Big Tech Earnings Defy CapEx Fears as Cloud Surges

云业务激增，科技巨头财报打破资本支出担忧 科技巨头凭借加速增长的云需求和强劲的核心营收，证明其数十亿美元的基础设施支出具有合理性，表明尽管投资者对成本感到担忧，但人工智能的商业化变现正在切实扩大规模。 本期节目深入剖析了涵盖 Google、Microsoft、Amazon 和 Meta 的重大科技财报日，分析了各超大规模云厂商如何通过云业务增长与广告收入来证明其巨额资本支出的合理性。主持人探讨了市场估值，将当前的市盈率与互联网泡沫时期进行了对比，同时探讨了企业级 AI 领域的动态变化以及 OpenAI 的合作格局。讨论内容还涵盖了科技领袖对 AI 创造就业机会的看法、正在进行的 Elon Musk 诉讼案庭审报道、Arc-AGI V3 基准测试，以及皮肤科自动化检查和 AI 生成电商播客等创新应用。 主持人 John Coogan 和 Jordi Hays 在参加完 Stripe Sessions 后回归，反思了科斯奇点（Coasean singularity）理论，指出随着 AI 降低交易成本，未来的企业规模可能会缩小。 (0:02) 在罕见的四大科技巨头同日公布财报中，Google 成为明确的赢家，其搜索业务营收达 604 亿美元，推动股价上涨 10%。 (0:36) Google Cloud 业务大幅提速，营收达到 200 亿美元（同比增长 63%），积压订单金额超过 4600 亿美元。 (0:36) Microsoft 报告总营收达 829 亿美元，新增 500 万个 Microsoft 365 Copilot 席位，使其在 4.5 亿总席位基数中的付费用户数达到 2000 万。 (0:40) 随着 OpenAI 拓展至 AWS，Microsoft 对 OpenAI 模型的独家专有权宣告结束，而近期重新谈判的合同取消了 AGI 条款，转为持续至 2032 年的营收分成机制。 (0:40)

TBPN hosts John Coogan and Jordi Hays argue that major tech firms are transitioning from a generic capital expenditure race into distinct monetization models anchored by resilient core cash flows. While frontier reasoning benchmarks show major technical limitations, the hosts contend that massive infrastructure spending is buffered by accelerating enterprise demand rather than speculative dot-com-style multiples. Diverging Hyperscaler Monetization and Infrastructure Models (0:36) A rare simultaneous earnings release highlighted diverging strategies among tech hyperscalers. Google demonstrated a vertically integrated full-stack model, dispelling fears that generative AI would erode core Search revenue ($60.4 billion, up 19% year-over-year) while Google Cloud surged 63% to $20 billion with an order backlog exceeding $460 billion. Amazon positioned AWS as neutral infrastructure for multiple foundation labs, re-accelerating cloud growth to 28% year-over-year alongside a custom chips business reaching a $20 billion run rate. Microsoft posted $82.9 billion in total revenue and expanded Microsoft 365 Copilot to 20 million paid seats out of 450 million enterprise seats. Microsoft also adapted its commercial model with OpenAI by removing Azure exclusivity—allowing OpenAI workloads onto AWS—and replacing the controversial AGI clause with a long-term revenue-sharing structure running through 2032. Meta grew revenue 33% to $56.3 billion through ad-tech automation, though management attributed a sequential dip in Daily Active People to regional service blockages in Iran and Russia. CapEx Scrutiny and Valuation Comparisons (15:51) Surging capital expenditure sparked divergent market reactions, particularly for Meta, whose shares dropped 10% after raising full-year CapEx guidance to $125 billion–$145 billion. Unlike enterprise cloud vendors that resell capacity, Meta must fund its frontier AI investments directly through advertising cash flows, raising investor debate over whether increased budgets reflect higher compute volumes or rising hardware supply chain costs. Despite mounting investor anxiety, the hosts noted that the primary hyperscalers trade at historically grounded forward price-to-earnings valuations (Meta at 16x, Google at 17x, Amazon at 24x, and Microsoft at 25x). This stands in sharp contrast to the dot-com bubble peak, where Microsoft traded at 73x, Cisco topped 200x, and Yahoo reached 800x. However, viewers highlighted that high valuation risks remain concentrated in other tech segments, citing elevated multiples for companies such as AMD (130x), Palantir (220x), Tesla (350x), and Intel (900x). Labor Economics: Task Automation Versus Occupation Loss (19:15) Addressing widespread labor concerns, the hosts cited technologist George Hotz and NVIDIA CEO Jensen Huang to challenge alarmist predictions regarding AI-driven unemployment. Huang dismissed forecasts that AI presents a 20% existential risk or will eliminate 50% of entry-level college positions as harmful and unfounded, arguing that efficiency encourages businesses to expand organizational scope rather than reduce headcount. Hotz framed AI progress through Jevons paradox, contending that making services cheaper and faster drives greater total consumption. The discussion emphasized the critical structural distinction between automating isolated tasks and replacing entire occupations. For example, while AI algorithms and robotic devices can automate specific medical image evaluations, radiologists perform wider diagnostic, interventional, and administrative duties and face record demand. Similarly, commercial truck driving requires cargo handling, mechanical troubleshooting, and security responsibilities that software alone cannot address. Reasoning Limits and Narrow Practical Deployments (27:56) Frontier models continue to face severe technical bottlenecks in generalized logic. On the Arc-AGI V3 benchmark—which tests abstract 2D reasoning—leading models scored poorly, with GPT-3.5 recording 0.43% and Opus 4.7 scoring 0.18%. Researchers attributed these failures to three primary flaws: susceptibility to local effects, reliance on false world models, and improper levels of abstraction. Because generalized frontier models remain constrained, commercial deployments are concentrating on narrow, structured tasks. Examples include automated robotic skin scanning systems awaiting FDA medical device clearance to supplement routine dermatological exams, as well as e-commerce experiments such as Amazon's synthetic dual-host podcasts designed to explain product SKUs to consumers. The hosts conclude that while modern tech giants have the core cash flows to absorb massive infrastructure outlays, the return on investment remains unequal. An unresolved question for the sector is whether hyper-scaler CapEx will generate meaningful new software platforms, or simply entrench established incumbents while frontier models continue to encounter basic reasoning plateaus.

本期节目梳理了科技巨头财报与AI资本支出的最新分化表现，指出尽管巨额基建投入引发市场对回报周期的审视，但稳固的现金流与差异化模式仍构筑了安全垫；同时讨论拓展至AI对就业结构、前沿推理评测以及垂直场景自动化的真实影响。 巨头财报分化与资本支出回报争议 (0:36) 在四家大型科技公司罕见同日发布财报的背景下，各平台展现出截然不同的增长韧性与市场反馈。Google成为最大赢家，季度搜索及其他收入达604亿美元（同比增长19%），Google Cloud收入达200亿美元（同比增长63%），营业利润6.6亿美元，其云业务积压订单超4600亿美元且半数以上将在两年内确认，推动股价两天内上涨10%。微软季度总营收达829亿美元（同比增长18%），Azure增长40%，其4.5亿付费企业套件用户中Copilot付费席位净增500万至2000万，同时Azure终止了OpenAI独家代销权并取消了AGI条款，转为延续至2032年的分成协议。亚马逊总营收达1815亿美元（同比增长17%），AWS增速重新加快至28%，自研芯片业务年化运行率达200亿美元，稳固的业绩消化了巨额资本支出。 相比之下，Meta季度营收虽达563亿美元（同比增长33%）且广告展示量增长19%，但由于将全年资本支出指引区间上调100亿美元至1250亿至1450亿美元，加之受伊朗和俄罗斯区域性网络阻断影响导致日活跃用户数首次环比下滑，其股价在财报发布后下跌近10%。主持人分析指出，Meta作为缺乏企业云分销渠道的纯广告平台，必须完全依赖核心应用产品与广告变现来消化高昂的算力成本，其巨额支出具有押注前沿通用人工智能（AGI）的高风险期权性质。 估值水平与互联网泡沫时期的对比 (15:51) 针对华尔街对科技巨头百亿美元级资本开支的焦虑，主持人援引分析师Ren的数据指出，当前四大巨头的估值体系显著不同于2000年互联网泡沫破裂期。目前Meta的前瞻市盈率为16倍，Google为17倍，Amazon为24倍，Microsoft为25倍；而在互联网泡沫巅峰期，微软市盈率曾达73倍，思科超过200倍，雅虎高达800倍，纳斯达克平均估值达200倍。主流巨头即便剥离AI炒作，其数字广告与SaaS云业务仍能贡献千亿美元级的实际自由现金流。 不过，讨论也指出了市场存在的估值结构分化。观众反馈显示，部分半导体与高估值科技股仍处于极端乘数状态，例如AMD（130倍）、特斯拉（350倍）、Palantir（220倍）以及英特尔（900倍），同时过去五年间诸多未盈利AI初创公司被推高至独角兽估值。此外，生成式AI工具并未如预期催生颠覆性的新兴消费级社交应用，反而强化了Instagram、WhatsApp和Roblox等成熟平台的竞争壁垒。 劳动力市场：任务自动化与杰文斯悖论 (19:15) 在探讨AI对就业市场的长期冲击时，科技从业者George Hotz提出了相对乐观的视角。他将AI技术扩散类比为劳动力引入，认为人类的需求几乎是无限的，技术赋能将引发经典的“杰文斯悖论”——当执行某项任务的效率提高、成本下降时，全社会的整体消费量反而会增加，从而做大整体经济蛋糕，而非净减少工作机会。 NVIDIA首席执行官黄仁勋在公开讲话中亦对渲染AI恐慌的言论提出反驳。他批评那些警告年轻人不要从事软件工程或放射科医学的观点，将“AI导致20%人类生存威胁”或“摧毁50%大学毕业生初级岗位”的主张斥为荒谬与不负责任，强调美国对软件工程师的需求正在扩大。节目主持人进一步指出，现实企业更多利用AI扩大业务野心与承接规模，且AI通常只能自动化单一特定任务（如扫描病灶），无法替代复合型职业的全套现实职责（例如放射科医生的综合诊疗或货车司机的安保与装载责任）。 前沿推理局限与垂直场景落地 (27:56) 在技术落地与基准测试方面，节目展示了AI从通用推理向专业硬件场景的演进动态。在医疗领域，新型自动化皮肤检查机器人设备通过标准化身体扫描检测潜在肿瘤病变，展现出类似机场安检般的一致性与高效率，在等待FDA审批的过程中，其定位是作为辅助工具提升医生筛查效率，而非直接替代皮肤科医师。 然而，前沿模型的通用推理水平仍显局限。根据Arc Prize发布的最新Arc-AGI V3基准测试（利用类2D游戏的全新环境考察推理泛化），领先模型的得分依然极低：GPT-3.5仅获得0.43%，Opus 4.7仅获得0.18%。研究人员归纳出AI的三大核心失效模式：受局部效应干扰、构建错误的世界模型，以及从训练数据中提取了错误的抽象层次（未能形成有效的奖励关联机制）。此外，针对亚马逊基于两名合成AI主播生成商品介绍播客的新功能，主持人对其是否真正契合日常消费决策需求表达了戏谑的怀疑态度。 本期讨论表明，科技巨头正从单纯的算力竞赛转向垂直差异化的商业变现阶段，庞大的主营现金流为当前的AI基础设施扩张提供了安全屏障；然而，前沿推理基准的平淡表现与垂直硬件落地的高标准，仍预示着高昂资本支出转化为全要素生产率的路径充满挑战。

## Background 该内容由科技播客频道 TBPN 于 2026 年 5 月 22 日发布在 YouTube 平台上，节选自其工作日直播精简版节目《Diet TBPN》。简介指出该节目由 John Coogan 和 Jordi Hays 主持，被《纽约时报》誉为“硅谷的新宠”，日常聚焦科技圈重磅人物对谈与实时行业热点解析。 ## TL;DR 本期节目聚焦四大科技巨头（Google、Microsoft、Amazon、Meta）极其罕见的同日财报发布，全面拆解各家截然不同的 AI 商业化叙事与激增的资本支出（CapEx）。主持人们对比了公有云业务的爆发式增速与市盈率估值，力驳“搜索已死”及“AI 毁灭就业”等末日论调，强调技术提升的是任务效率而非直接取代全套职业岗位。此外，节目还梳理了 Meta 激进投入前沿大模型的算力孤岛困境、最新发布的 ARC-AGI V3 基准测试中前沿大模型的惨淡得分，以及全自动皮肤癌筛查机器人和亚马逊 AI 商品播客等硬核与幽默兼具的科技新进展。 ## Atmosphere & style 节目氛围轻松幽默且极具极客对谈质感，两位主持人配合默契、互相抛梗，在严肃的万亿市值财务分析与荒诞的现场花絮间自如切换。演播室对话兼具第一线创投圈的敏锐洞察与接地气的即兴调侃，既有对 AGI 狂热和企业财报算力的硬核推演，也有对法庭庭审琐事、奇葩 AI 落地产品的戏谑吐槽。整场交流节奏明快、信息密度极高，给听众带来一种身处硅谷咖啡馆旁听两位资深老友畅聊科技前沿的沉浸感与鲜活感。 ## Bulletin - [01:15](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=75s) 四大科技巨头同一天集中发布财报，Google Cloud 季度营收同比狂飙 63% 突破 200 亿美元，待履约订单积压超 4600 亿美元，同时核心搜索业务强劲增长 19%，彻底粉碎了“AI 正在杀死传统搜索”的市场悲观论调。 - [07:43](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=463s) Meta 尽管一季度营收同比增长 33% 达到 563 亿美元，但因管理层将全年资本支出指引上下限粗暴上调 100 亿美元（修正为 1250 亿至 1450 亿美元），且缺乏对外转售算力的云业务承托，导致股价盘后重挫近 10%。 - [04:10](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=250s) 微软与 OpenAI 达成平稳的新协议重组，彻底废除了充满法律隐患的“AGI 解约条款”并锁定商业分成权益至 2032 年，虽失去了独家分销护城河并促成 OpenAI 登陆 AWS，但有效消除了长期悬置的治理风险。 - [29:12](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1752s) 最新发布的 ARC-AGI V3 基准测试如同一剂清醒剂，GPT-4.5 仅获 0.43% 的得分，Claude Opus 仅得 0.18%，在直观交互推理与世界模型构建上暴露出前沿大模型严重的底层能力短板。 ## Key takeaways - [01:15](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=75s) 谷歌搜索业务营收同比稳步增长 19%，以强硬数据打破了 ChatGPT 等生成式工具将颠覆其现金牛业务的叙事，为激进的云与模型研发提供了坚实的造血能力。 - [01:15](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=75s) Google Cloud 凭借 63% 的同比高增速与超 4600 亿美元的在手积压订单，展现出全栈 AI 基础设施的庞大需求，成为该季度表现最抢眼的公有云厂商。 - [04:10](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=250s) 微软 Copilot 付费席位达到 2000 万，但相比 4.5 亿 Microsoft 365 订阅底座，长线企业渗透率仍处在起步阶段，需要持续证明资本支出的投资回报率。 - [04:10](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=250s) 微软重构与 OpenAI 的合作协议，虽然 Azure 失去了排他性经销特权，但解除了模糊的 AGI 条款，并将分成权益锁定至 2032 年，平衡了资产增值与跨云竞争。 - [07:43](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=463s) 亚马逊 AWS 顶住巨大体量压力实现 28% 的超预期增长，自研 AI 芯片年化运行率越过 200 亿美元，凭借容纳 OpenAI 与 Anthropic 的中立基建生态稳居 CapEx 之王。 - [09:32](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=572s) Meta 的算力困境在于其缺乏对外商业化出租算力的公有云业务（RPO 储备），天量 CapEx 无法通过转租分摊风险，必须直接由内部广告转化与自研产品硬性兑现。 - [09:32](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=572s) 扎克伯格坚信超级智能与递归自我改进（RSI），但公开资本市场不愿为缺乏落地对齐的 AGI 宏大愿景买单，反而对激增的算力成本感到恐慌。 - [12:34](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=754s) 大模型突破并未如预期催生出颠覆性的新社交平台，反而通过降低内容创作门槛进一步巩固了 Instagram、WhatsApp 等成熟老巨头的流量壁垒。 - [14:33](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=873s) 科技巨头 AI 叙事出现结构性分化：谷歌走全栈平台闭环，微软依托企业分发，亚马逊扮演中立算力底座，Meta 则是广告现金牛捆绑前沿模型的看涨期权。 - [16:41](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1001s) 当前四大科技巨头的市盈率处在 16 至 25 倍之间，受数千亿美元真实自由现金流支撑，整体健康度远超 2000 年互联网泡沫时期的估值狂热，核心资产具备强防御性。 - [19:15](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1155s) 从杰文斯悖论的视角来看，AI 类似引入高效的外部劳动力，算力效率的大幅提升不会缩减整体经济规模，反而会激发此前因成本过高而受限的海量潜在需求。 - [22:20](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1340s) 黄仁勋痛批过度宣扬 AI 灭绝论与岗位消亡论的有害恐慌，指出盲目断言“软件工程师与放射科医生绝迹”会误导青年远离关键领域，造成现实的人才断层危害。 - [24:30](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1470s) AI 自动化接管的是特定“任务”而非整个“岗位”，多数企业引入 AI 的首要目标是借此拓展业务管线与抢夺市场份额，而非单纯缩减雇员编制。 - [29:12](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1752s) ARC-AGI V3 的极低得分揭示出前沿 AI 模型在抽象层级识别与世界模型建立上的致命软肋，证伪了当前大模型在通用推理上已经无所不能的过度炒作。 ## Detailed walkthrough ### [00:02](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=2s) Stripe Sessions 回归与科技股“四杀”财报日 节目录制前，两位主持人 John Coogan 和 Jordi Hays 刚从 Stripe Sessions 大会匆匆赶回演播室。他们对前一天与 Stripe 创始人考利森兄弟（Patrick Collison 与 John Collison）等人的深度对谈赞不绝口，并极力推荐观众去补看那场访谈。谈话中提及了一个颇具哲理的经济学术语——“科斯奇点”（Coasean singularity，源自经济学家罗纳德·科斯的企业理论），即随着交易成本因技术而骤降，未来企业规模可能会大幅缩水。虽然两人打趣说这种经济学行话对普通人来说简直像是一堆“天书（mumbo jumbo）”，但他们预言未来大众会越来越频繁地听到这些讨论，之后节目也打算邀请专业经济学家来深入拆解。 抛开抽象的经济学假说，当天科技界上演了现实版的高潮：四大科技巨头（Google、Microsoft、Amazon、Meta）在同一天集中发布季度财报，史称“四杀（quad kill）”。这种四巨头财报撞车极其罕见，上一次发生大约还要追溯到 2020 年，宛如一场“日食”般的天体巧合。两位主持人调侃这简直需要靠占星术来预测下一次发生的时间，不过现实原因往往只是因为季度结账窗口碰巧与节假日安排撞在一起；而作为 AI 硬件核心的英伟达（Nvidia）却从不凑这个热闹，它的财报发布日总是要比其他巨头晚两到三周。 ### [01:15](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=75s) 谷歌云全面提速与“搜索已死”神话的破灭 在当天的财报乱斗中，Alphabet（Google 母公司）成为毫无争议的大赢家，亮眼的成绩单直接推动其股价在接下来的几天内暴涨 10%。过去两年里，科技圈普遍弥漫着一种“ChatGPT 与 AI 聊天机器人将彻底颠覆并吞噬传统搜索收入”的悲观叙事，但这份财报用冰冷的数据给出了明确的反驳：谷歌搜索及其他业务营收达到 604 亿美元，同比强劲增长 19%。所谓“谷歌搜索已死”的论调不攻自破，这项超级现金牛业务不仅没有萎缩，反而持续扩大，为公司激进的 AI 研发提供了坚实的弹药库。 比起搜索的稳健，Google Cloud 的爆发式增长更是震撼了华尔街。科技科技资讯聚合网站 Techmeme 在梳理财报时，甚至直接略过了各大巨头的总体营收大盘，将所有目光锚定在各家的云业务增速上。Google Cloud 季度营收达到 200 亿美元，同比增速高达惊人的 63%，营业利润更是高达 66 亿美元，展现出强悍的利润率。更为夸张的是其云业务未完成订单（backlog）几乎翻倍，突破了 4600 亿美元，且其中超过一半的订单将在未来 24 个月内正式确认为收入，绝非遥不可及的十年长约。这种全栈式 AI 基础设施的惊人需求，让谷歌激增的资本支出（CapEx）显得极为合理与从容。 ```mermaid flowchart TD accTitle: 科技巨头最新季度云业务增速与大盘表现 accDescr: 对比 Google Cloud、Azure、AWS 的云端增速以及 Meta 的整体业务增速。 subgraph Clouds["三大公有云增速对比"] GC["Google Cloud<br/>营收 200 亿美元<br/>增速: +63% 同比"] AZ["Microsoft Azure<br/>增速: +40% 同比"] AWS["Amazon AWS<br/>增速: +28% 同比<br/>(超出 25% 预期)"] end subgraph Social["社媒巨头"] META["Meta 整体营收<br/>营收 563 亿美元<br/>增速: +33% 同比"] end ``` - **搜索与核心大盘**：Google 搜索及其他业务营收 60.4 亿美元，同比攀升 19%。 - **云业务积压订单**：Google Cloud 待履约订单达 4600 亿美元以上，50% 以上将在未来 24 个月内转化为实际营收。 - **模型预期**：分析师注意到谷歌在财报中对下一代模型保持克制，外界推测其可能选择在三周后的 Google I/O 大会上正式发布 Gemini 4 或专注代码能力的垂直模型。 ### [04:10](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=250s) 微软的企业 AI 现实：Copilot 渗透率与重组 OpenAI 契约 微软公布了稳健的财报，总营收达到 829 亿美元，同比增长 18%，全面超出华尔街分析师预期，但股价却微跌 2%。在聊到商业数据前，主持人还特意插播了一段演播室花絮，炫耀了他们身后等了六个月才到货安装的高清 LED 背景大屏，直呼“太值得等待”。回归正题，微软当前的挑战在于叙事重心的转移：市场需要看到 Azure 增长、Copilot 企业实际渗透率，以及对天文数字 CapEx 的合理回报证明。 微软目前的企业版 Microsoft 365 订阅坐拥 4.5 亿个付费席位（覆盖 Teams、Outlook 等标准组件，单人月费约 30 美元）。而在高附加值的 Copilot 席位上，本季度新增了 500 万个，总数达到 2000 万个。尽管增长数字本身可圈可点，但与 4.5 亿的基本盘相比，整体渗透率依然处于非常早期的阶段；微软的终极构想是让每一个办公席位都捆绑一份 Copilot 附加订阅，如何将渗透规模从 2000 万拉升到两亿乃至四亿，是其未来长期的核心考题。 与此同时，市场正在消化微软与 OpenAI 之间重新谈判后的新契约结构。这一变动对微软具有双重效应： ```mermaid flowchart LR accTitle: 微软与 OpenAI 重新谈判的影响权衡 accDescr: 展示 Azure 丧失独家经销权与 OpenAI 跨平台增长对微软资产估值正反两方面的影响。 subgraph Disadvantages["Azure 业务承压"] Loss["失去独占权"] --> AWS_Sale["OpenAI 登陆 AWS 平台"] AWS_Sale --> Cloud_Stick["无法单纯靠模型撬动企业客户迁移至 Azure"] end subgraph Advantages["资产与分成受益"] Expansion["OpenAI 跨云生态全面做大"] --> Valuation["推升微软所持股权的资本价值"] Expansion --> RevShare["确立持续收入分成 (延续至 2032 年)"] end ``` 过去 Azure 销售代表可以以“想要 GPT-5.5 只能来 Azure”为王牌要求客户把整套 IT 架构迁徙过来，但很多深度绑定 AWS 的大型企业根本不愿承担迁移成本；如今 OpenAI 模型正式登陆 AWS，微软虽然失去了独家垄断的护城河，但作为持有大笔股权的投资方，OpenAI 的全面扩张又会在账面上放大微软的资本利益。令人欣慰的是，萨提亚·纳德拉（Satya Nadella）完成了极具战略眼光的协议重组，双方没有对簿公堂，而是平稳过渡：协议彻底废除了模糊且充满隐患的“AGI 条款”（该条款原定 OpenAI 达成通用人工智能时即可单方面中断合作），改为白纸黑字明确商业收入分成延续至 2032 年到期，消除了长期悬在微软头顶的法律悬剑。 ### [07:43](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=463s) 亚马逊 AWS 强劲重燃与 Meta 激进 CapEx 引发的暴跌 亚马逊在此次财报季证明了自己依然是实打实的“资本支出之王（CapEx king）”，其资本支出指引再次拔得头筹。然而巨额开支并没有吓跑投资者，反而拉动股价小幅走高，核心就在于其云业务 AWS 的重新加速。AWS 季度营收同比增速达到 28%，击败了分析师普遍预期的 25%。对于 AWS 庞大的基数而言，能够顶住体量压力重现加速实属不易。亚马逊一季度总体销售额录得惊人的 1815 亿美元（同比增长 17%），几乎是在以迈向万亿美元年营收的狂暴姿态前进；其高利润的数字广告业务贡献了 172 亿美元，而自研云端 AI 芯片业务的年化运行率（run-rate）也正式越过了 200 亿美元关口。AWS 采取了极具包容性的平台策略，左右逢源同时吸纳 OpenAI 与 Anthropic，使得更庞大的算力集群投资显得极具说服力。 与亚马逊形成鲜明对比的是 Meta 的惨烈遭遇。财报公布后，Meta 股价遭遇断崖式下跌，盘后跌幅接近 10%。单看核心业务运营，Meta 表现得无可挑剔：一季度营收为 563 亿美元，同比大幅增长 33%；广告展示次数增加 19%，平均广告单价同样保持坚挺，广告商业引擎正开足马力轰鸣运转。 真正引发华尔街恐慌的是扎克伯格对资本支出的激进调高。Meta 原先预计的全年资本支出区间在 1150 亿至 1350 亿美元之间（即 1250 亿美元上下浮动 100 亿），但此次管理层直接将区间的上下限各自粗暴拔高了 100 亿美元，修正为 1250 亿至 1450 亿美元。市场瞬间神经紧绷，开始严重质疑如此规模的资金大跃进究竟能否买来成比例的实际有效算力，还是又一场陷入过度投入的烧钱狂欢。 ### [09:32](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=572s) Meta 的算力困境与“AGI 上头”的扎克伯格 在探讨科技巨头激增的资本支出（CapEx）时，主播们指出当下的核心分歧在于：巨额资本开支究竟反映了他们正在采购更多算力，还是仅仅因为底层算力供应链各环节的价格暴涨，导致企业不得不花更多钱去买与此前相同规模的算力。对于股东而言，答案并没有那么清晰乐观。为了凸显 Meta 的特殊性，主播开玩笑地设想：如果 Meta 现在是一家未上市的私营创业公司，给投资人讲的商业计划书写着“我们专门打造用于带货卖货的 AI Agent，当前年化营收（Run-rate）达到 2000 亿美元且保持 33% 的年增长率”，并拥有如此惊人的营业利润率，那么它的估值会被炒到多少？“20 万亿美元？很可能。”然而，Meta 终究是一家具备极高盈利能力的公开市场巨头，且缺少一块极为关键的拼图——它没有能够向外销售的公有云基础设施。 在这场财报季中，马克·扎克伯格（Mark Zuckerberg）展现出了最狂热、“彻底被 AGI 洗脑”（AGI-pilled）的超大规模算力企业 CEO 形象。他在财报电话会议上甚至直接大谈递归自我改进（Recursive Self-Improvement, RSI），强调 AI 的进化“绝不仅限于写代码”。扎克伯格发自内心地坚信超级智能的到来，并且决不允许自己沦落到“别人拥有了真正的超级智能，而他却没有”的落后处境。但这种宏大叙事在华尔街看来却过于模糊与虚幻——当前公开市场最买账的只是立竿见影的企业生产力工具，资本市场不仅无法消化扎克伯格的 AGI 愿景，甚至没有给予 Meta 现有核心业务应有的估值溢价。 这种尴尬源于 Meta 与亚马逊 AWS、微软 Azure 等云厂商的根本差异。AWS 投入海量资本支出极易向股东交待：无论 AI 赛道未来如何演变，这批算力都可以由 Anthropic 消化，或是转租给 OpenAI、其他新兴 AI 实验室，亦或是自用训练模型；模型无论是走向同质化平价还是保持高壁垒，算力底座都有明确的客户托底。然而 Meta 缺乏庞大的企业级 AI 合约或巨额履约义务（RPO）储备，巨额算力投入无法通过云出租来轻松获得市场标准的投资回报率（ROI）。这笔开支必须直接在 Meta 的广告转化和 AI 产品收入中兑现。即便 Meta 未来理论上可以转型成为“新型算力云”（Neo-cloud）对外转售算力，但以其体量而言，那将伴随一整套全新的运营挑战。 ```mermaid flowchart TD accTitle: 算力资本支出消化路径对比 accDescr: 对比传统云厂商与 Meta 消化 AI 资本支出的不同商业闭环逻辑 subgraph CloudHyperscalers["AWS / Azure 等云巨头"] direction TB A["巨额 CapEx 采购算力"] --> B["多重消化通道"] B --> C1["租给 Anthropic / OpenAI 等模型实验室"] B --> C2["外部企业客户云服务合同与高额 RPO"] B --> C3["自研模型与内部产品"] end subgraph MetaHyperscaler["Meta 的孤岛闭环"] direction TB D["巨额 CapEx 采购算力"] --> E["单一内部消化通道"] E --> F1["广告算法与推荐系统优化"] E --> F2["前沿模型与自研 AI Agent 研发"] E -.->|商业闭环压力| G["必须直接反映在当期广告业务回报中"] end ``` ### [12:34](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=754s) DAP 历史性微跌与“大模型强化老巨头”悖论 财报中引发市场一丝震颤的另一个数据，是 Meta 的日活跃自然人数（Daily Active People, DAP）出现了自公司披露该指标以来的首次环比下降。不过，Meta 管理层给出了合理的外部不可抗力解释：主要是由于伊朗地区的互联网中断以及俄罗斯对 WhatsApp 的限制所致。主播们认为，这并不代表全球网民的真实行为发生了根本性逆转，因为市场上根本没有任何新兴社交网络正在蚕食 Meta 旗下应用矩阵的蛋糕——曾经看似凶险的 TikTok 威胁、Snapchat 威胁早已常态化，甚至连 Sora 带来的颠覆性恐慌也是“来去匆匆”。虽然有调查显示年轻群体的社交媒体使用时间有所回落，但此前的基数实在太高，实际数据面上依然固若金汤。 两位主持人借此探讨了一个极具讽刺意味的技术悖论：人类迎来了大语言模型（LLM）这种犹如“外星科技”般的全新突破，它在数字形态下完全能让用户相信它是一个活生生的人；然而在这一划时代的技术浪潮下，全球竟然没有诞生任何一个全新的巨型社交平台。如果站在 2020 年回望，人们一定会断定拥有“数字分身”技术的时代必将催生出革命性的新型社交网络机制。但残酷的现实是，AI 并没有颠覆旧格局，反而极大地巩固了既有巨头的护城河。像 Roblox、Instagram、WhatsApp 和 Facebook 这样的成熟生态，只会承载更多由 LLM 生成或辅助增强的高质量内容；内容创作者们通过 AI 省去了繁琐枯燥的周边事务，将更多精力倾注于核心产出，从而进一步加固了老巨头的流量壁垒。 - **DAP 指标说明**：Meta 如今以 DAP（Daily Active People）取代传统的 DAU，旨在统计其家族应用（Facebook、Instagram、Messenger、WhatsApp）去重后的真实活跃自然人。 - **不可抗力因素**：本季度 DAP 下滑主要受中东（伊朗断网）及东欧（俄罗斯封锁 WhatsApp 服务）等区域性地缘限制影响，而非自然流失。 - **平台马太效应**：技术突破未能催生新社交产品，LLM 反而降低了创作者内容生产成本，强化了老平台的网络效应与内容充裕度。 ### [14:33](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=873s) 财报季分化：四大科技巨头的不同 AI 叙事 随着最新财报数据出炉，过去 12 个月里那种“谁报的 CapEx 数字大、谁签的合作大，谁就能无脑上涨”的一体化 AI 叙事正在发生破裂。如今，尽管每个超大规模计算玩家都在公布惊人的资本开支，但各自的底层商业逻辑已彻底分道扬镳： | 科技巨头 | 核心 AI 战略与叙事特征 | 竞争优势与潜在风险 | |:--- |:--- |:--- | | **Google** | **全栈平台（Full-Stack）** | 贯通自研芯片、底层模型、云基础设施到终端消费级应用的闭环生态。 | | **Microsoft** | **企业级落地与分发（Enterprise Adoption）** | 依托 Windows/Office 的无敌渠道力，聚焦 Copilot 渗透率与商业化客户变现。 | | **Amazon** | **算力底座与中立生态（Infrastructure & Partners）** | 拥有与 Anthropic、OpenAI 最纯粹的基建合作关系，无 DeepMind 级别的同业竞合包袱。 | | **Meta** | **广告优化 + 前沿模型高风险看涨期权** | 基本盘靠广告推荐变现，同时押注前沿大模型以期在未来寡头格局中跻身第一梯队。 | 主播特别指出，对于 Meta 而言，投资者面对的是一份“扎实的广告现金牛业务”附赠一份“前沿超级智能看涨期权（Call Option）”。如果其基础模型实验室（如 TBD Labs/MSL）取得类似 GPT-5.5 或 GPT-4.7 水平的重大突破，Meta 就能在这个由少数几家寡头瓜分的 AI 时代中，直接孵化出一个匹敌 OpenAI 或 Anthropic 规模的前沿企业业务。 从短期业绩来看，市场依然最渴求即时可见的真金白银收入，而核心公有云业务的爆发式增速成为本季最硬核的支撑： - **Google Cloud (GCP)**：同比暴增 63%，此前市场最乐观的预期也仅仅是 40%，堪称惊艳全场； - **Microsoft Azure**：保持了 40% 的高位强劲增长； - **Amazon AWS**：实现了 28% 的同比增长，虽略低于部分买方预期的 30%，但规模体量下依然稳健。 ### [16:41](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1001s) 估值之争：当下 AI 繁荣是否重演互联网泡沫？ 针对市场上蔓延的“AI 泡沫论”，社交媒体上有分析者（如用户 Ren）将当前科技巨头的市盈率（P/E Ratio）与 2000 年互联网泡沫顶峰进行了横向对比。数据显示，目前刚刚公布业绩的科技巨头估值出奇理性：Meta 的市盈率约为 16 倍，Google 为 17 倍，Amazon 为 24 倍，Microsoft 为 25 倍。反观 2000 年网络泡沫顶峰时期，微软的市盈率高达 73 倍，思科（Cisco）超过 200 倍，雅虎（Yahoo）甚至被炒至匪夷所思的 800 倍，纳斯达克整体市盈率曾攀升至 200 倍之巨。如今被称为“泡沫”的巨头们，其 16 至 25 倍的市盈率背后，是每年实打实产生数千亿美元真金白银的自由现金流支撑。即使在某种极端荒诞的假设下——AI 被证实全是一场空、人类重回纸笔时代——世界依然会退守到数字广告、云端电子表格和企业 SaaS 服务中，这些巨头依然处于极具防御性的赚钱机器地位。 ```mermaid flowchart LR accTitle: 互联网泡沫与当前 AI 估值结构对比 accDescr: 对比2000年网络泡沫与当前科技巨头在市盈率与现金流支撑上的核心差异 subgraph Bubble2000["2000 年网络泡沫峰值"] direction TB A1["Nasdaq 整体 P/E: 200x"] A2["微软: 73x | 思科: 200x+"] A3["雅虎: 800x (商业化微弱)"] end subgraph TodayMarket["当今科技巨头估值"] direction TB B1["Meta: 16x | 谷歌: 17x"] B2["亚马逊: 24x | 微软: 25x"] B3["坚实支撑: 数千亿美元真实自由现金流"] end ``` 然而，主播们也指出这一对比存在的局限性。如果只挑选人类商业史上最伟大的四家印钞机公司，自然很容易得出“毫无泡沫”的结论。但若将目光转向过去 5 年成立的一批未上市一级市场 AI 公司，其估值从 2000 万被击鼓传花推高至数亿、数十亿美元，大部分却几乎毫无利润乃至营收可言，它们的状态更接近当年的雅虎。同时，公开市场其他板块同样存在令人咋舌的估值： - **AMD** 市盈率约 130 倍； - **Palantir** 市盈率约 220 倍； - **Tesla** 市盈率约 350 倍； - **Intel** 市盈率甚至飙升至 900 倍。 即便如此，看多科技股的核心逻辑依然稳固：推动美股整体市值大头的并非 Intel 或 Palantir，而是市盈率仅十几到二十余倍的四大现金流核心巨头。即便某些中细分领域存在透支与狂热，整个科技市场的绝对重心依然牢牢锚定在非常稳健、且具备强大变现能力的盈利引擎之上。 ### [19:15](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1155s) George Hotz 与黄仁勋的乐观论调：AI、移民与杰文斯悖论 乔治·霍茨（George Hotz）在 5 月 1 日公开发布了一剂“乐观药丸”（white pill），直言“AI 将创造更多就业机会”。他提到英伟达（NVIDIA）首席执行官黄仁勋（Jensen Huang）也在谈论同一观点，并指出只要深入思考，其中的逻辑就非常显而易见：AI 的本质与移民极为相似——它们都代表着系统中新增了劳动力。人类的欲望和需求在实质上是无限的，这符合经典的“杰文斯悖论”（Jevons paradox），即当我们大幅提高某种资源的利用效率时，最终的消耗量反而会爆炸式增长。霍茨还引用了他在 Facebook 实习时学到的一句名言：“如果你把网站速度提升 10%，用户在上面停留的总时间就会增加 5%。” 霍茨尖锐地讽刺了那些哭诉“移民压低了美国本土劳动力工资”的保护主义者。他指出，AI 确实会在某些具体工种上击败部分人类，但“保护主义是属于失败者的哲学”（protectionism is for losers）。真正关键的并不是通过二次分配去调节，而是通过底层设计扩大整体经济蛋糕，同时在结构上控制不平等，未来的社会将迎来“比以往任何时候都更多的事情可做”。 - **核心类比**：将 AI 劳动力涌入类比为外来移民，两者都能通过提升生产力来做大经济增量。 - **杰文斯悖论与效能转化**：效率提升往往不会导致总需求缩水，反而会解锁更多原本因成本高昂而未被满足的需求。 ### [20:20](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1220s) 庭审微小困扰手记：Mike Isaac 的推特现场花絮 在连线《纽约时报》科技记者迈克·艾萨克（Mike Isaac）前，两位主持人饶有兴致地讨论起法庭每周四天的工作制，笑称司法系统提前进入了未来工作形态。艾萨克近期一直在现场报道埃隆·马斯克（Elon Musk）的诉讼案，而他发在社交平台 X 上的直播推文别具一格：除了记录马斯克的证词、律师的唇枪舌剑、法官与陪审团的互动外，他还详尽记录了自己在庭审旁听过程中遭遇的一连串琐碎的个人不便，甚至被网友整理成了一份按分钟记录的“小烦恼信息图”。 这些细微的现场花絮充满了荒诞感：从午餐吃的烤番茄玉米披萨、匆忙抢座时忘了装水导致在激辩时刻口渴难耐，再到旁听席上突然有人笔记本电脑意外外放起油管视频，大声播放着关于“花园水管喷嘴”的解说。这种极具烟火气的法庭侧写让严肃的高科技诉讼报道多了一层诙谐的真人秀质感。 - **法庭花絮焦点**： - 庭审旁听的窘迫细节：没装水的空水壶、缺少垫靠的硬座椅。 - 突发法庭事故：旁听人员电脑意外外放的花园水管喷嘴介绍视频。 - 主持人预告：后续将邀请 Eric Seufert（“Seufernator”）与 Mike Isaac（“Isaacanator”）展开深度拆解。 ### [22:20](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1340s) 批驳 AI 毁灭论：黄仁勋痛斥末日言论阻碍人才发展 主持人随后播放了黄仁勋的一段公开发言视频。黄仁勋在发言中直言不讳地批评了当前盛行的 AI 警示论调，认为许多耸人听闻的言辞不仅毫无建设性，而且对社会造成了实质性的伤害。某些科学家可能自以为是在发出善意预警，宣称“AI 将彻底接管放射学，因此放射科医生将不复存在”；但现实是，如果全社会的年轻人都因此被吓退而不去学习放射医学，导致现实中急需放射科医生时出现人才断层，这种后果对人类社会而言极其有害。同样的道理，过早断言软件工程师即将绝迹，导致年轻人不敢进入计算机专业，也是完全错误的误导。 黄仁勋强调，行业应当负责任地宣传技术能力、倡导合理的监管政策与护栏，而不是用荒谬的言辞制造恐慌。诸如“AI 有 20% 的概率对人类带来灭绝性危机”、“AI 即将消灭 50% 的大学应届生岗位”、“AI 会彻底摧毁民主”等说法毫无根据，完全是虚张声势。主持人调侃黄仁勋这是在现实中直接对某些 AI 巨头进行不点名的“隔空喊话”（subtweeting），并指出所谓“50% 概率”预测是最狡猾的骑墙话术——无论未来发生什么，预测者都能自圆其说。 ```mermaid flowchart TD accTitle: 盲目警示论调对人才市场的负面反馈循环 accDescr: 展现过度夸大 AI 风险如何导致年轻人才流失并最终反噬现实行业需求的逻辑 subgraph Alarmism["非理性警示与末日论"] A["学者/专家宣称：AI 摧毁放射学 / 淘汰程序员"] B["兜售 20% 灭绝论与 50% 就业消亡论"] end subgraph MarketReaction["人才市场连锁反应"] C["年轻人产生恐慌心理"] D["学生放弃相关专业选择"] end subgraph RealWorldImpact["现实社会恶果"] E["行业实际需求不降反增"] F["专业人才断层，造成实质社会伤害"] end Alarmism --> MarketReaction C & D --> RealWorldImpact ``` ### [24:30](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1470s) 任务自动化而非岗位消失：从放射科医生到卡车司机的多维现实 两位主持人深入探讨了“岗位消亡（Jobs）”与“具体任务自动化（Tasks）”之间的根本区别。他们与大量打造或销售 AI 产品的公司交流后发现，各行业的雄心不仅没有缩水，反而在持续膨胀。金融科技公司 Rogo 的创始人 Gabe 曾指出，企业应用 AI 的主流选择并非削减成本，而是借此激进地去抢夺市场份额、拓展业务管线并承接更多客户。 以常常被杰弗里·辛顿（Geoffrey Hinton）等人当作自动化典型的放射科医生为例：虽然视觉模型能够出色地识别 CT 或核磁共振图像中的肿瘤，但“看图找肿瘤”仅仅是医生一天日常中众多任务的其中一项。现实情况是，放射科医生当下的收入创下新高，市场需求比以往更加旺盛。类似的道理也适用于卡车司机——卡车司机绝不仅仅是在握方向盘，调查显示 70% 的卡车司机在行车时随身携带武器，因为他们实际上兼具了货运安保的职责，同时还要负责车辆日常检修、货物配重平衡以及路线突发状况处理；甚至路上有坐着孩子的小汽车打招呼时鸣笛互动，也是自动化系统无法替代的人性温度。 - **岗位与任务的分离**：特定任务（如阅片、路径驾驶）能被算法接管，但综合性职业包含安保、沟通、突发协调与整体交付。 - **商业意图的转变**：多数企业利用 AI 是为了做大营收、抢占市场份额，而非单纯压缩人力成本。 ### [26:22](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1582s) 数据中心美化构想与洛杉矶的城市伪装艺术 随着大规模算力设施引发周边社区对能耗与噪音的抵触，如何对数据中心进行“审美美化”（beautification）成为了网络热议的话题。网友们脑洞大开地提出了各种争取公众支持的改造方案：既然数据中心会产生大量废热，为何不将其改造成面向社区免费开放的恒温泳池或大型水上滑梯？甚至有人提议在数据中心屋顶修建半管滑板场，或是常年举办大脚怪越野车集会（Monster Truck Rally）与“极限马戏团”（Nitro Circus）摩托特技秀，打造全民娱乐地标。 主持人还提及了投资人 Julie Young 分享的一个冷知识：洛杉矶市内长期存在着精巧的伪装建筑。例如在洛杉矶闹市区正中央，有几处正在运转的采油机（oil derricks）被直接伪装成了普通的犹太教堂建筑，绝大多数市民日常经过时根本无从察觉。洛杉矶市中心也隐藏着不少用于边缘计算与内容分发网络（CDN，如用于缓存 Netflix 视频）的数据中心，它们通常外观毫不起眼，唯一的外部特征就是整栋大楼没有任何窗户。 ### [27:55](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1675s) 皮肤科 AI 检查：自动化筛查进入临床健康领域 节目的后半段转向了一款刚刚发布的全新医疗硬件产品——全自动化皮肤检查设备。主持人展示了相关的演示视频，并分享了普通人去皮肤科进行全身黑素瘤或皮肤癌排查时的普遍心理体验。当患者面对全身皮肤病变扫描时，传统人工检查往往伴随着某种尴尬与繁琐，而自动化的全身扫描与图像识别技术的引入，正在让这种预防性筛查变得高效而无创。 ### [28:21](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1701s) 自动化皮肤检查机器人与医疗审批挑战 主持人形象地将现有的皮肤科检查比作过安检时的敷衍搜身：安检人员往往只是快速拍两下就放行，很多皮肤科医生也是匆匆扫上一眼，随口说句“看起来挺好”，背后的原因可能是医生太忙或状态疲惫。相比之下，专门的医疗机器人能够极其精准、一致且高效地完成全身体表扫描，不仅能极大提升医疗资源的可及性，还并非取代医生，而是作为辅助工具帮医生大幅提高筛查效率。 不过，这项技术的落地仍面临漫长的监管考验。主持人探讨了这家创业公司目前所处的阶段，特别是其是否需要经历美国食品药品监督管理局（FDA）严格的医疗器械审批流程。这类审批往往可能耗时数年之久，但一旦顺利通过并投入临床使用，极有可能在商业和医疗市场上取得巨大的成功。 - **检查对比**：传统医生容易受疲劳和时间紧迫影响，而机器人能做到高精度、高一致性筛查。 - **定位与角色**：该机器人并非完全取代医生，而是通过自动化体检流程加快诊疗速度。 - **监管关卡**：商业化落地亟需面对耗时可能长达数年的 FDA 医疗器械审批挑战。 --- ### [29:12](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1752s) ARC-AGI V3 基准测试发布：前沿大模型的惨淡得分 ARC Prize 针对 ARC-AGI V3 发布了最新基准评测更新。主持人首先调侃团队成员 Tyler 是否还在人类排行榜上，Tyler 则自嘲表示自己可能早被挤下去了，因为现在测试题目的体量已经增加。ARC-AGI V3 对当前最前沿的 AI 模型来说依然是一道难以逾越的天堑，测试机制本身就像是一款人类能轻松过关、但 AI 极难理解的 2D 视频游戏。 评测团队对 OpenAI 和 Anthropic 的顶级模型进行了基准测试，结果惨不忍睹：GPT-4.5 仅获得 0.43% 的得分，而 Claude 3.7 Sonnet / Opus 也仅仅拿到了 0.18%。主持人指出，ARC-AGI 测试的奇特之处在于模型的初始基准总是从接近零（0% 到 1%）起步，随后随着模型能力的代际跃迁逐步抬升。在当下铺天盖地宣称“AI 已经无所不能”的宏大叙事中，这类测试如同一记清醒剂，清晰划定了当前模型的智能边界。 ```mermaid flowchart TD accTitle: ARC-AGI V3 核心测试结果与三大失效模式 accDescr: 展示主流前沿模型在 ARC-AGI V3 中的极低得分以及测试揭示的三种关键失效模式 subgraph Scores["评测基准表现"] M1["GPT-4.5: 0.43%"] M2["Claude Opus 4.7: 0.18%"] end subgraph Failures["模型三大核心失效模式"] F1["真实局部效应 (True local effect)"] F2["错误的世界模型 (False world model)"] F3["训练数据抽象层级错误 / 解决关卡但未强化奖励"] end Scores --> Failures ``` - **GPT-4.5 得分**：仅为 0.43%。 - **Claude Opus 4.7 得分**：仅为 0.18%。 - **测试本质**：规则极简的 2D 交互环境，普通人类能够直观理解，但对大模型来说极度困难。 - **三大失效模式**：评测揭示了模型的三种典型崩溃场景——真实局部效应（True local effect）、错误的世界模型（False world model），以及来自训练数据的抽象层级错误（例如虽然通关了该层级，但未能建立并强化正确的内在奖励机制）。 --- ### [30:48](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1848s) 亚马逊推出全自动商品带货播客与节目收尾 节目进入尾声，主持人幽默地宣布要介绍“过去 100 年里诞生的最令人激动的全新产品”。根据 Business Insider 记者 Katie 的报道，亚马逊将允许为其在售商品自动生成 AI 播客。在这类播客中，两位由 AI 驱动的主持人会围绕商品展开讨论，甚至模拟接听听众的热线电话来答疑解惑。 主持人忍不住打趣这种功能荒谬可笑，调侃道大家只可能给最奇葩、最滑稽的产品生成播客，很难想象普通消费者在买东西前真的愿意听完一整期针对具体商品的播客。随后，主持人调侃搭档之前在选购厨房纸巾时的偏执挑剔，笑称他恨不得买 1950 年代老配方的纸巾，原本用户只想要一个精准好用的商品筛选过滤器，而亚马逊给出的“解决方案”却是：“给你生成一段长达一小时的播客深入分析每一个 SKU，让你做出最坚定的购买决定”。两位主持人在欢声笑语中互道周末愉快，结束了当天的直播。 ## Notable quotes > [04:10](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=250s) 想要 GPT-5.5 只能来 Azure — 主持人 > [07:43](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=463s) 资本支出之王（CapEx king） — 主持人 > [09:32](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=572s) 我们专门打造用于带货卖货的 AI Agent，当前年化营收（Run-rate）达到 2000 亿美元且保持 33% 的年增长率 — 主持人 > [09:32](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=572s) 绝不仅限于写代码 — 马克·扎克伯格 > [19:15](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1155s) 如果你把网站速度提升 10%，用户在上面停留的总时间就会增加 5%。 — George Hotz > [19:15](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1155s) 保护主义是属于失败者的哲学 — George Hotz > [22:20](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1340s) AI 将彻底接管放射学，因此放射科医生将不复存在 — 黄仁勋 > [22:20](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1340s) AI 有 20% 的概率对人类带来灭绝性危机 — 黄仁勋 > [30:48](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1848s) 过去 100 年里诞生的最令人激动的全新产品 — 主持人 > [30:48](https://www.youtube.com/watch?v=KHoWIm9e4Ew&t=1848s) 给你生成一段长达一小时的播客深入分析每一个 SKU，让你做出最坚定的购买决定 — 主持人 ## Mentioned John Coogan, Jordi Hays, TBPN, Stripe, Patrick Collison, John Collison, 罗纳德·科斯, Google, Alphabet, Google Cloud, Microsoft, Microsoft Azure, Microsoft 365, Copilot, Amazon, AWS, Meta, 马克·扎克伯格, 萨提亚·纳德拉, Nvidia, 黄仁勋, OpenAI, Anthropic, Techmeme, George Hotz, Mike Isaac, Eric Seufert, Geoffrey Hinton, Rogo, Julie Young, ARC-AGI V3, GPT-4.5, GPT-5.5, Claude 3.7 Sonnet / Opus, Gemini 4, TikTok, Snapchat, Sora, Roblox, Instagram, WhatsApp, Facebook, Business Insider, FDA ## Video description Diet TBPN delivers the best of today’s TBPN episode in 30 minutes. TBPN is a live tech talk show hosted by John Coogan and Jordi Hays, streaming weekdays 11–2 PT on X and YouTube, with each episode posted to podcast platforms right after. Described by The New York Times as “Silicon Valley’s newest obsession,” the show has recently featured Mark Zuckerberg, Sam Altman, Mark Cuban, and Satya Nadella. Follow TBPN: https://TBPN.com https://x.com/tbpn https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231 https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235 https://www.youtube.com/@TBPNLive [Source](https://www.youtube.com/watch?v=KHoWIm9e4Ew)

[4:10] 主持人: 想要 GPT-5.5 只能来 Azure [7:43] 主持人: 资本支出之王（CapEx king） [9:32] 主持人: 我们专门打造用于带货卖货的 AI Agent，当前年化营收（Run-rate）达到 2000 亿美元且保持 33% 的年增长率 [9:32] 马克·扎克伯格: 绝不仅限于写代码 [19:15] George Hotz: 如果你把网站速度提升 10%，用户在上面停留的总时间就会增加 5%。 [19:15] George Hotz: 保护主义是属于失败者的哲学 [22:20] 黄仁勋: AI 将彻底接管放射学，因此放射科医生将不复存在 [22:20] 黄仁勋: AI 有 20% 的概率对人类带来灭绝性危机 [30:48] 主持人: 过去 100 年里诞生的最令人激动的全新产品 [30:48] 主持人: 给你生成一段长达一小时的播客深入分析每一个 SKU，让你做出最坚定的购买决定

[8:54] Can you imagine just take 1 second and imagine that Meta's private company and their pitch is that we make AI agents for selling products. They're at a $200 billion run rate growing 33% a year. How's it priced? 20 trillion? [19:15] George Hotz: AI and immigration are fundamentally the same... Wants are effectively unlimited. It's classic Jevons paradox. [19:15] if you make the site 10% faster, people spend 5% more total time on it. [19:15] Jensen Huang: It is hurtful if we convince all the young college graduates to not be software engineers and it turns out [the] United States need more software engineers than ever.

[0:02] We had to rush back from Stripe [0:03] sessions, fantastic interview with the [0:05] Collison brothers, many others [0:07] yesterday. Go check it out if you [0:08] haven't already. Fun stuff. We didn't go [0:10] nearly deep enough on the whole Coasean [0:12] singularity. There's a debate on the [0:14] timeline about it. We can get into it [0:15] later. We need to have some more [0:16] economists on the show to break it down [0:18] for us. But this is their big pitch. [0:20] Firms get smaller and the future the the [0:24] the Coase theory of the firm, anyway. If [0:26] you studied economics, you might be [0:27] familiar. But it's jargon and you're [0:29] going to be hearing a lot more of it. [0:31] >> A whole lot of mumbo jumbo. [0:34] >> I'm not even going to try. [0:36] I'm not even going to try, Jordan. Well, [0:37] here's some here's some less mumbo [0:39] jumbo. Tech earnings, quad kill recap. [0:43] Four in one day. Last time this [0:45] happened, I think it was 2020. Four big [0:48] tech companies all same day. It is [0:50] random. No one really knows why. [0:52] It just sort of lines up. It's like a [0:54] solar eclipse. [0:55] >> astrology to predict the next one? [0:57] >> Potentially. Potentially. [0:59] >> You would have to imagine that it's some [1:00] something around the alignment of the [1:02] the planets that causes the whole [1:05] >> More like the holidays because books [1:06] close, then certain things line up. Also [1:09] like Nvidia is never in the conference. [1:11] They're always like two or three weeks [1:12] later. But anyway, let's go through what [1:15] actually happened. Google was the big [1:16] winner. They absolutely crushed. The [1:18] stock's up 10% in the last couple days. [1:21] Really successful. And there's a few [1:23] reasons. So the core business still [1:24] growing. Google search. This was the [1:26] question of is AI or chatbots going to [1:29] be eating into the core Google search [1:31] revenue, the core business that drives [1:33] all the other investments. The answer is [1:35] unequivocally no. [1:37] Google search and other revenue was 60.4 [1:40] billion. Yeah. Up 19% year-over-year. I [1:45] mean, two years ago, there was the whole [1:46] like Google search is dead thesis. That [1:49] has not been born out, at least this [1:50] year. Very, very strong results there. [1:53] Then Google Cloud is like the major [1:54] story. So much so that Tech him when he [1:58] uh did his earnings recap, he didn't [2:00] even talk about top lines at Google, [2:03] Amazon, and Microsoft. He just says [2:05] Google Cloud revenue 63% year-over-year, [2:08] Amazon Web Services 28% year-over-year, [2:10] Azure 40% year-over-year, and then Meta [2:13] overall revenue, which we'll get into, [2:15] 33% year-over-year. So, he's not even [2:17] looking at the at the rest of the [2:18] business. All that matters for the three [2:20] major public clouds um are those cloud [2:23] businesses, not even uh anything's [2:26] anything specific within the overall [2:28] business. So, uh long story short, [2:30] higher capex at Google totally justified [2:32] in a world of cloud acceleration. [2:35] Backlog growth is huge. Uh cloud backlog [2:38] uh nearly doubled to more than $460 [2:41] billion. [2:42] And importantly, this isn't like a [2:44] 5-year, 10-year contract type thing. [2:46] Half of that, more than half of that [2:48] backlog is expected to be recognized in [2:50] the next 24 months. So, really good news [2:52] there. Uh Google Cloud hit 20 billion, [2:55] up 63%. Now, it is smaller than Amazon, [2:58] and that's why AWS is People are hunting [3:00] for high 20 growth rates, maybe begins [3:03] with a three. Uh in Google, it's [3:05] smaller, so it can grow 63%, but that's [3:07] still fantastic. And uh cloud operating [3:09] income was 6.6 billion, which is very [3:11] strong margin. So, everything is looking [3:13] good. Search durability, demand for AI [3:16] infrastructure, it's the full stack AI [3:18] play right now. Uh Microsoft showed [3:21] solid execution, but they didn't [3:23] dramatically change any particular [3:25] narrative. I think a lot of this goes to [3:26] like the the the just naturally slow [3:29] deployment of enterprise AI, but this is [3:31] the focus for Microsoft. They are the [3:33] enterprise AI uh play. So, [3:36] >> one thing was notable is they weren't [3:38] they weren't heavily teasing anything [3:40] related to the next Gemini release. [3:43] >> Yeah, I I they wouldn't tease that at at [3:45] earnings. [3:46] >> No, no, they have they have [3:47] historically. So, uh maybe they maybe [3:49] they want to let the model speak for [3:51] itself. [3:51] >> Yeah, probably announced Gemini 4, I [3:54] would imagine. Uh [3:56] announced it uh [3:57] >> Or a coding specific model. [3:59] >> Yeah. Yeah, something there. I mean, [4:01] there's certainly demand for it. They [4:03] have the capacity. There's lots of [4:04] opportunity. Uh we'll probably be [4:06] following that in 3 weeks when Google [4:08] I/O happens. So, Microsoft. Uh investors [4:11] wanted to see Azure growth, [4:14] Copilot adoption, and a solid [4:16] justification for all the CapEx. Uh the [4:19] stock's down 2%, nothing crazy, but uh [4:21] Microsoft beat on headline numbers. [4:23] Revenue was 82.9 billion, up 18% [4:27] year-over-year. Uh and from an analyst [4:29] perspective, it was a clean beat. Um [4:32] but, there's more nuance here. So, in [4:34] terms of AI adoption, Microsoft added [4:36] >> Look at this beautiful LED wall behind [4:38] us. [4:38] >> million. Yeah, let's give it up to the [4:40] >> Still underrated. [4:41] We ordered this bad boy [4:43] maybe like 6 months ago. [4:45] >> ago. Took a while. Took a while. [4:46] >> But, it is fun. We we need to figure out [4:48] >> Worth the wait. [4:48] >> more things to do with it. On Copilot [4:51] adoption, this is the big question. For [4:53] Microsoft 365, like total seats, you [4:55] know, you think about Microsoft Teams, [4:57] Outlook, like the standard Microsoft [5:00] enterprise seat. I think it's about $30 [5:02] per user per month. Competes with uh [5:05] Google Apps enterprise, right? But, [5:07] Microsoft's been doing this for much [5:08] longer. They have 450 million paid [5:11] seats. So, in terms of Copilot, they [5:13] added 5 million, which is great. They're [5:15] at 20 million now. So, that's a is solid [5:18] adoption, but it's still small compared [5:20] to the 450 paid seats. Ideally, like [5:23] every seat would have a Copilot [5:25] alongside of it. That I mean, that's the [5:26] idea, right? Uh and so, the question is [5:29] how quickly can they get that 20 million [5:31] Copilot seats up to [5:33] 200 million, 400 million, you know, [5:35] ideally everyone has one of these add-on [5:38] subscriptions, and that's lifting the [5:40] overall business. Uh market reaction was [5:42] a bit choppy as the new OpenAI [5:44] relationship gets digested, and so [5:46] there's two sides to the Microsoft [5:48] OpenAI deal right now. Of course, OpenAI [5:52] is no longer limited to selling just [5:55] through Azure. Azure no longer has [5:57] exclusive access to OpenAI models. [5:59] OpenAI is now available on AWS as well. [6:02] And it's And it cuts both ways. So, on [6:04] one hand, Microsoft, you know, they used [6:07] to be the exclusive provider. If you're [6:08] a sales guy at Microsoft on Azure, [6:10] you're probably really excited to call [6:11] somebody up and say, "Hey, you want to [6:13] use GPT 5.5? We're the only place in [6:15] town. Like Like you should come over [6:16] here. You should migrate more of your [6:17] infrastructure. You should be using [6:19] Azure." But, it was probably frustrating [6:21] because a lot of people say, "No, we're [6:22] locked into AWS. We are stuck in that [6:25] cloud, and we'd love to, but we'll just [6:27] use the API directly or we'll do We'll [6:29] figure out some other workaround, but [6:30] we're not moving [6:31] >> Yeah, or people could be like, "Yeah, [6:32] we're we're we're open to trying it, but [6:34] it'll take, you know, a couple quarters, [6:37] realistically, to get to get ramped up." [6:39] >> Yeah. And so, from an Azure perspective, [6:42] losing exclusivity is a negative, but [6:45] from a business perspective, [6:47] the more growth for OpenAI is good for [6:49] Microsoft's equity stake in the company. [6:52] And so, these two tensions, like the the [6:54] capital balance sheet side of the [6:56] business versus the Azure sales and [6:59] acceleration side of the business, like [7:01] >> Not to mention Copilot versus Codex. [7:04] >> Sure. Yeah. And so, these two these two [7:06] sides of the business are sort of [7:08] butting up against each other, and Satya [7:10] Nadella ultimately has to go and un- un- [7:12] you know, renegotiate the contract, and [7:14] they did. And it seems like they're in a [7:16] good place because [7:17] a lot of people were were sort of saying [7:19] like, "Oh, well, is this going to get [7:20] messy? Is there some going to be some [7:21] lawsuit here?" It seems like there was a [7:23] very clean negotiation, renegotiation, [7:25] and they Open AI got rid of the AGI [7:28] clause that allows them to to stop [7:30] sharing models, but they are sharing [7:32] revenue and [7:33] >> Basically, the the I think 2032 end date [7:36] >> Yeah. Is just the end of [7:37] >> share instead of this like arbitrary [7:40] >> Yep. [7:40] >> AGI. [7:42] >> Uh so, over at Amazon, Amazon still, of [7:44] course, the CapEx king. They top the [7:47] CapEx forecast. Uh, but the as scary as [7:50] that would be, AWS is re-accelerating, [7:53] and so that's very good news. So, uh, [7:54] the stock moved slightly upwards. Uh, Q1 [7:57] 2026 sales for Amazon overall were 181.5 [8:01] billion for the quarter. For the [8:03] quarter, they're they're like on track [8:05] to make a trillion in revenue soon. Up [8:07] 17% year-over-year, beating [8:09] expectations. AWS is expected to see 25% [8:12] growth. That was the expectation, 25%. [8:14] That's sort of where they've been. They [8:15] came in at 28%. So, they beat on AWS [8:18] growth, and that's really, really [8:19] important, especially for their scale. [8:21] This was very good news. Uh, the ads [8:23] business is still cooking. Uh, churned [8:26] out 17.2 billion dollars in revenue, and [8:29] the chips business just crossed 20 [8:31] billion run rate, which is very, very [8:32] good. And they have big deals. AWS is in [8:35] this interesting position, less of this [8:37] like Google pure play, full stack, and [8:39] more focused on, uh, working with both [8:42] OpenAI and Anthropic at this point. And [8:44] so, that all justifies more investments [8:46] in training and AI infrastructure [8:48] broadly, and the market seems to like [8:50] it. Huge CapEx numbers, but reasonable [8:52] response from the market and the stock. [8:54] So, Meta Meta had a rough go. Huge drop, [8:56] nearly 10% down following earnings. It's [8:58] sort of jumping up and down now, but, [9:00] uh, the core business, uh, was extremely [9:02] strong. So, Q1 revenue was, uh, 56.3 [9:06] billion, up 33% year-over-year. So, high [9:09] growth for the overall business. Um, ad [9:11] impressions rose 19%, average price per [9:13] ad rose [9:15] They raised the CapEx outlook from, uh, [9:18] this they they have this range, it's [9:19] within 10 billion dollars of 125. They [9:22] added 10 billion to both. So, instead of [9:24] saying, "Hey, we'll be between 115 and [9:26] 135." Now they're saying, "We're going [9:27] to be between 125 and 145." Will they [9:31] actually get more compute for that, or [9:32] is that just a reflection of rising [9:33] prices for all the inputs that go into [9:35] compute spending? That's a big debate, [9:37] because there's one bull case where oh, [9:39] they're optimistic, they're going to buy [9:40] more compute. But, it's like no, there's [9:41] a possibility they're just spending more [9:43] to get the same amount of compute as [9:45] they thought. But, the big problem for [9:47] shareholders, the big answer is it's not [9:49] as clean of a [9:52] >> Can you imagine just take 1 second and [9:54] imagine that Meta's private company and [9:57] their pitch is that we make AI agents [9:59] for [10:01] uh selling products. They're at a $200 [10:03] billion run rate growing 33% a year. [10:06] How's it priced? [10:08] 20 trillion? [10:09] >> Probably. [10:10] Probably. [10:11] >> With those operating margins? [10:12] >> Yeah. [10:13] >> I'm joking. Uh it it truly is one of the [10:15] most incredible businesses [10:17] >> Yeah. [10:17] >> of all time. But, [10:19] >> They don't have the cloud ball. [10:20] >> are Yeah, yeah, they don't have they [10:22] don't have the cloud thing. He's kind of [10:24] acting like the most AGI-pilled CEO, [10:27] like like hyper scalar CEO, right? [10:30] >> Yeah. [10:30] >> He was talking about like basically [10:32] talking about like RSI on the earnings [10:35] call. He's like it's not just about like [10:37] code you know, coding. Like he really [10:39] seemingly believes and he doesn't want [10:42] to be in a position where [10:44] someone else has actual super [10:46] intelligence and he does not. [10:48] >> And so [10:48] >> And and that's just like a little bit [10:50] too fuzzy, because when you look in the [10:51] rest of the market right now, it's like [10:53] productivity tools, right? [10:56] And so, the market just like can't [10:57] really read into it. [10:58] >> Yeah. [10:59] >> And they're just kind of like not even [11:00] giving him full credit for the business [11:02] that he has. [11:03] >> Yeah. And I forget if he's ever [11:04] mentioned the idea of of selling tokens, [11:07] selling AI infrastructure, getting in [11:10] the cloud market, but it's way easier [11:12] for an AWS [11:13] >> of loosely mentioned that last year when [11:16] he was like starting to ramp stuff up, [11:19] like basically saying like maybe we [11:21] could do this at some point. [11:23] >> Yeah. Cuz it's way easier to justify the [11:25] the CAPEX at Amazon when you say, oh, [11:27] well, like, you know, [11:29] if [11:30] you know, regardless of how this market [11:32] plays out, maybe Anthropic takes some of [11:33] the compute, maybe OpenAI takes the [11:35] compute, maybe there's a whole bunch of [11:37] new [11:38] AI labs that take compute, maybe we'll [11:40] use it for our services, our own models. [11:43] If models commoditize, it's still [11:44] valuable. If they don't, it's still [11:46] valuable. There's a whole bunch of [11:48] different ways to to shift that CapEx [11:51] around and justify it. Meta, it's got to [11:53] show up in the AI figures. It's got to [11:55] show up in the ad business, at least [11:57] right now. And so, there's there are [12:00] questions about how the big CapEx spend [12:02] feeds back into cash flow when the [12:04] company doesn't have a platform with big [12:05] enterprise AI contracts, massive RPO [12:08] figures. There's not an easy place to [12:10] just like stuff compute necessarily, and [12:12] at least get you know, a market standard [12:15] return on investment. Of course, they [12:18] could spin up reselling. They could [12:20] become a neo cloud or something like [12:21] that, but at their scale, it's a it's a [12:23] whole new host of challenges if they go [12:25] that direction. And I think that that's [12:27] not the plan. I think you're dialing [12:29] into it that they are in fact recursive [12:32] self-improvement pill at this point. The [12:34] other sort of [12:35] tremor in the system was the fact that [12:37] daily active people, which is something [12:40] that they were it's not users anymore, [12:41] it's DAP, daily active people, [12:44] declined sequentially since the first [12:46] time that Meta began reporting the [12:47] metric. So, less people were using Meta [12:49] platforms [12:51] this quarter than previously. [12:53] But, they had some good explanations for [12:56] it. They said that there were internet [12:57] disruptions in Iran and WhatsApp [13:00] restrictions in Russia. And so, this was [13:02] not representative of a real change in [13:04] internet user behavior, and I think we [13:06] all [13:07] recognize that that is reality in the [13:10] sense that there's not some [13:12] up-and-coming social network that's [13:14] eating Meta platforms family of apps [13:16] lunch, right? Like, the TikTok threat [13:19] came and sort of went. The Snapchat [13:21] threat came and sort of went. Sora came [13:24] and went, right? And so, there there was [13:27] a different post I saw that social media [13:30] use is declining especially under young [13:32] people, but it was already so high it's [13:35] not really showing up in the data. [13:37] >> it will truly be insane if we get this [13:39] new alien technology, we get LLMs which [13:42] in digital form can [13:44] convince you that they're a person. [13:47] >> Yeah. [13:47] >> And we end up getting no new big social [13:49] platforms. [13:50] >> But just strengthen the current ones. [13:52] >> Well, yeah. [13:53] No, I mean I think that's the base case [13:54] right now, but it's still wild. I think [13:56] you would expect I think again if you [13:57] rewound to like 2020 and you said, "Hey, [14:00] we're going to have this technology [14:01] where you can effectively have a digital [14:03] twin of yourself." You would think that [14:05] somebody would figure out some new [14:07] >> Yeah. Maybe. [14:08] >> mechanism. [14:09] >> I I I I continue to think it just [14:10] strengthens the incumbents. Roblox, [14:12] Instagram, WhatsApp, Facebook, like [14:15] these platforms will just get more [14:17] content on them that will probably be [14:19] LLM generated, LLM enhanced. Even if [14:22] it's not LLM written or developed, it's [14:25] it's you know, a creator taking less [14:28] time doing other boring things and [14:30] spending more time on their core output. [14:31] So, [14:33] uh we'll keep tracking it. So, overall [14:35] what what did we learn this earning [14:37] season? One possible takeaway is that [14:39] there's a little bit of a fracturing in [14:41] the overall AI narrative. So, the AI [14:45] narrative over the last 12 months has [14:47] basically been like say the biggest [14:48] number, biggest CapEx number, biggest [14:51] deals, just grow grow grow. Now every [14:54] hyper uh every hyperscaler reports a big [14:57] CapEx number, but each company does have [14:59] a different story. So, Google's the [15:00] full-stack platform, Microsoft's focused [15:03] on enterprise adoption and distribution, [15:05] Amazon's most aggressive on [15:06] infrastructure and partnerships with [15:08] OpenAI and Anthropic. They have the the [15:10] cleanest less uh it's a less competitive [15:13] relationship with Anthropic and OpenAI [15:16] because they don't have the DeepMind [15:17] equivalent necessarily. Meta is all [15:19] about ad optimization, and Meta's also [15:22] this interesting like high-risk option [15:24] on Frontier AI. Like it's possible that [15:27] if TBD Labs MSL, if that pays off, you [15:30] wind up getting an OpenAI-sized or [15:32] Anthropic-sized business out of it in [15:34] the oligopoly world where they have a [15:36] breakthrough, they get to something [15:37] that's 5 5 or 4 7 level, and then [15:41] they're competing at at at an equal [15:43] level, and they're playing in that [15:44] oligopoly. Um and so you like the Meta [15:47] traders are sort of or the investors are [15:49] sort of weighing both of those sides. [15:51] You have a very solid ads business, and [15:52] then you have this call option on uh [15:54] potential Frontier AI, uh which is still [15:57] early but showing some really solid [15:59] signs. So, uh the market likes [16:02] short-term revenue evidence right now, [16:04] and there's still many strategies that [16:06] will play out over the rest of the year, [16:07] but uh this quarter was focused on can [16:10] you justify your CapEx this quarter? And [16:13] so, we'll keep following it. Um let's go [16:15] over the timeline and see what people [16:18] were saying. The headline numbers, of [16:20] course, were huge. [16:23] Cloud growth, AWS 28%, GCP 63%, Azure at [16:26] 40%, uh [16:28] ab- absolutely massive numbers. Uh [16:31] remember, people were hoping for GCP [16:33] just be at 40%, and it went up 63%, [16:36] absolutely massive. Uh AWS, people were [16:39] hoping for 30, they got 28, not too bad. [16:41] >> This poster Ren is comparing the [16:44] price-to-earnings multiples today versus [16:47] the dot-com boom. So, today's mad four [16:50] PE ratios, the companies that reported [16:52] earnings, uh Meta is at 16x [16:54] price-to-earnings, Google's at 17x, [16:56] Amazon's at 24, and Microsoft's at 25 [16:59] price-to-earnings ratio. Compare that to [17:02] during the dot-com. Microsoft was at [17:04] 73x, Cisco was at 200x plus, Yahoo was [17:08] trading at 800x earnings. Uh the Nasdaq [17:11] as a whole traded at 200 times earnings [17:14] during the peak of of dot-com bubble. Uh [17:16] today's bubble is trading at 16 to 25 X [17:20] earnings on companies generating [17:21] hundreds of billions of dollars in real [17:22] cash flow. Now, they're drawing down on [17:24] that cash flow, but even in some bizarre [17:27] world where, you know, AI is a bubble [17:29] and it's fake and we all go back to [17:31] pencil and paper, uh we'll probably go [17:33] back to digital advertising and [17:36] spreadsheets in the cloud and, you know, [17:38] SAS and uh all of these companies will [17:40] still benefit and be in fantastic [17:41] position. [17:42] >> the only thing that that's in you know, [17:44] or one thing that's imperfect with these [17:46] comparisons is when you look at like the [17:48] peak, [17:49] you can say like, okay, Meta, Google, [17:51] Amazon, and and Microsoft have [17:53] relatively reasonable PE ratios, yeah, [17:56] but if you look at the rest uh you look [17:58] at all the companies that have been [18:00] started in the last like 5 years, [18:02] especially in the private markets that [18:03] have been marked up from 20 to 100 to [18:06] 500 to a billion and beyond, many of [18:09] them have absolutely no earnings. [18:11] >> Yeah. [18:11] >> And so, you have to look at like maybe [18:13] something more like a Yahoo [18:15] >> Yeah. [18:15] >> as like a decent comp, right? Where [18:17] Yahoo was trading at peak at 800X. And [18:20] so, [18:21] >> And there is some there is some pushback [18:22] in the in the replies here. Reasonus 4 [18:24] says AMD's at 130, Tesla's at 350, [18:28] Palantir's at 220, and Intel is at 900 [18:31] times earnings. [18:33] >> I know. So, it's like very very [18:35] convenient to just [18:36] >> Yeah. [18:36] >> take the four four of the greatest [18:38] companies ever in the history of the [18:39] world and say like, there's no bubble. [18:42] Like, they're they're reasonably priced. [18:44] >> Yeah. I think it is it is it is useful [18:47] to go back to the fact that like there [18:50] there's there's a lot of hype in tech, [18:51] but what's driving the vast majority of [18:54] the market cap right now is still fairly [18:57] low PE companies. Like, Intel is so much [19:00] smaller on a market cap basis than [19:03] Microsoft, Amazon, Google, Meta, that [19:05] even if you're overpaying on earnings [19:07] for that, the overall market is anchored [19:10] to a pretty decent earnings engine [19:13] still. And that's like the bull case [19:15] here. [19:15] >> George Hotz is back with a white pill. [19:18] >> he say? [19:19] >> May 1st, he says, "AI will create jobs." [19:23] Says it's nice to see Jensen talk about [19:25] this, and it's super obvious when you [19:26] think about it. AI and immigration are [19:28] fundamentally the same. There's new [19:29] people showing up, and hopefully [19:30] everyone understands how and why [19:32] immigration creates jobs. [19:34] Wants are effectively unlimited. It's [19:36] classic Jevons paradox that if we make [19:37] something more efficient, we end up [19:38] using more of it. Or cool aphorism I [19:41] learned at Facebook, if you make the [19:42] site 10% faster, [19:44] people spend 5% more total time on it. [19:47] >> at Facebook for He was like an intern. [19:49] >> He's part of the machine. Now, just like [19:50] you get the wah-wah crying people about [19:52] how immigration lowers wages for [19:54] native-born Americans, and we got to [19:55] keep the hard-working immigrants out [19:57] because you have some right to be lazy [19:59] or something, you'll get this about AI. [20:01] AI will outcompete some humans at some [20:03] jobs. But protectionism is for losers. [20:05] The important thing is that the overall [20:07] pie grows, and inequality stays somewhat [20:09] in check, not by redistribution, but by [20:12] design. There will be more to do than [20:14] ever before. [20:16] >> More to do than ever before? [20:17] >> pull up this video of Jensen? [20:19] >> Yeah, let's do it. And I [20:21] >> In the production chat for you. [20:24] >> Yo, you have it? [20:25] >> Yeah, we can play it. [20:25] >> Great. [20:27] I'm excited for Eric Seufert's deep dive [20:28] here. We'll get into that later. [20:30] >> The Seufernator? [20:31] >> The Seufernator. [20:33] And reminder, we have the Isaacanator [20:35] coming on in just 7 minutes. Mike Isaac [20:38] has been covering the Elon Musk versus [20:41] >> Do they do trial? Is it like a 4-day [20:43] work week with trials? [20:45] >> It is. [20:45] >> Are they living in the future? [20:46] >> It is. Wait, wait, wait. So, can can can [20:48] we pull up this screenshot? Because did [20:51] you read Mike Isaac's full thread on [20:55] on his experience covering the trial? [20:58] Did you read the whole thing? Cuz he's [21:00] been live-tweeting it, and it's very [21:01] funny because he will post So, we got to [21:04] zoom in here. This is a [21:05] >> Did you turn this into an info [21:06] >> it into an infographic. So, this is a [21:09] this is a minute-by-minute chronological [21:11] log of off-topic/personal/minor [21:15] inconvenience moments because Mike Isaac [21:18] will talk about oh, Elon Musk just said [21:20] this on the stage. Oh, the lawyer fired [21:22] back with this. The jury said this. The [21:24] judge said that, right? But then in [21:26] between that, he has been logging very, [21:29] very minor inconveniences that he's that [21:32] he's uh experienced. So, he talks about [21:34] his his lunch, the roasted tomato and [21:37] corn pizza. He says he forgot a pillow [21:40] for his bird to sit on. [21:42] >> wait. [21:43] >> He said he [21:43] >> Corn pizza? [21:44] >> I don't know. Uh he said he bought he [21:46] brought [21:46] >> Auckland thing. [21:47] >> At at at 16:07:45, [21:50] this is transcribed from X, so I think [21:52] the times are [21:53] times are in UTC or something. Uh he [21:55] brought a water bottle, but forgot to [21:57] fill it in the rush for a seat. Now, [21:59] he's increasingly thirsty during an [22:01] intense moment of testimony. Someone's [22:04] laptop suddenly blasted a loud YouTube [22:07] video in the gallery. This is a garden [22:09] hose nozzle, it said. Courtroom mishap. [22:12] So, he's going through it all. Uh it's [22:14] very fun to to track his And we'll And [22:16] we'll dig into his experience both in [22:17] the court and actually with what's going [22:19] on with the case. But let's play this [22:20] Jensen clip. This sounds fun. [22:21] >> There's the things that the things that [22:23] are said are very counterproductive and [22:26] in fact hurtful. [22:27] I [22:28] On the one hand, maybe [22:30] maybe a scientist thinks that by warning [22:32] people that AI is going to completely [22:37] permeate and proliferate across [22:39] radiology and therefore radiologists are [22:42] going to get wiped out. Um on the one [22:44] hand, that might be considered warning [22:47] and therefore helpful, but in fact, the [22:49] counter would have been hurtful. If we [22:51] convinced everybody not to be [22:52] radiologists and we now need [22:55] radiologists, that actually is hurtful [22:57] to society. It is hurtful if we convince [22:59] all the young college graduates to not [23:01] be software engineers and it turns out [23:03] United States need more software [23:04] engineers than ever. That's hurtful. And [23:06] so, we have to be mindful of how we [23:09] communicate [23:11] the importance of this technology and [23:12] what it's able to do [23:14] to advocate for policy and advocate for [23:17] guardrails on the one hand, on the other [23:19] hand, scaring people with things like [23:21] saying nonsensical things which are not [23:24] going to happen, that this is an [23:26] existential threat to humanity. There's [23:28] 20% chance that this is existential. [23:31] That's ridiculous. [23:32] That it's going to wipe out 50% of [23:37] new college grad jobs. That it's going [23:39] to completely destroy democracy. I mean, [23:42] these kind of comments are not helpful. [23:45] They're not based [23:46] They're They're made by [23:50] >> He's subtweeting in real life. He's not [23:52] vague posting, he's subtweeting. [23:54] I I do think the 50% numbers are so [23:56] funny because it's like 50/50 is like [23:58] the most neutral you can be about some [24:00] prediction. It feels like I mean, I get [24:02] that it's not in that case, it's like [24:03] 50% of jobs, but whenever you issue like [24:06] a 50/50 proclamation, it sort of tells [24:09] you nothing. Like we were looking at [24:10] like the prediction markets on the trial [24:12] and you're like, "Oh, it's 50/50. Okay, [24:14] so that actually doesn't give me any [24:16] information about what's going to happen [24:17] in the future." And it's very easy to [24:19] make a prediction that's like, "Oh, I [24:20] predict this will happen. 50/50." And [24:22] then if it doesn't happen, you can be [24:23] like, "Well, yeah, 50% chance it wasn't [24:24] going to happen." Like I had it really [24:26] high that it wasn't going to happen. And [24:27] then you can also say, you know, "I [24:29] predicted it." [24:30] >> Yeah, so it is it is truly universal [24:33] when we talk to these companies that are [24:35] building AI products or selling AI [24:38] products. It seems like the ambition [24:40] level in every single industry is just [24:42] going up. Even even Gabe from Rogo was [24:44] saying, "Yeah, there's an opportunity to [24:45] cut costs or there's an opportunity to [24:47] get aggressive and like take market [24:49] share." [24:50] And And he said like a lot of firms are [24:52] saying like let's just do more work. [24:54] Let's grow Let's grow our revenue. Let's [24:56] take on more clients, right? It's harder [24:58] and harder to to doom about the job [25:01] market. [25:01] >> Yeah. [25:03] Do you need to doom about the task [25:04] market? [25:05] If you're [25:06] >> It's over for task. [25:07] >> Tasks are getting automated. No, it's [25:09] serious. Like there are tasks that like [25:12] the reading of the x-ray or whatever it [25:15] was, the the radiologist scan. What What [25:17] do they actually read? Is that an x-ray [25:20] that they're looking for tumors in? You [25:21] know what I'm talking about, right? The [25:22] famous Hinton [25:23] >> Yeah, I don't know if it's actually an [25:24] x-ray. [25:25] >> It's not, right? It's a cat scan or [25:26] something. [25:26] >> It's some scan where you can look at the [25:28] >> do a scan and of course image generation [25:30] and image models are very good at [25:31] detecting cancers, but that is just one [25:34] of several tasks, many tasks that a [25:36] radiologist does throughout the day. And [25:38] so radiologists are currently making [25:39] more money than ever and they are in [25:40] more demand than ever. And I was telling [25:42] you before the show about truck drivers, [25:45] they don't just drive the truck. [25:47] There was a survey that said 70% of them [25:49] carry weapons because they are [25:51] effectively they are security guards for [25:53] the truck [25:54] as well as probably a bunch of other [25:56] things that you don't think about light [25:58] repairs on the truck, you know, stopping [26:00] for different things, rearranging things [26:02] in the back, making sure that the load [26:04] is like balanced and what not. [26:06] >> with kids in the car? [26:07] >> just yeah. [26:07] >> If you're driving with kids in the car [26:09] and you drive by a big truck on a road [26:10] trip, you can go like this [26:12] >> Yes. [26:12] That's another thing that the that the [26:14] self-driving car has to do. [26:16] >> A computer could never look out the [26:17] window, see some kids in the car and and [26:20] honk. [26:20] >> But well, yeah, I mean well, we'll see. [26:22] Data center beautification [26:25] and he summed it up perfectly. He says [26:27] make data centers aesthetically [26:29] beautiful and so people have been quote [26:32] tweeting, posting, riffing on this, [26:34] sharing different ideas. More people [26:36] would be pro data center if every data [26:39] center had a beautiful open to the [26:41] public heated pool. That's an [26:44] interesting twist. [26:46] Uh pool water slide. I think half pipe [26:49] is a big idea, monster truck rally [26:52] constantly going on. These are ways to, [26:54] you know, win people over. Sort of a [26:56] nitro circus going on at every data [26:58] center. [26:59] >> rally on the roof of a data center? [27:01] >> Yeah. Sort of a universal basic nitro [27:03] circus. So, you get Are you familiar [27:05] with nitro circus? Are you not [27:08] Is this a three fingers moment for you? [27:10] Are you like, "Ah, I'm familiar." [27:11] >> No, I'm just imagining they're building [27:12] a wall. [27:13] >> never actually been to nitro circus. [27:15] >> imagining they're building the like dirt [27:17] bike ramps over the data center. [27:19] >> Exactly. [27:19] >> Backflip thing. [27:20] >> Exactly. Julie Young shares a trick, [27:23] this one neat trick that Los Angeles [27:25] uses to hide oil derricks. Uh, we have a [27:28] few oil derricks in LA that are [27:31] disguised as synagogues. They are [27:33] literally just sitting in the middle of [27:34] the city. Lol, surprisingly few people [27:36] are aware. There's a number of data [27:38] centers that are disguised in Los [27:39] Angeles as well. Usually, you can tell [27:41] because there's no windows in the [27:42] building, but these exist in downtown. [27:44] Usually for edge computing, for uh, like [27:47] you need a content delivery network. You [27:49] put store on your Netflix videos there, [27:51] that type of stuff. Not doing training [27:53] runs, not gigawatt clusters. We're not [27:55] >> Just in, skin exams are getting [27:57] automated. [27:58] >> Oh, yes. Skin [27:59] >> This is very cool product. [28:00] >> Okay. Break it down for us. Let's pull [28:02] it up. [28:02] >> we have video here. [28:03] >> Yes. [28:04] >> Uh [28:05] >> Dravens [28:06] >> anybody that's been to a dermatologist [28:08] >> Mhm. [28:08] >> it's like looking for potential, um, [28:12] you know, skin cancers, things like [28:14] that, that abnor- abnormalities, has had [28:16] the experience of like the, you know, [28:18] the when you're, you know, that video or [28:20] the meme where the the guy's like [28:21] walking through security and the and the [28:23] security guy's just like, you know, does [28:25] this like really quick [28:26] >> yeah. [28:26] >> pat down. It's like that's basically [28:28] what like a lot of dermatologists are [28:30] doing where they just like take a quick [28:31] glance and they're like, "Oh, yeah, look [28:33] look looks good." [28:34] >> Yeah, they might be busy. They might be [28:35] sleepy. [28:36] But this robot [28:37] >> a robot that just does it like really, [28:39] really precisely, consistently, ideally, [28:42] uh, hopefully, hopefully a lot more [28:44] accessibly. I can see this making a lot [28:46] of sense. And then you actually doesn't [28:48] even replace the job of doctor. It just [28:51] helps them [28:52] speed up the process. [28:53] >> Uh [28:54] where is this company? What stage are [28:56] they? Do they have to go through FDA [28:58] medical [28:59] device approval process? That could be a [29:00] couple years maybe, but this seems like [29:02] once you get it through it's going to be [29:04] massively successful. [29:06] I wonder I wonder where they are in the [29:09] FDA process. Hopefully they can move [29:10] along quickly. [29:12] Arc prize put out an update on Arc-AGI [29:16] V3 which Tyler are you still on the [29:19] human leaderboard for that? Or if [29:21] >> I don't think I'm on the leaderboard [29:22] anymore. That was still early on. Like I [29:23] think there's actually more games now. [29:26] >> Roasted. [29:28] Uh [29:28] well, Arc-AGI V3 has been very tough for [29:33] even the most frontier AI models, but [29:36] the good folks over at Arc prize have [29:39] benchmarked the two latest and greatest [29:42] AI models from OpenAI and Anthropic. [29:44] GPT-3.5 scored 0.43% [29:49] Opus 4.7 scored 0.18% [29:53] >> It is It is state of the art. This is [29:55] the weird thing about these these [29:57] Arc-AGI tests is that they they start [29:59] with such a low baseline 0% 1% and then [30:02] they start they start shooting up as the [30:04] models get more capable, but it is very [30:07] interesting and reassuring to see these [30:09] powerful models that you sometimes hear [30:11] these narratives it can do everything. [30:13] And then there's something that you know [30:15] to to continue working for to look [30:17] forward to in the future. Uh there's [30:19] more applications here. And so they [30:21] found three total failure modes. True [30:24] local effect, false world model, wrong [30:26] level of abstraction from training data. [30:28] Solved the level but didn't reinforce [30:29] the reward. And so the Arc prize account [30:32] goes through [30:33] a little a bit of what's going on when [30:36] an AI model tries to play Arc-AGI V3 [30:38] which is basically a 2D video game. [30:41] Simple enough for any human to progress [30:44] through, but increasingly difficult for [30:46] AI models. [30:47] >> Okay. [30:48] >> Yes. [30:48] >> We got to talk about one of the most [30:50] exciting new products created [30:52] >> What's that? [30:52] >> in the last 100 years. [30:53] >> Tell me. [30:54] >> According to Katie at Business Insider, [30:56] Amazon will now create an AI podcast [30:59] about their products where two AI hosts [31:01] discuss the products and take your [31:03] questions as if it's a calling show. [31:05] >> Yes. [31:06] >> Let's play. [31:07] >> We don't need to play this one. This one [31:08] is so rude, but you can imagine the type [31:11] of products that people are generating [31:14] AI podcasts for. Only the silliest [31:17] things will be generated. I don't know [31:19] if people want to listen to a full [31:21] podcast about every product they buy on [31:22] Amazon, but you have had some very [31:24] strong opinions about [31:26] paper towels. You wanted paper towels [31:28] from like the 1950s or something. [31:30] Remember this whole thing? You [31:31] >> Well, I I just wanted a filter to be [31:33] able to buy only shop for things on [31:35] Amazon. [31:36] >> instead of a filter, you can listen to [31:37] an hour-long podcast about every [31:39] possible skew, and then you can make the [31:42] most determined [31:43] decision available. Go enjoy your [31:46] weekend. Enjoy your Friday. We will see [31:49] you on Monday. [31:50] >> Goodbye. [31:52] >> We love you. [31:52] >> Flash bang.