Airtable Sells at 90% Discount as AI Coding Disrupts Software Models

AI编程冲击软件模式，Airtable折价90%出售 生成式AI编程工具对低代码软件供应商构成威胁，而企业安全与合规要求则保护了老牌基础设施平台。 嘉宾团队分析了Google的AI管理层更迭与工程人才离职，以及其重注基础设施资本支出而非冒险开发模型的策略。他们深入探讨了SpaceX单季78亿美元的营收、Starlink庞大的现金流创造能力，以及扩展AI算力集群的资金需求。最后，主持人们以Airtable折价出售给Bending Spoons为例剖析了软件行业的估值缩水，并探讨了美国企业向中国科技公司出售AI训练数据的地缘政治影响。 Demis Hassabis 转任 DeepMind 董事长兼 Google 首席科学家，效力 27 年的功勋工程师 Jeff Dean 离职创立 Discovery Loop，导致 Google 股价应声下跌约 4%（市值缩水约 2000 亿美元）。 (2:16) Google 承诺今年投入 2000 亿美元资本支出用于 AI 基础设施，借助加速折旧税收优势锁定可预测回报；同时 Google Cloud 营收同比增长 82%，Gemini 月活跃用户数已达 9.5 亿。 (3:43) David Sacks 指出前沿大模型市场已收敛为 OpenAI 与 Anthropic 的双寡头垄断，其中 Anthropic 的 ARR 已从年初的 100 亿美元激增至超 800 亿美元，年底有望突破 1100 亿至 1200 亿美元。 (9:14) SpaceX 第二季度营收达 78 亿美元，估值站上 1.4 万亿美元并提前至 2030 年冲击 1 万亿美元 ARR 目标；其中 Starlink 单季进账 43 亿美元充当造血核心，后续 Starship V3 卫星单次部署带宽将跃升超 20 倍至 60 Tbps。 (20:39) Airtable 以 12.8 亿美元企业价值（含现金总计 22.5 亿美元）出售给欧洲买家 Bending Spoons，估值较 2021 年峰值缩水约 90%，且其核心 AI 业务 HyperAgent 已在交易前剥离为独立公司。 (48:01)

The All-In panel examines structural shifts across the technology sector, arguing that massive capital expenditures and rapid AI advancements are disrupting traditional software models, straining hyperscaler capital allocation, and consolidating frontier intelligence into a high-stakes capital race. Google's Capex Dilemma and Frontier Research Departures (2:16) Google faced major leadership changes as Demis Hassabis was named Chair of DeepMind and Chief Scientist at Google, alongside reports of low morale and development delays on Gemini 3.5 Pro. The departure of 27-year veteran Jeff Dean and three senior researchers to launch Discovery Loop coincided with an estimated 4% decline in Google's market cap, wiping out roughly $200 billion in value. David Friedberg highlights that Google committed $200 billion in capital expenditures this year, noting that physical data center infrastructure yields predictable returns on invested capital bolstered by immediate accelerated depreciation tax benefits (~26% offset). In contrast, Friedberg characterizes frontier model development as high-beta and increasingly vulnerable to commoditization from open-weight models. Brad Gerstner and Friedberg emphasize an internal channel conflict within tech giants: cloud units seek to rent compute externally to third-party labs like Anthropic—with Microsoft reporting over 30% returns on invested capital for token infrastructure—while internal research teams compete for those same GPU clusters, encouraging researchers to depart for fully capitalized independent ventures. Frontier Duopoly vs. Enterprise Hybrid Model Adoption (9:14) David Sacks argues that the frontier intelligence market has compressed into an OpenAI and Anthropic duopoly capable of commanding premium pricing, citing Anthropic's annual recurring revenue expanding from $10 billion to over $80 billion, with projections between $110 billion and $120 billion or more. Under this framework, open-source or lagging models become a commoditized tier competing strictly on compute costs and inference speed. Conversely, Jason Calacanis argues that open-source models already satisfy 95% of typical tasks, pointing to Google Cloud's 82% year-over-year revenue growth and Gemini reaching 950 million monthly active users in Q2. David Friedberg contends that enterprises will avoid choosing exclusively between open-source and frontier models, instead selecting a hybrid blend. While everyday internal tasks rely on cheap open-weight models, enterprises will buy specialized closed models where companies possess proprietary domain data, such as Google's life sciences and video assets. Brad Gerstner adds that closed models can prove more economical for enterprises once the hidden expenses of fine-tuning, hosting, and maintaining open-source systems are factored in. SpaceX Financial Momentum and Capital Expenditure Constraints (20:39) SpaceX settled at a $1.4 trillion valuation following its public earnings report showing second-quarter revenue of $7.8 billion, up 92% year-over-year. Compute rental revenue from Elon Web Services reached $2.6 billion, while quarterly AI capital expenditures climbed to $18.4 billion. Brad Gerstner highlights Elon Musk's aggressive projections targeting $100 billion in ARR by year-end and advancing a $1 trillion ARR target to 2030, contrasting with Morgan Stanley's 2030 estimate of $325 billion. Starlink serves as the primary cash engine, generating $4.3 billion in quarterly revenue and $2.6 billion in adjusted EBITDA across 12 million subscribers, with projections toward $30 billion in free cash flow within 12 to 18 months. Sacks and Friedberg argue Starlink cash flow subsidizes experimental initiatives like the Terafab semiconductor project, enabled by Starship V3 launches that expand added bandwidth capacity from 2.6 Tbps to 60 Tbps per launch. However, Sacks warns that expanding compute capacity from 2 GW to 8 GW requires $300 billion in financing ($50 billion per GW), exacerbated by high-bandwidth memory supply projected to grow only 20% against demand growth exceeding 200%. Airtable's Valuation Collapse and Software Defense Moats (48:01) Airtable's acquisition by Bending Spoons for $1.28 billion enterprise value ($2.25 billion including cash) represents a roughly 90% plunge from its peak $11.7 billion valuation in 2021, despite generating $480 million in revenue with 20% growth. Sacks notes Airtable spun off its AI agent initiative, HyperAgent, prior to the transaction, and points out that only 30% of Airtable's enterprise sales team hit quota. Sacks and Gerstner observe that venture-backed startups struggle to shift into aggressive operational cost-cutting due to cap table friction from high liquidation preferences, while private equity buyers can trim overhead and rely on modern AI tools to maintain legacy codebases. The panel differentiates between vulnerable software categories and defended incumbents. Low-code and no-code platforms face direct disruption because natural language coding allows teams to build custom internal applications rapidly at a fraction of commercial SaaS costs. However, Sacks cautions against anticipating a universal software collapse, noting that core enterprise platforms like Microsoft and Salesforce maintain defensible positions secured by identity systems, deep organizational integration, and regulatory compliance accreditations such as FedRAMP High and DoD Impact Level 5. Geopolitical Risks in Cross-Border AI Data Commerce (65:58) US data annotation providers such as Scale AI and Surge AI are reportedly selling expert-annotated and human alignment datasets to Chinese technology firms, including Tencent, ByteDance, Alibaba, and Moonshot AI, with Chinese labs spending an estimated $500 million annually on Western training data. Jason Calacanis notes that these specialized data pipelines have driven notable performance and reasoning improvements in Chinese models such as Qwen and Kimi. The panel examines the geopolitical tradeoffs of these transactions. David Sacks argues that curated datasets function as commercial commodities rather than dual-use military technologies, warning that export restrictions could invite reciprocal Chinese retaliations on strategic supply chains such as rare earth elements. Brad Gerstner contends that if American intelligence assesses that foreign models are erasing the US lead, commercial data sales to foreign competitors will attract intense political and regulatory scrutiny in Washington. The discussion leaves open whether massive compute expansions can generate sustainable returns before debt and seller-financing leverage strain balance sheets, and whether enterprise software can maintain historic pricing power as automated coding and commodity open-source models advance.

All-In 节目嘉宾分析指出，人工智能浪潮正步入资本开支与现实回报的严峻博弈期：谷歌等大厂面临基础设施投资与前沿模型高风险研发的战略分歧，SpaceX 依托 Starlink 的强劲现金流大举押注算力网络，而无代码软件在自然语言编程冲击下遭遇估值崩溃，同时中美 AI 竞争也正从硬件层蔓延至数据要素流动的地缘审查。 谷歌人才动荡与资本配置两难 (2:16) 谷歌近期迎来重大人事调整，Demis Hassabis 履新 DeepMind 董事长兼首席科学家，为公司效力27年的核心元老 Jeff Dean 则携团队出走创立 Discovery Loop，消息引发谷歌市值单日下挫约4%（市值蒸发约2000亿美元），同时 Axios 披露 Gemini 3.5 Pro 研发进度受内部士气影响延期数月。David Friedberg 分析指出，谷歌今年承诺投入2000亿美元用于数据中心和 AI 算力基建，在美国加速折旧税法下享有约26%的即时税收抵免，投资回报率可预期性强（如微软报告代币化算力云服务的投资回报率超过30%），属于高 Alpha、低 Beta 资产；相反，动辄耗资百亿美元训练前沿模型则面临开源模型快速追赶的贬值风险，属于高 Beta 投机。 关于大模型竞争格局，David Sacks 主张前沿智能领域正在形成 OpenAI 与 Anthropic 的双头垄断，其中 Anthropic 年化经常性收入（ARR）已由年初的100亿美元飙升至800亿美元以上，并在向年底千亿美元目标迈进，促使非前沿落后模型沦为只能拼推理速度与电费的同质化资产。Jason Calacanis 则提出反向论据，指出谷歌云营收同比强劲增长82%，且 Gemini 在二季度月活跃用户达到9.5亿，庞大的产品矩阵与分发网络仍是其最核心的竞争壁垒；而在大多数日常任务中，低成本开源模型已逐步逼近商用及格线，促使企业未来更多采取混合调用策略。 SpaceX 业绩爆发与巨额算力融资困局 (20:39) SpaceX 首份公开财报显示其二季度营收达到78亿美元（同比增长92%），公开市场估值稳定在1.4万亿美元左右。其中出租 Colossus 集群算力的 Elon Web Services（EWS）营收环比激增超两倍至26亿美元。Elon Musk 将2030年实现1万亿美元 ARR 的目标进一步提前（远超摩根士丹利预期的3250亿美元）。节目指出，Starlink 季度营收达43亿美元、调整后 EBITDA 达26亿美元，用户翻倍至1200万户，未来12至18个月内有望产生高达300亿美元的自由现金流，成为资助太空探索与地面半导体基建的关键资金引擎。 然而，激进的算力扩张也带来了空前的流动性考验。Musk 计划将算力基础设施从现有的 2 吉瓦（GW）扩展至约 8 GW，新增的 6 GW 按当前每吉瓦500亿美元的基准建设成本计算，意味着未来一年需要高达3000亿美元的资本开支。Sacks 与 Brad Gerstner 警告，目前虽然每瓦50美元的现货算力支撑了低于一年的回报周期，但全球高带宽内存（HBM）产能明年预计仅增20%，远落后于超200%的需求激增；加之行业广泛采用卖方融资与表外特殊目的载体（SPV）加杠杆，一旦终端需求不及预期，供应链可能面临剧烈回调风险。 Airtable 暴跌与企业软件分化 (48:01) 米兰投资机构 Bending Spoons 以12.8亿美元企业价值（含账面近10亿美元现金总计22.5亿美元）收购协同办公平台 Airtable，较其2021年117亿美元的估值峰值缩水近90%。Sacks 透露，Airtable 在出售前已将核心 AI 智能体业务 HyperAgent 剥离给原团队，而其传统销售团队的业绩指标达成率仅有30%。在零利率时期（ZERP）过度推高的估值和复杂的优先清算权架构，使得原管理层与风投无法在内部实施残酷的裁员重组；而外部私募买家可通过削减85%至90%的成本并剥离销售人员，直接榨取3亿至4亿美元的 EBITDA 利润。 嘉宾深入探讨了“无代码/低代码”软件在 AI 时代遭遇的技术性颠覆。用户曾需花费高昂学习成本掌握专有软件工具，如今通过 Claude Code 等模型的“自然语言编程”（vibe coding），小团队仅需几周即可搭建过去需要数十万美元 SaaS 授权及百万美元集成费用的定制系统。不过，Sacks 强调传统成熟 SaaS 仍享有稳固的合规与身份护城河：微软 Active Directory 深度绑定组织账号，Azure 具备美军 Impact Level 5 和 FedRAMP High 认证，Salesforce 更是深度嵌入全美15个内阁级联邦机构，极高的迁移成本和合规审计壁垒使其免于被轻易替换。 美中 AI 数据贸易与地缘政治审查 (65:58) 媒体披露包括 Scale AI 和 Surge AI 在内的美国顶尖数据标注企业，长期向腾讯、字节跳动、阿里巴巴和月之暗面等中国科技实验室出售由西方专家及博士标注的高级训练与 RLHF 数据，中国科技公司每年在此类高质量训练数据上的采购开支预估达5亿美元。这批高质量数据显著促进了 Qwen、Kimi 等中国模型在推理与速度维度的性能跃升。 这笔跨国商业交易在嘉宾间引发了监管与安全层面的激烈分歧。Sacks 认为，基础与标注数据属于广义的商业商品而非双用途军工技术，且中国本土庞大的理工科（STEM）毕业生规模赋予其天然的长期标注能力；若美国切断数据贸易，可能招致中国对稀土等关键原材料的反制。Gerstner 则反驳称，随着中国开源大模型迅速逼近西方顶尖水平，白宫与国会对向战略竞争对手输出高质量逻辑对齐数据的敏感度正急剧上升，该领域可能面临与过去先进制程光刻机禁运类似的严格国家安全审查。 本期节目展现出 AI 浪潮正在由前期的纯技术狂热迈向残酷的资本重构期：前沿模型开发面临收益不确定性与开源蚕食，非核心 SaaS 遭遇大模型自然语言编程的直接挤压，而数百亿美元级别的基础设施扩建则高度依赖 Starlink 这类成熟业务的造血能力。随着算力融资杠杆持续累积与跨境数据流动引发监管警觉，如何在资本高消耗与实际商业落地之间维系平衡，成为摆在科技巨头面前的核心课题。

## Background 本期内容由知名科技与商业播客《All-In Podcast》于 2026 年 8 月 13 日发布在 YouTube 平台上。视频简介显示，本次节目由 Altimeter Capital 创始人 Brad Gerstner 替补 Chamath Palihapitiya 出席，几位主持人重点探讨了 Google 的 AI 高管人事震荡与资本支出抉择、SpaceX 亮眼的季度业绩及算力数据中心扩张计划、协作软件独角兽 Airtable 遭遇 90% 估值跳水贱卖，以及美国初创企业向中国 AI 实验室出口训练数据等焦点议题。 ## TL;DR 本期节目深入剖析了科技巨头在 AI 军备竞赛中的资本重构与行业洗牌。嘉宾们首先探讨了 Google 顶尖科学家离职潮背后的战略逻辑，指出巨头正从高风险的前沿模型研发转向回报确定性更高的模型中立算力基建；随后拆解了 SpaceX 的首份季度财报，分析 Starlink 强劲的现金流如何反哺其在 AI 算力租赁与 Terafab 超级芯片厂上的百亿美元豪赌。在探讨协同办公平台 Airtable 以 1 折估值被 Bending Spoons 收购的案例时，嘉宾深入揭示了无代码赛道遭受 AIAgent 降维打击与传统 SaaS 泡沫破裂的残酷现实；最后围绕美国数据公司向中国 AI 巨头出售博士级微调数据的报道，就大宗商品化数据与国家安全管制红线展开了激烈交锋。 ## Atmosphere & style 整场节目延续了《All-In》标志性的富豪风投沙龙风格，在随性幽默的插科打诨与极其硬核的商业推演之间无缝切换。开场对 Chamath 工地迷因与着装的调侃奠定了轻松熟络的基调，但进入核心议题后，对话迅速展现出顶级资本操盘手极具压迫感的洞察力与专业度——不仅信手拈来税收折旧法案、资本回报率（ROIC）与资本结构表清算优先权等深层逻辑，还在中美大模型竞争和无代码软件灭绝等议题上数度针锋相对。整场对谈既充溢着对马斯克式工程硬科技奇迹的英雄主义崇敬，又透出对资本泡沫出清、企业战略误判与地缘博弈暗流的冷静审视，为观众带来了犹如亲历顶级投资机构闭门合伙人会议的真实感。 ## Bulletin - [02:16](https://www.youtube.com/watch?v=muRIXCDw-k0&t=136s) 谷歌遭遇核心技术元老出走冲击，效力27年的传奇架构师 Jeff Dean 携团队离职创立科学AI公司 Discovery Loop，Demis Hassabis 转任 DeepMind 董事长，推动谷歌加速向资本回报更确定的算力基础设施平台转型。 - [20:39](https://www.youtube.com/watch?v=muRIXCDw-k0&t=1239s) SpaceX 首份上市季报营收达78亿美元，其中将超级算力出租给 Anthropic 和谷歌的 AI 云业务季度营收突破26亿美元，单季资本支出狂飙至184亿美元以布局2到8 GW规模的算力集群。 - [48:01](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2881s) 无代码巨头 Airtable 以12.8亿美元超低估值被 Bending Spoons 收购，较2021年117亿美元估值缩水近90%，交易前其将 AI Agent 业务 HyperAgent 独立剥离，印证了传统 SaaS 估值与商业模式在 AI 浪潮下的残酷洗牌。 - [01:05:56](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3956s) 《福布斯》调查披露 Scale AI 和 Mercor 等美国高估值数据公司正向腾讯、字节、阿里、月之暗面等中国实验室供应博士级高质量专业数据，引发关于顶尖训练数据是否应受地缘政治管制的激烈交锋。 ## Key takeaways - [02:16](https://www.youtube.com/watch?v=muRIXCDw-k0&t=136s) 谷歌高层管理与技术骨干剧烈震荡，Gemini 3.5 Pro 发布延期，Jeff Dean 离职创办 Discovery Loop，引发市场对模型研发团队士气及巨头核心竞争力的担忧。 - [03:41](https://www.youtube.com/watch?v=muRIXCDw-k0&t=221s) 巨额 AI 投资促使科技巨头聚焦资本回报率（ROIC）：在税收加速折旧和算力饥渴驱动下，搭建算力中立平台的收益确定性远高于吞噬数百亿美元的前沿模型研发。 - [07:11](https://www.youtube.com/watch?v=muRIXCDw-k0&t=431s) 云服务巨头内部普遍面临渠道冲突：云部门倾向将高溢价 GPU 租给第三方以锁定利润，而内部前沿大模型团队却面临算力配额被挤压的困境。 - [09:26](https://www.youtube.com/watch?v=muRIXCDw-k0&t=566s) 最顶尖大模型市场正演变为 OpenAI 与 Anthropic 主导的双头垄断，落后6至12个月的模型迅速沦为微利的大宗商品，形成类似 iPhone 与 Android 的利润与规模分化格局。 - [14:46](https://www.youtube.com/watch?v=muRIXCDw-k0&t=886s) 企业在模型选型上正快速走向混合编排（Blend of Models）：高敏复杂领域采购顶级闭源前沿模型，常规低成本工作流则部署开源权重。 - [20:39](https://www.youtube.com/watch?v=muRIXCDw-k0&t=1239s) SpaceX 的 AI 算力出租业务“Elon Web Services”单季收入环比激增至26亿美元，显示出其依托 Colossus 等超级算力中心变现的惊人爆发力。 - [21:52](https://www.youtube.com/watch?v=muRIXCDw-k0&t=1312s) 马斯克设定 SpaceX 于2030年达成1万亿美元 ARR 的激进目标，计划将算力从年底的 2 GW 扩建至 8 GW，但单年高达3000亿美元的潜在资本开支引发华尔街对现货价格可持续性的忧虑。 - [24:46](https://www.youtube.com/watch?v=muRIXCDw-k0&t=1486s) Starlink 成为 SpaceX 旗下强大的造血机器，单季创收43亿美元且 EBITDA 达26亿美元，充沛的现金流持续反哺深空探索与 Terafab 超级晶圆厂等高风险硬科技资产。 - [36:23](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2183s) Starship 搭配 V3 卫星单次发射可带来超过 20 倍的总网络容量跃迁（达 60 Tbps），直接为手机直连卫星及承载全球半数互联网流量的构想铺平技术路径。 - [40:22](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2422s) 业内警惕由硬件厂商兜底引发的过度“卖方融资”与“循环收入”隐患，一旦需求或现货价格松动，建立在透支预期上的高杠杆算力基建将面临资产重估风险。 - [49:26](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2966s) Airtable 估值从117亿美元跌落至12.8亿美元贱卖，暴露了强行用传统大企业直销团队替代自驱动 PLG 模式导致的效率崩塌与利润被蚕食。 - [57:57](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3477s) 自然语言 Agent 助手正在对无代码（No-Code）赛道实施降维打击，普通用户无需再投入高昂学习成本掌握小众平台的伪编程规则。 - [01:00:20](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3620s) 基于 AI 的轻量化“氛围编程（Vibe Coding）”可快速低成本开发内部工具，但大型企业及政府机构由于深度嵌入的合规、安全与认证壁垒，依然牢固维系着传统核心 SaaS 的护城河。 - [01:02:32](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3752s) Airtable 交易中后期高估值基金依靠1倍不参与分配清算优先权拿回本金，早期投资者实现盈利套现，体现了硅谷在估值神话破灭后的典型软着陆退出生态。 - [01:05:56](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3956s) 中国前六大 AI 实验室每年斥资约5亿美元采购美国由顶尖学者参与的高难度 RLHF 与领域真值数据，引发该类数据属于大宗商品劳务还是关键战略资产的政策博弈。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=muRIXCDw-k0&t=0s) 开场寒暄：夏日缺勤与查马斯的“沙丘”数据中心梗 节目在八月初录制，主持人 Jason Calacanis 开场打趣说夏天凑齐四位“好基友”（Besties）真不容易。本次节目由著名投资人、Altimeter Capital 创始人 Brad Gerstner 替补 Chamath Palihapitiya 出席。Jason 沿袭传统为 Brad 编了一段押韵的打油诗介绍，戏称他是“如果你热衷市场博弈，他就是带你发财的布鲁斯·韦恩（Bruce Wayne）”。随后，稍稍迟到的 David Sacks 穿着一身格外抢眼的夏日马甲（Gilet）现身，引发众人调侃他是不是穿得太暖和，以及模仿 Chamath 经典名言“让赢家继续奔跑”（Let your winners ride）的搞笑段子。 现场还展示了网络上疯传的 Chamath 迷因图。据称 Chamath 正奔波在外视察其在内华达或科罗拉多州投资的数据中心建设项目，网友们却纷纷将他在工地的照片恶搞为科幻电影《沙丘》（Dune）中的剧照，甚至连 Chamath 的妻子都在社交媒体上公开调侃。众人一致认为，当一个人的家人都开始加入调侃时，说明这个网络迷因已经彻底火出圈了。 ### [02:16](https://www.youtube.com/watch?v=muRIXCDw-k0&t=136s) 谷歌管理层巨震：AI 大脑流失与战略转向 Jason 迅速切入当周重磅科技新闻：Google 内部发生重大 AI 人事震荡。DeepMind 联合创始人兼负责人 Demis Hassabis 转任 DeepMind 董事长兼 Google 首席科学家。外界与媒体普遍将此举解读为 Demis 被“明升暗降”（kicked upstairs），尽管 Google 官方口径坚称这是一次晋升。与此同时，根据 Axios 报道，Google 的旗舰模型 Gemini 3.5 Pro 的发布进度已落后数月，部分内部人士透露这是由于团队士气低落；更严重的是，多位顶级研究主管（包括 Gemini 的联席负责人）已相继离职投奔竞争对手。 更为轰动的是，Google 的功勋元老、传奇架构师 Jeff Dean 携手另外三位顶级 AI 专家宣布离职，创办了一家名为 Discovery Loop 的新公司，专注于利用 AI 推动深层科学突破。Jeff Dean 早在 1999 年作为第 30 号员工加入 Google，连续任职长达 27 年，是计算机系统与 AI 领域的殿堂级人物。受此离职消息冲击，Google 股价应声下跌 4%，市值蒸发约 2000 亿美元。Jason 由此向曾任职于 Google 的 David Friedberg 提问：这究竟是一场优胜劣汰的“创造性破坏”，还是在 AI 资本狂潮面前，即便功勋元老也无法抵挡离职自立门户的诱惑？ - **人事变动要点**：Demis Hassabis 转任 DeepMind 董事长兼首席科学家；Gemini 3.5 Pro 延期且内部士气低迷。 - **Jeff Dean 离职**：在 Google 效力 27 年（1999 年入职，第 30 号员工），离职联合创立 Discovery Loop。 - **市场反应**：Google 股价单日下挫 4%，对应市值缩水近 2000 亿美元。 ### [03:41](https://www.youtube.com/watch?v=muRIXCDw-k0&t=221s) 资本配置的抉择：买铲子还是淘金子？ 针对 Google 的人事变动，David Friedberg 给出了极具深度的资本配置分析视角。他指出，这并非单纯的人才流失，而是 Google 董事会与管理层在资本回报率（ROIC）上做出的战略抉择。Google 今年承诺在 AI 基础设施与数据中心建设上投入高达 2000 亿美元的资本支出（Capex）。得益于美国现行的税收与加速折旧政策，这类硬核算力投资具有极大的税收优势，并且由于全球对算力的极端渴求，该业务的投资回报路径非常清晰、确定性极高。 与稳赚不赔的基础设施相比，研发前沿大模型则需要吞噬数百亿美元，但商业变现前景却充满变数。随着开源模型与开放权重（open weights）技术的迅猛追赶，各大顶尖实验室之间的差距被迅速抹平，继续砸重金自研单一前沿模型的边际效益正在锐减。Google 拥有全球首屈一指的企业级与消费级客户群，其最优解是转变为“模型中立”（Model-agnostic）的算力平台——既可以托管 Anthropic、OpenAI 的模型，也能支持各类开源生态，还能凭借对 Anthropic 和 SpaceX 的投资获利。在这种背景下，原本一心追求模型前沿突破的科学家们发现内部算力资源正向云基础设施倾斜，自然会选择带上光环离开体制，凭借顶尖声誉用几页 PPT 去风险投资市场轻松募集数十亿美元。 ```mermaid flowchart TD accTitle: AI投资回报路径对比 accDescr: 对比资本支出投入计算基础设施与投入前沿模型研发的风险收益特征差异 subgraph Infra["基础设施路线 (算力中心 / 云服务)"] A1["2000亿美元 Capex 投入"] --> A2["税收加速折旧优势 (约26%直接抵扣)"] A2 --> A3["全球算力极度渴求 / 模型中立托管"] A3 --> A4["低 Beta、高 Alpha: 确定性高额 ROIC"] end subgraph Model["前沿模型路线 (Gemini 等基础模型)"] B1["数百亿美元模型研发与训练"] --> B2["开源/开放权重模型极速追赶"] B2 --> B3["各实验室差距抹平、变现承压"] B3 --> B4["高 Beta、回报不确定: 资本配置意愿下滑"] end Infra -.->|"挤压模型研发资源导致"| Push["顶尖科学家 (如 Jeff Dean) 外出创业"] ``` ### [07:11](https://www.youtube.com/watch?v=muRIXCDw-k0&t=431s) 渠道冲突与巨头分化：基础设施与模型实验室的宿命 Brad Gerstner 完全认同 Friedberg 的判断，并补充指出微软目前也呈现相同态势。摩根士丹利的最新报告显示，微软在“代币即服务”（Tokens as a Service，即 AI 算力与基础设施租赁）上的已投资资本回报率（ROIC）超过 30%。这种“建好机房租给所有人”的铲子生意极其暴利，但直接导致了巨头内部不可调和的“渠道冲突”（Channel Conflict）：Google Cloud 希望将所有算力高价租给 Anthropic 等第三方客户以兑现利润，而内部从事前沿模型研发的团队却急需这些算力去与 Anthropic 竞争。 这种渠道冲突同样发生在 SpaceX、微软甚至正在进军算力服务业的 Meta 身上。反观 OpenAI 和 Anthropic 等纯粹的前沿实验室，完全没有算力外租的渠道冲突，目标纯粹而聚焦。Brad 进一步算了一笔财务账：如果按美国 26% 的企业所得税率计算，得益于当期资本支出全额折旧扣除的法案，企业每向数据中心投入 1 美元 Capex，当年就能直接省下 26 美分的税款，这等于政府直接给硬件基础设施投资打了 74 折。这进一步促使 Google 和微软等巨头从模型先锋蜕变为基础设施提供商。随后，Jason 将话题引向 Polymarket 预测市场上关于“谁将在年底拥有全球第一 AI 模型”的下注情况。 - **基础设施的高利润率**：根据投行报告，微软等云服务商在 AI 代币基础设施业务上的 ROIC 达到 30% 以上。 - **不可避免的渠道冲突**：云基础设施部门（追求租售算力获得确定性现金流）vs 内部模型研发团队（争抢内部算力攻克前沿）。 - **税收红利杠杆**：在 26% 适用税率与资本支出一次性折旧政策下，硬件与数据中心投资在首年即可获得约 26% 的税收抵免效应。 ### [09:26](https://www.youtube.com/watch?v=muRIXCDw-k0&t=566s) 前沿大模型走向双头垄断与市场分层 针对语言模型市场究竟是前沿模型胜出还是正在迅速商品化、未来是该卖 Token 还是深耕应用层的问题，大卫·萨克斯（David Sacks）提出了关于当前市场结构的判断。在看到谷歌近期的人事与战略调整后，他的第一反应是整个赛局“只剩下两家”（"and then there were two"）。一年前尚有五大主流竞争者在全力争夺前沿模型实验室的头把交椅，而如今真正的最顶尖前沿模型市场已逐渐演变为由 Anthropic 和 OpenAI 主导的双头垄断格局。尽管埃隆·马斯克（Elon Musk）的 xAI 仍在追赶，谷歌也坚称自己仍在前沿梯队，但谷歌本身存在相互冲突的激励机制——即便仅凭借强大的算力基础设施，谷歌也能赚得盆满钵满。 萨克斯认为前沿智能并未被商品化，而是演化出了清晰的双层市场结构：一层是真正的“前沿智能”，另一层则是落后 6 到 12 个月的“大宗商品化智能”。对于落后的模型，厂商无法对模型权重本身收取溢价，只能赚取基础算力、推理调用费以及配套的工程咨询服务费；唯有真正处于最顶尖前沿的实验室，才能为模型层享有丰厚的品牌溢价与定价权。萨克斯将这种格局类比为苹果（Apple）与安卓（Android）的对抗：虽然开源的 Android 占据了全球绝大部分的用户体量，但绝大部分利润却被提供极致溢价体验的 Apple 收入囊中。同样，虽然非前沿模型体量巨大，但人们只愿意支付算力成本；而对于真正的最前沿智能，客户甘愿支付高昂溢价。这一趋势从两家领头羊迅猛的 ARR 增速中得到了有力印证。 ```mermaid flowchart TD accTitle: AI市场双层分化架构 accDescr: 前沿智能双头垄断收取溢价，大宗智能依靠底层算力与推理微利竞争 subgraph FrontierLayer["前沿智能层（高溢价模式）"] Duopoly["OpenAI & Anthropic 双头垄断"] Duopoly --> Premium["模型层高溢价定价权"] Duopoly --> EnterpriseScale["前沿探索与关键高价值业务"] end subgraph CommodityLayer["通用/落后模型层（薄利多销模式）"] Laggards["开源/落后 6-12 个月的模型"] Laggards --> CostOnly["模型权重免费，仅赚算力与推理成本"] Laggards --> HighVolume["常规自动化与低敏工作流"] end ``` - **市场梯队预测份额**：OpenAI 占 32%、Google 占 20%、阿里巴巴占 14%，Moonshot、xAI、Meta、字节跳动各占约 10% 左右。 - **Anthropic 业绩爆发**：从年初的 1000 万/百亿级别基数飞速攀升，最新统计已突破 800 亿美元/千万级 ARR（节目口述为 "80 billion ARR"），原定年底冲刺 1000 亿的目标有望提前数月完成，华尔街对其年底出境 ARR 预期已被上调至 1100 亿至 1200 亿美元或更高。 - **OpenAI 增长再提速**：同样呈现强劲的加速上扬态势，进一步夯实了双头垄断的定价壁垒。 --- ### [12:32](https://www.youtube.com/watch?v=muRIXCDw-k0&t=752s) 谷歌生态反扑与开源可用性的论战 杰森·卡拉卡尼斯（Jason Calacanis）对此持不同看法，他从分发渠道与规模化落地角度为谷歌站台。谷歌云（Google Cloud）刚刚录得 82% 的惊人同比营收增速，创下主流云服务商历史新高；同时马斯克的 xAI 与 SpaceX 协同建立的算力服务（被戏称为“埃隆网络服务” Elon Web Services）也在迅猛起量。杰森坚信谷歌依然会成为头号 AI 巨头，其杀手锏在于深厚的消费者产品壁垒：谷歌旗下目前拥有 5 款月活跃用户（MAU）突破 30 亿的超级应用（Android、Google Search、Gmail、Chrome、YouTube），这些产品正在全面向“AI 优先”转型。此外，谷歌跨过 10 亿 MAU 门槛的产品总数已达 13 款，其中 Gemini 在第二季度的月活跃用户已突破 9.5 亿，实现同比三倍暴增。 在模型选型上，杰森透露自己日常工作中已几乎完全使用非前沿的开源模型，并强调在 95% 的实际应用场景中其表现已经“足够好”（"good enough"）。杰森在社交平台发推表示开源模型与前沿模型的差距在实际感知中已微乎其微，随即引发埃隆·马斯克的回击。马斯克反驳称两者存在“天壤之别”（"a world of difference"），前沿模型在复杂推理上不可或缺；而杰森则坚持认为，除非是要克隆一款复杂的电子游戏或对超低延迟有严苛要求，否则普通人根本不需要前沿模型。当前大量企业采购前沿模型只是因为自建开源部署门槛过高，随着开源工具链愈发成熟，开源大模型与 Gemini 必将统治绝大部分市场。 - **谷歌超级应用矩阵**：Android、Search、Gmail、Chrome、YouTube 均坐拥 30 亿以上月活，且均在加速植入前沿 AI 功能。 - **Gemini 爆发增长**：第二季度月活达 9.5 亿，较去年同期暴涨 300%，在消费级 AI 用户覆盖量上遥遥领先。 - **马斯克与杰森论战**：核心争议在于日常企业任务中开源与前沿模型是否存在不可替代的断代差距。 --- ### [14:46](https://www.youtube.com/watch?v=muRIXCDw-k0&t=886s) 极致性能与模型混合编排的博弈 萨克斯回应了杰森的观点，指出廉价手机固然能满足基本通话需求，但重度用户依然会为顶级旗舰机买单。对于对冲基金等处于高度对抗性赛道的企业而言，稍逊一筹的智能水平就意味着巨大的财务损失，竞争动力会倒逼其不计成本采用最前沿的模型。引用 Decagon 博客的技术洞见，在大量探索性、未成熟的 AI 落地场景中，早期根本无法预知价值释放点，寻找最佳商业模式带来的潜在回报远超微小的 Token 溢价差额；因此只要员工不搞“恶意刷榜”（"token maxing"）浪费算力，采购最顶尖前沿模型是绝对性价比最高的策略。 大卫·弗里德伯格（David Friedberg）则给出了兼具技术与商业视角的折中框架：未来的终局并非在开源与顶尖闭源之间进行非此即彼的二选一，而是“混合编排”（blend of models）。在生命科学与基因组学建模等高精尖领域，企业必须采购最顶级模型；在影视媒体的多模态渲染上，视频生成首选 Gemini 或 Sora；而在海量常规工作流与个人轻量级工具中，部署廉价的开源权重则是最优解。弗里德伯格强调不可低估谷歌在垂直领域的护城河：谷歌在视频多模态数据与生命科学领域耕耘已久，德米斯·哈萨比斯（Demis Hassabis）领导的 Isomorphic Labs 在蛋白质折叠等前沿生物学领域遥遥领先。企业客户更倾向于选择能够同时提供全栈混合模型调度的云平台（如 Google Cloud Platform），这倒逼 OpenAI 与 Claude 发起剧烈的 Token 价格战，促使模型编排框架迅速渗透，最终在这场残酷的军备竞赛中，整个美国科技生态与广大企业消费者成为了最大的赢家。 ### [18:41](https://www.youtube.com/watch?v=muRIXCDw-k0&t=1121s) 开源与闭源的博弈：前沿模型真的遥遥领先吗？ Brad Gerstner 指出，当前市场正处于极其健康的超级竞争状态：不仅有中国开源大模型、美国本土开源项目，还有各家前沿国家实验室（Frontier National Labs）不断发力，从而对模型价格形成了持续的下行压力。针对此前 David 提到的“双寡头”（Duopoly）格局，Brad 认为目前断言为时尚早，毕竟行业才发展了几年，且亚马逊、微软、谷歌等巨头仍在重金押注；但不可否认的是，纯模型公司在钱包份额（Share of Wallet）和营收上正不断拉开身位。 对于近期市场上的主流声音，Brad 提出了两个非共识（Non-consensus）观点。首先是 Elon Musk 对 Jason Calacanis 的反驳：过去两周市场普遍鼓吹“中国开源模型在智力水平上已经赶超闭源模型，且成本低廉得多”，但 Elon 直接回击“没那么快（Not so fast）”，并直言“我们正在迈入技术奇点，前沿模型的技术代差远比大众想象的要大”。Brad 赞同这一判断，认为普通用户的常规场景感知不深，但在最严苛、最复杂的训练和工程任务上，前沿闭源模型仍然稳居制高点。其次，英伟达的黄仁勋（Jensen Huang）也在本周提出，闭源模型对企业而言实际综合成本更低——因为企业无需承担庞大的自行预训练算力、复杂的微调（fine-tuning）工程、专业运维以及持续的安全护栏（guardrails）成本。 这两点解释了为何头部前沿实验室在商业变现层面持续狂飙。Jason 将这种竞争态势比作移动互联网时代的经典对决：“这完全就是 Android 与 iPhone 模式的重演——一个生态赚走了全行业的绝大部分利润，另一个则在全球范围内赢得了最广泛的用户规模和调用量。” ```mermaid flowchart TD accTitle: 开源与闭源 AI 生态的分工竞争格局 accDescr: 展示开源模型与闭源模型在商业模式、成本结构和市场占有上的分化路径 subgraph OpenSource["开源生态 (类似 Android)"] OS_Cost["低 Token 成本 / 自由定制"] OS_Burden["高昂维护、微调与安全护栏成本"] OS_Market["赢得绝大多数长尾开发者与调用规模"] OS_Cost --> OS_Burden --> OS_Market end subgraph FrontierClosed["前沿闭源生态 (类似 iPhone)"] FC_Lead["奇点级能力代差 / 严苛复杂任务表现"] FC_TCO["开箱即用 / 节省企业全栈基础设施开销"] FC_Econ["攫取绝大多数行业利润与经济价值"] FC_Lead --> FC_TCO --> FC_Econ end ``` ### [20:39](https://www.youtube.com/watch?v=muRIXCDw-k0&t=1239s) SpaceX 惊艳财报与百亿美元 AI 算力赌局 话题转向 SpaceX 转型为上市公司后的首份季度财报。财报发布后，SpaceX 股价当日下挫 13%，相较于 6 月上市初期逾 2 万亿美元的高点已累计回撤约 30%，目前市值在 1.4 万亿美元附近筑底。Jason 强调，股价下跌主要是因为资本市场对飙升的 AI 资本支出（Capex）产生顾虑。然而抛开股价波动，第二季度的基础业绩堪称惊艳：单季营收达到 78 亿美元，同比暴增 92%，环比增长 67%。 尤为引人瞩目的是其被称为“Elon Web Services”的 AI 算力出租业务。该板块季度营收环比激增超 3 倍，达到 26 亿美元——这还是在未计入对 Cursor 收购（该交易尚未交割）的情况下，仅靠将 Colossus 超级计算机集群的算力出租给 Anthropic 和 Google 等机构所实现的收益。但在投入端，SpaceX 单季度资本支出飙升至 184 亿美元，同比暴增 6 倍，年化 Capex 跑道（Run rate）已逼近 750 亿美元大关。 - **市值走势**：IPO 时估值超过 2 万亿美元，现回调至约 1.4 万亿美元，回撤幅度符合高成长科技股 IPO 后 6 个月内常见的中枢回踩。 - **业务核心引擎**：传统航天火箭发射与 Starlink 稳健推进的同时，依托 Colossus 超级算力中心的 AI 基础设施租赁成为爆发增长点。 - **资本开支压力**：单季 184 亿美元的 Capex 规模让华尔街部分投资者对巨额资金的现金流承压产生疑虑。 ### [21:52](https://www.youtube.com/watch?v=muRIXCDw-k0&t=1312s) Brad Gerstner：从算力分销到千亿 ARR 路线图 Brad Gerstner 对此表示，从无到有创造出 1.4 万亿美元市值的企业本身就是奇迹，IPO 后的回调在预期之内。他特别强调了 Elon 在电话会议上极为激进的业绩指引：SpaceX 预计到今年年底将实现 1,000 亿美元的年经常性收入（ARR），并将原定于 2031 年达成的“1 万亿美元 ARR 目标”整整提前了一年到 2030 年。相比之下，摩根士丹利对 SpaceX 2030 年的营收预期仅为 3,250 亿美元，而该公司去年全年营收才刚刚达到 180 亿美元。无论是按机构预期还是 Elon 自身的宏图，市场目前的估值定价都远远没有反映这一预期。 针对市场眼下的担忧，Brad 梳理了核心争论点： 1. **算力出租（Compute Rental）的可持续性与倍数陷阱**：SpaceX 将大批算力打包租给 Anthropic 等公司，但市场在质疑，未来是否能持续找到有足够资本消耗 10 至 20 吉瓦（GW）算力的超级买家？更重要的是，像 CoreWeave 这类纯 GPU 租赁业务在公开市场上一直只能享有极低的估值倍数。 2. **前沿模型资产的估值爆发力**：Elon 透露 Grok 的 Token 调用量在 7 月份实现翻三倍，加之快速增长的 Cursor（原本预期在年底前 ARR 从 30 亿飙升至 100 亿美元），两者的组合到年底可能创造 100 亿至 200 亿美元的 ARR。这一纯模型和高附加值软件资产在资本市场上的估值乘数将显著高于数据中心资产。 3. **流动性与解禁抛压**：硅谷各大风险投资基金近期开始对持有的 SpaceX 原始股份进行分红派息（Distribution），不少长线机构选择落袋为安，给盘面带来了可预测的短期抛压。但 Jason 与 Brad 均表示自己将坚定长期持有，“留给孙辈”。 ### [24:46](https://www.youtube.com/watch?v=muRIXCDw-k0&t=1486s) David Sacks 与 Chamath：Starlink 的飞轮与终极现金奶牛 David Sacks 同样对财报电话会持有极其乐观的态度，甚至对股价的下跌感到意外。他指出，SpaceX 的多项底层核心能力正在形成闭环：Starship（星舰）近期的测试飞行大获成功，隔热盾成功抵御返航并在海面软着陆，这意味着 Starship 发射频次将显著加快；这将直接加速 V3 代大型卫星的部署，极大扩充 Starlink 的网络带宽，进而引爆 Direct-to-Cell（星链直连手机）业务，全面颠覆传统电信运营商。 在数据中心业务的量化拆解上，Sacks 指出，SpaceX 的算力规模预计将从目前的 1.4 GW 提升至年底的 2 GW。按照 Elon 给出的每瓦 30 至 50 美元的算力现货价格测算（1 GW = 10 亿瓦，折合每吉瓦年化 300 亿至 500 亿美元产值），若以高端 50 美元/瓦满产运行，仅仅 2 GW 的算力租赁业务就足以支撑 1,000 亿美元 ARR 的指引，甚至完全无需将 Starlink、火箭发射及 Grok 等业务计算在内。 随后嘉宾们进一步拆解了 SpaceX 的三大业务分部（航天发射、网络连接 Starlink、AI 基础设施）： - **航天发射板块**：季度亏损约 2 亿美元，基本处于盈亏平衡边缘。 - **AI 算力板块**：贡献约 11 亿美元利润，但市场仍存疑其高溢价是否因当前全市场算力紧缺带来的短期暴利。 - **Starlink 网络连接**：无可争议的超级造血机器（Juggernaut cash machine）。单季度创造 43 亿美元营收，调整后 EBITDA（几乎等同于经营性现金流）高达 26 亿美元。 Starlink 的付费用户数已突破 1,200 万，同比实现翻倍增长；月度每用户平均收入（ARPU）达 66 美元，且季度环比增速维持在 20%。如果加上航空机载 Wi-Fi 等企业级客户的放量，Starlink 很快将向 2,400 万订阅用户的年化水平迈进，单体业务即可贡献超过 400 亿美元的顶层收入，并转化出约 300 亿美元的强劲自由现金流。正是 Starlink 庞大的自我造血能力，为 Elon 旗下狂飙的 AI 与深空探索提供了最稳固的底层弹药库，若按高确定性订阅业务享受 30 倍估值倍数计算，其单体价值就已蔚为壮观。 ### [28:15](https://www.youtube.com/watch?v=muRIXCDw-k0&t=1695s) Starlink 估值逻辑与对传统卫星运营商的颠覆 嘉斯汀与嘉宾们首先拆解了 Starlink 惊人的基本面表现，并给出了极其激进的看多逻辑。凭借极高的用户续约率和极低的边际资本支出（capex），单凭 Starlink 这一项业务，按照约 30 倍的市盈率计算，完全有可能在未来 18 个月到 2 年内成长为市值达 1 万亿美元的超级巨头。其源源不断的现金流不仅能实现自给自足，甚至可以为 SpaceX 旗下的其他深空探索与前沿项目提供充足的研发弹药。回顾过往，如 HughesNet 和 Viasat 这类老牌卫星通信公司，以及 AT&T、Verizon 等地面电信巨头的部分传统业务，已经在这个市场中遭受重创；过去在加州索诺马县（Sonoma County）等偏远牧场，用户必须忍受每月花费 200 美元却因地球同步轨道极度延迟而体验糟糕的高轨卫星网络，而 Starlink 的低轨星座彻底重塑了这片市场。 目前 Starlink 在消费端每个季度净增约 200 万用户，若未来手机直连卫星（Direct-to-Cell/Handset）功能全面普及，季度新增用户数极有可能进一步加速到 400 万至 500 万。仅在美国本土就有多达 4 亿个移动订阅用户基数，全球渗透空间更为广阔。核心讨论点在于：Starlink 作为“现金奶牛”所释放的海量多余资本，交到埃隆·马斯克（Elon Musk）手里去投资 Starship、AI 算力以及超级晶圆厂（Terafab），无疑是最具想象力也是最值得信赖的资本配置方式。 - **高续约与低 CapEx**：Starlink 业务单体估值有望在 18 至 24 个月内冲击 1 万亿美元。 - **颠覆传统高轨网络**：彻底击溃 HughesNet、Viasat 等延迟高、资费昂贵的旧一代卫星服务商。 - **用户增长提速**：当前单季新增 200 万消费者用户，随着手机直连技术落地，单季增速预计可达 400 万至 500 万。 - **国内存量空间**：仅美国一地就拥有 4 亿移动订阅用户，蕴含巨大的卫星直连渗透潜力。 --- ### [29:35](https://www.youtube.com/watch?v=muRIXCDw-k0&t=1775s) 豪赌未来：从 Starlink 现金流到半导体自主的“Terafab” 对于马斯克将 Starlink 赚取的充沛利润大举投向高风险领域的做法，布拉德·格斯特纳（Brad Gerstner）与全场嘉宾表达了高度赞赏。如果换作绝大多数上市公司的传统职业经理人或 CEO，守着 Starlink 这样一个估值 1 万亿甚至奔向 2 万亿美元的印钞机，绝不会轻易去涉足极高风险的超级半导体制造厂（Terafab）、超大规模 AI 数据中心或自研前沿大模型。正如同科幻小说照进现实，Terafab 的构想被视为美国摆脱对海外半导体供应链极端依赖的关键战役；如果马斯克能够兑现这一愿景，这里将崛起为全球规模最大的半导体制造基地。 嘉宾们感叹，这种对前沿创新的“奔放热情”（unbridled enthusiasm）堪称英雄主义之举。反观当前资本市场，许多传统上市科技公司因避险情绪浓厚，宁愿选择股票回购或分红，也不愿把钱砸向未来。近期 DoorDash 和谷歌因加大资本支出（capex spend）而遭到华尔街无情抛售，但马斯克却始终拒绝妥协于安全牌，坚持把从卓越业务中赚到的每一分钱，重新回流到对美国国家战略与科技命脉至关重要的前沿硬科技领域。 ```mermaid flowchart TD accTitle: Starlink 现金流反哺硬科技创新循环 accDescr: Starlink 业务产生充沛现金流，重新投资于超级晶圆厂、AI数据中心和星舰系统 subgraph CashCow ["核心现金牛业务"] SL["Starlink 卫星互联网业务"] REV["高续费率与全球用户规模扩张"] SL --> REV end subgraph Reinvestment ["高风险战略资本再投入"] TF["Terafab 超级芯片工厂<br/>(破解海外半导体依赖)"] DC["超大规模 AI 算力中心<br/>(突破算力瓶颈)"] SS["Starship 运载平台<br/>(下一代深空基础设施)"] end REV -->|"海量自由现金流反哺"| TF REV -->|"海量自由现金流反哺"| DC REV -->|"海量自由现金流反哺"| SS ``` --- ### [31:18](https://www.youtube.com/watch?v=muRIXCDw-k0&t=1878s) 市场疑虑应对：特斯拉生态融合与估值倍数消化 针对市场对 SpaceX 长期增长的质疑，嘉宾们重点剖析了“算力天花板”与“Starlink 目标市场受限”两大反方论点。有人担忧端侧计算、开源小模型是否会抑制未来对集中式云端算力与 Token 的需求，但讨论认为按需智能（on-demand intelligence）的上限不可限量；针对“光纤覆盖成熟的城市居民不需要也无法安装 Starlink 卫星天线”的质疑，嘉宾指出特斯拉与 SpaceX 业务协同将带来破局点：未来每一辆特斯拉（包括庞大的 Robotaxi 车队）都将内置集成 Starlink 模块，届时车辆不仅化身为移动 Wi-Fi 热点，支持行人手机直接连接，还能配合下一代 Starlink 卫星构建起无缝视距直连网络。 在二级市场估值层面，正如本杰明·格雷厄姆所言，私人市场风投像“投票机”，而公开市场则是“称重机”。SpaceX 在市场情绪过热时曾一度达到 160 倍的市销率（P/S），引发了大量估值焦虑；然而随着营收规模快速攀升，其市销率已健康回落至约 45 倍。目前公司处于 1.4 万亿美元估值区间，站在 3 到 4 年的长周期来看，依托其扎实的营收增速，投资者依然有很大机会实现三倍的回报。格斯特纳回顾自己在 CNBC 上的表态强调：“入场价格至关重要”（entry price matters），虽然长期极度看好马斯克的资本分配能力，但决不能在 IPO 首日冲破 2 万亿美元情绪溢价时无脑追高。 --- ### [34:15](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2055s) 科技 IPO 展望与马斯克无与伦比的“物理制造”护城河 顺着 SpaceX 的上市表现，格斯特纳进一步对比了下半年备受关注的 Anthropic IPO 预期。外界传闻其估值可能落在 1.5 万亿至 2 万亿美元之间，而年化运行营收（Run-rate revenue）预计年底突破 1000 亿美元；对于一家增长 10 倍且传闻在第二季度实现盈利的公司而言，10 到 15 倍的市销率其实相对温和。市场在经历 7 月份半导体板块普遍回调 30% 左右的震荡筑底之后，恐慌情绪已经释放，为接下来的科技巨头上市潮奠定了扎实基础。 更为关键的是，嘉宾们深入探讨了马斯克在物理世界层面的绝对壁垒。在这场全行业对数据中心与半导体晶圆厂的残酷争夺战中，软件层必须深深植根于物理硬件设施。无论是对比 Dario Amodei、Sam Altman 还是拥有 27 年基础设施底蕴的 Alphabet，马斯克在建设物理工厂上的执行力无人能及。马斯克在业绩电话会上曾形象地调侃道：“建设数据中心绝非易事，但与发射火箭的难度相比，它根本不值一提；这又不是造火箭（creating data centers is not rocket science）”。SpaceX 将航天工程级别的精细化硬件与制造能力降维输出给算力设施，正是其数据中心落地速度和体量远超所有对手的底层核心能力。 --- ### [36:23](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2183s) Starship 对 Starlink 的倍增跃迁：从 V2 到 V3 的量级突破 讨论随后聚焦到 Starship（星舰）对于 Starlink 商业飞轮的决定性意义。SpaceX 研发的新一代 V3 卫星，其网络带宽达到了现有 V2 卫星的整整 10 倍，而两者的运载发射能力展现了压倒性的代际差距： | 对比维度 | 猎鹰 9 号（Falcon 9）+ V2 卫星 | 星舰（Starship）+ V3 卫星 | 性能跃迁倍数 | |:--- |:--- |:--- |:--- | | **单次发射卫星数量** | 约 27 颗 | 约 60 颗 | **~2.2x 运载量** | | **卫星单机带宽性能** | 基准带宽（V2 标准） | 相当于 V2 的 10 倍 | **10x 吞吐能力** | | **单次发射增加总网络容量** | 约 2.6 Tbps | 约 60 Tbps | **>20x 总带宽跃迁** | 在此前最新的 Starship 试飞中，不仅成功验证了隔热瓦（heat shield）的极端再入性能，团队还在测试中一次性部署了 20 颗 V3 试验卫星，并成功与地面建立通信链路验证其功能。马斯克团队甚至秉持着“人生苦短，放手一搏”（YOLO）的精神，在这些测试卫星上安装了高清摄像头，从而为全球观众传回了令人震撼的 Starship 太空实况视角。尽管这批测试卫星最终受控在大气层中烧毁，但整个系统的可行性已得到里程碑式的实证，宣告着未来每单次发射即可为全球星座注入 60 Tbps 容量的低成本互联时代正式拉开序幕。 ### [37:44](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2264s) Starship 发射 V3 卫星与 Starlink 承载全球半数流量的野心 讨论首先聚焦于 SpaceX 的技术里程碑：当下一代重型运载火箭 Starship 能够一次性将大约 60 颗 Starlink V3 卫星送入预定轨道并建立连接时，整个卫星互联网网络的可用带宽将迎来质的飞跃。随着带宽提升 10 倍乃至 100 倍，Starlink 将能够真正解锁直连手机（Direct to Cellular）等颠覆性应用。 主持人们讨论了总裁兼首席运营官 Gwynne Shotwell 此前关于地面站演进的暗示，甚至开玩笑称 SpaceX 凭借其庞大的财力与野心，未来完全有能力直接收购像 T-Mobile 这样的传统电信运营商。此外，Elon Musk 曾预计 Starlink 最终可能承载全球大约一半的互联网流量，这意味着其商业版图将远远超越目前 1200 万订阅用户的范畴。 ```mermaid flowchart LR accTitle: Starlink网络升级与流量承载路径 accDescr: 展示Starship发射V3卫星如何带动带宽数十倍增长并最终实现直连手机与全球流量承载。 subgraph 部署发射 A["Starship 一箭60星"] --> B["Starlink V3 卫星入轨联网"] end subgraph 能力扩展 B --> C["网络带宽提升 10x - 100x"] C --> D["直连手机 / 地面站扩展"] end subgraph 终局愿景 D --> E["承载全球 50% 互联网流量"] end ``` ### [38:38](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2318s) 算力数据中心扩张：巨额资本开支与现货价格的博弈 话题随后转向了 AI 数据中心的建设规模与资本开支。David Sacks 抛出了一组关键数据：Elon Musk 提到其算力规模到年底将达到 2 GW，并计划在次年扩张至 5 到 10 GW（偏向 10 GW，保守折中按 8 GW 计算）。这意味着单年净新增算力高达 6 GW。按照目前建设 1 GW 算力数据中心至少需要 500 亿美元资本支出（CapEx）的标准计算，新增 6 GW 意味着未来一年需要惊人的 3000 亿美元资金投入。 David Sacks 提出了两个核心疑问：首先，目前每瓦 50 美元的高昂现货价格（Spot Price）能否持续？Elon Musk 似乎认为由于存储（Memory）产能受限（预计明年存储增产仅 20%，而算力需求增长超 200%），现货价格仍会坚挺，但其中蕴含着巨大风险。其次，面对 3000 亿美元的天文数字，如何用非稀释性手段进行融资？尽管据称投资回收期（Payback Period）不到一年，但该测算完全建立在当前极端高昂的现货价格之上，届时 NVIDIA 是否会提供过桥担保成为关键悬念。 - **算力扩建规模**：从年底的 2 GW 跃升至 8 GW，单年净增 6 GW 计算能力。 - **建设成本门槛**：每 GW 算力中心的建设成本最低为 500 亿美元，6 GW 对应 3000 亿美元 CapEx。 - **供需瓶颈**：当前整个算力供应链受制于存储瓶颈，存储供给增长 20% 对比算力需求超 200% 的爆发式增长。 ### [40:22](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2422s) 循环收入隐忧、卖方融资与前沿实验室的算力饥渴 代班 Chamath 的 Brad Gerstner 进一步剖析了数据中心建设的融资困局。他指出，除了市场借贷或股权稀释（两者都不是理想选择）之外，业内正越来越多地寻求 NVIDIA 等硬件巨头提供兜底支持。然而，NVIDIA 的股东并不愿意无限制兜底，因为一旦现货价格回落，投资回收期就会迅速被拉长。几个月前业界普遍预期回收期为 4 到 5 年（年化收益 100 亿至 150 亿美元，6 年期年化回报率可超 20%），但目前极度短缺的供给让前沿实验室（如 Anthropic 与 OpenAI）甘愿支付高达常规市场价 3 到 5 倍（即每瓦 30 至 50 美元）的溢价，以换取确定的规模算力优势。 Brad 强调，今年年初 OpenAI 与 Anthropic 两家的初始总算力加起来仅约 5 GW，而 Elon 计划一年内增建的规模就超过了这两家巨头的起步总量。尽管 Anthropic 实现了年化 10 倍增长、OpenAI 也保持着 4 倍以上的扩张速度，但市场暗流涌动。7 月份中国模型（如 Kimi）的出现引发了市场对前沿模型收入被低价侵蚀的担忧，导致算力租赁商（如 CoreWeave）及相关半导体股票大跌 40%。投资人 Bill Gurley 多次警示这种过度的“卖方融资”（Seller Financing）与“循环收入”（Circular Revenues）潜藏着系统性杠杆风险：一旦需求端发生轻微松动，建立在提前透支收入之上的巨额杠杆将引发极度剧烈的资产抛售与信用利差走阔。 > "Jensen said to me on the pod, he's like, 'Nobody comes close. Microsoft doesn't come close. You know, Google doesn't come close in terms of standing it up in that time frame.'" —— Brad Gerstner 引述黄仁勋对马斯克团队算力基建部署速度的评价。 > "As our good friend Bill Gurley likes to remind us, he's like, 'I can't believe that we're all just taking in stride this level of seller financing, right?' He would call it circular revenues..." —— Brad Gerstner 谈及算力市场中的过度杠杆与循环收入。 ### [45:44](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2744s) 第五届 All-In 峰会重磅官宣 节目的这一段落最后公布了定于 9 月举行的第五届 All-In 峰会（All-In Summit）的嘉宾阵容与活动细节。David Friedberg 牵头筹办了此次峰会，并确定了一批科技与商业领域的重磅演讲嘉宾： - **NVIDIA 创始人兼 CEO 黄仁勋（Jensen Huang）**：分享全球 AI 发展方向的最前沿洞察； - **微软 CEO 萨蒂亚·纳德拉（Satya Nadella）**：第二次做客播客线下峰会； - **NASA 宇航员兼商业航天先锋 Jared Isaacman**； - **投资搭档 Brad Gerstner 与 Bill Gurley（BG2 组合）回归**； - **SpaceX 总裁兼首席运营官 Gwynne Shotwell**； - 知名网红拳手 Jake Paul、创作者 Nick Shirley，以及可能到场的 Martin Shkreli 等多界名人。 本次峰会再度包下了好莱坞环球影城（Universal Studios）作为专属私人活动场地，并策划了创纪录规模的“赌场之夜”（Casino Night）与待揭晓的重磅音乐会。来自全球 60 多个国家的创业者、投资人和热心听众将汇聚于此，围绕全球最核心的商业与技术议题展开深入探讨。 ### [47:12](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2832s) All-In 峰会筹备与“SaaS 末日”的到来 Jason 与嘉宾们在节目中探讨了即将到来的 All-In 峰会。他们强调，举办峰会的核心目标不仅是在舞台上展示精彩的内容，更重要的是让与会者在几天的时间里享受沉浸式体验。Jason 总结了峰会聚焦的三大核心要素，称其为“三合一的完美组合”（trifecta）：第一，在舞台上向顶尖人物学习新知识；第二，结识新朋友并拓展人脉；第三，获得独一无二的现场体验。 当 Brad Gerstner 询问峰会具体时间时，Jason 提醒他“看下日历，你可是发言嘉宾之一”，并正式宣布峰会将于 9 月 13 日至 15 日在洛杉矶举行。主持人们对此兴奋不已，认为这个时间点极为关键且敏感——距离美国中期选举不到 60 天，甚至可能在 Anthropic 等 AI 巨头 IPO 预期的 30 天窗口期之内。然而，市场的宏观情绪正处于复杂博弈中：所谓的“SaaS 末日”（SaaSpocalypse，Jason 打趣澄清不是 Sacks 带来的末日）正在持续蔓延，虽然资本市场的“消化不良”似乎正在出清，但残酷的洗牌已经不可避免。 ### [48:01](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2881s) Airtable 惨遭一折贱卖：从百亿明星到 Bending Spoons 的战利品 随后，话题切入本周轰动硅谷的重磅并购案：协同办公与无代码平台 Airtable 以远低于其历史融资总额的价格出售。Airtable 曾是一家令人艳羡的 SaaS 独角兽，年营收接近 5 亿美元（约 4.8 亿美元），且保持着 20% 的稳健年化增长，账面上还趴着近 10 亿美元现金。如果放在公开市场，这本应是一家体面的公司，但其最终售价仅为 12.8 亿美元（计入现金储备的总交易价值约为 2.25 亿美元至 22.5 亿美元区间），相比其在 2021 年牛市顶峰时的 11.7 亿美元估值，缩水了整整 90%。 收购方并非传统科技巨头，而是来自意大利米兰的软件收购集团 Bending Spoons。这家公司以收购“陷入困境但拥有独特黏性资产”的科技业务而闻名，其投资组合包括 Evernote（印象笔记）、Eventbrite、Vimeo、Meetup.com 以及 AOL 的传统遗留资产。Bending Spoons 刚刚完成上市，其股价在宣布收购 Airtable 的消息后应声暴涨 15%。Jason 由此发问：Airtable 做对了那么多事，坐拥巨额资金储备，最终创始人与高位入局的投资人却似乎双双精疲力竭，Bending Spoons 是否已经成为整个 SaaS 行业“最后的清道夫与接盘侠”？ - **Airtable 财务数据**：年营收约 4.8 亿美元，年增速 20%，账面拥有近 10 亿美元现金。 - **估值断崖**：2021 年峰值估值 117 亿美元，本次主体交易价格仅 12.8 亿美元，较顶峰缩水近九成。 - **Bending Spoons 资产版图**：Evernote、Eventbrite、Vimeo、Meetup.com，并以苛刻削减成本著称。 ### [49:26](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2966s) 拆解交易内幕：分拆 HyperAgent、销售团队折戟与 PE 模式的接盘 David Sacks 深入剖析了这笔交易的底层结构与商业逻辑。他指出，这笔收购前夕发生了一个极具战略意义的关键动作：Airtable 在交易完成前，将其前沿的 AI Agent 业务“HyperAgent”剥离成了一家独立的初创公司。这意味着创始团队和顶级技术人才已经对拯救原有成熟业务失去了耐心，彻底认清原有业务已经沦为“私募股权（PE）游戏”，而他们选择抽身去专注充满想象力的全新 AI 创业项目。“人才集中精力去追求能够带来超额回报的风险投资模式，而把沉重的 PE 盘子甩卖给了 Bending Spoons。” ```mermaid flowchart TD accTitle: Airtable 资产剥离与并购架构 accDescr: 展示 Airtable 分拆 AI 业务与将原有 SaaS 业务出售给 Bending Spoons 的双轨路径 subgraph 分拆前 AT["Airtable 原业务矩阵<br/>4.8 亿美元 ARR / 20% 增速"] end subgraph 剥离与分流 AT --> AI["AI Agent 业务<br/>HyperAgent"] AT --> Core["成熟核心 SaaS 业务<br/>传统协同数据库"] end subgraph 终局归宿 AI --> VC["风投赛道 (VC Play)<br/>原班技术人才独立创业"] Core --> BS["私募并购赛道 (PE Play)<br/>出售给 Bending Spoons 削减成本"] end ``` Sacks 进一步援引了一个极具破坏性的运营数据：Airtable 的销售团队中，仅有 30% 的员工能够完成业绩指标（quota attainment）。这一数据揭示了内部极大的战略扭曲——Airtable 原本凭借极度成功的产品驱动增长（PLG）机制实现了有机裂变，但背负着 110 亿美元超高估值的董事会不甘心只有 20% 的自然增长，强行向团队施压，要求违背产品基因强行嫁接传统“销售驱动”（sales-led）模式。结果数以百计的高薪销售人员如同“推绳子前行”（push on a string）般吃力，根本无法加速增长，反而大幅拉垮了利润率。 Sacks 认为，Bending Spoons 的绝妙机遇正是效仿埃隆·马斯克接管 Twitter 时的激进做法：直接砍掉 80% 至 90% 的冗余成本，全面废除低效的企业直销团队，重归纯粹的 PLG 模式。即便有人预测其 EBITDA 利润率只能达到 30%，Sacks 坚信其实际利润率完全可以冲上 80% 到 90%，每年沉淀 3 亿至 4 亿美元自由现金流，两三年内即可收回收购成本。对于早期和后期的风险投资人而言，面对原本需要翻十倍才能解套的死局，能在保本边缘收回资金并“弃牌走人”，已经是当下能够争取的现实妥协。 ### [53:42](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3222s) 估值倍数坍塌、人才逃亡与 AI 赋能的“维护模式” Jason 提出疑问：既然转入现金流模式如此诱人，为什么创始团队和董事会不能自行实行这种重组？Sacks 解释称，风投基金与创始人从心理机制和组织架构上根本不适合做这种“血腥拆解”；他们对一手建立的团队抱有深厚的情感忠诚，无法狠心裁撤绝大多数员工。此外，错综复杂的股权结构与庞大的优先清偿权（Liquidation Preference）彻底锁死了团队激励，唯有交由不受人情束缚的专业接盘机构处置。 Brad Gerstner（Altimeter Capital 创始人）结合其管理公开与私募资产的亲身经历发表了严谨的看法。他警示称，硅谷长期习惯用极度粗糙的“市销率”（PS）倍数给企业估值，这种估值倍数只有在公司维持 50% 以上高复合增长时才有效；一旦增长失速，市销率坍塌的速度令人措手不及。外界惊叹 Airtable 仅以约 2 倍市销率贱卖，但如果穿透其实际自由现金流，这笔收购的估值实际上可能高达 30 倍自由现金流。 Brad 补充道，陷入增长瓶颈的软件公司往往面临着剧烈的组织崩溃：“一旦增长停滞，士气便瞬间跌入谷底，客户流失率飙升并形成恶性循环，每天上班都痛苦不堪。”雪上加霜的是，当前几乎所有优秀的软件销售与研发工程师都在疯狂跳槽投身 AI 行业，传统 SaaS 根本留不住核心技术骨干。然而 Sacks 强调，这恰恰是 Bending Spoons 这类收购者的天然护城河：他们拥有统一的底层中台架构，可以无缝挂载这些软件。更革命性的是，过去进入“维护模式”（maintenance mode）必须保留懂老旧代码的资深员工来维系“组织记忆”，而现在借助 AI 工具，AI 可以在瞬间学会并完全接管庞大的代码库，使得维系老牌 SaaS 运转的边际成本骤降至历史最低点。讨论随后引向了零利率时代（ZIRP）与当前 AI 狂热周期之间的历史对照。 ### [56:29](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3389s) AI 资本开支与 SaaS 泡沫的本质区别 关于当前 AI 领域高昂的算力成本以及部分公司高达 100 倍市销率（P/S）的现象是否与曾经的 SaaS 泡沫存在相似之处，嘉宾们给出了明确的否定。这两种资本配置模式的底层范式完全不同：当前的 AI 资本支出（Capex）和模型训练投入，核心在于争夺基础设施与技术演进的主导权，而非 SaaS 时代那种单纯通过极高收入乘数套利的游戏——当时只要能以 20 倍市销率把 1 美元估值放大到 20 美元，热钱就会无脑涌入。 在零利率时代（ZERP），SaaS 公司的估值泡沫主要由两大因素推动：一是人为压低的利率催生了投机性超级资产泡沫；二是市场将这些公司视为能够永续增长的“确定性年金”。当时投资者看到 120% 的净金额留存率（Net Dollar Retention, NDR），便理所当然地认为它们每年至少能维持 20% 的稳健复合增长。然而，AI 带来的颠覆彻底打破了这种年金幻觉，许多 SaaS 公司的客户流失率（Churn Rate）显著攀升，迫使市场给予它们大幅折价。 ```mermaid flowchart LR accTitle: ZERP 时代的 SaaS 与 AI 时代的资本范式对比 accDescr: 展现零利率时代高估值 SaaS 模式与当下 AI 资本开支建设的底层逻辑差异。 subgraph SaaS["零利率时代 SaaS 模式"] A["人为低利率"] --> B["年金化幻觉 (120% NDR)"] B --> C["高估值倍数套利 (20x P/S)"] end subgraph AI["AI 范式转换"] D["AI 技术颠覆"] --> E["客户流失率上升"] E --> F["SaaS 估值大幅折价"] G["基础设施与算力 Capex"] --> H["不依赖单纯收入乘数驱动"] end ``` ### [57:57](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3477s) Airtable 的困境与无代码（No-Code）的覆灭 虽然市场情绪低迷，但不能将 Airtable 的重挫简单推导至整个 SaaS 行业。与 Salesforce、Workday 这类核心系统相比，Airtable 自诞生起就带有某种小众而奇特（quirky）的产品属性。在估值泡沫的顶峰时期，它曾被吹捧为“新一代 Excel”、“新一代 Google Sheets”甚至是“文字版电子表格”，但它从未真正兑现这种普适性承诺。大众本能地理解数字表格的用法，却很难直观明白什么时候必须用 Airtable；尽管它拥有忠实的死忠追随者（cult following），但缺乏清晰明了的杀手级用例，营销策略始终难以破圈。 更为致命的是，Airtable 所处的“无代码（No-Code）”赛道恰恰是遭受 AI 冲击最为严重的重灾区。像 Airtable、Retool 这类工具，虽然免去了写代码的门槛，但用户仍需投入大量精力去学习它们特有的规则，相当于掌握了一套套冷门的小众伪编程语言。而如今面对 Claude Code、Lovable、Perplexity 等 AI Agent 工具，用户只需用自然语言描述需求，就能自动生成专属的数据看板或口头化表格。从学习成本到构建效率，AI 彻底降维打击了无代码工具。 - **产品定位落差**：曾被誉为“文字版电子表格”和“新一代 Office”，但大众认知度始终未能突破极客圈层。 - **替代品冲击**：Claude Co-work Agent 与现代编码助手让用户无需学习任何无代码平台的中间语法，直接以自然语言生成所需工具。 ### [01:00:20](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3620s) “Vibe Coding”热潮与深层企业护城河的防御 AI 编程对软件内部自研的门槛带来了毁灭性降低。现场嘉宾分享了各自团队在内部“Vibe Coding（氛围编程）”的真实案例：David Sacks 的团队仅耗时一个月，就自主开发出了一套令人惊叹的基金投资组合管理系统，全面掌握被投公司、竞对格局和创始人动态；而在过去，购买现成的商用 SaaS 软件需要花费 25 万美元许可费，外加持续两三年、耗资上百万元的系统集成成本。 然而，这种基于 AI 的内部轻量替代并不意味着所有 SaaS 都会被颠覆。知名投资人 Leopold 之所以遭遇惨败爆仓，正是因为他在重仓芯片股回调的同时，激进做空了 Adobe 等一系列 SaaS 公司，结果全面踩错节奏。盲目断言“SaaS 将走向彻底灭绝”显然过于武断，大型企业与政府部门的核心护城河不在于代码本身，而在于深度的合规性与组织嵌入： > “没人买微软是因为微软写的代码最好。大家买微软，是因为微软是所有业务运行的底层轨道。Active Directory 承载着员工的身份认证，Excel 是董事会财务数据的来源，Teams 记录着满足合规要求的会议，Azure 拥有 FedRAMP High 授权和美国国防部 Impact Level 5 安全许可——这意味着国防承包商绝不可能为了省钱而随意把它换掉。” 同样，Salesforce 首席执行官 Marc Benioff 刚刚发文表示，美国联邦政府全部 15 个内阁部门均在 Salesforce 上运行。任何大型机构都不可能冒险将关乎合规、ERP、HR 或 CRM 的核心命脉，替换成某个临时由 AI 捏造出来的未验证系统。 ### [01:02:32](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3752s) SaaS 行业分化与 Airtable 的清算收场 SaaS 行业目前呈现出剧烈分化的态势，而非整体溃败。拥有长期主义创始人、深厚用户黏性并全力向“AI 原生应用”转型的企业（如 Figma），依然具备强大的跃迁能力。反映高成长软件股的综合指数 IGV 在过去 6 个月上涨了 20%，过去 5 年也录得 20% 的涨幅；Snowflake 在过去 6 个月内暴涨约 90%，其强劲势头甚至能与领涨的半导体 AI 股票分庭抗礼，Databricks、ClickHouse 等数据底座企业亦表现卓越。相反，缺乏深层护城河的顶层应用软件和无代码平台正在加速认清现实，选择套现离场。 在 Airtable 以九折跳楼价甩卖的交易中，最终结果呈现出硅谷特有的退出生态： - **后期成长型基金**：曾以 20 亿美元、50 亿美元甚至 110 亿美元极高估值进入的知名后期机构，虽然颗粒无收但凭借标准的“1 倍不参与清算优先权（1x Non-Participating Liquidation Preference）”拿回了全部投资本金。 - **早期投资人与团队**：早期低估值入场的投资者依然斩获了丰厚的账面回报。 - **行业定性**：虽然 90% 的估值缩水堪称巨额蒸发，但由于各方资金基本实现保本或盈利退出，这在硅谷语境下被戏称为一场“极其体面的成功失败（pretty good failure）”。 ### [01:05:04](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3904s) 投资条款的本质：干净条款与参与清算优先权之争 在上一话题的尾声，几位投资人深入探讨了资本结构表（Cap Table）中清算优先权的设计逻辑。关于并购案中利息计算被移除的现象，David Sacks 指出，硅谷最标准、最健康的“干净条款”（Clean Terms）实际上就是简明的基础 1 倍不参与分配清算优先权（1x non-participating liquidation preference）：在公司出售时，优先股股东先拿回本金，随后普通股股东再参与剩余收益的分配。这一机制完全合情合理。Jason Calacanis 随即点出反面例子：“参与型优先股（Participating preferred）本质上就是在双重得利（Double dip）吧？” Sacks 坦言自己的基金从未采取过参与型优先股这种惩罚性条款，因为投资人并非要故意惩罚创业者。核心诉求在于，股权结构表上绝不能出现“一部分人亏得血本无归，另一部分人却坐收暴利”的荒谬局面。那种情况纯粹是将一部分持股人的价值强行转移给另一部分人。合理的标准操作应当是优先保障投资人本金得以安全退出，随后所有人再平分企业成长的上行红利。 --- ### [01:05:56](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3956s) 福布斯重磅曝光：美国数据公司正在“喂饱”中国 AI Jason Calacanis 抛出了当天的第四个焦点议题：中国顶级大模型实验室正大肆从美国供应商手中采购训练数据。《福布斯》发表了一篇题为《这些美国初创企业正在让中国的 AI 变得更聪明》的深度调查报道，直指美国顶尖数据标注与合成数据初创公司正在向中国实验室出售极其宝贵的训练数据集，加速协助中国团队缩小与美国前沿实验室的差距。估值超过 200 亿美元的行业领头羊（如 Scale AI 和 Mercor 等）一方面向 OpenAI、Anthropic 以及美国联邦机构供应数据集，另一方面却把同样的数据集转手卖给了腾讯、字节跳动、阿里巴巴、月之暗面（Moonshot AI）等中国头部大模型公司。 ```mermaid flowchart TD accTitle: 数据流向与中美大模型竞争机制 accDescr: 展现美国专业数据初创公司向中美双方AI实验室同时供应高价值专业数据并引发的地缘政治争议 subgraph US_Data_Suppliers["美国数据与RLHF供应商 (Scale AI, Mercor, Micro1 等)"] Raw["西方顶尖专业人才 / 博士群体"] --> ExpertData["高难度领域真值数据 (编程、生物科学、RLHF)"] end subgraph US_Ecosystem["美国顶尖 AI 阵营"] ExpertData --> US_Labs["OpenAI / Anthropic / 联邦机构"] end subgraph China_Ecosystem["中国顶级 AI 阵营"] ExpertData -->|"年耗资5亿美元采购'秘方'"| China_Labs["腾讯 / 字节跳动 / 阿里巴巴 / 月之暗面"] China_Labs --> China_Models["Kimi / Qwen (通义千问) / GLM"] end subgraph Policy_Debate["华盛顿与投资界争论"] China_Models -.-> Friction["地缘摩擦 & 议会审查隐忧"] Friction -.-> BanCheck{"是否具备军民两用 / 真正专有属性？"} end ``` 根据该报道，中国排名前六的 AI 实验室每年合计斥资约 5 亿美元，大肆采购被《福布斯》称为“核心秘方”（Secret Sauce）的训练资料——涵盖博士级专家撰写的专业内容、强化学习（RLHF）管线以及高阶认知流水线数据。Jason 透露自己也是若干数据公司的投资人（例如 Micro1），而 Micro1 的创始人出于原则考量，主动放弃并拒绝向中国客户出售数据。当前公开互联网上的高质量文本已被各大模型厂商抓取殆尽，甚至出现了把物理实体图书拆开书脊送入扫描仪的极端做法。Jason 抛出核心质问：在数据已成为绝对瓶颈的当下，允许本国企业将这些高质量数据出售给中国竞争对手，是否合乎战略利益？ - **数据采购规模**：中国前六大 AI 实验室每年在购买美国高级训练数据上的支出高达 5 亿美元。 - **数据供给巨头**：以 Scale AI 和 Mercor 为代表的高估值初创团队，同时服务于美国前沿军工、联邦机构、Anthropic、OpenAI 与中国头部科技巨头。 - **采购核心内容**：涵盖 PhD 级别专业学者验证的专业学科知识、强化学习纠错流水线与复杂编码基准数据。 --- ### [01:07:36](https://www.youtube.com/watch?v=muRIXCDw-k0&t=4056s) 萨克斯与杰森的交锋：是战略泄密还是普通商品？ 面对这道战略考题，David Sacks 主张首先必须明确国家战略目标：美国究竟是打算与中国打一场全面的经济战、全面切断一切商业往来，还是恪守传统的国家安全边界？从历史规则来看，出口管制的红线一贯在于是否属于具有军事用途的“军民两用技术”（Dual-use technology）。在 Sacks 看来，数据标注本质上是高度同质化、缺乏壁垒的“大宗商品”（Commodity），中国拥有庞大的人力储备，根本不缺能够执行数据标注的技术劳动力。如果美国盲目出台一刀切的禁令，势必招致对等的报复行动——例如中国在关键稀土供应上切断对美出口。Sacks 强调自己虽然支持国家主权与供应链完全独立，但现实中美两国仍存在千丝万缕的相互依赖，必须审视该数据是否具备不可替代的独占性与明确的军事应用价值。 Jason 立即反驳称，这绝不是低端廉价的数据标注劳务，而是重金聘请西方顶尖博士与高精尖专业人才，专门针对大模型输出中出现的高难度技术错误（如复杂的编程缺陷、前沿生物学与自然科学问题）进行精确纠错、提供最新见解并反复双重验证的“绝密流水线”。Jason 指出，这正是当下大模型智力突飞猛进的核心驱动力，中国大模型突飞猛进的背后正是因为源源不断流入的西方核心知识包。但 Sacks 依旧持怀疑态度，质问凭什么断定中国无法组织自己的高学历团队完成复现：“中国每年毕业的理工科（STEM）毕业生人数比全世界其他国家的总和还要多，他们最不缺的就是聪明绝顶的人才。”随后 Jason 讽刺地补充了美国移民政策的痛点：“我们把全世界最聪明的人招来培养毕业，转头又把他们赶出境，这才是我们必须解决的大问题。” - **历史管制对比**：Sacks 高度赞扬 2019 年特朗普第一任期内促成的 ASML 极紫外光刻机（EUV）对华出口禁令，认为唯有像 EUV 这样真正形成“强力痛击”（Pack a punch）的关键卡脖子技术，才值得实施精准的战略管制。 - **中国人才基数**：中国每年的数学与自然科学类毕业生总数规模已远超全球其他地区总和。 --- ### [01:09:46](https://www.youtube.com/watch?v=muRIXCDw-k0&t=4186s) 布拉德的研判：身处领先优势下的监管平静与暗涌 Brad Gerstner 加入讨论，认可在当前阶段美国整体依然稳占胜势。不仅在 Anthropic、OpenAI 等前沿闭源实验室处于绝对领跑地位，在开源技术生态上也保持主导，且目前监管环境相对宽松，双边高层接触依然在有序推进。然而 Brad 发出警示，这种“向中国对手转售博士级训练数据”的商业行为，必然会深度激怒华盛顿的强硬派官员。在政界看来，芯片出口管控是一回事，而知识提炼（Distillation）与高级训练数据的毫无防备输出，让外界严重质疑美国是否正在主动给中国实验室抹平差距提供“绿色通道”。 Brad 犀利指出，华盛顿目前之所以尚未痛下杀手，仅仅是因为美国仍在领跑这场比赛。一旦未来半年内相关评估报告显示中国大模型已经反超或彻底追平，这一领域的商业往来势必会遭遇前所未有的雷霆审查。Jason 顺势感叹，自己在过去两个月里密集深度试用了来自中国的 Kimi、阿里通义千问（Qwen）以及智谱 GLM 等开源与商用模型，直言其能力已经“好得不可思议（My lord, these things are good）”。在 Jason 看来，把凝结了全西方顶级科学家智慧的高纯度微调数据打包转卖给潜在对手，不仅在商业上短视，甚至“一点都不爱国”。 讨论在双方对核心数据是否真正具备“军民两用”不可替代性的追问中渐入尾声。Jason 在总结中表达了虽然具体数据集细节仍未完全公开，但其高度专有且直接缩减了追赶周期的事实无可争议。随后 Jason 向代班 Chamath 出席节目的 Brad Gerstner 表达了诚挚感谢，并向正在全球旅行休假的 Chamath 送上祝福，正式结束了本期播客的深度录制。 ### [01:13:53](https://www.youtube.com/watch?v=muRIXCDw-k0&t=4433s) 节目尾声、Chamath 签名毛衣调侃与片尾音频集锦 在节目的最后，主持人们借着当季白色高领毛衣到处脱销的梗开起了玩笑。Jason 调侃说，如果大家在外面买不到白色高领毛衣，可以前往节目的官方商城 allin.com/store；节目组即将上线 1,000 件带有 Chamath 亲笔签名的标志性白色高领毛衣，观众现在就可以提前登记排队。随后 Jason 又抖出了节目的经典包袱，宣称这批毛衣的“所有销售收益都将捐赠给慈善机构”，但紧接着就笑着改口揭晓谜底：“我说的慈善机构，其实是指游艇基金（yacht fund）。”在轻松逗趣的告别声中，主持人和嘉宾互道下周再见，完成了正片录制。 录制结束后，节目切入了标志性的片尾混剪音频。伴随着经典口号 "Let your winners ride"，背景音播放了 Besties 历来被粉丝疯狂二创的各种搞笑名场面与日常互怼剪辑：从调侃把内容开源给观众后社区陷入疯狂，到诸如狗在自家车道上拉屎、找裁缝量身定制服装等无厘头生活琐事，再到几位主持人之间关于“简直需要找个房间开派对来释放彼此之间张力”的荒诞暴论。在一连串标志性的原声碎片与 "I'm going all in" 的收尾旋律中，整期播客正式落下帷幕。 - **官方周边戏谑**：推荐前往 allin.com/store 登记预购限量 1,000 件的 Chamath 签名款白色高领毛衣，收益戏称为几位亿万富翁的“游艇基金”。 - **经典片尾混剪**：包含 "Let your winners ride"、各期互损名场面音频切片以及收尾主题曲 "I'm going all in"。 ## Notable quotes > [09:26](https://www.youtube.com/watch?v=muRIXCDw-k0&t=566s) and then there were two — David Sacks > [12:32](https://www.youtube.com/watch?v=muRIXCDw-k0&t=752s) good enough — Jason Calacanis > [12:32](https://www.youtube.com/watch?v=muRIXCDw-k0&t=752s) a world of difference — Elon Musk > [18:41](https://www.youtube.com/watch?v=muRIXCDw-k0&t=1121s) 这完全就是 Android 与 iPhone 模式的重演——一个生态赚走了全行业的绝大部分利润，另一个则在全球范围内赢得了最广泛的用户规模和调用量。 — Jason Calacanis > [31:18](https://www.youtube.com/watch?v=muRIXCDw-k0&t=1878s) 入场价格至关重要 — Brad Gerstner > [34:15](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2055s) creating data centers is not rocket science — Elon Musk > [40:22](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2422s) Jensen said to me on the pod, he's like, 'Nobody comes close. Microsoft doesn't come close. You know, Google doesn't come close in terms of standing it up in that time frame.' — Brad Gerstner > [40:22](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2422s) As our good friend Bill Gurley likes to remind us, he's like, 'I can't believe that we're all just taking in stride this level of seller financing, right?' He would call it circular revenues... — Brad Gerstner > [49:26](https://www.youtube.com/watch?v=muRIXCDw-k0&t=2966s) 人才集中精力去追求能够带来超额回报的风险投资模式，而把沉重的 PE 盘子甩卖给了 Bending Spoons。 — David Sacks > [53:42](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3222s) 一旦增长停滞，士气便瞬间跌入谷底，客户流失率飙升并形成恶性循环，每天上班都痛苦不堪。 — Brad Gerstner > [01:00:20](https://www.youtube.com/watch?v=muRIXCDw-k0&t=3620s) 没人买微软是因为微软写的代码最好。大家买微软，是因为微软是所有业务运行的底层轨道。Active Directory 承载着员工的身份认证，Excel 是董事会财务数据的来源，Teams 记录着满足合规要求的会议，Azure 拥有 FedRAMP High 授权和美国国防部 Impact Level 5 安全许可——这意味着国防承包商绝不可能为了省钱而随意把它换掉。 — David Sacks > [01:07:36](https://www.youtube.com/watch?v=muRIXCDw-k0&t=4056s) 中国每年毕业的理工科（STEM）毕业生人数比全世界其他国家的总和还要多，他们最不缺的就是聪明绝顶的人才。 — David Sacks > [01:09:46](https://www.youtube.com/watch?v=muRIXCDw-k0&t=4186s) My lord, these things are good — Jason Calacanis ## Mentioned Google, DeepMind, Demis Hassabis, Jeff Dean, Discovery Loop, Gemini 3.5 Pro, Anthropic, OpenAI, SpaceX, Microsoft, Meta, Elon Musk, xAI, Colossus, Starlink, Starship, Cursor, CoreWeave, Grok, NVIDIA, Jensen Huang, Satya Nadella, Gwynne Shotwell, Jared Isaacman, Bill Gurley, Airtable, Bending Spoons, HyperAgent, Evernote, Salesforce, Marc Benioff, Claude Code, Lovable, Perplexity, Figma, Snowflake, Scale AI, Mercor, Micro1, Kimi, 通义千问 (Qwen), 智谱 GLM, 月之暗面 (Moonshot AI), 字节跳动, 腾讯, 阿里巴巴, ASML, Jason Calacanis, David Sacks, David Friedberg, Brad Gerstner, Chamath Palihapitiya ## Video description (0:00) Bestie intros! Brad Gerstner fills in for Chamath (2:16) Major shakeups at Google: AI brain drain or better strategy? (20:39) SpaceX's big quarter: Terafab, AI Capex, $1T revenue projection? (45:44) All-In Summit Speaker Announcements! (48:01) Airtable sells for a 90% discount: SaaSpocalypse? (1:05:56) Chinese AI labs are buying US training data to catch up Apply for Summit 2026: https://allin.com/events Follow Brad: https://x.com/altcap 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://x.com/the_ai_investor/status/2084687703707361429 https://x.com/Tesla/status/2085365278276284803 #allin #tech #news [Source](https://www.youtube.com/watch?v=muRIXCDw-k0)

[9:26] David Sacks: and then there were two [12:32] Jason Calacanis: good enough [12:32] Elon Musk: a world of difference [18:41] Jason Calacanis: 这完全就是 Android 与 iPhone 模式的重演——一个生态赚走了全行业的绝大部分利润，另一个则在全球范围内赢得了最广泛的用户规模和调用量。 [31:18] Brad Gerstner: 入场价格至关重要 [34:15] Elon Musk: creating data centers is not rocket science [40:22] Brad Gerstner: Jensen said to me on the pod, he's like, 'Nobody comes close. Microsoft doesn't come close. You know, Google doesn't come close in terms of standing it up in that time frame.' [40:22] Brad Gerstner: As our good friend Bill Gurley likes to remind us, he's like, 'I can't believe that we're all just taking in stride this level of seller financing, right?' He would call it circular revenues... [49:26] David Sacks: 人才集中精力去追求能够带来超额回报的风险投资模式，而把沉重的 PE 盘子甩卖给了 Bending Spoons。 [53:42] Brad Gerstner: 一旦增长停滞，士气便瞬间跌入谷底，客户流失率飙升并形成恶性循环，每天上班都痛苦不堪。 [60:20] David Sacks: 没人买微软是因为微软写的代码最好。大家买微软，是因为微软是所有业务运行的底层轨道。Active Directory 承载着员工的身份认证，Excel 是董事会财务数据的来源，Teams 记录着满足合规要求的会议，Azure 拥有 FedRAMP High 授权和美国国防部 Impact Level 5 安全许可——这意味着国防承包商绝不可能为了省钱而随意把它换掉。 [67:36] David Sacks: 中国每年毕业的理工科（STEM）毕业生人数比全世界其他国家的总和还要多，他们最不缺的就是聪明绝顶的人才。 [69:46] Jason Calacanis: My lord, these things are good

[3:43] David Friedberg: Capex is high alpha, low beta in data center infrastructure... Model development theoretically could be high alpha, but it's very high beta, it's a very risky way to deploy capital. [9:14] David Sacks: The market for frontier intelligence has become a duopoly... Anthropic is now over 80 billion of ARR, started the year at 10. It had forecast 100 billion as exit ARR... now it looks like they're going to do it with a couple of months to spare. [15:48] David Friedberg: On the consumer side... they can pay 20, 40 bucks a month and get ChatGPT or Gemini or Claude and be very happy... For enterprise, I think the enterprise is going to be very active in selecting a blend of models. [15:48] David Friedberg: It is way too early to count Gemini out on building incredible specialized models. They have the best video data. They have the best life sciences data. [15:48] Brad Gerstner: America's winning. This is exactly what you want. We have a massively competitive market. [15:48] Brad Gerstner: Closed models are actually cheaper, you know, if you don't have to build it for yourself, if you don't have to... [pay] the training costs and a lot of expertise to fine-tune and maintain and guardrail. [20:39] Jason Calacanis: Q2 results were spectacular... 7.8 billion in revenue, up 92% year-over-year. Let that sink in. [20:39] Brad Gerstner: He pulled forward the $1 trillion target in ARR by a year from 2031 to 2030. [20:39] Brad Gerstner: Cursor plus Grok could be at 10 to 20 billion by the end of the year. That would be an extraordinarily valuable asset. [24:46] David Sacks: Starlink's just an unbelievable juggernaut cash machine... That alone provides the cash flow to fund much of what Elon's doing. [31:18] Brad Gerstner: In the short run, the market is a voting machine, but in the long run, it is a weighing machine. [38:38] David Sacks: If there is a scare about demand, the whole sector trades down together because of the leverage you're pumping into the system. [48:01] David Sacks: This was a company that had a successful PLG motion... growing about 20% a year. But that was not good enough for its board... So the board pressures the founders to bolt on a traditional sales-led motion to get growth up faster. Does that work? No. [52:48] David Sacks: It's almost impossible for a software company today to keep any decent sales people or product development people, because they all want to go work on AI. [61:06] David Sacks: Nobody buys Microsoft because Microsoft writes the best code. They buy Microsoft because Microsoft is the rail that everything else runs on. Active Directory is where employee identities live... Azure holds a FedRAMP High authorization and DoD Impact Level 5 clearance. [65:58] Brad Gerstner: If the president asks his advisers... 'Are we winning against China?' And all of a sudden he gets a response, 'No, we're no longer winning, they've caught up,' then these things will get a lot more scrutiny.

[0:00] All right, everybody. Welcome back to [0:02] your favorite podcast. It's the All-In [0:03] podcast. It's the summer. It's August 6. [0:06] Having a hard time getting a quorum here [0:09] on the podcast, but David Friedberg is [0:12] here. David Freeberg is back. Our Sultan [0:14] science. How you doing, brother? [0:15] >> Great to be with you. [0:17] >> It's great to be with you. And everybody [0:20] loves when Brad Gersonner is here. He's [0:23] your Bruce Wayne. If markets are your [0:25] game, he brings that namaste to your [0:28] payday. Yes. [0:29] >> Buy his glasses at discount and he'll [0:31] get you one of those fancy Trump [0:33] accounts. All right. Welcome back to the [0:34] program, Brad. [0:36] >> I love it. I love it. You bring the [0:37] rhymes back. [0:38] >> I bring a little intro back. We've been [0:40] trying Chimath is on the road right now. [0:42] Chimath is on the road, but we will get [0:44] a field report from Chimath and uh I I [0:46] call Daniel somehow Sax is going to be [0:50] here, but you know how he is. He's [0:51] always late because you know get a phone [0:53] call from very important people, but he [0:55] will break in at some point. Oh, wait. I [0:58] see in the text here. Oh, there he is. [1:00] He made it. [1:00] >> Hey guys, [1:01] >> you made it. [1:02] >> How do you like my beautiful summer [1:04] jallet? [1:05] >> Uh, it's incredible. It fits perfectly. [1:08] You look warm. [1:09] >> I don't know how Jimoth does this. [1:14] >> Let your winners ride. [1:16] >> Rainman David [1:21] and [1:21] >> we open source it to the fans and [1:23] they've just gone crazy with it. Love [1:25] you. We can walk [1:28] in. [1:29] >> Well, here's the report, everybody. As [1:30] everybody knows, Chimath is on the road. [1:33] He uh Oh, here he is. He um this is a [1:35] photo sax. He went he went to check his [1:38] data center progress. Uh I think that's [1:40] in Colorado or Nevada where he was [1:42] building a data center. [1:43] >> I think that's on Dune. [1:44] >> Oh, it's on Dune. Yes. Dune 4. Oh, yes. [1:47] Here he is admiring himself. Oh, look. [1:50] Here's Nat. You know when a meme has [1:54] reached its peak? when your wife starts [1:56] dunking on you. There it is. And here we [1:58] are. This was at the Christmas party, I [2:00] think. Oh, you were on CNBC with Andrew [2:03] Ross working. Oh, there you go. [2:05] >> I wasn't sure if it was my Twitter feed [2:07] that was just selecting into it, but it [2:09] hit everyone, right? This was a viral [2:11] thing. [2:11] >> This has hit everything. All right, [2:12] listen. We got a lot to get to. Enough [2:14] with the shenanigans and small talk. [2:16] Google had uh two major shakeups to its [2:18] AI staff on Wednesday. Deis Hassabis [2:22] has moved to chair of Deep Mind and [2:24] chief scientist at Google. Reports [2:26] describe this as Demis stepping down or [2:28] being kicked upstairs. Uh we'll get into [2:31] that, but Google framed it as a [2:33] promotion uh and says he was stepping [2:34] up. Here's Axios's quote explaining the [2:36] shakeup. Quote, "Google's Gemini 3.5 Pro [2:39] is months behind with some company [2:42] sources telling Axios that it's in part [2:44] due to low morale." Interesting. Several [2:47] top researchers, including Gemini's [2:49] co-lead, have left the firm for [2:50] competing AI labs. Chef Dean, plus three [2:53] other AI superstars, are leaving Google [2:55] to start a company called Discovery [2:56] Loop. Dean is a legend. Freeberg, uh, [2:58] and I think you worked with him at [3:00] Google, one of the world's great AI [3:01] engineers. He was employee number 30, [3:03] joined in 1999, and has worked there for [3:06] what I understand continuously for 27 [3:08] years. Discovery Loop's going to be [3:10] focused on deep scientific breakthroughs [3:12] in AI. Google Share is down 4% on the [3:15] news of Dean leaving. So 200 billion in [3:18] lost market cap if you want to correlate [3:21] those two things. Freeberg, this is your [3:23] alma mada. What are your thoughts here? [3:25] Is this creative destruction? Maybe [3:27] these people weren't delivering and they [3:29] wanted fresh blood or is this just the [3:31] siren call of doing a startup in an age [3:34] of unlimited capital for AI and [3:35] unlimited opportunity just being too [3:38] much for the OGs at Google to not take [3:41] advantage of. Maybe it's the third [3:43] bucket, which is if you're the board and [3:45] the management, you're having a debate [3:48] about how to best deploy capital. Google [3:50] has made a commitment to deploy $200 [3:53] billion in capex this year in AI [3:56] infrastructure data center buildout. [3:58] Because of the capex and accelerated [4:00] depreciation, [4:01] making an investment in AI compute in [4:05] the US right now is hugely tax [4:07] advantaged. And because of the extreme [4:09] demand for compute, it's a pretty [4:13] obvious kind of ROIC model, return on [4:16] invested capital. So if you make this [4:18] sort of an investment, you have [4:19] significant demand for that compute [4:21] infrastructure. You're very good at [4:23] running the compute infrastructure. That [4:25] capital can deliver massive profit [4:28] returns for you with very high [4:29] confidence in some forecasted period. [4:32] Building the most advanced frontier lab [4:36] driven model also takes tens of billions [4:39] of dollars of capital. And the question [4:41] really is can you deliver the profits [4:44] from the model? And in a world where [4:47] open-source is becoming so good and open [4:49] weights models are catching up so [4:51] quickly and all the frontier labs are [4:53] catching up to each other so quickly, [4:55] does it really make as much sense to [4:57] deploy tens of billions of dollars [4:59] against building a model? And I think [5:01] that the scientists that we're seeing [5:02] transition out are the scientists that [5:04] have been at the core of model [5:06] development of making these frontier [5:08] models. And they were certainly first [5:10] out the gate. [5:11] >> You can look at some of the early [5:12] interviews with Jeff Dean from a couple [5:13] years ago where they actually had a chat [5:15] GPT equivalent internally a year before [5:18] chat GPT came out from OpenAI. Google [5:21] chose not to release it for fear of [5:23] cannibalizing search and so on. That's [5:25] when Sergey stepped in and there was [5:27] this whole kind of revitalization. But [5:28] as time has gone on and as everyone has [5:31] competed on models, as we've talked [5:33] about many times on the show, I think [5:35] it's pretty obvious that it is very hard [5:37] to get the same sort of return on [5:40] capital invested in model development as [5:42] it is in capital invested on compute [5:44] infrastructure and being model agnostic. [5:47] What Google has is probably one of the [5:49] greatest install enterprise bases in the [5:52] world for compute. So they have the most [5:54] enterprise customers. They have the most [5:56] consumers. And in both cases, they don't [5:58] necessarily need to have the best model [6:00] to make an incredible business. They can [6:02] be model agnostic. They can work with [6:04] anthropic. They can work with OpenAI. [6:06] They can work with SpaceX. They have a [6:08] significant ownership stake in SpaceX [6:10] and in Anthropic and they can work with [6:12] all the open weights models. They can [6:14] host them all. So now if you're one of [6:16] the great computer scientists, you're [6:17] Demis, you're Jeff Dean, you're this [6:20] whole crew, and you're inside at Google [6:22] and they're allocating capital not to [6:23] your models, not to the things that [6:25] you're most interested in, but they're [6:26] allocating capital to infrastructure and [6:28] data centers and supporting the broad [6:30] ecosystem of models, you start to say, [6:33] well, given the fact that I can go down [6:35] the road and visit Brad Gersonner and a [6:36] couple other people and raise a couple [6:37] billion dollars at a multi-billion [6:39] dollar pre- money with a PowerPoint deck [6:41] because I'm the greatest in the world at [6:42] doing this, that might be a better path. [6:44] for me and I think that that's the [6:46] moment. So I the way I would frame it is [6:48] capex is high alpha low beta in data [6:52] center infrastructure that capital and [6:55] model development theoretically could be [6:57] high alpha but it's very high beta it's [7:00] a very risky way to deploy capital so so [7:02] if I'm the board I'm the management I'm [7:04] deploying more capital in computing [7:05] infrastructure less capital into model [7:07] development that's what I think's going [7:09] on [7:09] >> Brad what's your take on this [7:11] >> I think David nails it I mean listen the [7:13] same thing's going on at Microsoft right [7:16] is out this week saying, you know, [7:17] citing Morgan Stanley's report and [7:19] saying they're seeing over a 30% return [7:22] on invested capital in tokens as a [7:24] service, right? So, in the [7:26] infrastructure business. So, I think [7:27] David's exactly right. Those are such [7:30] good businesses, right? You you you you [7:32] deploy capital, everybody's running it [7:35] from you, but the scientists who want to [7:37] be involved in super intelligence, who [7:39] want to cure cancer, who want to be on [7:41] the frontier of these models, right? [7:43] They're sitting there dealing with this [7:44] channel conflict at Google because you [7:47] know uh uh Google cloud wants all of the [7:50] compute in order to rent it out to [7:51] Anthropic and those those building the [7:54] frontier models internally want that [7:56] compute in order to compete with [7:58] Anthropic. So you have this inherent [7:59] channel conflict uh between those [8:02] wanting to build the models. I think [8:03] David said it really well. Um and I [8:06] think that's a that that's a big [8:07] challenge for them. It looks like it's [8:09] being resolved in favor of being more of [8:12] an infrastructure company. So where does [8:14] you know telescope out for a second [8:17] SpaceX also reported this week they also [8:19] have channel conflict. They're renting [8:21] out their compute to anthropic at the [8:23] same time they're trying to build their [8:24] own model with Grock and Cursor. You [8:26] have that channel conflict at Google. [8:28] You have that channel conflict at [8:29] Microsoft although I don't even really [8:31] see them pushing the frontier anymore in [8:33] terms of models. Meta's talking about [8:35] getting into the infrastructure as a [8:37] service game. And then at Anthropic and [8:39] Open AI, you don't have any of that [8:40] channel conflict. They say we're not in [8:42] the infrastructure business. We're only [8:43] in the model business. So, I think it's [8:45] a, you know, uh a a clarifying view as [8:50] we look forward that we may in fact not [8:52] have those companies on the frontier of [8:54] of model development if all these people [8:56] leave. By the way, thanks to the law [8:58] passed on capex depreciation, if you [9:00] assume a 26% corporate tax rate, every [9:03] dollar you deploy in capex, because you [9:05] get to write it off in this year, you're [9:07] basically getting 26% off. You know, [9:10] that's money you get right back. Yeah. [9:12] >> Yeah. Hey, uh Sax, let me have you [9:14] comment on uh this as well. Poly market, [9:16] which companies will have the number one [9:18] AI model by the end of this year on [9:20] December 31st? Um, now of course in the [9:24] last time they did this anthropic one, [9:26] so they're not on the list. They're the [9:27] winner, but who will have it uh going [9:29] forward? OpenAI 32%, Google 20%, Alibaba [9:32] 14, and then you got Moonshot, XAI, [9:34] Meta, Bite Dance, all at about 10%. So [9:38] Sax, your thoughts here on what's the [9:41] better business? Is the better business [9:43] being in the language model, frontier [9:47] model, or is that getting quickly [9:48] commoditized? and really you want to be [9:49] in the token sale business or is that [9:52] also going to be a commodity and you [9:53] just need to be on the application [9:54] layer? [9:54] >> Here's what I think is going on in terms [9:56] of the the market structure is when I [9:58] saw this Google news, my reaction was [10:02] and then there were two because like [10:04] Brad was saying, we used to have five [10:06] major companies in the hunt to be the [10:09] leading frontier lab, the leading [10:11] frontier model just a year ago. Now [10:13] we're really down to just anthropic and [10:16] open AI. So the market for frontier [10:19] intelligence has become a duopoly. Now [10:21] Elon is still on the hunt. I'm sure [10:23] Google would say they're still on the [10:24] hunt, but like Brad is saying, they may [10:26] have contradictory incentives there [10:29] because they can actually do quite well [10:31] just with their compute. So I think that [10:34] the market for frontier intelligence has [10:36] become a duopoly. I think it's a very [10:37] powerful duopoly. I don't think it's [10:40] being commoditized. I think that what [10:41] we're evolving to is a two-tier market [10:44] structure where there's a market for [10:46] frontier intelligence and there's a [10:47] market for let's call it kind of [10:49] commodity or lagging intelligence [10:51] whatever you want to call it that's 6 to [10:53] 12 months behind. There is a market for [10:56] those tokens those models but the [10:58] reality is you can't charge anything for [11:00] the weights. You can charge for the [11:02] compute you can charge for the inference [11:04] that you're providing. You can charge [11:06] for essentially consulting services to [11:08] help put the whole thing together. [11:10] But if you're not at the frontier, you [11:12] can't charge for the model layer itself. [11:16] If you are at the frontier, you can [11:18] charge a premium. And that's where [11:20] Anthropic and Open AI are. And I think [11:23] the proof for this is just you look at [11:25] the growth rates of these companies. The [11:27] latest we heard is Enthropic is now over [11:29] 80 billion of ARR. Started the year at [11:32] 10. It had forecast 100 billion as exit [11:36] ARR for the year. And most people said [11:38] that that would be impossible to [11:39] achieve. Now it looks like they're going [11:41] to do it with a couple of months to [11:43] spare. So their estimates are going up. [11:46] I mean 110 120 or higher for end of year [11:49] ARR. Open AAI seeing acceleration. So I [11:52] think what you're seeing now is a very [11:54] clear bifurcation in the market. You've [11:57] got a frontier model duopoly that can [11:59] charge a premium. I think of it like [12:02] Apple. You know Apple's competing [12:04] against Android. It's open source. [12:07] Android actually has more users in the [12:08] world, but all the monetization goes to [12:11] Apple because people are willing to pay [12:12] for the premium experience. I think in a [12:15] similar way, people are willing to pay a [12:17] premium for true frontier intelligence [12:19] if it's really at the leading edge. But [12:22] if you're not the leading edge, there's [12:23] a huge market for that, too. But it's [12:26] highly commoditized. People are just [12:27] willing to pay you for the compute. [12:29] >> So, I mean, that's what I see happening [12:31] right now. [12:32] >> Yeah. [12:32] >> Jason, what do you think? Jason, what do [12:34] you think? thing. Uh, well, if you look [12:35] at Google Cloud, they posted 82% [12:37] year-over-year revenue growth, which is [12:39] something we've never seen uh, in the [12:41] history of these cloud providers. Elon [12:44] Musk and XAI just had the SpaceX [12:47] earnings. We're going to get into that, [12:48] but they also had massive uptick in [12:50] their Elon web services, as I've dubbed [12:52] it. And if you look at Google, I still [12:55] think Google will be the number one uh [12:57] AI company because they have so many [13:01] people using AI inside of their products [13:04] already. They have five products now [13:06] with over three billion monthly users [13:08] each. Android search, Gmail, Chrome, [13:10] YouTube all have over three billion. If [13:13] you've used any of these products [13:15] recently, uh they are becoming AI first [13:18] products. YouTube especially, but [13:20] obviously Chrome and Gmail, you're [13:22] seeing um tools pop up there for AI. And [13:25] then Freeberg, you kind of alluded to [13:28] this. They have 13 products total with [13:30] over a billion. And that now includes [13:32] Gemini. In Q2, Gemini had over 950 [13:36] monthly active users, tripling [13:38] year-over-year. They will be the number [13:40] one AI company in terms of consumer [13:43] usage by far, I think, this year. That [13:46] doesn't mean that the um frontier models [13:49] are not great businesses. They obviously [13:51] are. But I have been using exclusively [13:54] non-frontier models. And for 95% of the [13:57] jobs I'm doing, Sachs, [14:00] it's good enough. And I just posted [14:02] about this, you know, um and Elon and I [14:04] got into it a little bit here, and I [14:06] think you referenced this in our group [14:07] chat. Uh I I tweeted just the other day, [14:09] the difference between the open source [14:10] models I'm using and Frontier is [14:12] negligible already. I believe that to be [14:14] a true statement for the work I'm doing. [14:16] And he said, Elon responded back to me, [14:17] it's actually a world of difference. You [14:19] know, if you're doing something other [14:22] than making a copy of a video game or [14:26] you have incredible speed needs, the [14:28] Frontier models are not necessary [14:30] anymore. They're just not necessary. The [14:32] people using the Frontier models are [14:34] doing it because their company set it up [14:37] and they it's too hard to implement open [14:39] source right now, but it's going to get [14:40] easier and easier to implement it. So, [14:42] I'm still going with open source and [14:43] Gemini being the leaders in this. [14:46] >> Look, I think it it's true for your use [14:48] cases that let's say the cheaper [14:50] commodity intelligence that middle of [14:52] the market is good enough. Look, an [14:55] Android phone would be good enough for [14:57] me. I could get by on a cheap Android [14:59] phone. You know what? I still pay a [15:01] premium for this because I use it so [15:03] much. So if you're a business that let's [15:06] say you are a hedge fund and you're in a [15:08] highly competitive industry, you don't [15:10] want to take the chance that you're not [15:11] getting the best intelligence to power [15:14] your models, you know, and there's a lot [15:16] of industries like that where the [15:17] competitive dynamics will drive you to [15:20] pay for the best intelligence. There's [15:21] also situations, this goes back to the [15:23] blog post that Decagon posted, which is [15:26] if you're looking for use cases, you [15:28] also want to use the true frontier [15:30] because again, when you're dealing with [15:32] immature use cases, you don't know where [15:34] the value is going to be and you're [15:35] searching for opportunity to use AI. You [15:38] just want to use the best because again, [15:40] the return on finding those use cases is [15:43] going to be so much greater than the [15:44] small premium you're paying at the token [15:46] level. So I think there's a lot of [15:48] examples like that when you know the use [15:49] case is immature where you're in a [15:51] competitive industry where you're just [15:53] deploying AI you want the convenience of [15:55] the full [15:56] >> so why not go frontier model [15:59] >> you know unless your employees are doing [16:00] something stupid like you create a [16:02] leaderboard and they're token maxing I [16:04] don't think the cost is that great and [16:06] again the benefit that you're getting is [16:07] huge so a lot of people are just like [16:09] give me the best I'm willing to pay a [16:10] premium for the best [16:11] >> I'll take a slightly different take I [16:13] think that it's not necessarily do you [16:15] take the best model or the open source [16:17] model. I think that there's a blend [16:19] that's happening. At least that's what I [16:20] see. For example, we'll use open- source [16:23] open weights for a vast majority of [16:26] simple workflow applications. But when [16:29] it comes to specialized applications [16:30] where we really need to have high [16:33] quality model proficiency, for example, [16:36] in life sciences, in genomics modeling, [16:38] I am going to go for the premium model. [16:40] If I'm working at a media company and [16:43] I'm trying to do AI rendering of video, [16:46] I'm going to use Gemini's model that [16:48] does video. It is the best model or Sora [16:50] or whatever the best model is for that [16:52] particular application. So, I think the [16:54] idea that there's kind of a model that [16:56] you pick for everything, I think is the [16:58] false assumption. On the consumer side, [17:00] it is likely the case that the consumers [17:02] are not going to be using some [17:03] openweight model because they can pay 20 [17:05] 40 bucks a month and get chat GPT or [17:08] Gemini or Claude and be very happy [17:10] paying 40 bucks a month and they'll [17:12] basically be able to minimize their cost [17:14] to run that for consumers. For [17:16] enterprise, I think the enterprise is [17:17] going to be very active in selecting a [17:20] blend of models that are going to make [17:21] the most sense. Very cheap openweight [17:24] model for simple workflow applications, [17:26] individual employees spinning up an app, [17:28] whatever. and then more complex models [17:30] for those really key workflow tasks and [17:32] then specialized models. And I will say [17:34] it is way too early to count Gemini out [17:36] on building incredible specialized [17:39] models. They have the best video data. [17:41] They have the best life sciences data. [17:43] They've been working on this for far [17:44] longer than anthropic or open AI in the [17:46] life sciences side. They're very well [17:48] ahead on that front. I mean Demis is [17:50] still going to be running isomeorphic [17:51] labs. So when it comes to these [17:53] specialized models, verticalized [17:54] specialized models like video, life [17:56] sciences, protein folding, I think these [17:58] are the things where you're really going [17:59] to see Gemini shine and then every [18:01] enterprise is going to have a mixture. [18:03] But hey, if you can be the cloud service [18:04] provider with that mixture of models, [18:06] which is what Google GCP can now be, I'm [18:09] going to sign up for working with GCP [18:11] versus working just with Anthropic. One [18:13] of the things this has created is [18:15] downward pressure on the pricing. We saw [18:17] OpenAI and Claude uh do massive price [18:20] cuts uh for tokens. So they are [18:22] reacting. They're not taking it sitting [18:24] down and the orchestration between these [18:26] models is being built into a lot of [18:29] harnesses inside of enterprises. So [18:31] what's your take on the downward [18:33] pressure on token pricing or is this [18:35] just great for consumers and [18:36] enterprises? Because [18:37] >> listen, we're got massive competition. [18:40] >> That's the thing. America's winning. [18:41] This is exactly what you want. We have [18:43] massively competitive market. We have [18:45] Chinese open source, domestic open [18:46] source, Frontier National Labs that are [18:49] doing what they're doing. We have [18:50] downward pressure on pricing. You know, [18:52] David re referenced a duopoly. You know, [18:54] I think it's hard to call it a duopoly [18:56] when you're, you know, only a few years [18:58] into this and you have giants like [19:00] Amazon, Microsoft, and Google. I do [19:02] think he's right. I do think they've [19:03] emerged, you know, as the pure plays. [19:06] Their revenues would suggest that [19:07] they're, you know, they're gaining share [19:09] of wallet. But there are two points I [19:11] want to make here because I think [19:12] they're non-conensus views that were [19:15] spoken this week. One was Elon's [19:17] response to you, Jason. Right. Over the [19:19] last two weeks, everybody's been saying [19:21] that the Chinese have caught up, that [19:23] open source tokens have caught up in [19:25] intelligence, that they're much cheaper, [19:27] etc. And Elon comes out and says, "Not [19:29] so fast. We're entering the singularity [19:32] and the frontier models are way further [19:34] ahead than people think." I believe that [19:35] to be true. I think for your use case [19:38] they're very similar but I don't think [19:40] that's the most sophisticated use case [19:42] that people are trying to train on and [19:44] trying to experience. And then Jensen [19:46] came out this week and said closed [19:47] models are actually cheaper you know if [19:49] you don't have to build it for yourself [19:51] if you don't have to uh you know the [19:54] training costs and a lot of expertise to [19:56] fine-tune and maintain and guard rail [19:58] and keep it safe. So he's basically [20:00] making the argument that not only are [20:02] the the the frontier models further [20:04] ahead, but that the cost differential [20:06] between the two is not what everybody's [20:07] making it out to see to to be, which I [20:10] think explains why they continue to run [20:12] away with it on the revenue side of the [20:14] equation. Um, but I think we have [20:17] healthy competition. I you're right JCL [20:20] you know for the vast majority of use [20:22] cases I think token consumption is going [20:24] up for the open source guys while share [20:26] of economics is going up for the [20:28] frontier labs I think that's what we [20:30] want to see yeah and it's just Android [20:32] versus iPhone all over again one [20:35] platform makes the profit one gets the [20:37] majority of users at least globally in [20:38] usage all right let's talk SpaceX here [20:41] uh they had their first earnings report [20:43] as a public company shares dropped 13% [20:47] uh I I think because people were a [20:48] little concerned about the surging AI [20:50] capex. It's down 30% since going public [20:53] in June, but it's now trading at, it [20:55] seems to have settled in at a 1.4 [20:57] trillion valuation. Went public [20:59] obviously above 2 trillion. Q2 results [21:02] were uh spectacular is the only way to [21:04] put it. 7.8 billion in revenue, up 92% [21:08] year-over-year. Let that sink in. Uh and [21:11] 67% quarter over quarter. AI revenue. [21:14] Elon Web Services more than tripled [21:16] quarter over quarter to $2.6 billion. [21:20] That's not cursor. That hasn't closed [21:22] yet. Uh but that's going to be one of [21:24] the great purchases in history. This is [21:26] from Elon Web Services uh renting out [21:28] compute specifically to Anthropic and [21:30] Google from the Colossus uh collection [21:33] of servers. But capex was up 18.4 [21:37] billion in the quarter. That's 6x [21:39] year-over-year. Obviously, you can do [21:40] the math there for a run rate of about [21:42] 75 billion dollar. I'll stop there and [21:45] get your reaction, Brad, to the SpaceX [21:48] IPO. I know you've been tracking this [21:50] and commented on it heavily. [21:52] >> I mean, listen, I think that one, first, [21:53] let's start off. $1.4 trillion of value [21:57] creation for this company is [21:58] extraordinary. So, the fact that from [22:00] peak to trough, it's down 40 or 50% from [22:02] the IPO. We had that chart out a few [22:04] weeks ago. Remember that within 6 months [22:07] of the IPO almost all these tech stocks [22:09] are down 50% peak to trough. We see it [22:11] again here with SpaceX. I thought it was [22:13] a really solid quarter. I thought his [22:15] guides were pretty extraordinary. 100 [22:17] billion in ARR by the end of the year [22:19] and he pulled forward the $1 trillion [22:22] target in ARR by a year from 2031 to [22:25] 2030. Now to just put that in [22:27] perspective, Morgan Stanley's 2030 [22:30] revenue estimate is 325 billion which is [22:33] also extraordinary. Remember, this [22:35] company did 18 billion in revenue last [22:37] year. So whether you're taking Morgan [22:39] Stanley's numbers or Elon's numbers, [22:41] clearly the market is not pricing that [22:43] in. At 2 trillion, we were pricing ahead [22:46] a couple years. I think now it's, you [22:49] know, the the value reflects kind of [22:50] where we are. The market has questions [22:53] about a few things. Here's what they [22:54] are. Number one, on the rental business, [22:56] the rental of compute business, he [22:58] rented out a huge block of compute to [23:00] Anthropic. It's the question that we've [23:02] been talking about here. Are you going [23:04] to use the compute to build your own [23:05] frontier model or are you going to rent [23:07] it out? And if you rent it out, are you [23:10] going to be able to find those people [23:12] who have the capital to offtake that [23:14] compute? He's talking enormous numbers, [23:17] 10 to 20 gigs, and people are wondering [23:19] how they're going to be able to finance [23:20] that. And remember, those businesses, [23:22] the GPU rental businesses tend to trade [23:25] at very low multiples. Look at [23:26] Coreweave, etc. On the frontier model [23:29] business, I think this is the sleeper. I [23:31] think he said on the call that Grock [23:33] tripled tokens in the month of July. [23:36] That doesn't include cursor. Curser was [23:38] already on a path to go from 3 billion [23:40] to 10 billion by the end of the year. [23:43] Curser plus Grock could be at 10 to 20 [23:46] billion by the end of the year. That [23:48] would be an extraordinarily valuable [23:50] asset going to trade at a much higher [23:51] multiple than the data center business. [23:53] And then of course we haven't even [23:54] talked about Starlink and what he's [23:55] going to do uh you know I think going to [23:58] run the table on mobile. So this is the [24:00] normal consolidation. We have funds like [24:03] uh across Silicon Valley that are [24:05] distributing their shares. The stock has [24:07] traded down a bit. Nothing surprising to [24:09] me here. Now it's all about execution. I [24:11] think the most important thing to watch, [24:13] the two most important things to watch [24:14] are number one, how do the Grock and [24:17] Cursor revenues end the year? [24:19] >> And number two, um you know, the [24:21] traction they get on, um you know, [24:24] continuing to replace traditional mobile [24:26] carriers with Starling. The distribution [24:28] started I think today or yesterday. I [24:30] got my first distribution from a fund [24:32] I'm in. I'm in a couple of funds that [24:34] are in SpaceX. Seems like everybody's in [24:35] that and that will obviously create [24:38] downward pressure if you are amongst the [24:40] people who want to cash out and been in [24:42] it for a long time but I'm holding these [24:43] for my grandkids. Saxs your take on [24:46] these spectacular [24:49] I guess is the only way to describe them [24:50] results coming from a vertical that [24:53] wasn't part of SpaceX's business but [24:56] nine months ago. Yeah, look, I thought [24:58] it was a very bullish earnings call. I [25:00] was a little bit surprised that the [25:01] stock went down after the earnings call [25:03] because not only was it a beaten raise, [25:06] but also I think Elon spoke to a lot of [25:08] their plans. The only thing I would add [25:10] to to what Brad said was around [25:13] Starship, Elon basically said, we all [25:15] saw it, right, that the Starship test [25:18] flight was successful. The Starship's [25:20] floating in the ocean, the heat shield [25:21] worked. That's going to enable more [25:23] flights of Starship now at a more [25:25] accelerated rate. that paves the way for [25:27] the V3 satellite which enables much more [25:30] bandwidth for the Starlink network which [25:32] then powers the whole direct to cell [25:34] play. So you had that piece of it. I [25:36] mean just the whole telecom aspect [25:38] seemed very on track and they're very [25:40] bullish about that and then you've got [25:41] the whole AI data center play. Now on [25:44] the data centers I think what they said [25:46] is that they expected to go from 1.4 [25:49] gawatts of compute to about two by the [25:52] end of the year. And Elon said that the [25:55] spot pricer computes in the $30 to $50 [25:57] per watt range. So, you know, you do the [26:00] math, a gigawatt is a billion watt. So [26:03] 30 to$50 per watt means 30 to50 billion [26:07] per gawatt. And I think they're at the [26:09] high end of that range right now. So [26:11] when Elon says, look, we're going to end [26:13] the year at 100 billion of ARR, all you [26:16] have to believe is that they're at 2 [26:17] gawatt of compute running for $50 a watt [26:21] to hit that. That doesn't include [26:23] Starlink or the launch business or the [26:25] Grock cursor piece or any of these [26:27] things. So I think that's why they're [26:30] >> multiple ways to win is what you're [26:32] saying, Sax. There's multiple ways to [26:33] win with the stock. [26:34] >> I think Starlink's just an unbelievable [26:36] juggernaut cash machine. If you look at [26:38] the financials, their segment reports, [26:40] space, connectivity, and AI. And on the [26:43] connectivity side, the Starlink side, it [26:45] they generated $2.6 billion in adjusted [26:48] EBA. You can kind of approximate that to [26:50] be kind of operating cash flow. Space [26:53] was kind of, you know, negative 200 [26:55] million. So call it break even and AI [26:58] was plus 1.1 billion. But AI, to Brad's [27:01] point, it's unclear whether the pricing [27:04] they're getting on compute rental today [27:06] is temporary and at a premium because of [27:09] the lack of compute available in the [27:12] market today. And people that need [27:13] comput are paying Elon a premium for [27:14] that comput. So I think there's a [27:16] question mark where that goes. But the [27:18] connectivity piece on Starlink, 4.3 [27:22] billion in the quarter and 2.6 billion [27:26] in adjusted EBIT. He's got 12 million [27:28] subscribers. That's doubled [27:29] year-over-year. $66 arpoo per month uh [27:33] what people are paying per month. And he [27:36] grew 20% quarter over quarter. So if you [27:38] extrapolate this out, he's pretty close [27:40] to being at a 24 million subscriber run [27:43] rate on this multiple. Assuming this [27:46] enterprise stuff which is like airlines [27:47] and other things scale which they seem [27:49] to be scaling with the consumer [27:51] business, Starlink alone could be [27:55] generating on the order of $40 billion [27:57] of revenue topline with a huge amount of [28:00] that flowing to free cash. That could be [28:02] a $30 billion free cash flow within the [28:05] year. That alone provides the cash flow [28:07] to fund much of what what Elon's doing. [28:10] And if you just put a 30x multiple on [28:12] that, which I think you can because [28:13] these subscription businesses are very [28:15] high renewal rate, very low capac, I [28:18] think you could probably get a 30x just [28:20] on the Starlink business. The Starlink [28:22] business alone could be a trillion [28:24] dollar market cap within 2 years, within [28:27] 18 months, let's say. That I think funds [28:29] all of the rest of this as kind of [28:31] science projects and upside. So, I'm [28:33] kind of making a bull case. It's crazy [28:35] to me how well the Starlink business [28:37] performs and you can see it in AT&T and [28:40] Verizon, Huenet, Viaat. I mean, these [28:43] companies have been decimated. I used to [28:44] have a Huenet satellite dish on my [28:46] Sonoma County ranch in order to get [28:48] internet. That's what we had to use. It [28:50] was like, you know, 200 bucks a month or [28:51] something [28:53] orbit, right? And they take forever to [28:55] >> terrible service. And that market got [28:57] decimated by Starlink. And if he [28:59] launches the handset thing, that [29:01] subscriber growth is going to go right [29:03] now. He's adding 2 million subscribers [29:05] on the consumer side a quarter. You [29:07] could see that going to four to 5 [29:09] million a quarter. You could actually [29:10] see an acceleration in the consumer [29:13] million mobile sub 400 million mobile [29:15] subs just in the United States. [29:18] >> I think you can make you can make the [29:19] bull case on Starlink alone and then the [29:21] rest of it is like, hey, is Elon going [29:23] to do well with investing the excess [29:24] capital that's spitting off of Starlink? [29:26] How's Elon going to do with that money? [29:28] Well, I don't know who else I give it to [29:30] to like, you know, do what he's doing [29:31] with Starship and with AI compute and [29:34] the terra fab. [29:35] >> Oh my god, this is a science fiction uh [29:39] county text [29:41] >> how the US gets off of this dependency [29:43] with Taiwan and China from [29:44] semiconductors. If Elon takes this on [29:46] his shoulders and he delivers what he's [29:49] showing as a vision here today, this is [29:50] going to be the greatest semiconductor [29:52] fabrication site on planet Earth. Well, [29:55] you know, I I would say something, you [29:57] know, David, to your point, you know, [30:00] how many CEOs or founders would just [30:02] take that Starlink business, which is [30:04] such an exceptional business, [30:06] trillion dollar business going to two [30:08] trillion, and they would not take any of [30:09] these other risks. They would not do [30:12] terapab. They would not try to build out [30:14] the data center. They would not try to [30:15] build their own model. That's highly [30:16] risky but highly important investments [30:19] that are being made. I mean, it is [30:21] heroic and important that we have this [30:24] level of I just think unbridled [30:27] enthusiasm for innovation uh on the [30:29] frontier that Elon's doing. And I wish [30:32] we saw more CEOs, more public companies [30:35] willing to take this level of risk. We [30:37] just got done talking about, you know, [30:38] some CEOs maybe that were taking less [30:40] risk because the safe bet was was easier [30:43] to make. Elon refuses just to take the [30:47] safe bet. He's taking all the dollars [30:49] from this thing where he has an [30:50] extraordinary business and plowing them [30:52] back into these things that are [30:54] critically important to the United [30:55] States. And by the way, Brad, such a [30:57] good point because if you look at other [30:58] CEOs and other management teams, they're [31:01] getting in on this. They're starting to [31:02] realize that buying back your shares, [31:04] giving dividends is not as important as [31:07] betting on the future. Door Dash got [31:09] taken to the woodshed because they're [31:11] investing too much in capex. Obviously, [31:13] Google got smacked with their capex [31:14] spend. So, that keeps happening over and [31:16] over again. And just on the headwinds [31:18] that SpaceX is going to face, the [31:21] arguments that I think will turn out to [31:24] be wrong, but they're valid to talk [31:25] about here are, hey, is this demand for [31:28] tokens and compute going to keep up? Or [31:30] does onrem and desktops and open source [31:33] models getting smaller, better, does [31:35] that actually mute at some point demand? [31:37] I don't think it does. I don't know [31:39] there's an upper uh I don't know there's [31:41] an upper bound for on demand [31:43] intelligence. The second one obviously [31:44] is Starlink is for people who are in a [31:47] rural neighborhood. If you've got [31:49] Verizon fiber to your building or [31:51] Spectrum, you're not putting nor can you [31:54] put a Starlink on your building. So, the [31:56] piece there that's going to be um uh [32:00] explained probably in the next year or [32:01] two is every single Tesla sold is going [32:04] to have Starlink in it when they get [32:05] that merger done. What that means is [32:07] you're going to have Wi-Fi networks uh [32:10] connecting any phone to any Tesla, say [32:13] all those robo taxis out there. You'll [32:16] be able to connect also directly with [32:18] the next generation of Starlink. So your [32:19] phone will be able to direct if it's got [32:21] clear line of sight. It's going to be [32:22] able to connect to any Tesla on the [32:24] road, which there are many that all [32:26] future ones will have a Starlink built [32:28] into them. So those are super promising. [32:30] And then finally, you know, there's been [32:32] a lot of speculation about the [32:33] valuation. Brad, you brought it up. I [32:35] think at liquidity when people were [32:37] asking you and I heard you talk about it [32:38] hey [32:40] private companies venture capital we are [32:43] a voting mechanism and then when it goes [32:45] public it becomes a weighing mechanism [32:47] and sometimes you'll have this moment in [32:50] time uh where there's hand ringing about [32:52] those valuations and the hand ringing [32:54] peaked in the last quarter you had 160 [32:58] times uh price to sales ratio for Tesla [33:01] when it first came out 160 times right [33:04] you take their this 2 or3 trillion [33:06] market cap and you put it against a [33:09] smaller revenue number. Well, if you [33:10] look at the revenue number increasing, [33:12] now we're down to a 45 times price to [33:14] sales ratio. So, some kind of uh balance [33:18] is occurring here. Yeah, Brad, between [33:20] these private and public markets as well [33:22] as the increase in revenue. [33:24] >> Yeah, I I I mean, honestly, I think this [33:26] is all super healthy. I think the SpaceX [33:29] IPO was extraordinary. I think the [33:31] consolidation here is perfectly [33:32] predictable. And now you have a company [33:34] at 1.4 trillion that I think if you take [33:36] a three or a fouryear view, you can see [33:39] yourself tripling your money in this [33:40] business at a very reasonable valuation [33:43] on the Morgan Stanley numbers or on the [33:45] Elon numbers or whatever. But that's [33:48] always been the bet. Do you believe that [33:49] Elon is the greatest innovator and a [33:51] great allocator of capital? But the [33:53] price of entry matters, right? When you [33:55] get carried away on day one of an IPO [33:57] and you buy this thing over two [33:59] trillion, you got to know that this is [34:01] going to happen. And I was on CNBC the [34:02] day of the IPO and I said I would want [34:04] to own this company but I'm not sure [34:06] today is the day I would buy the [34:07] company. Right? And so I you know [34:10] >> entry price matters. I mean fundamental [34:13] >> but but let me give you another one. You [34:15] know like we've talked about the [34:16] anthropic IPO uh or a lot of people have [34:18] talked about it later this year. I hear [34:19] a lot of people saying 1.5 or$2 [34:21] trillion. David just talked earlier that [34:24] it's going to be run rating over a [34:25] hundred billion maybe by the end of the [34:27] year. That's like 10 to 15 times [34:29] revenue. That is not that much for a [34:32] company that just grew 10x and is [34:34] rumored to be profitable in Q2. And so I [34:37] look at the market, the consolidation we [34:39] saw in the month of July, you know, we [34:41] put in the Leopold bottom hopefully in [34:43] July that, you know, a lot of semi [34:45] stocks were down, you know, 30. Hey, [34:48] hey, listen. The guy's doing great. He's [34:52] apparently still up 80% for the year. [34:54] Just made another big private [34:55] investment. I I I I think he's done an [34:57] extraordinarily good job building a firm [34:59] in a short period of time, but the [35:00] market did panic around that. Yeah. Um [35:03] as as as he had to cover, I think all of [35:05] that is really good. So, as I look ahead [35:07] marching to these IPOs later in the year [35:10] on the back of the SpaceX IPO, I think [35:13] we're in in in really good shape. Um you [35:15] know, particularly if these revenues [35:17] continue a pace. You know what Elon's [35:19] really good at is just building stuff [35:22] like factory specific physical physical [35:25] sites. That is such a core advantage in [35:27] this world where everyone's competing [35:29] for data centers and fabs. The software [35:32] layer needs hardware in the physical [35:35] world in order to deliver their software [35:36] services. And there is no one better [35:38] than Elon at actually doing that. Look [35:40] at how gigafactories have been stood up [35:42] around the world. This is his core [35:45] competency. So Brad, like when you put [35:46] Elon up against Adaio and a SAM and even [35:50] an Alphabet which has 27 years of doing [35:53] this, I I mean man, Elon's got a core [35:56] advantage if this is what this world [35:57] comes down to. [35:58] >> He says something like that on the call [36:00] where he said, "Look, putting up data [36:02] centers is nothing compared to the [36:05] difficulty of putting up a rocket, [36:07] right? It's like, you know, creating [36:08] data centers is not rocket science." So [36:11] they take some of those hardware [36:12] expertise that they have from SpaceX and [36:14] they put them into data centers and [36:15] that's why they've been able to stand [36:16] up, you know, more data centers or or [36:18] bigger data centers faster than all the [36:21] competitors. A couple points there. I is [36:23] it clear why Starship is so important to [36:26] Starlink? Okay, let me just explain this [36:28] quickly. So basically SpaceX has [36:31] developed a new V3 satellite that has [36:34] 10x the bandwidth of its V2 satellite. [36:37] So currently the Starlink network is [36:40] powered by V2 satellites. They deploy [36:44] them on the Falcon 9 rocket and they [36:47] launch about 27 satellites per launch [36:50] and that adds about 2.6 terabits per [36:53] second of total network capacity. [36:56] Starship deploys 60 of these V3 [37:00] satellites per launch. That would add 60 [37:03] terabytes per second of total network [37:05] capacity per launch. So over 20 times [37:08] more capacity per launch. That's the [37:12] power of it. So if they get Starship [37:15] working and by the way the last test, [37:16] not only did it prove that the heat [37:18] shield worked, my understanding is they [37:20] actually launched or rather they [37:22] deployed 20 V3 satellites as a test and [37:27] they were able to make connection with [37:28] those satellites and prove that it [37:29] worked. They even had cameras on them. [37:32] The reason we were able to see the [37:34] Starship was because they're like, [37:36] "YOLO, let's put some cameras, HD [37:38] cameras on them." [37:39] >> Right now, I think those satellites [37:40] basically it was just a test and they [37:41] they burn up. [37:42] >> They they burned up. So, I think the [37:44] next big milestone here will be when [37:46] they launch Starship with, let's say, 60 [37:49] of these V3 satellites, put them in the [37:51] correct orbit, make connection with [37:52] them, add the bandwidth to the network. [37:54] That's going to be a big milestone. But [37:57] you play this out to its logical [37:58] conclusion and the bandwidth available [38:01] to the Starlink network goes up 10x or [38:04] eventually 100x times and that's when [38:06] they can do all the interesting things [38:08] like direct to cellular. There were some [38:10] interesting hints that Gwen Shaw talked [38:12] about about with ground stations about [38:15] what they could potentially do there and [38:16] I think [38:17] >> and they might buy T-Mobile or something [38:19] like that sack. it's easily within their [38:21] range of purchases [38:22] >> and and I think Elon mentioned something [38:24] about potentially the the Starlink [38:26] network could eventually handle roughly [38:29] half of internet traffic. So I mean this [38:32] this thing could get so much bigger than [38:34] just 12 million subscribers to your [38:36] point Freeberg. But look I I want to [38:38] actually talk about the data centers for [38:40] a second Brad. I do have a couple of [38:41] questions about this. So Elon mentioned [38:44] that okay we're going to be at 2 gawatt [38:46] by the end of the year. He said that we [38:48] will be at five to 10 next year. Closer [38:50] to 10 than five. So let's just say [38:53] eight. Okay. So I'm just making that up, [38:55] but it's in their range. [38:56] >> So let's just say that's an ad of 6 [38:58] gawatt. So they go from two to eight. [39:02] Okay. To me, there's two questions [39:03] there. One is how do you know that the [39:05] spot price is going to stay where it is? [39:08] You know, can it stay at $50 per watt? [39:12] How do we know? How do we track that? [39:14] How much risk is there around that? I [39:16] got the sense on the call that Elon [39:18] thinks that number is going up because [39:20] the market is memory constrained right [39:22] now. I think he mentioned that we might [39:25] see a 20% increase in memory production [39:27] next year but the demand is going up [39:29] 200% plus. So the market is constrained [39:32] by whatever the bottleneck is at that [39:34] time. Right now the bottleneck is [39:36] memory. So where do you see the spot [39:38] price going? How do we know how much [39:40] risk is there around that? And then the [39:42] other question I would have is if you go [39:44] from two to eight gigawatts that you [39:47] have net is six. We know that a gigawatt [39:51] power data center is you know 50 billion [39:54] of capex. [39:55] >> So six incremental gigawatts of compute [39:58] would be 300 billion of capex next year [40:01] assuming they build that right. I mean [40:03] they have optionality around that I'm [40:04] sure. So how do you finance that? You [40:07] know what's the most non-dilutive way? [40:09] They said their payback is a year or [40:11] less. I'm sure that's tied to the spot [40:13] price. So, you only have to finance it [40:15] for a year. And the question, do you [40:17] think Nvidia gives them that financing [40:19] or how will this play out? I guess is my [40:21] question. [40:22] >> It's a great framing, David. First, it's [40:24] $50 billion per gigawatt to build [40:26] minimum. Okay. So, you're $300 billion. [40:30] So, in order to finance that, it seems [40:32] to me you either have to go into the [40:34] market and borrow the money or you have [40:36] to do a dilutive equity raise. neither [40:38] of which they want to do. Um or you get [40:42] Nvidia to backs stop it. Um which [40:44] they've indicated that they're going to [40:46] do more of. But the problem there is [40:48] Nvidia shareholders don't want them back [40:50] stopping unlimited because the fear in [40:51] the world is that that spot price at [40:54] some point right may go against you and [40:57] when it does the payback period changes. [40:59] Nobody thinks that the payback period is [41:01] going to be one year even though the [41:02] spot price is suggesting that it is that [41:05] today. Right? Just a few years ago, [41:07] people thought you would get paid or not [41:09] few years ago, a few months ago, people [41:11] thought you'd get payback over four [41:12] years. So, you basically spend 50, you [41:15] then earn 10 to 15 per year. You get [41:19] payback over four to five years and then [41:21] hopefully you get the sixth year, which [41:23] really takes you up well above 20% in [41:26] terms of your returns. Um, right now the [41:29] shortage is so acute and the willingness [41:32] to pay from the front Frontier Labs is [41:35] so high because they all recognize [41:37] they're on the verge of some massive [41:39] breakthroughs that they're willing to [41:41] pay three, four, 5x market pricing in [41:45] order to get at scale compute. And [41:46] that's what happened with the Anthropic [41:48] deal uh with SpaceX. I think Anthropic [41:50] would buy a lot more of that today if [41:52] they could. Same with OpenAI. You're [41:54] saying Brad they would be willing to [41:56] overpay by a factor of up to 5x. [41:59] >> Well, that's the 50 that you know that's [42:01] the $50 per watt that David was [42:03] referencing. They would be willing to [42:05] pay this 30 to 50 if they could get [42:07] atscale compute that would give them a [42:08] competitive advantage over the other [42:11] people in the market. And remember, [42:13] there aren't a lot of people who have [42:15] the offtake revenue that can afford to [42:17] buy compute at this scale, right? It [42:19] wasn't the Chinese open source companies [42:21] that were buying, you know, SpaceX's [42:23] excess compute or building the 10 [42:25] gawatt, you know, plant in in Ohio. [42:28] That's open AI and anthropic. So, the [42:29] vast majority of the offtake commitments [42:32] are coming from Anthropic, Open AAI, and [42:34] NVIDIA, right? When you hear about the [42:36] hyperscalers building all of this, you [42:39] know, this compute out, they're building [42:40] it out to sell to the people that we uh, [42:43] you know, that we just mentioned. So [42:44] David, Netnet, if he builds 6 gigawatts [42:48] next year, and by the way, probably only [42:50] Elon um, you know, can actually stand up [42:53] that much in that time frame. Like [42:55] Jensen said to me on the pod, he's like, [42:57] "Nobody comes close. Microsoft doesn't [42:59] come close. You know, Google doesn't [43:01] come close in terms of standing it up in [43:03] that time frame." Um, I think that he's [43:06] going to have a challenge, you know, [43:07] getting all of the componentry, right? I [43:10] know he can stand it up, but can he get [43:11] the memory? Can he get the chips? Can he [43:13] get the land powered shell all in time? [43:15] I think the offtake is there, right? But [43:19] to put it in perspective, this year [43:21] Anthropic and Open AAI combined, their [43:23] starting total compute was like 5 GW. [43:27] So he's talking about incrementally [43:29] adding more than they had as combined [43:31] companies, right? [43:32] >> That's not that much of an increase when [43:34] Anthropic is growing 10x year-over-year [43:37] and OpenAI is maybe at what 4x or maybe [43:39] higher now. So [43:41] >> the demand exists in the world. The [43:42] demand exists in the world today. I [43:44] think it will exist in the world for [43:46] well, you know, the next 12 to 24 [43:48] months. But there is a wall of worry in [43:50] the market. The reason we saw the [43:52] pullback in July is Kimmy scared people [43:55] into thinking, oh my gosh, they're going [43:57] to undercut the Frontier's revenues. And [43:59] if they undercut the Frontier Labs [44:01] revenues, who the hell is going to pay [44:02] for all this compute? That's why you saw [44:05] a 40% trade down in the core weaves of [44:08] the world and you know the the all of [44:11] the the semiconductor stocks and [44:13] semiconductor related AI [44:14] >> in some ways the fact that there's a [44:16] discussion going on that this next 10 [44:18] gigawatts is going to cost you know [44:20] Sachs$500 billion dollars and you asked [44:23] the question where does that come from a [44:24] secondary offering does Nvidia put it on [44:26] their books do they create SPVS off [44:28] their books like some people are doing [44:30] you know the fact that we're having this [44:32] conversation everybody's is aware of it. [44:34] The market has been educated on it means [44:36] I think people will be able to change in [44:39] real time if it doesn't come to pass or [44:42] if it slows down which I suspect this [44:45] cannot keep up at this pace you know [44:48] more than another two years or so. So a [44:51] as our good friend Bill Gurley likes to [44:52] remind us, he's like, I can't believe [44:55] >> that we're all just taking in stride [44:57] this level of seller financing, right? [45:00] He would call it circular revenues, [45:02] right? But the market has gotten [45:04] comfortable with this. And remember, [45:06] like we saw in July, if there is a scare [45:08] about demand, the whole sector trades [45:11] down. [45:12] >> Yeah. [45:12] >> Everything will trade down, you know, [45:14] together because that's just the [45:16] leverage that you're pumping into the [45:18] system. you're effectively backstopping [45:20] people's ability to build ahead of their [45:22] revenue. So, it becomes much more [45:24] violent if you ever see demand slippage. [45:27] Um, you know, famous last words, I don't [45:30] see it today over the course of the next [45:31] 12 to 18 months. Um, but you know, you [45:33] have these unknown unknown moments that [45:36] certainly causes people to be fearful. [45:38] >> Credit spreads are blowing, you know, [45:40] have continued to to stay wide on these [45:42] deals. So, there is fear in the market [45:44] about them. [45:44] >> All right, everybody. The fifth annual [45:47] if it's September, you know it's time [45:49] for the All-In Summit. The fifth annual [45:51] is happening. Yes, that's right. Uh [45:53] David Freeberg's been at work and we [45:56] have an allstar allstar [45:59] list of people joining us. Jensen Wang, [46:01] founder and CEO of Nvidia. If you care [46:03] about where AI is headed, you won't want [46:05] to miss this conversation. [46:07] >> The best the Oracle [46:08] >> Satia Nadella, CEO of Microsoft, fan of [46:10] the pod will be coming on for the second [46:12] time. Jared Isaacman from NASA, the one, [46:14] the only Brad Gersonner and Bill Gurley, [46:16] BG2 coming back. SpaceX's Gwen Shotwell, [46:19] my guy Jake Paul, Nick Shirley. Lot of [46:23] incredible people coming. Martin Shreley [46:25] maybe is even coming. He's that's going [46:27] to be fun. Go to the allinssummit.com to [46:30] apply today. Allin.com or [46:32] theallinssummit.com. [46:33] Any of those will get you there. And [46:36] we're taking over Universal Studios [46:39] again. We'll have our own private [46:41] playground. Dave Friedberg, great job on [46:43] the summit. Casino night, too. I heard [46:45] it's going to be a big casino night. [46:47] >> Biggest yet. And the concert to be [46:50] announced who will be performing at the [46:51] concert, but it is going to be [46:52] incredible. So, I'll just say one of the [46:54] things about the summit, we've had [46:56] people come to the summit from over 60 [46:58] countries. It's really incredible to [47:00] meet all these people, entrepreneurs, [47:01] investors, people that are just really [47:03] interested in the topics that we talk [47:05] about. We try and have the world's most [47:06] important conversations, but it's really [47:08] this amazing community experience. [47:10] That's what brings folks back. So, we [47:12] try and invest more and more every year [47:15] in making it an amazing experience, not [47:17] just cool content on a stage, which I [47:19] think is what a lot of these other shows [47:21] really deliver, but it's like how do you [47:22] actually come and have a have an [47:24] experience for a couple days. It's going [47:25] to be awesome. So, [47:26] >> it really is those three things that we [47:28] focus on. One, you're going to learn [47:29] something, right? You got these great [47:31] people on stage. You're going to learn [47:32] something from them. You're going to [47:33] meet new people. You're going to [47:34] network. And then you're going to have [47:35] these great experiences. It's the [47:37] trifecta. Folks, you're excited, Brad? [47:39] You excited to be back? What? What are [47:40] the dates? What are the dates again? [47:42] >> Look at your calendar. You're speeding. [47:45] >> September 13th through 15th in LA. This [47:48] couldn't be better dates for the summit. [47:50] I mean, we're we're going to be within [47:52] 60 days of an election, midterm [47:54] election. We're going to be within 30 [47:56] days of an IPO, you know, potentially of [47:58] anthropic. I mean, like, it it's going [48:00] to be heated. The SAS apocalypse, not [48:03] the Saxp apocalypse. This is the SAS [48:05] apocalypse is, I guess, winding its way [48:08] out. Uh, the indigestion might be [48:10] clearing. Air Table just got acquired [48:12] for less than it raised. It's a [48:14] profitable SAS company, a great product, [48:16] $480 million, half a billion dollars in [48:20] annual revenue, growing 20% a year, [48:22] respectable if it was a public company [48:24] with almost a billion dollar in cash has [48:27] been sold. It's been sold for $1.28 [48:30] billion, about 10% of its peak [48:32] valuation, which was 11.7 billion in [48:34] 2021. Now, they did have a bunch of [48:36] cash. showed me include the cash [48:37] position sale was 2.25 billion. They [48:40] were acquired by a firm called bending [48:42] spoons. This is an Italian company, [48:44] Milanbased company. They buy challenged [48:48] but you know interesting businesses. [48:50] AOL's legacy business, Evernote, [48:52] Eventbrite, Vimeo, Meetup.com and they [48:55] just went public last month. Shares uh [48:57] that is uh bending spoons went public [48:59] last month. Shares sh 15% on the air [49:02] table news sacks. When we look at this, [49:05] this was a company that had done a lot [49:07] of things right, had a massive amount of [49:09] cash in their war chest, but rumors were [49:13] maybe the founders were a little [49:15] exhausted. Maybe some of the investors [49:17] were exhausted who bought in at a high [49:19] level. What can we take away from this [49:21] transaction in bending spoons? Are they [49:23] the buyer of last resort now? [49:26] >> Well, I think they're creating a great [49:27] business for themselves because I think [49:29] this will end up being a fairly [49:30] profitable acquisition for them. Let me [49:32] just add a piece to this which is Air [49:34] Table spun out its AI agent business uh [49:38] which is known as hyper agent into a [49:40] separate independent company prior to [49:42] this acquisition. So I think what's [49:44] going on here is that the founders and [49:47] talent of the company they said look we [49:49] don't want to have to make this legacy [49:51] product work that's basically a private [49:52] equity play. I'll explain what that [49:54] means in a second. We want to focus on [49:56] the new thing, the AI company. That's [49:58] where the big value creation is going to [49:59] be in the future or the potential for [50:01] it. So essentially, the talent is going [50:03] to focus on the venture play and then [50:06] they're selling the private equity play [50:08] to Bending Spoons. Now, why do I think [50:10] this could be a good acquisition for [50:12] Bending Spoons. I think there was a [50:13] really interesting data point that I saw [50:15] in the commentary on this, which is only [50:17] 30% [50:19] of Air Table sales team was making [50:22] quota. They had a 30% sales attainment [50:26] number and that told me a lot about this [50:28] business. Okay, what it told me is, and [50:31] I'm reading between the lines here, but [50:34] this was a company that had a successful [50:36] PLG motion, in other words, organic [50:39] growth, productled growth, and they were [50:41] growing about 20% a year. But that was [50:44] not good enough for its board. You know, [50:47] these are investors, some of whom [50:49] invested at an 11 billion peak [50:52] valuation. So, they're looking for a [50:54] venture type outcome. So, what happens? [50:56] The board pressures the founders to do [50:58] something that frankly is unnatural for [51:01] them, which is they say, "Look, you [51:03] should bolt on a traditional salesled [51:06] motion here to get the growth up [51:08] faster." Does that work? No. They [51:11] probably get a little bit of growth out [51:12] of it, but they only get 30% attainment. [51:14] So they've got hundreds and hundreds of [51:15] sales reps here trying to push on a [51:18] string and it's not making it grow [51:20] faster. So now what's the opportunity [51:22] for the acquirer here? Bending spoons [51:24] can go in here and do what Elon did at [51:26] Twitter. Eliminate 85 90% of the cost [51:30] structure. Don't do this salesled [51:32] motion. Just go back to your productled [51:35] growth roots. You'll probably keep most [51:38] of that 20% growth and it'll be a very [51:41] profitable company. you'll be able to [51:43] >> 80% profitable probably right [51:45] >> probably. I mean [51:46] >> 400 million to the bottom line pays for [51:47] the acquisition in a couple years. [51:49] >> People are saying they're only going to [51:50] generate 30% ebidom margin. I think like [51:52] you're saying it could be 80 90%. I [51:55] don't think you need to keep most of [51:57] this business or most of the cost [51:59] structure associated with this business. [52:01] Um Air Table is a company that has its [52:03] fans. Um I think they will probably [52:06] stick with it and you know you'll you'll [52:08] be generating I don't know you could [52:09] probably generate 300 million of Ibida a [52:11] year or 400 million uh while growing you [52:14] know 10 to 20%. So that's a play for [52:17] bending spoons [52:18] >> and the venture investors here Sachs [52:20] they're happy to get their money back [52:22] and move on to the next thing. It's a [52:23] bit of a push for them, you know, in [52:25] terms of at the blackjack table rather [52:28] than they've got to go 10x just to catch [52:30] up and then they would have to go 10x [52:32] again to make their LPs happy. It's not [52:34] going to happen. [52:35] >> I think the question is if Bending [52:37] Spoons can basically take this business [52:39] that's not making money and probably [52:41] generate 400 million a year of EBA and [52:44] pay for the acquisition in just three [52:46] years. [52:47] >> Amazing. [52:48] >> Why isn't that something that the [52:49] company could do on its own? And I think [52:51] that's the structural problem is I think [52:54] it's very hard for both VCs who are on [52:58] the board and the founders to shift into [53:00] private equity mode. Why? Because [53:02] they're going to have to demolition what [53:04] they've built, right? They've got all [53:06] this loyalty to the team. They don't [53:08] want to think about how do I eliminate [53:09] 80 90% of the cost structure. It's just [53:12] not what they do. I mean, what what [53:14] founders want to do and and the outcome [53:16] that the board members are going for is [53:18] a venturebacked outcome. And I think [53:21] they could have done this. They could do [53:22] what Bending Spoons does, but [53:23] >> they're not built for it, Sax. [53:24] >> They're not built for it. And moreover, [53:26] the structure of the cap table is all [53:28] wrong because they're sitting behind [53:29] this giant liquidation preference. All [53:32] these investors who have to get paid [53:34] back who invested at this 11 billion [53:36] valuation and and you know, all the way [53:38] up [53:39] >> the the incentives are broken, Brad. And [53:42] you you yourself at your firm, Alimter, [53:45] you were pretty frisky in this period. [53:46] You made a lot of bets. So uh I don't [53:49] know if air table was one of them uh but [53:51] you made some sass bets there. Some of [53:53] them were at high valuations. How are [53:54] you looking back at that time period? [53:56] Any lessons that you take going forward? [53:59] Multiples of revenue can compress very [54:01] quickly. Right? It works great when the [54:04] company's growing greater than 50%. But [54:06] remember it's just a heristic. It's just [54:08] a very rough estimate used almost [54:10] exclusively in Silicon Valley you know. [54:12] So people are saying oh my god this [54:14] thing sold for two times revenue. But [54:16] when you actually look at it on a look [54:18] through basis, probably sold for maybe [54:20] 30 times free cash flow. I don't think [54:22] it's easy to get it to 400 million in [54:23] ibita. I think if it was, the board [54:25] would have done that. I'm on, you know, [54:27] we're involved in some of these [54:28] companies. Once they slow down, the [54:30] company morale goes to hell. Turnover [54:33] among your customers, uh, you know, [54:35] begins to spike. Um, it starts to feed [54:38] on itself. So I sucks to go to work [54:40] every day. What do you need to keep? [54:42] What do you need to think about? Very [54:44] tricky. I don't know the core product [54:45] and what's happening in terms of [54:46] turnover in the core product, David, but [54:49] my hunch is that the core product has [54:52] started uh to really fizzle as the [54:55] advances in the core product has slowed [54:57] down. You're seeing a bunch of churn out [54:59] of it on the product side and now people [55:01] are saying listen it's almost impossible [55:04] for a software company today to keep any [55:06] decent sales people to keep any [55:08] different decent product development [55:10] people because they all want to go work [55:11] on AI. [55:12] >> Agreed. But you don't need them for this [55:13] product. [55:14] >> I mean the market the market's being [55:16] efficient. I mean look this is where I [55:17] think bending spoons has an advantage [55:19] that the company's board and founders [55:22] wouldn't have which is they already have [55:23] an infrastructure right they have a core [55:26] team at bending spoons that's managing [55:28] now I don't know dozens of these [55:30] properties and so they can plug this in. [55:33] I think AI in a way makes their job [55:35] easier because in the past the reason [55:36] why you couldn't eliminate like all of [55:39] the talent the infrastructure is because [55:41] you needed the institutional memory. You [55:43] needed people who knew the codebase. Now [55:45] AI can learn the codebase instantly. [55:48] >> That's an interesting insight. [55:49] >> And so yeah [55:50] >> maintaining is easier with AI. [55:51] >> I think maintenance mode becomes way [55:53] easier with AI because you don't need [55:54] the historical knowledge anymore. The AI [55:56] can go in and sort of reconstitute that [55:59] that historical knowledge. Let me get [56:02] you in here freeird uh if I may. when we [56:05] look at the lessons from peak zerp and [56:07] SAS and then we look at you know this [56:10] moment in time this surging AI market [56:13] any parallels that we might find here uh [56:16] or lessons uh between the two [56:18] >> between Zerp and AI era [56:21] >> the ZER SAS era we had a lot of very [56:24] high valuations a lot of enthusiasm a [56:27] lot of suspending disbelief we're here [56:29] in the AI era we just talked about you [56:31] know the price of compute and all these [56:33] companies being at a 100x uh price to [56:36] sales ratio. Any parallels here or not? [56:38] It's a kind of a softball question for [56:40] you. [56:40] >> No, this is a very different paradigm. [56:42] Uh the AI capex buildout and model [56:45] training which is where the predominance [56:48] of the capital is flowing is not about [56:50] some high multiple on revenue which is [56:54] where capital was flowing into SAS. It's [56:56] like oh you get a 20x multiple turn a [56:58] dollar into 20 that's great let's do it [57:00] all day long. this is a very different [57:02] structure and strategy and capital um [57:05] allocation process. So I don't think [57:08] that I I would look at them as being [57:10] linked. [57:10] >> It was a softball question to be honest. [57:12] I was letting you hit it out of the [57:14] park. [57:14] >> Look, I mean obviously SAS companies [57:16] were overvalued during the Zer era for [57:19] two reasons. One is that we had [57:20] artificially low interest rates. So we [57:22] had a kind of a a speculative asset [57:25] super bubble. But the other is that [57:27] people were treating these things like [57:29] guaranteed annuities. and actually [57:31] growing annuities, they'd look at it and [57:33] see, oh, 120% net dollar retention, so [57:36] this thing will just grow 20% [57:37] year-over-year forever as a base case, [57:39] right? And they were then priced that [57:41] way. But what we've seen with AI is [57:43] obviously [57:45] there's disruption and you can't to Brad [57:48] said, I'm sure they're seeing elevated [57:49] churn right now and it's not an annuity. [57:52] Things can change. So obviously now [57:54] these things are trading at a much [57:55] greater discount. All of that being [57:57] said, let me just say I don't think you [57:59] can extrapolate to the entire SAS space [58:02] based on this one company, Air Table. I [58:04] think there's some things about Air [58:06] Table that make it very different than, [58:08] I don't know, let's say a Salesforce or [58:10] a Workday is, you know, Air Table was [58:13] always a little bit of a quirky product. [58:15] I remember at the peak hype for this [58:18] company, people were saying like, oh, [58:19] this is like a new Excel or a new Google [58:22] Sheets. It's [58:23] >> new Microsoft Office. Yeah. [58:24] >> Yeah. It was basically a spreadsheet for [58:26] words. That's how people were were [58:29] viewing it as this like this new kind of [58:30] spreadsheet for for words as opposed to [58:33] numbers. And it never achieved that kind [58:35] of promise. It never achieved that kind [58:37] of ubiquity. People understand how to [58:39] use spreadsheets. Everyone uses them. [58:42] Air Table never got to that point. Most [58:44] people still don't know what Air Table [58:45] is. Again, it had its dedicated fans, [58:48] but it was a hard product to explain to [58:50] people. When do you use it? [58:52] >> It had a cult following. It had a cult [58:54] following but but it never it never [58:55] achieved that sort of level of [58:57] acceptance. It was never [58:58] self-explanatory in terms of why you [59:00] should use it, what the use cases are, [59:02] they never were able to kind of get the [59:04] marketing right because of that. [59:06] >> And to be honest, if you look at claude [59:08] co-work [59:10] agents, those things are now doing what [59:12] air tableable did. So [59:13] >> it never carved out, I think, a niche [59:15] where it was super clear when you were [59:18] always supposed to use error table and [59:19] and really it was part of this [59:21] hodgepodge of of this grab bag you [59:23] should you could say of no code tools. [59:26] This is the category it was put in and [59:28] no code has to be the most impacted the [59:31] most disrupted area of SAS right now [59:33] because I mean what is clawed code [59:35] really good at? I mean that's the [59:36] ultimate no lovable claw code [59:39] perplexing. The thing with yeah the [59:42] thing with air tableable or retool [59:44] things like this is it's true you didn't [59:45] need to be a coder to use them but you [59:47] had to learn how to use air tableable [59:49] you had to learn how to use retool all [59:51] these it was kind of these you know [59:53] alternative programming languages in a [59:55] way and you just don't need to learn any [59:57] of that anymore I mean you use claude [59:59] and you just tell it what you want it to [60:01] create and so you know if you do want to [60:03] create a some sort of new dashboard some [60:06] sort of I don't know like a verbal [60:08] spreadsheet or whatever you just tell [60:09] Claude what you want. You don't have [60:11] this learning curve. Look, all of SAS is [60:14] being impacted right now, but this has [60:15] got to be the most impacted area. So, I [60:18] don't know that you can totally [60:18] extrapolate based on what's happening to [60:20] Air Table. I don't necessarily think [60:22] that you want to replace your CRM, your [60:24] ERP, your HR system with something [60:27] that's been vioded. You want the [60:29] certainty, you know, for anything that [60:31] involves compliance. I got I got to be [60:32] honest. My team Sachs made I don't do [60:36] you use like a portfolio uh offthe-shelf [60:39] SAS tool for managing crafts like um uh [60:43] portfolios and everything. [60:44] >> Well, we we vibe coded something [60:46] actually. [60:47] >> So yeah, we just did the same too. So my [60:49] team just built something that is so [60:51] mind-blowing that to buy it with [60:53] off-the-shelf software would have been a [60:55] quart million dollars in software and [60:57] like a million dollars in integration [60:58] over two or three years. and we built it [61:00] in a month and and now we have complete [61:02] insight into the whole portfolio, the [61:03] competitive set, the founders, [61:05] everything going on. [61:06] >> Keep in mind that one of the reasons why [61:08] Leopold got blown out, okay? I mean, is [61:11] because he bet on the SAS apocalypse. [61:14] Remember, it wasn't just that he was [61:16] super long these chip stocks that had a [61:18] correction. [61:18] >> Oh, is that right? He was short. [61:19] >> He was super he was short Adobe and a [61:22] whole bunch of other SAS companies. And [61:24] those trades also moved the wrong way on [61:27] him. So again, I just think that it's [61:29] painting with too broad a brush to say [61:31] that all of SAS is going to get [61:32] obliterated here. [61:34] >> Yeah. [61:34] >> And there was a really good post about [61:36] this. Let me just quote from this where [61:37] they said, "Nobody buys Microsoft [61:40] because Microsoft writes the best code. [61:41] They buy Microsoft because Microsoft is [61:43] the rail that everything else runs on. [61:45] Active Directory is where your employee [61:47] identities live. Excel is where your [61:49] boardex numbers come from. Teams is [61:51] where the compliance recorded [61:52] conversation happens. Azure holds a Fed [61:55] ramp high authorization and Department [61:57] of Defense impact level 5 clearance, [62:00] which means a defense contractor cannot [62:02] casually swap it out for something [62:03] cheaper and so on down the line. So, [62:06] there's a lot of really good compliance [62:07] reasons why if you're a large [62:09] enterprise, you're not going to want to [62:10] spend tens of millions of dollars [62:12] ripping out something that costs you a [62:14] million dollars a year. It just that [62:16] just doesn't make sense. And I noticed [62:17] that Ben off just tweeted 5 minutes ago [62:21] that 15 out of 15 cabinet agencies run [62:24] on Salesforce. Look, the government is [62:26] not going to rip and replace Sal with [62:28] somebody vibe coded. So look, not all [62:30] assass is equal in this dimension. [62:32] >> I just want some uh Figma. I just think [62:34] some of these SAS companies with great [62:35] founders who are in it for the long term [62:37] and they have like passionate user [62:39] bases. I think they will make the jump [62:41] to AI first products and I I put Figma [62:43] in that bucket. Just to wrap this this [62:45] section, please. IGV is up 20% in the [62:48] last six months. It's up 20% in the last [62:50] five years. [62:51] >> Explain IGV, please. [62:52] >> So, the high growth software stock [62:55] index, [62:56] >> right? Snowflakes up 88% in the last 6 [62:59] months. [63:00] >> That's it's an IGV is an ETF of [63:02] >> IGV is an ETF of of growth software [63:05] companies. Right. So, to to David's [63:07] point, there was a panic about software [63:09] companies. There was a big trade out. [63:11] you know, honestly, they they performed [63:13] pretty well and and as he mentioned in [63:16] the month of July, they were up when a [63:17] lot of the semiconductor AI stocks were [63:20] down. And some of these companies, data [63:22] bricks, snowflake, click house, etc. are [63:24] doing extraordinarily well. As I just [63:26] mentioned, Snowflake's up 90% in the [63:28] last 6 months, which puts it in the same [63:30] category as the semiconductor AI stocks. [63:32] So, to David's point, you can't throw [63:34] them all in the same bucket. But I do [63:36] think that for these no code a lot of [63:37] these application software companies [63:39] they're realizing like that that you [63:42] know the game is up sell the company get [63:44] what you can get you know importantly [63:46] here in the air table story all the [63:49] latestage investors right we passed on [63:51] this in the last three funding rounds [63:53] right which I think we're at 2 billion 5 [63:55] billion and 11 billion but all those [63:56] latestage investors which were the most [63:58] venerable of growth firms they all got [64:00] their money back and the early stage [64:02] investors ended up making a lot so if [64:04] this is a failure This is a pretty good [64:06] failure for Silicon Valley. [64:07] >> This is one of the points that was made [64:08] at that time which is hey we this is a [64:12] strong enough company and team and [64:14] revenue base that if we just get our [64:15] money back with the optionality hey [64:17] maybe this would be a good investment. [64:19] You could say the same thing about some [64:20] AI bets. [64:21] >> Well this is one of those cases where [64:23] the liquidation preference actually [64:24] mattered. You know normally it doesn't [64:26] matter but [64:28] >> straight money here. I my understanding [64:30] is this wasn't like they had like a 7% [64:32] you know interest rate or they didn't [64:34] have like a participating preferred [64:35] where you get two times your money back [64:37] and then they do the trade. Does anybody [64:39] know? Because I looked deeply into this. [64:41] >> I was just saying I think that net of [64:43] cash they may have come in a little bit [64:45] less than the total cash raised but it [64:46] seemed like everybody got made whole. [64:48] >> Yeah. But if they had the I guess sachs [64:51] they we live through moments in time [64:53] where companies had to guarantee a 1x [64:55] right you know [64:57] >> 1x liquidation preference is standard it [64:58] just means you get your money back [65:00] before other people start to profit [65:03] which is appropriate [65:04] >> but the interest rates were taken out [65:07] right of these deals uh I think during [65:10] the standard terms you know what's known [65:12] as clean terms is just a simple 1x [65:14] liquidation preference the preferred [65:16] just gets their money back before the [65:18] common starts to participate in a [65:21] successful sale of the company. That [65:22] just makes sense, right? [65:24] >> Participating preferred is the double [65:26] dip, right? [65:27] >> Yeah. And look, we've never done that. [65:29] You know, we believe in clean terms. No [65:31] one's trying to be punitive towards [65:32] founders. It's just it doesn't make [65:34] sense for some people on the cap table [65:37] to be making money while other people [65:39] are losing money. It just doesn't make [65:41] sense, right? [65:43] Well, that that that's just a transfer [65:45] of value from some people on the cap [65:47] table to other people on the cap table. [65:49] So, the standard thing you do is you [65:51] make sure that the investors get paid [65:53] back and then everybody is participating [65:55] in the upside. [65:56] >> Okay. Fourth story here. China is [65:58] training on US data from US providers. [66:01] Forbes published an investigation called [66:04] these American startups are making [66:06] China's AI smarter. And I think this [66:09] relates to a lot of your work in the [66:11] early part of the administration. Sachs, [66:13] they claim US data labeling startups are [66:15] selling valuable training data to [66:18] Chinese labs which in turn is helping [66:20] them catch up with the US frontier ones. [66:22] Two startups, Sergei and Meror are both [66:25] valued at over $20 billion. They sell [66:28] training data sets to people like OpenAI [66:30] Anthropic federal agencies. Um they all [66:34] sell the same data sets to top Chinese [66:36] AI companies. According to this report, [66:38] like Tencent, Bance, Alibaba, Moonshot, [66:41] etc. Top six AI labs in China, according [66:44] to this report, uh are spending $500 [66:46] million a year buying what Forbes calls [66:49] secret sauce, uh PhD written content, [66:53] reinforcement learning, knowledge [66:55] pipelines, uh all that kind of great [66:57] stuff. I have investments in a couple of [66:59] these companies, including Micro One. [67:02] The founder of Micro One didn't [67:04] participate in selling to China. He made [67:06] that decision. effects. What do you [67:08] think here about this new wrinkle in [67:10] terms of really the secret sauce behind [67:13] a lot of these models is the data? We've [67:16] run out of uh open data on the web. [67:19] Obviously, we talked last week about the [67:21] books being uh you know having the [67:23] spines taken off of them and scanned in. [67:25] I mean, people are looking for data. [67:27] Mira, micro one, all these companies are [67:29] providing it. Should they be providing [67:31] the same data and selling it to Chinese [67:34] open source companies or not? Well, [67:36] look, I think we got to decide what our [67:37] objective is here. Are we trying to just [67:39] get in like a full-blown economic war [67:41] with China? Are we just trying to [67:42] prevent all of our companies from doing [67:44] business over there? If that's our [67:46] objective, then you can take that [67:48] position. Historically, the rules have [67:51] been that you want to be careful about [67:53] technology transfer of technology that [67:56] has a dual use, right? That it has a [67:58] military application. My sense of data [68:01] is that it's largely a commodity. I [68:03] mean, data labeling certainly is. If you [68:06] basically tell them that they can't use [68:09] data labeling, I guarantee you there's [68:11] no shortage of labor in China that they [68:14] can use to do the data labeling. In [68:16] fact, they probably are. What I'm saying [68:18] is there's a lot of ways to get this [68:19] data. So look, if we basically ban these [68:23] companies from selling to China, we [68:25] should expect reciprocal actions taken [68:28] by China to ban companies over there [68:32] selling to us. Maybe rare earths. These [68:34] two countries are not completely [68:37] independent of each other. By the way, I [68:39] want us to be as independent and [68:41] sovereign as possible. I don't want to [68:43] have any dependencies. No dependencies. [68:45] >> But we still at this moment in time do [68:47] have some dependencies. So I think you [68:48] have to ask the question is this data [68:51] really proprietary? Does it have a dual [68:53] use? Does it have a military [68:55] application? [68:55] >> Yeah, I don't think it has military. [68:57] It's definitely not data labeling. This [68:59] is like hiring PhDs, hiring super [69:01] professionals to, you know, create [69:03] unique data sets. So it's [69:06] >> China can do that too and I guarantee [69:07] you they are. I don't think this is [69:09] going to give us a decisive advantage in [69:10] the AI race. It's going to annoy it's [69:13] going to create annoyance. It's going to [69:15] create friction. and how bad do you want [69:17] our relationship with them to be? Do you [69:18] want to risk starting another trade war? [69:20] Look, I'm not against restrictions when [69:22] I think they're going to pack a punch. [69:24] For example, I'm really glad that the [69:26] first Trump administration limited the [69:29] export of EUV lithography machines to [69:31] China. You know, that was all the way [69:33] back, I think, in 2019. [69:36] So, that was a really important [69:37] decision. And so, look, I think targeted [69:40] strategic controls make sense. I would [69:43] just make sure that this one actually [69:45] meets that bar. [69:46] >> Brad, any thoughts here on this [69:50] open-source catchup, the data being sold [69:52] to China and our adversaries? Are you [69:54] concerned about these open source models [69:55] and then us providing data to them? [69:58] >> First, you know, I'm in absolute [69:59] agreement with David that we want [70:02] maximum competition at as we sit here [70:05] today, the US is winning. We talked [70:07] about it at the start. Our Frontier Labs [70:08] are winning. Our open source is winning. [70:10] and we have fairly limited regulations, [70:13] right? She's coming here in September in [70:15] a bilateral meeting to meet with the [70:17] president. We're advancing relations on [70:19] a variety of fronts. So, I think [70:21] everything looks good and you want to [70:22] continue down that path. With that said, [70:25] I will tell you that this will irritate [70:27] people in Washington who feel that this [70:30] along with distillation and other things [70:33] um could be the export of chips. All of [70:35] which at a certain level make sense [70:38] cause people to wonder whether or not [70:40] we're making it too easy on the Chinese [70:42] labs to catch up with American labs uh [70:45] you know in the race to frontier [70:46] intelligence. So it you know it's the [70:48] type of story Jason that I think will [70:50] continue to muddy the waters that will [70:52] continue uh to be monitored. The reason [70:54] I don't think it will cause us to change [70:56] our stance with respect to China is [70:59] because we're winning. But if the [71:01] president asks his adviserss, you know, [71:03] one of these days, six months down the [71:04] line, are we winning against China? And [71:06] all of a sudden he gets a response, no, [71:08] we're no longer winning, they've caught [71:09] up, they've passed us, etc., then these [71:12] things will get a lot more scrutiny uh [71:14] than they're getting today. I think the [71:16] only reason they pass muster today is [71:18] because we're still leading the race. I [71:20] got to say, using Kimmy and Quen and, [71:24] you know, GLM 52 for the last 60 days, [71:26] my lord, these things are good. And I [71:28] don't think it's very patriotic to be [71:30] giving them an advantage. I wouldn't do [71:31] it. I'm glad the company [71:33] >> Sorry. What's the advantage? What's the [71:34] data set that you you're worried about [71:36] that's so proprietary? [71:37] >> Any of these data sets are um created by [71:40] experts here in America who are given [71:43] like the queries that have errors in [71:45] them. So when you give um you know a [71:47] thumbs down to a query that's highly [71:49] technical, it could be code, it could be [71:51] biology and science, these are you know [71:54] PhDs going in there and putting in the [71:56] latest and greatest content and then [71:58] verifying it, double verifying it. And [72:00] that's why we're getting better and [72:02] better results out of the LLMs. So [72:04] essentially, you're just helping them [72:06] catch up. And this could be a big [72:07] advantage for America if we weren't [72:09] sending it there. I think a big reason [72:11] these models are getting better is [72:13] because data is being leaked to them. [72:15] >> But what makes you think that China [72:16] can't do this? They have tons of PhDs [72:18] over there. [72:19] >> They would have to hire No. No. If they [72:21] were to do it at this scale, they would [72:24] need to hire the best and brightest uh [72:27] scientists and experts in the west. So [72:30] basically all the knowledge of the West [72:32] is being um you know put into packages [72:35] for our LLMs to get better. They're [72:38] sending those same packages and [72:40] reselling them to Chinese companies, [72:42] which means they catch up just as quick. [72:44] I think it's a big part of why they're [72:45] catching up in line with distillation. [72:48] You know, they're it's it's really very [72:50] similar process. [72:51] >> Look, if there's something truly [72:52] proprietary here, I don't want us to [72:54] sell our secret sauce to China. So, you [72:57] know, I'd have to look into that and see [72:59] like is there some real secret sauce [73:00] here. But this idea that it would [73:03] seriously disadvantage China, you know, [73:04] they're graduating more math and science [73:07] graduates every year than the rest of [73:09] the world combined. I mean, they don't [73:13] have a shortage of smart people, [73:14] especially [73:14] >> and we're graduating and kicking them [73:16] out of the country. That's the other [73:17] problem. We got to get that fixed. [73:19] >> Well, it's like this a lot of different [73:21] issues here. I don't know how many you [73:22] want to conflate, but I this idea that [73:24] they can't but the this idea that they [73:27] can't recreate those data sets. I mean, [73:28] look, if there's something truly [73:29] proprietary here, if it has a dual use, [73:32] if it's military related, but I don't [73:34] know that that's what this is. [73:35] >> Well, they're all proprietary by design, [73:37] but I don't know about the dual use cuz [73:40] I don't have the data sets here. All [73:42] right, folks. That's another amazing [73:44] episode of your Allin podcast. Thank you [73:46] so much, Brad, for joining us, Chimath. [73:49] Good luck on your world tour. Hope [73:51] you're enjoying a little rest and good [73:53] luck um trying to buy a white turtleneck [73:57] this season. that are sold out [73:58] everywhere. So, go to the all-in.com [74:01] store, allin.com/store. [74:03] We have 1,000 signature Chimath [74:07] autographed white sweaters coming. You [74:09] can sign up in advance for those. All [74:10] proceeds go to charity. By charity, I [74:13] mean yacht fund. All right, we'll see [74:16] you next week, everybody. Bye-bye. [74:19] >> Let your winners ride. [74:27] We open sourced it to the fans and [74:28] they've just gone crazy with it. [74:30] >> Love you. Queen of [74:34] >> yours. [74:39] >> Besties are [74:42] that is my dog taking a notice in your [74:43] driveway. [74:47] >> Oh man, my habitasher will meet me up. [74:49] We should all just get a room and just [74:51] have one big huge orgy cuz they're all [74:53] just useless. It's like this like sexual [74:54] tension that we just need to release [74:56] ourselves. [75:03] >> We need to get mercy. I'm going all in. [75:12] I'm going all in.