核心资本趋势**：地面电冷建设正面临强烈的通胀阻力，而航天入轨成本处于极速通缩通道。 --- ### [21:31](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1291s) 主权 AI 的现实局限与主流演进路径 面对各国竞相建立“主权 AI（Sovereign AI）”的热潮，主持人质疑：即使拥有供电厂房和最顶级的 GPU，缺乏顶尖科研人才和海量真实调用吞吐，多数国家是否已经错过了角逐前沿模型（Frontier Models）的末班车？贝克坦言，出于国家防御底线，每个国家都必须制定某种形式的主权 AI 战略，但他断言绝大多数国家不可能在前沿层面上具备竞争力。网络上关于欧盟限制开源模型 Mistral 导出、或所谓通过 Mythos 蒸馏屠榜的传闻，多半只是社区玩梗，即便模型优秀也难以撼动前沿格局。 在贝克看来，除中美两强之外（且他认为中国在前沿领域会进一步落后），绝大多数国家的主权 AI 终局形态将非常标准化：依赖开源生态中的最优模型，利用专业服务商结合本国语言、历史文化与价值观做强化学习（RL）微调，配上一套严密的系统提示词（System Prompt）和监督微调（SFT），最后将其部署在本国境内的本地数据中心运行。这样做的核心诉求是建立“安全感”——确保涉及国防机密、情报分析、公共安全及日常警务巡查的自主 Agent 能够在全天候 24 小时运转时完全可控，不被外部力量窥探或切断，但这绝不等同于拥有自主研发前沿基础模型的能力。 - **主权 AI 的标准技术栈**：全球最强开源基础模型 + 本土文化/语言监督微调（SFT）与强化学习（RL）+ 严密系统提示词。 - **物理落地场景**：本地主权数据中心托管，专供国防军工、国家情报与 24 小时治安执勤 Agent。 - **前沿模型竞争格局**：真正的前沿突破将高度集中在少数顶级实验室，主权 AI 更多是应用层与控制权防御，而非前沿科研冲刺。 --- ### [23:43](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1423s) 中美大模型代差扩大与“工业化蒸馏”的终结 谈及中国开源大模型与美国闭源前沿模型之间日益扩大的差距，贝克直言中国犯了一个战略误判：在前期政策环境一度允许采购 NVIDIA H200 或 P30 等降规/合规芯片时，中国过度自信于“自主研发芯片已足够好”，但现实并非如此。这种错觉很大程度上源于中国实验室在“工业化规模模型蒸馏”上的极致技巧——曾有业内人士透露，通过从 OpenAI 的 o1 和 o3 抓取仅仅 16 万条推理链轨迹（reasoning traces），就足以支撑起早期 DeepSeek 的惊艳表现。 贝克形容这种蒸馏操作就像中国那些壮观的“iPhone 手机刷单墙”一样，演变成了一种工业级流水线：利用成千上万个终端节点通过各种 API 渠道广泛刺探、抓取前沿模型的输出与思考路径。然而，这种套利的窗口期正在迅速关闭。随着美国顶级闭源实验室（如 Anthropic 及传闻中的 Mythos 模型）全面收紧安全防线，严格封锁底层推理链输出和防范数据间谍行为，单纯依靠 API 蒸馏走捷径的方式将难以为继。如果源头的前沿能力不再公开泄漏，缺乏自主算力底座和原创突破的蒸馏路线就会彻底撞墙。 - **推理蒸馏的惊人杠杆**：早期 DeepSeek 仅依靠提取自 o1 和 o3 的约 16 万条推理轨迹，便实现了极高的推理性能跃迁。 - **工业化蒸馏模式**：设立多达 10 万个分布式 API 节点同时轮询海外接口，进行全天候逆向提炼。 - **断流风险**：随着海外前沿实验室对思考链日志（CoT）实施全链路加密和防抓取限制，蒸馏红利即将消耗殆尽。 --- ### [25:17](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1517s) 早期重仓传奇与回击对超级富豪的政治偏见 主持人感慨投资人生涯中能投中一家 SpaceX 这样的公司便已是莫大幸运，甚至担心未来是否还会出现同等伟大的企业。贝克对此虽然感慨，但也淡然列出了自己惊人的早期投资战绩：他曾在英伟达（Nvidia）市值不足 20 亿美元时买下了其 15% 的流通股份，也曾在特斯拉（Tesla）市值不到 20 亿美元时重仓买入 10% 的股份。但他依然强调，参与 SpaceX 的历程是最非同寻常的体验，SpaceX 已经、且极有可能是人类商业史上最重要的一家企业。 贝克借此对当下西方左翼针对马斯克等超级富豪的口诛笔伐进行了犀利反击。他引用好友凯文·马菲（Kevin Maffey）的话调侃道：外界非要把马斯克塑造成一个“007电影里的超级反派”，然而这位“反派”却在干着全世界左翼曾经最推崇的事业——推动全球深度脱碳转型，通过星链（Starlink）以超低成本把低收入国家贫困偏远地区的学校和医院接入现代互联网，旗下 Neuralink 甚至已经能让盲人重新看见世界、与现实重新交互。不仅如此，SpaceX 还让大量一线蓝领技工获得了前所未有的股权财富。有人尖锐点评：“左翼其实已经赢了，可他们甚至都意识不到——世界上第一个万亿富翁一个人为解决地球环境危机所做的贡献，比全球其他人加起来还要多。” - **加文·贝克的超早期持仓记录**： - 英伟达（Nvidia）：在市值低于 20 亿美元时持有约 15% 股份； - 特斯拉（Tesla）：在市值低于 20 亿美元时持有约 10% 股份； - SpaceX：长期重仓并见证其迈向历史最伟大企业。 - **针对万亿富翁叙事的荒谬性**：针对曾有机构宣称“50 亿美元就能解决全球饥饿”，马斯克公开承诺“只要给出确切账目我就卖股票捐款”，事实证明全球系统性饥饿绝非砸几片股票市值就能解决，反而会破坏经济基本面和数百万就业岗位。 --- ### [28:00](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1680s) 火箭发射的感官震撼与“太阳系东印度公司”的终局 抛开冷冰冰的投资回报率，贝克认为没有任何一家商业实体能像 SpaceX 这样带来如此直击灵魂的生命震撼。他极力建议所有听众一定要带孩子亲临一次火箭发射现场，而且根本不需要费尽心机去求所谓的 SpaceX 贵宾通行证——只要去到德克萨斯州博卡奇卡（Boca Chica）的公共海滩，或者加州范登堡太空军基地周围的汽车旅馆，你站的位置甚至比官方贵宾观礼台离发射台还要近。那种发动机点火瞬间撕裂空气的超低频巨响、滚滚扑面而来的灼热气浪，以及亲眼目睹人类智慧将数千吨钢铁推向虚空的视觉奇观，往往会让现场无数成年人当场热泪盈眶。 贝克进一步描绘了 SpaceX 无法被复刻的历史定位：如果 SpaceX 最终成功在月球乃至火星建立起永久性人类城市，它实际上将成为横跨整个太阳系的“不列颠东印度公司（British East India Company of the solar system）”。这背后虽然还横亘着极其艰难的工程挑战，但 SpaceX 必将成为全人类历史上最重要的一家商业实体。在对这一宏大叙事的赞叹中，对话最后切入到了更广泛的资本市场生态：在“瓶颈炒作者（bottleneck bros）”已经把几乎所有具备概念的 20 亿美元市值以下标的炒上天的当下，包括生物医药（Bio）、材料科学以及像 Neuralink 这样长期处于 20 亿以下估值的底层赛道，是否正在孕育 AI 扩散效应下的下一轮超级机遇。 - **观礼实用建议**：无需特权门票，博卡奇卡公共海滩或范登堡基地周边酒店均具备极佳的第一视角视距。 - **终极终局愿景**：从近地轨道物流巨头进化为星际地缘商业纽带——“太阳系的东印度公司”。 - **下一代未被透支的领域**：在公募与私募市场上寻找尚未被基建热钱完全溢价推高的领域，特别是 AI 深度赋能的生物学、先进材料与脑机接口等底层技术创新。 ### [30:22](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1822s) 一级市场与二级市场小盘股的反差：社交媒体上的 AI 炒作乱象 在谈及当下科技创业生态时，讨论转向了一级风险投资（VC）与公开二级市场之间的鲜明对比。加文·贝克（Gavin Baker）指出，在一级市场中，估值低于 20 亿美元的优秀初创公司（包括大量具有极高价值的 SaaS 软件企业）可以说是数不胜数，他自己每周至少要与这些团队深入交流数次。然而在公开市场上，情况却大相径庭：在所谓的“瓶颈兄弟”（Bottleneck Bros，指在公开市场狂热寻找各类 AI 供应链瓶颈概念股的投资者群体）和生成式 AI 浪潮爆发之前，公开市场的小盘股领域（Small Cap Land）长期缺乏活力、几乎无人问津。 如今，伴随 AI 概念的扩散，二级小盘股领域虽然重新被激活，却衍生出了一种令人不安的投机生态。加文对此表达了强烈的警惕与不适：现在的社交媒体上普遍存在这样一种怪相——某些在网上拥有庞大粉丝群体的账号，先是号召大家买入某只冷门股票，随后在社交平台上发布一份针对市值仅有 3500 万美元左右的微盘股的“投资逻辑分析”（thesis），随即就能推动该股暴涨超过 1000%。加文特别澄清，自己的批评并非针对市场上那些真正聪明的知名分析者（例如像 Serenity 这类账号），“我认为那个人确实很聪明”，但他强调，当下围绕 AI 小盘股的各种操作中，已经充斥着大量非常可疑、形迹可疑的操纵行为（"a lot of shady stuff happening"）。 - **一级市场繁荣度**：估值 20 亿美元以下的初创企业和优质 SaaS 标的层出不穷，加文每周均保持极高频次的调研对接。 - **微盘股流动性操纵风险**：市值仅 3500 万美元级别的微型标的极易被社交媒体声量拉抬，轻易出现单边暴涨 1000% 的非理性泡沫行情。 ```mermaid flowchart TD accTitle: 社交媒体杠杆推动微盘股炒作链条 accDescr: 展示具有庞大粉丝量的匿名账号如何通过发布研报催化 3500 万美元微盘股暴涨千倍的过程。 subgraph Influence["社交媒体杠杆"] A["高粉社交媒体账号"] --> B["低位吸筹 / 选定标的"] B --> C["公开发布投研逻辑 (Thesis)"] end subgraph MarketImpact["公开市场流动性冲击"] C --> D["散户跟风涌入 3500 万美元微盘股"] D --> E["股价非理性暴涨 1000%"] end E --> F["缺乏合规监管与实质业绩支撑"] ``` --- ### [31:36](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1896s) 从互联网泡沫到匿名时代：合规真空与投资风险 播客主持团队随即将当下的狂热与加文早年在富达投资（Fidelity）的经历进行了对照。在 2000 年互联网泡沫（Dot-com bubble）走向尾声的周期里，当时还只是一名年轻分析师的加文亲眼见证了极其相似的市场极度狂热（euphoria）。主持人指出，周期的高潮过后往往伴随着残酷的清算：随之而来的几年里，许多人不得不面临严厉的民事处罚甚至法律制裁，大批明星公司轰然破产倒闭，整个市场经历了剧烈的动荡和阵痛；当然，在那场惨烈洗牌中，最终也淬炼出了一批跨越周期的世代级伟大企业。 然而加文与主持团队一致认为，今天的局面比当年更加复杂且充满隐患。在当年的互联网泡沫时期，即便卖方投行的研究报告（Sell-side Research Report）可能吹嘘得有些过火，但其背后依然存在完整的企业合规部门（Compliance Department）和监管问责机制。而如今在社交媒体上呼风唤雨的许多吹鼓手，往往隐藏在完全匿名的马甲之后，登录社交账号用的是随手捏造的临时邮箱，网络链路则通过 VPN 进行层层伪装，既无合规约束，也难以追究法律责任。主持人坦言，公开市场已经进入了一片从未经历过的“未开垦的法外之地”（uncharted territory），因此当下的普通投资者必须格外注意资金安全，牢记“自行做好独立调研”（Do your own research）。 | 维度对比 | 2000 年互联网泡沫晚期 | 当前社交媒体 AI 概念炒作 | |:--- |:--- |:--- | | **主要传播媒介** | 传统卖方投行研究报告、财经媒体 | 社交媒体（如 X / Twitter 等）、散户网络论坛 | | **机构问责与约束** | 拥有完整的公司合规部门审核机制 | 账号普遍匿名化、虚构邮箱注册、挂载 VPN | | **违规行为后续** | 破产清算、面临监管民事罚款与调查 | 难以追溯真实身份、追责难度极高 | | **共同点** | 市场充斥非理性繁荣，终将经历周期淘汰洗牌 | 虽有投机乱象，但真实革命中亦会孕育代际巨头 | --- ### [32:40](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1960s) 节目收尾与未来邀约：全员合伙人会议式的对谈展望 在经历了一场涵盖 SpaceX 长期愿景、AI 基建算力工厂、主权大模型竞争态势以及二级市场投机生态的长谈之后，对话进入了轻松融洽的收官环节。TBPN 团队向加文由衷致谢，直言整期对话聊得非常过瘾，并热切期待他能尽快再次返场作客。 加文热情地接受了邀约，并积极提议下次不仅是他个人参与，更可以拉上他创立的 Atreides Management 投研团队的所有成员一起来做一期联名节目。主持团队兴奋地赞同这一想法，笑称加文完全可以在公司找一间配备了超广角摄像头的会议室，把整个投资核心团队全部拉到镜头前坐在一起，直接把播客录制打造成一场生动的“投资合伙人会议”（Partner Meeting）。在愉快的道别与祝愿声中，全场深入而跨越多个科技前沿领域的访谈圆满落下帷幕。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=0s) 20% pop almost to a tee > [00:00](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=0s) 马拉松就是 26 次 1 英里连着跑 — Gavin Baker > [07:19](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=439s) installed atoms on earth — Gavin Baker > [09:53](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=593s) So you're saying there's a chance. > [09:53](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=593s) If you're not first, you're last. > [09:53](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=593s) 你是国王，你想做什么都可以，但我谨代表投资者提出建议 — Brad Gerstner > [09:53](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=593s) 过几个小时再来看看 — Gavin Baker > [12:41](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=761s) Stupid is as stupid does — Gavin Baker > [25:17](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1517s) 左翼其实已经赢了，可他们甚至都意识不到——世界上第一个万亿富翁一个人为解决地球环境危机所做的贡献，比全球其他人加起来还要多。 > [28:00](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1680s) British East India Company of the solar system — Gavin Baker > [30:22](https://www.youtube.com/watch?v=PW5n3ZnEJN0&t=1822s) a lot of shady stuff happening — Gavin Baker ## Mentioned Gavin Baker, SpaceX, Goldman Sachs, Morgan Stanley, Tesla, Amazon, AWS, Foxy, Atreides Management, Cursor, Altimeter, xAI, Google, Jensen Huang, Composer 2.5, Colossus 2, Fortune 500, Anthropic, Nvidia, Rubin, Blackwell, Groq, LPU, Meta, Fireworks, Mark Zuckerberg, Instagram, OpenAI, Codex, Talladega Nights, Brad Gerstner, Elon Musk, Forrest Gump, Cloudflare, Cerebras, Powerwall, Ajinomoto, Starship, Mistral, Mythos, DeepSeek, o1, o3, Kevin Maffey, Starlink, Neuralink, Serenity, Fidelity, TBPN ## Video description This is our full conversation with Gavin Baker. We discuss the aftermath of the SpaceX IPO, why he believes SpaceX could become the most important company of all time, how AI infrastructure and "token factories" are reshaping capital markets, why sovereign AI strategies may struggle to compete with frontier labs, and what the future of data centers, orbital compute, and the AI endgame could look like. TBPN is made possible by: Ramp - https://ramp.com Public - https://public.com Cisco - https://www.cisco.com Console - https://www.console.com CrowdStrike - https://www.crowdstrike.com Figma - https://www.figma.com MongoDB - https://www.mongodb.com NYSE - https://www.nyse.com Railway - https://railway.com Shopify - https://www.shopify.com/ Codex - http://openAI.com/codex Sign up for TBPN’s daily newsletter at TBPN.com Follow TBPN: https://TBPN.com https://x.com/tbpn https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231 https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235 https://www.youtube.com/@TBPNLive [Source](https://www.youtube.com/watch?v=PW5n3ZnEJN0)

[0:00] 20% pop almost to a tee [0:00] Gavin Baker: 马拉松就是 26 次 1 英里连着跑 [7:19] Gavin Baker: installed atoms on earth [9:53] So you're saying there's a chance. [9:53] If you're not first, you're last. [9:53] Brad Gerstner: 你是国王，你想做什么都可以，但我谨代表投资者提出建议 [9:53] Gavin Baker: 过几个小时再来看看 [12:41] Gavin Baker: Stupid is as stupid does [25:17] 左翼其实已经赢了，可他们甚至都意识不到——世界上第一个万亿富翁一个人为解决地球环境危机所做的贡献，比全球其他人加起来还要多。 [28:00] Gavin Baker: British East India Company of the solar system [30:22] Gavin Baker: a lot of shady stuff happening

[4:46] Gavin Baker: Being a token factory is more than enough for the next 5 to 10 years, man. We need more token factories. [12:41] Gavin Baker: People say retail in a pejorative way, and I would just say stupid is as stupid does... Retail is probably outperforming the overwhelming majority of professional money managers. [25:18] Gavin Baker: The left won and they don't even appreciate it. The world's first trillionaire has done more to kind of solve the environment than like everyone else on Earth combined.

[0:00] Gavin Baker, welcome to the show. How [0:02] are you doing? [0:03] >> Doing great, man. How are you guys? [0:05] >> Fantastically, right? Uh SpaceX IPO. It [0:09] seems like everything went just [0:11] flawlessly to a tea. Like to the point [0:14] where it wasn't even dramatic. [0:15] >> A round of applause for the banker. [0:16] >> The banker. It's like 20% [0:18] >> 20% pop almost to a te. [0:20] >> Yeah. What I mean by that is like is [0:22] like it would have been more dramatic if [0:24] it had like popped up 70% and then went [0:26] down 30 and then went up 50 and it was [0:28] drama but it was just like perfect [0:30] execution from start to finish. Was that [0:32] how you interpreted it? [0:34] >> I thought Goldman Sachs and Morgan [0:35] Stanley did a very good job. [0:37] >> Yeah. Um what do you think the market is [0:40] looking for SpaceX next? Is it all eyes [0:42] on the first earnings report? Like where [0:44] do we go from here? Because there's so [0:46] much of the SpaceX story preipo that was [0:49] a decade out, 5 years out, two years [0:52] out. But is the market going to be [0:54] processing quarter by quarter plays like [0:58] a like many other companies? [0:59] >> Well, I think the market is always, you [1:01] know, quarters matter. [1:03] >> Yeah. [1:03] >> Um, you know, the runners on X, you [1:06] know, it's like they're these reply [1:07] people on X for any topic or experts. [1:10] >> Yeah. [1:10] >> But, um, somebody once told me a [1:12] marathon is 26 onem runs. Yeah. [1:15] >> And so, you know, and then all the [1:17] runners were like, "No, that's not true [1:18] at all." Running, you know, obviously [1:20] running a mile is different than running [1:21] a marathon. But, you know, every quarter [1:24] matters. Uh but at the same time I do [1:26] think if you look at how [1:29] um the market has has looked at Tesla [1:31] over the last five six years you know [1:34] Amazon during kind of the days of you [1:37] know when they were building out AWS and [1:40] then um you know kind of their their [1:43] retail distribution infrastructure when [1:44] they'd go through these investment [1:45] cycles. I do think the public market has [1:48] a much greater tolerance [1:50] um for investment and a much longer time [1:53] horizon than a lot of people in the [1:55] venture ecosystem give it credit for. By [1:58] the way, this is Foxy in the back. Um [2:00] he's our he's um he's one of the [2:03] analysts here at Trades. [2:04] >> I was just listening to him on the Brad [2:06] Gersonner podcast. I think he chimed in [2:08] on that, right? You brought him in. [2:09] There we go. [2:10] >> Yes, he did. But he did the work on [2:12] SpaceX. So, I thought you guys might [2:13] want to fantastic about that. Um, [2:16] but I do think that I'm not sure the [2:20] story is as far out as you made it seem [2:23] to be. [2:24] >> I think there are two variables that are [2:26] going to matter a lot over the next [2:27] year. [2:28] >> Mhm. [2:29] >> The first is just how quickly can they [2:30] bring on terrestrial compute. [2:32] >> Yeah. [2:33] >> It does seem like they monetize [2:35] gigawatts at a higher rate than anyone [2:37] else. [2:37] >> Yeah. [2:38] >> And we know from Jensen that they bring [2:40] on data centers faster. [2:42] >> Yeah. [2:42] >> Than anyone else per Jensen's words. [2:44] Yeah. [2:44] >> Um [2:46] and everyone in the ecosystem is really [2:48] incented to get um land and power in the [2:52] hands of people who can energize it [2:54] because everybody starts making more [2:56] money when the GPUs are energized. Um so [3:01] you know if they can what what are what [3:03] what was the altimeter figure for how [3:05] what they were monetizing gigawatts at [3:07] Foxy? Do you remember for which deal? [3:09] Well, I think it was like two to three, [3:11] wasn't it? Like at least two times [3:13] higher than like average Neocloud [3:15] pricing and that's partly in due to [3:16] scale and customer quality and things [3:19] like that, but it was meaningful. [3:21] >> Yeah. So, they're doing 50 billion a [3:22] gigawatt on the Google deal. [3:24] >> Wow. [3:24] >> So, how quickly they can add gigawatts [3:26] really matters. [3:27] >> Yeah. [3:28] >> Right. If you can energize two, three, [3:30] four gigawatts [3:32] >> y [3:32] >> in the next year, [3:34] >> that's a lot of revenue. Now you know [3:36] that prices may go up, they may go down, [3:39] but that's really going to matter. [3:40] >> So all eyes on Colossus 3, 4, 5, 6, 7, [3:42] 8, 9, 10, etc. At least in the short. [3:45] >> Yeah. Or macro hard, macro harder, macro [3:48] >> hardest [3:49] >> to that name. Yeah. Macro hardest. [3:51] >> Yeah. [3:52] >> Um [3:53] >> and then I think cursor is another big [3:56] variable. [3:57] >> Yeah. Um, you know, we we talked about [3:59] it with Brad about how composer 2.5 [4:02] after three weeks of, you know, RL and [4:06] supervised fine-tuning Colossus 2 is [4:09] kind of paro dominant. So, what's going [4:12] to happen when it's applied to a bigger, [4:14] better base model? Um, I do think cursor [4:17] being in, you know, half the Fortune 500 [4:20] >> is interesting. So, I think those two [4:23] things, [4:25] you know, while I mean, I'm super [4:27] excited for going to Mars and asteroid [4:29] mining and [4:30] >> and you know, a city and mass drivers on [4:32] the moon, like all that stuff is awesome [4:35] and I think there's a, you know, decent [4:37] chance it happens in my lifetime. I just [4:40] think there is there's much more [4:42] tangible near-term [4:45] drivers here. [4:46] >> Yeah. How do you think about the tension [4:48] of the cloud business versus their own [4:50] application and and model business? Uh I [4:54] think a lot of people were surprised [4:55] when the initial anthropic deal happened [4:57] purely because uh knowing Michael Truel, [5:00] he would have been very excited. [5:02] >> That's my compute. [5:03] >> Yeah. Yeah. Get access [5:06] only child around here. [5:07] >> And it and it feels like uh you can [5:09] imagine SpaceX, you know, bringing on a [5:11] lot more compute very quickly. Uh but at [5:14] the same time, every lab has had to go [5:16] through this tension of like how much [5:17] how much compute do we allocate to [5:19] training versus inference. [5:22] >> A way to interpret that is that maybe [5:25] the team at SpaceX is pretty confident [5:27] in their ability to bring gigawatts [5:29] online. [5:30] >> Yeah, that makes sense. [5:32] >> And the altimeter figure that they've [5:34] ordered 20% of the Reubins and the [5:36] Reubin is an epic chip. Like Blackwell, [5:40] it was really hard to get it online. [5:42] Ruben is kind of a more drop in [5:44] replacement. [5:45] >> It's a really really good chip. You'll [5:47] have the um you know the Grock LPUs [5:50] integrated at some point in the next six [5:52] nine months. [5:53] >> Um [5:55] >> so one way to interpret that is they're [5:57] pretty confident that they can bring [5:59] data centers online pretty quickly. Um, [6:02] but I also [6:05] I I don't remember all all of the [6:07] details, but I do think there there was [6:09] an out in the anthropic agreement that [6:12] if they [6:12] >> Yeah. And the and the Google and the [6:14] Google D could be short-term. Well, and [6:16] that's that's what I just meant. It's [6:17] like it is there is like a tension there [6:19] where if you you know there's a moment [6:21] where maybe you want the compute, but [6:23] you don't necessarily have all of the [6:24] revenue yet and so you could see [6:26] Anyways, I'm I'm very confident they can [6:28] figure it out. [6:29] >> Yeah. on on on the topic of like Elon [6:31] Web Services, is there a world where uh [6:34] SpaceX winds up building out more of [6:36] what looks like AWS? Like we've seen [6:39] between OpenAI and Anthropic, there's [6:41] this tension of like you got to be on [6:42] AWS because it or you know the ecosystem [6:46] or enterprises just want to be on AWS. [6:48] They won't even go over to Azure or [6:50] Google Cloud for whatever. And you could [6:52] imagine like it might is it enough to [6:55] just be a token factory or do you need [6:57] to have you know a database and an [6:59] in-memory storage and you know cloud [7:02] storage and cold storage and all the [7:04] things that ABS provides. [7:07] >> My thought is that being a token factory [7:08] is more than enough for the next 5 to 10 [7:11] years man. We are [7:12] >> yeah we need more token factories. Um [7:15] but we'll we'll see and if they need [7:17] that stuff like they will build it. [7:19] >> Yeah. How are you thinking about the [7:21] other hyperscalers, the other big tech [7:23] companies? Uh we we were sort of [7:25] puzzling over uh Meta's strategy at this [7:28] point. Tons of resources, cash flow, [7:31] data centers. They have experience [7:33] there. Maybe they're not as fast as as [7:35] SpaceX, but there's a lot there, but [7:37] it's a completely different motion to go [7:40] into token factory. Um and yet there's [7:42] still a lot of demand. So, how do you [7:44] think all of that plays out? [7:46] >> Well, I think there's many ways to go [7:48] into the token factory business. You [7:49] could if you're if you're meta, [7:51] >> yeah, [7:52] >> you could just say to someone like [7:54] Fireworks, um, [7:56] >> hey, [7:57] >> here are 100,000 GPUs. We'll we'll take [8:00] a revenue split and we'll see what [8:01] happens. Y [8:02] >> um I am increasingly looking at EV to [8:06] net PP&E as a valuation metric [8:09] just with the idea being installed atoms [8:11] on earth I think are going to appreciate [8:14] in value and this is a little bit that [8:16] halo trade that um I think maybe Goldman [8:19] Sachs came up with you know high asset [8:21] value low obsolescence [8:23] >> but Meta [8:25] you know Meta's EV to net PP& multiple [8:27] is in an interesting place it just [8:29] suggests that the market has an immense [8:31] amount skepticism about their ability to [8:34] monetize their own asset base. [8:37] >> And and I think that's I think that's [8:39] warranted given that all we've really [8:42] got out of management is this concept of [8:45] personal super intelligence. [8:47] >> And it would be different if Meta AI was [8:49] like at the top of the app store and [8:51] everyone you were talking to was like, [8:52] "Yeah, you know, I'm using Meta AI [8:54] >> where there even there was like a killer [8:55] feature within Instagram. It's like, oh, [8:57] people are shopping in Instagram now." [8:58] And I think no one no one doubts that [8:59] like once Zuck has an AI product with a [9:02] billion dollars of revenue, he can take [9:04] it to to tens of billions, [9:06] >> but it's unclear. I feel like, you know, [9:09] going into enterprise, like, you know, [9:11] what is this API business going to going [9:13] to look like? We know they can use the [9:15] GPUs just on making the ads product [9:18] better at the end of the day. And um [9:20] >> but there's a lot of questions. [9:22] >> Look, I think a few things. I do think [9:25] it [9:26] if I look at the way different players [9:28] are behaving, it feels like everyone [9:31] thinks we're entering the end game. [9:32] Everybody, you know, loves to use these [9:34] chess analogies and it looks like the [9:37] endgame maybe here sooner than we think. [9:39] I think that's why OpenAI is cutting [9:42] codeex pricing [9:43] >> simply because coding, you know, coding [9:46] tokens are so valuable [9:48] >> and if you don't have enough of them, [9:50] you may not be able to get into this RSI [9:53] loop that that everyone is, you know, [9:56] >> pushing for. [9:58] >> Um, [9:59] and I I think if you're Zuckerberg, you [10:02] feel that acutely. I thought muse, what [10:05] did it? [10:06] >> Muse spark. It it was to me it was an [10:10] upside surprise and it made me feel like [10:14] you know what's the uh what's what's [10:15] what's from Tallaladega Nights or what [10:17] is it? So you're saying there's a [10:19] chance. [10:19] >> Oh yeah. [10:20] >> Like yeah so you're telling me there's a [10:22] chance. Um [10:23] >> I mean the other quote from Tallaladega [10:25] Nights is if you're not your f if you're [10:26] not first you're last. [10:28] >> And and you know that's how there are so [10:31] many good quotes from Talladega Nights. [10:33] >> Sold the windshield. [10:35] >> Yeah. 100%. [10:37] But I think if um [10:41] I'm pretty confident Mark Zuckerberg has [10:43] proven to be a good entrepreneur over [10:46] time and has been willing to pivot, you [10:48] know, and it's just it's so funny, you [10:50] know, right now Facebook is, you know, [10:52] Meta has tell investors there's like no [10:55] chance that we're going to, you know, [10:57] you know, monetize GPUs externally. [10:59] Well, like 10 days before they announced [11:03] all those layoffs. Um, which I think was [11:06] at the end of 22, if not I don't [11:09] remember when they started to announce [11:10] the layoffs. It was short shortly after [11:12] Brad, maybe at two months after Brad [11:14] wrote that letter to Mark where he said, [11:15] "You're the king. You can do whatever [11:17] you want, but this what I might [11:19] respectfully encourage you to do." Um, [11:21] you know, they went from saying, "We're [11:23] not going to do that. we're gonna keep [11:24] investing to, you know, super focused on [11:29] OPEX in a couple of days, man. So, it's [11:32] like people they meet with Meta and they [11:34] hear, "Oh, you know, we're not um we're [11:37] not going to monetize our GPUs [11:38] externally." Well, I don't know, check [11:41] back in a few hours. [11:43] >> Yeah. Yeah. No, it's like everyone it [11:45] comes down to like everyone has a price. [11:46] You look at you look at Elon and [11:48] Anthropic relationship four months ago. [11:51] uh and then you look at kind of kind of [11:54] as the opportunity unfolded [11:57] >> it made sense for them to partner. I [11:59] wanted to ask how [12:00] >> on the meta thing I think there is one [12:02] point that's important what ultimately [12:04] drives behavior for all these big cap [12:06] companies like getting to AI I think [12:09] that does feel um existential for all of [12:11] them but the stock price also really [12:14] matters sure [12:15] >> because if you don't if the stock goes [12:17] down you know you gave people a grant at [12:19] this price stock goes down you know your [12:22] engineers are unhappy and it's tougher [12:23] to get to where you want to go if you [12:26] don't have you know these you know Oh, [12:28] forget the 10x engineer. People are [12:30] talking about the 10,000x engineer. [12:33] >> You know, you're not going to get there [12:34] if you don't keep those 10,000x [12:36] engineers. [12:37] >> Yeah. [12:37] >> So, just I think [12:39] >> like we'll see we'll see with Meta. [12:41] >> Yeah. [12:41] >> Yeah. Uh how did this IPO uh update your [12:46] your general kind of framework for the [12:48] role that retail is playing in capital [12:51] markets? Uh a lot of uh it was very [12:54] funny last week to see people like uh [12:56] say like oh that the IPO is only 2x [12:59] overs subscribed by retail all these [13:02] other you know three billion IPOs were [13:04] we're 10x over subscribed and it's like [13:06] no we're we're talking about you know a [13:09] very you know order of magnitude [13:10] difference here but it still feels like [13:12] retail is going to [13:15] >> play a very very important role over the [13:18] next 12 months maybe more so than ever [13:20] given that the the the hyperscalers are [13:23] are now running um you know cash flow [13:26] negative. You have the conflict in the [13:28] Middle East, right? There's rebuilding [13:30] efforts that needs to needs to happen. [13:32] And so I feel like retail this is this [13:34] is kind of um retail's moment now maybe [13:38] more than ever. [13:39] >> Yeah. Well, so a few things. One, people [13:42] say retail in a porative way, and I [13:44] would just say stupid is stupid does. [13:46] And like you know we track you know [13:48] there's all these indices of retail [13:50] favorites and man they're up a lot and [13:54] they're up a lot [13:55] >> 25 and they're up a lot in 24 [13:58] >> you know so kind of stupid as a stupid [14:00] does. So I'm not um you know [14:04] like I think retail is probably [14:06] outperforming the overwhelming majority [14:08] of professional money managers whether [14:09] they're you know private equity managers [14:11] venture [14:12] >> you know public equity managers [14:14] whatever. So like retail, I don't think [14:16] it's pjorative. I do actually think in [14:18] the case of um [14:21] >> SpaceX, I'm not sure, you know, you [14:23] process so much information. I'm not [14:25] sure where this comes from, [14:26] >> but I actually think after the IPO [14:29] um you know, for whatever reason, it [14:32] some of that retail stock may have may [14:34] have gotten allocated to people who were [14:36] flipping it because my understanding is [14:38] there was more institutional buying. I [14:41] don't know where this after the IPO was [14:43] retail buying. Now, I'm sure you know [14:46] retail is very sensitive to momentum. So [14:48] maybe what's happening today changes [14:49] that. But I think another really [14:51] underappreciated thing about the SpaceX [14:53] IPO is that 10 over 10,000 SpaceX [14:57] employees bought on the IPO and that's [15:01] just you know everybody does all these [15:03] lockup analyses and oh how much stock is [15:06] going to come come to market and you [15:08] know for sure there's all these triple [15:10] layer SPVS or whatever you know or maybe [15:12] there are there are people speculate [15:14] there are um and I'm sure some of that [15:17] will get unwound [15:19] >> but the reality reality is is that [15:21] SpaceX employees own a lot of this. And [15:24] if you're an employee or even if you're [15:25] an investor who has an S an SPV or [15:28] you're directly on the cap table, you've [15:31] had a chance to sell every six months [15:33] for the last 10 years, [15:34] >> if you wanted to sell, you could have [15:37] sold in, you know, whatever when they [15:39] did that offering in in the second half [15:41] of last year. So I think the supply [15:44] demand dynamics here are going to be [15:46] very interesting. But with retail in in [15:49] general um [15:52] yeah man they're [15:54] they've been a more powerful force in [15:58] the market over the last three and a [16:01] half years than at any other time in my [16:03] career including the year 2000 bubble [16:06] which I think this is nothing like by [16:07] the way. Um but yeah, they they matter [16:11] and stupid is and stupid does. [16:14] >> Uh I'm interested to go a click deeper [16:16] on the idea of like value will acrue to [16:20] companies that are in the token path. I [16:22] think that phrase is uh extremely [16:24] powerful right now. But I want to sort [16:26] of expand it and understand the fringes [16:29] of the token path because uh we talked [16:32] to Matthew Prince at Cloudflare stocks [16:34] at all-time highs. It's delivering [16:37] tokens at the edge and obviously there's [16:40] an immense amount of AI agents that are [16:42] interfacing with different websites. [16:44] They need to be distributed. So there's [16:46] a lot of value there. But is Cloudflare [16:49] in the token path if they're the one [16:53] finally sending you the packet of tokens [16:56] at the edge [16:57] >> or does that not count? How does that [16:59] how do you think about that? [17:00] >> Every CDN has has this business. You [17:03] know, they signed [17:04] >> Yeah. what was it a$ 1.8 billion dollar [17:06] with anthropic and if you have all these [17:08] points of presence yeah [17:09] >> and you can deliver really low latency [17:12] tokens to highv value users [17:14] >> I think what's interesting for for cloud [17:16] for all these CDNs [17:17] >> is they're they're getting a massive [17:20] premium for that latency [17:22] >> you know it turns out you know one of [17:23] the lessons from from cerebrus is people [17:25] are willing to pay for speed [17:27] >> and you see that in the prices these [17:29] guys are commanding but I do think I [17:32] mean my gut is in terms of tokens [17:34] consumed [17:36] um on planet earth. These CDNs are [17:39] delivering less than 1% of them and I [17:43] mean maybe less than 10 basis points. [17:45] >> They're all most of the time they're [17:46] happening internally and then you just [17:48] get the result. [17:49] >> Um [17:51] >> would I say that these guys are in the [17:53] token path? I mean I would say they have [17:54] a path towards being in the token path. [17:57] Yeah. [17:57] >> And they have part of their they have [17:59] part of their business that's in it [18:00] today. [18:00] >> Yeah. And they're trying and like [18:02] everyone they're trying to move it more [18:03] and more there. [18:04] >> Yeah. Yeah. I mean uh Matthew was [18:06] talking about uh h colllocating Nvidia [18:09] servers on the edge specifically for the [18:12] delivery of voice models because you [18:13] want those to be very low latency. Maybe [18:15] you want your coding model to go off and [18:17] cook in Virginia for 20 minutes or an [18:20] hour and then come back to you with the [18:21] final result. But uh if you're trying to [18:23] actually talk to a model, you probably [18:25] want it as fast as possible. What about [18:27] on the opposite end? Wasn't there that [18:28] startup that Nvidia just invested in [18:30] that like will [18:31] >> Oh yeah. on your house [18:32] >> money [18:33] >> that put a GPU you know four a four GPU [18:36] server on the outside of your house. [18:39] >> I mean [18:40] >> GPUs are going everywhere going in your [18:42] bed eventually. [18:43] >> I mean it's full circle too. Wasn't [18:45] Jensen talking about stuff like that [18:46] back during like the original crypto [18:49] cycle being like everyone's going to be [18:51] you know mining. [18:52] >> Yeah. And I mean the the Tesla Power [18:54] Wall is like a version of that for [18:55] storing energy. It doesn't store [18:57] compute, but it stores energy locally on [18:59] the edge, off the grid, and uh there's a [19:02] lot of powerful things that get unlocked [19:03] by that. What about land? Is land in the [19:06] token path at some point on the extreme [19:08] other end? Uh because people are [19:10] obviously searching for go going further [19:13] and further upstream. They found a a [19:15] toilet company that makes a particular [19:17] material that goes into semiconductor [19:19] supply chain. People are hunting for [19:20] like the final the final ingredient. [19:23] Maybe it's just sand that turns into [19:24] silicon. Yeah, I do think the bottleneck [19:27] bros as they are called are this like [19:30] bottleneck trade is kind of nearing its [19:32] end. [19:33] >> Okay. [19:33] >> And [19:35] you know I did um [19:37] >> you know the Wall Street Journal had the [19:39] story about this Japanese company that [19:40] makes um I forget what they make [19:45] a [19:47] >> Yeah. [19:48] >> Aimoto. Um, and you know, everybody's in [19:51] this because they thought they were [19:52] going to raise prices and then they just [19:54] said, "We're actually not going to raise [19:55] prices. This is the wrong thing for us." [19:57] And I almost posted yesterday on X, [20:00] welcome to Japan to all of the uh [20:02] bottleneck people. But, you know, [20:04] everybody's, you know, having Claw and [20:06] Run, what's the next bottleneck? [20:08] >> Yeah, [20:08] >> that was that was the game for the last [20:11] year. The next game is what has enduring [20:14] franchise value. [20:16] >> Yeah. kind of on the other side of these [20:18] bottlenecks whenever whenever that is [20:21] >> as far as land is land a bottleneck man [20:24] I do think the math for orbital compute [20:26] to me it once they can reuse Starship [20:30] >> I think the math for orbital compute [20:31] becomes pretty compelling you know it's [20:34] just 60 billion to bring on a gig [20:36] terrestrially 25 is power and cooling [20:38] you don't need that in space [20:40] >> and so the right comp for that 35 [20:42] billion of kind of you know IT equipment [20:45] GPUs CPUs switches [20:47] um you know memory storage um [20:52] is is that 25 billion versus the cost of [20:54] launch and one starship is reusable I [20:57] think the cost of launch is 5 billion so [21:00] 20 that means you can put a gig into [21:03] space for 30 billion and a gig on earth [21:05] is 60 billion and that 25 the power and [21:08] cooling [21:09] >> feels reasonably inflationary to me [21:11] >> sure [21:12] >> um so I don't know that I would say and [21:16] is in the tokens token path. [21:17] >> Maybe beachfront property, but that's [21:19] about it. [21:20] >> Yeah, beachfront property is beachfront [21:24] property, airplanes. [21:26] >> Yeah. [21:26] >> Really nice cars. [21:27] >> Yep. Firmly in the token path. [21:30] >> All that stuff is in the token path. [21:31] >> I gota uh how do you think the events of [21:35] of last week with Anthropic and DC [21:38] update uh different countries on their [21:41] own sovereign AI play? Because there's [21:43] been there's been attempts [21:45] It doesn't feel like, you know, many of [21:47] the attempts have, you know, you can [21:49] have a powered shell and you can have [21:51] the best GPUs, but if you don't have the [21:54] talent, it's really, it's going to be [21:55] really hard to get to the frontier. And [21:57] now, if you don't have massive usage, [21:59] it'll also be really hard to get to the [22:00] frontier. And so, do you think the ship [22:02] has sailed? [22:04] >> No. Well, look, look, I think every [22:06] country is going to want to have their [22:07] own sovereign AI strategy at a minimum [22:08] for national defense. I think where that [22:11] ends up is um I saw today that the EU [22:15] had restricted exports of mrols. [22:19] >> I think that's a joke. I think that's a [22:21] meme. So people have gone so far on the [22:25] MRL leashhat thing that that they're [22:28] saying it's like mythos is distilled on [22:30] it and it's breaking all the benchmarks. [22:32] I think it's a good model, but I don't [22:34] know that it's quite there yet. [22:36] >> Yeah. [22:36] >> But you can imagine that we get there [22:38] for sure. [22:39] >> Yeah. So I think what sovereign AI will [22:41] look like for essentially all countries [22:43] really other than the United States and [22:45] and China, although I think China is [22:47] going to fall further and further behind [22:50] >> is just some sort of you do you use um [22:54] you know you use one of these providers [22:56] to do some reinforcement learning on [22:58] your language, your culture, your [23:00] values. [23:01] >> Um you have a system prompt. um you do [23:05] some supervised fine-tuning and you run [23:07] that on whatever the best open source [23:09] model is. I think that's where sovereign [23:11] AI ends up. And then you run it in your [23:13] own data centers so you feel like you [23:16] know it's safe and whatever defense [23:18] questions you're asking it or whatever [23:20] forget questions whatever defense and [23:23] intelligence um and maybe um [23:27] you know policing [23:29] activities your sovereign AI agents are [23:32] doing you know 24 hours a day. You feel [23:35] like they're safe. [23:36] >> Yeah. So I think that's where sovereign [23:37] AI ends up and I think you know there'll [23:39] still be a big buildout for that but man [23:41] sovereign AI at the frontier I don't see [23:43] it. [23:43] >> Yeah. What do you think the drivers are [23:45] of the widening gap between Chinese open [23:48] source models and American closed source [23:50] models is? [23:50] >> China's made a terrible mistake not you [23:53] know taking it feels like the [23:56] administration has been willing to let [23:58] them buy H200s or P30s. [24:00] >> Oh yeah. And I think they're on this [24:02] like, you know, they have this crazy [24:05] belief that, oh, you know, our own [24:07] internal chips are good enough. They're [24:09] not. [24:10] >> Yeah. [24:10] >> And then, you know, I think what's [24:13] making them think that is that Chinese [24:15] labs are very, very good at [24:18] distillation. Somebody told me it only [24:19] took 160,000 reasoning traces [24:22] >> from 01 and 03 to get the original [24:25] deepseek. You know, they're very clever [24:28] at industrial scale distillation, [24:30] running it through multiple APIs, [24:33] >> pinging each API from, you know, we've [24:35] all seen those iPhone farms in China. [24:38] You know, they've got they've got the [24:39] same thing for distillation and they've [24:42] got, you know, 100,000 endpoints, [24:44] distilling these models across every API [24:47] available. They've gotten really really [24:48] good at that, but man, all that goes [24:51] away if people stop releasing these [24:53] models at the frontier. And I think [24:55] mythos is a side of things to come [24:57] there. [24:58] >> Yeah. Yeah. It seems like they're [24:59] getting much better at locking down the [25:01] reasoning traces, locking down the the [25:03] espionage. It feels like uh back during [25:05] the 01 days, the labs weren't even aware [25:09] that distillation was so possible that [25:12] that was uh that that was something that [25:13] was such a such a potential threat. For [25:16] sure. [25:17] >> George, please. [25:18] >> Is any part of you worried that you'll [25:20] never see another another company like [25:22] SpaceX? [25:24] Um, [25:27] >> worried is the wrong word. [25:29] >> Not not worried, not worried, but but [25:30] like you have to appreciate you have to [25:33] appreciate that a moment like that you [25:36] it's possible that you have an [25:37] opportunity to invest in a company like [25:38] that one time in your career. And thank [25:40] God you did it because there's plenty of [25:42] people that that had the opportunity and [25:44] didn't. And uh [25:46] >> well well I did own 15% of Nvidia when [25:49] it was a sub2 billion market cap but I [25:51] own 10% of Tesla when it was a sub$2 [25:53] billion market cap. Um [25:57] but um but SpaceX certainly it's been a [26:02] special experience for me [26:05] >> and I think there is a chance that it is [26:08] you know one of the most important [26:10] iconic companies of all time. I mean it [26:12] already is. Uh maybe there's a chance [26:14] it's the most. Um [26:17] and I think in general it's just [26:20] like I I kind of want to take a step [26:22] back and like [26:24] all this trillionaire hate from the [26:26] left. It's like the you know there was a [26:29] great post from my friend um [26:32] Kevin Mafy saying like oh this is a bond [26:35] villain who's decarbonizing the world. [26:37] Last time I checked, people on the left [26:40] >> were like the environment [26:42] >> connecting u you know poor people in [26:45] low-income countries, schools and [26:47] hospitals to the internet at very low [26:49] cost and he's helping blind people see [26:53] today and you know interact with the [26:55] world. [26:57] That's a villain. Um and then just like [27:00] there's so many parts of the SpaceX [27:01] story like all of the you know [27:03] bluecollar workers who've gotten super [27:05] wealthy. um if you're opposed to data [27:08] centers on Earth because the [27:10] environment, well, SpaceX is your is [27:12] your solution. So, um you know, there [27:15] was another great post that just said, [27:17] you know, the left won and they don't [27:19] even appreciate it. The world's first [27:20] first trillionaire, [27:22] >> you know, has done more to kind of solve [27:23] the environment than like everyone else [27:25] on Earth combined. [27:27] >> Y [27:29] >> I saw there was another good one. It's [27:30] like, yeah, Elon could solve world [27:33] hunger by, you know, market selling [27:35] every, you know, uh, piece of stock that [27:38] he has and send the world into a global [27:41] economic downturn, losing, you know, [27:43] millions of millions of jobs in the [27:45] process. [27:46] >> Yeah. That uh, somebody said that Elon [27:48] that the World Bank could solve hunger [27:50] with five billion. And he and he said, [27:52] "If you prove it to me, I'll do it." [27:54] >> Yeah, of course. [27:54] >> And turns out you can't solve world [27:56] hunger for $5 billion. [27:58] >> Yeah. Um, but yeah, SpaceX is certainly [28:01] I would say, you know, SpaceX, Nvidia, [28:04] Tesla, they've all been very special to [28:07] me [28:09] in in kind of a different way, [28:11] but man, there is nothing like a rocket [28:13] launch. And so, I just don't think as a [28:16] visceral experience, I highly recommend [28:17] everyone listening, you guys, if you [28:19] have kids, take your kids to a lodge. [28:21] You don't want to go with SpaceX. [28:22] Everybody's always like, "Oh, get me the [28:24] special tickets." You just literally you [28:26] could go to a beach at Bokhica. Yeah. [28:29] >> And be where the special SpaceX viewing [28:31] area is. You could literally be closer [28:33] to the lodge. Yeah. [28:34] >> You could there's hotels near [28:36] Vandenberg. Take your kids [28:38] >> see a launch. It's super inspirational [28:41] that humans are capable of this and it's [28:44] much more visceral than you realize. [28:47] >> You know, the sound, the blast of hot [28:49] air, a lot of people cry. So go see a [28:52] lodge. there will be a ne never I don't [28:55] think a I'm pretty skeptical [28:58] that there will be another um company [29:02] that delivers a visceral experience like [29:04] that and then listen if SpaceX puts a [29:08] city on on the moon and then Mars and [29:11] enables you know becomes the British [29:13] East Indian company the solar system I [29:15] think this will and those there's a lot [29:17] of hard engineering that needs to happen [29:19] to go into all that this will probably [29:21] be the most important company um of [29:24] certainly my lifetime of maybe maybe of [29:26] all time. Um and so from that sense, I [29:28] don't think there will ever be anything [29:30] else like it. And yeah, I guess that is [29:32] a little bit sad, but you know what? [29:34] There's so much awesome stuff happening [29:36] in the world that it's it's it's tough [29:38] for me to get too down, man. [29:41] >> I love it. [29:41] >> Yeah. Hard hard to hard to be hard to be [29:44] >> I have I have one last question. We have [29:46] uh um you mentioned owning Nvidia sub [29:49] two billion. What is the health of the [29:51] sub2 billion uh market right now? Is [29:54] there is there anything that you're [29:55] excited about? [29:56] >> Is there anything that the bottleneck [29:57] bros haven't ran up? [29:59] >> Haven't run up to pros have taken [30:01] everything under two billion above. [30:04] >> Maybe not even maybe not even uh like uh [30:07] like public companies but private areas. [30:09] I mean we're starting to see the [30:10] knock-on effects of AI in bio and uh [30:14] material science and like Neurolink was [30:16] sub two billion for a long time, right? [30:18] And so there are more like not in theme [30:22] but accelerated by. So I don't know [30:24] there's also just like [30:25] >> SAS company [30:26] >> there's loads of incredible venture [30:29] companies you know startups under two [30:31] billion. Yeah. [30:32] >> Um like loads of them. Yeah. Um you know [30:36] I I I I [30:38] talked to them you know several times a [30:40] week minimum. [30:41] >> Yeah. in public, you know, until the [30:44] bottleneck bros came along, until AI ca [30:47] came along, there was not a lot [30:49] happening in public small cap land. And [30:53] by the way, on this like all these small [30:56] caps, these bottlenecks, this AI, [30:58] >> it does feel a little weird to me that [31:00] you could have like a large following [31:04] on social media, post about, you know, [31:07] buy buy stock at something. [31:09] >> Yeah. and then post about, you know, [31:12] outline your thesis in a $35 million [31:16] market cap company and send it up like a [31:19] thousand%. [31:20] Um, and there's a lot of that going on. [31:23] And by the way, this not this is not [31:24] like a serenity comment. I think that [31:26] guy's that person [31:28] >> is smart, [31:30] >> but I mean there's there's there's [31:32] there's a lot of shady stuff happening [31:34] it feels like to me. [31:35] >> Sure. [31:35] >> Well, yeah. And it you look back at uh [31:38] you know you were a young uh young [31:40] analyst at at Fidelity right during the [31:42] the doc you know the end of the the.com [31:45] cycle and there was a lot of just like [31:47] euphoria stuff happening and in the [31:50] years that followed there was [31:52] consequences right there there were [31:54] people ultimately had to to face um [31:58] >> sometimes civil penalties all sorts of [32:00] different things companies went bankrupt [32:01] there were all sorts of different uh you [32:03] know perturbations in the market. And of [32:06] course, we got a co we got a bunch of [32:08] really enduring generational companies [32:10] through that period, too. [32:11] >> Although a lot of these accounts are [32:13] anonymous and I'm sure they're logged [32:15] into social media with a madeup email. [32:17] >> Yeah. [32:18] >> You know, gone through a VPN. And so [32:21] >> I don't I don't know. We'll see. [32:23] >> It is different than a research than a [32:25] cellside research report that maybe goes [32:28] a little bit too far, but there's still [32:29] a whole compliance department around it. [32:31] Uh we are in Yeah. uncharted territory. [32:34] So, uh, everyone needs to stay safe out [32:37] there and do your own research, I [32:38] suppose. [32:40] >> Yeah. [32:40] >> Anyway, [32:40] >> well, guys, man, this was super fun. [32:42] >> This was great. Let's do it again soon. [32:43] >> Come on more. [32:44] >> Thank you so much. Anytime. Let's get [32:46] the whole team on, too. [32:47] >> Yeah. You know what? That's actually [32:49] fun. [32:50] >> Let's do that with the whole team. [32:52] >> That' be amazing. [32:53] >> Awesome. [32:53] >> I would love it. Yeah. Yeah. You got a [32:55] conference [32:57] room with a with a nice wide angle [32:58] camera. You can just have the whole crew [33:00] there. That's great. [33:01] >> Partner into a partner meeting. Exactly. [33:03] Awesome. Great to hang with you. [33:05] >> Have a great rest of your day. Thank you [33:06] so much for coming on.