Anthropic, AMD Sign $5B AI Chip Deal as OpenAI Targets $750B Spend

Anthropic与AMD签署50亿美元AI芯片协议，OpenAI目标支出达7500亿美元 随着领先的人工智能开发者迅速加大数十亿美元级的基础设施承诺，Anthropic将部署高达两吉瓦的AMD Instinct MI450加速器。 主持人在旧金山AMD Advancing AI大会现场带来报道，重点介绍了AMD全新的Helios机架级AI系统、其与Anthropic达成的50亿美元Instinct MI450合作项目，以及与Cerebras的协作。他们分析了Google第二季度财报——重点关注激增的Google Cloud收入并驳斥了有关搜索业务人为虚假增长的说法，同时讨论了OpenAI扩大至7500亿美元的算力支出预期以及DeepSeek的投资者电话会议。此外，本期内容还涵盖了Cognition收购Interaction、年轻创业者利用搜索基金收购具备稳定现金流的小型企业、Meta的AI乐观宣传活动，以及OpenAI的桌面语音和健康功能。 AMD 在 Advancing AI 大会上发布了搭载 Instinct MI450 加速器的全新机架级 AI 系统 Helios，并与 Anthropic 达成部署高达 2 GW 算力、价值约 50 亿美元的战略合作，同时还联合 Cerebras 推出专为 Agentic AI 打造的解耦推理方案。 (0:01) Alphabet 公布 Q2 财报，总营收同比增长 24% 至 1198 亿美元，其中 Google Cloud 暴增 82% 达 248 亿美元；其模型 API 每分钟处理 220 亿 Token，但沉重的数据中心资本支出导致季度自由现金流罕见转负。 (2:33) OpenAI 将截至 2030 年的云算力支出预算提升至 7500 亿美元，并斥资 200 亿美元启动佐治亚州的 Project Camila 数据中心项目；与此同时，DeepSeek 凭借约 85% 的推理毛利率和 10 个月的 GPU 投资回报期实现 10 亿美元 API 营收并达成现金流为正。 (7:38) OpenAI 推出可实时语音操控多个 Agent 的桌面端 Catch Voice 与 ChatGPT Health 健康数据集成功能；而在近期 OpenAI 网络安全事件后，美国国会也正式提出了一项两党联合的“AI Kill Switch”监管法案。 (22:34)

TBPN reports on a massive surge in AI infrastructure commitments alongside strong earnings from tech giants, while examining how young operators are turning to search funds as an alternative to corporate finance paths. Big Hardware Bets and AI Capital Commitments (0:01) Hardware providers and model builders are accelerating multi-billion-dollar infrastructure plans. At its Advancing AI event, AMD unveiled its Helios rack-scale architecture powered by the Instinct MI450 accelerator, alongside an expanded partnership with Cerebras for disaggregated inference. AMD also announced a deal valued at around $5 billion with Anthropic to deploy up to two gigawatts of Instinct MI450 accelerators, leveraging Anthropic's own models to optimize hardware drivers and bridge the historical software gap against Nvidia's ecosystem. Simultaneously, The Wall Street Journal reported that OpenAI raised its projected cloud computing budget through 2030 to $750 billion, up from $600 billion, to support frontier coding agents and foundation models. As part of this buildout, OpenAI committed $20 billion to launch Project Camila, a flagship data center complex in Effingham County, Georgia. In contrast, China's DeepSeek showcased extreme capital efficiency during a four-hour roadshow led by founder Liang Wenfeng. According to Zephr at Kattrini, DeepSeek generated $1 billion in API revenue with approximately 85% inference gross margins on roughly 20,000 Nvidia H100 equivalents, claiming a 10-month GPU payback period while ruling out consumer applications to focus solely on AGI. Google Cloud Expansion and Search Quality Debates (2:33) Alphabet reported Q2 revenue up 24% year-over-year to $119.8 billion, with Google Cloud surging 82% to $24.8 billion. Heavy data center expenditures pushed Alphabet into its first negative quarterly free cash flow in recent memory, offset by over $100 billion in cash and non-marketable equity gains, including a multi-billion-dollar stake in SpaceX. Infrastructure demand reached 22 billion tokens processed per minute across model APIs, largely driven by Gemini Flash, while SEC filings indicated growing Cloud revenue from direct TPU system sales to external AI labs. Search advertising revenue reached $63 billion, up 17% year-over-year, sparking debate over monetization sustainability. Tech commentator Max Anderson argued that search growth is artificial, alleging Google forces higher ad prices to offset traffic lost to non-monetized AI queries. However, ad-tech analyst Souvenirator and Eric Seufert rejected these claims, citing SEC data showing search queries reached record highs during Q1 and the Q2 World Cup, ads in AI Overviews monetize at parity with legacy search, and auction mechanics remain based on standard generalized second-price models. OpenAI Features, Meta Messaging, and Governance (11:58) On the consumer and enterprise application front, Cognition acquired Interaction, the team behind the AI device Poke. OpenAI expanded desktop capabilities by introducing Catch Voice powered by GPT Live, allowing hands-free computer control while directing multiple concurrent agents in ChatGPT and Codex. OpenAI also announced ChatGPT Health, enabling users to integrate Apple Health metrics and medical records for personalized context. These product rollouts coincide with heightened regulatory scrutiny, as a bipartisan AI kill switch bill was introduced in Congress following an OpenAI cybersecurity incident. Meanwhile, Axios reported that Mark Zuckerberg launched a public relations push focused on AI optimism. The campaign positions Meta's AI investments as empowering tools designed to deepen human connection, explicitly rejecting fear-centric narratives promoted by competitors, though Meta continues to face public skepticism regarding the quality of social media feeds. The Search Fund Boom and Business Acquisition (13:00) Beyond software and hardware, young professionals are increasingly bypassing traditional corporate ladders for Entrepreneurship Through Acquisition. Prompted by stalled career paths and difficulty exiting assets in mega-cap private equity, mid-level workers and MBA graduates are raising investor-backed search funds to acquire stable, fragmented service firms such as HVAC outlets, machine shops, and plumbing providers. A 2025 Stanford study identified 77 new search funds raised that year, though the actual count of acquisition vehicles is likely higher. The hosts highlighted the case of Bari Akil, who transitioned from homelessness in 2015 to acquiring two cash-flowing companies through SBA loans, seller notes, and search capital. The segment also examined an alternative model proposed by Sahil Lavingia, who suggested that groups of prospective college students could pool roughly $1 million in cash to acquire a small business, paying themselves living salaries between ages 18 and 22 while developing operational skills rather than spending equivalent sums on traditional tuition. The discussions illustrate an escalating divide in scale: foundation model developers and cloud giants are committing hundreds of billions to long-term computing capacity and agent workflows, while individual operators are using search capital and small-business cash flows to sidestep traditional institutional finance.

TBPN本期节目重点聚焦于人工智能基础设施的剧烈资本扩张与软硬件重构，指出科技巨头与新兴研究机构在数千亿美元级算力投入、芯片生态及变现效率上出现深层分化；同时，商业领域亦兴起绕过传统大学与金融体制、借助“搜索基金”收购实体小企业的新型创富路径。 AMD发布新硬件并深化大模型合作 (0:01) 在旧金山举行的Advancing AI大会上，AMD推出了搭载Instinct MI450加速卡的全新机架级AI系统“Helios”，并证实其芯片正为复古掌机Mod Retro M64提供算力支持。最具行业轰动性的是AMD与Anthropic达成的重大战略合作：Anthropic计划部署高达2吉瓦（GW）的MI450加速卡，该交易估值约为50亿美元。节目强调，这笔交易不仅是算力采购，更涵盖软件栈的底层联合工程研发；Anthropic正利用自身AI模型反哺并加速新硬件的性能调优与爬坡，形成软硬件协同的飞轮效应，以此突破英伟达长期以来的CUDA生态壁垒。此外，AMD进一步扩展了与Cerebras的合作，推出面向智能体AI（agentic AI）的解耦推理方案，通过将专用引擎分配至推理的不同阶段来提升实时性能。 AI基建资本开支激增与商业效率分化 (2:33) 科技巨头在AI基建上的财务表现与资本路径呈现鲜明反差。Alphabet第二季度财报显示总营收同比增长24%至1198亿美元，受向外部AI实验室直销TPU系统的推动，谷歌云营收同比暴增82%至248亿美元，搜索收入达630亿美元，其模型API每分钟处理220亿个Token（主要由Gemini Flash驱动）。尽管巨额数据中心开支导致其近年来首次出现季度自由现金流为负，但公司拥有超1000亿美元现金与股权储备。针对评论员Max Anderson关于谷歌搜索因大模型蚕食而“人为推高劣质广告价格”的指控，分析师Souvenirator援引SEC文件与财报数据予以反驳，指出Q1及Q2世界杯期间谷歌搜索查询量均创历史新高，AI概览广告变现率与传统搜索持平，且继续采用标准广义次高价拍卖机制。 与此同时，据《华尔街日报》报道，受前沿编程智能体和基础模型扩展需求驱动，OpenAI已将截至2030年的云计算预算大幅上调至7500亿美元（此前为6000亿美元），并斥资200亿美元在佐治亚州启动大型数据中心“Project Camila”。与之形成对比的是，中国AI机构DeepSeek创始人梁文锋在长达4小时的投资人路演中展示了截然不同的高资本效率模式：据第三方分析，DeepSeek截至5月仅依托约2万张英伟达H100/Hopper等效芯片，便实现了约85%的推理毛利率与10亿美元API营收，已实现现金流转正，GPU投资回收期短至10个月（硬件按3至5年折旧）。梁文锋明确表示公司绝不开发消费级聊天软件、视频或图像生成以及世界模型，将完全聚焦于实现通用人工智能（AGI）。 终端智能体落地、产品整合与监管介入 (11:58) 在终端软件与产品应用方面，AI软件工程初创团队Cognition（Devin开发方）宣布收购消费级AI硬件Poke的开发团队Interaction，市场普遍关注Poke是保持独立品牌还是被完全融入企业级智能体生态。OpenAI则在桌面端上线了基于GPT Live的“Catch Voice”功能，允许用户通过实时语音调度ChatGPT或Codex中的多个自主智能体协同工作，并同步面向全球推出“ChatGPT Health”，支持安全接入Apple Health数据与电子病历以提供情境化健康解答。在产业文化层面，Meta首席执行官扎克伯格近期启动以“AI乐观主义”为主题的公关宣传，主张AI应赋能人际连接，以抗衡行业中的末日焦虑叙事；传统媒体如派拉蒙也开始试水微短剧，但节目主播测试AI生成的真实犯罪解说内容后指出，AI视频与配音虽已达标，但文本内容扁平空洞，仍缺乏人类讲述者的深度与细节洞察。 随着前沿应用快速铺开，监管审查也随之收紧。在OpenAI近期发生一起网络安全事件后，美国国会两党议员正式提出了一项“AI断路器（kill switch）”法案，旨在建立联邦层面针对先进AI系统的紧急关停机制，该法案目前正引发两党关于监管尺度的激烈争论。 收购传统小企业与搜索基金热潮 (13:00) 除前沿人工智能外，年轻创业者正转向通过“搜索基金”（Search Funds）收购传统现金流实业的“收购创业”（ETA）模式。《华尔街日报》报道了Bari Akil的经历：他曾是一名大学辍学者并经历无家可归，后通过自学哈佛商学院案例与金融播客，利用美国小企业管理局（SBA）贷款、卖方借票及专业投资者资金，收购了暖通空调（HVAC）和管道工程等碎片化传统服务企业，37岁时已拥有两家数百万美元企业并实现数字游民生活。 Gumroad创始人Sahil Lavingia进一步提出替代传统大学教育的设想：四名年轻人可将共计约100万美元的大学学费预算转为资本，通过加杠杆收购一家成熟的现金流小企业，在18至22岁期间以经营者身份自发工资，既能获取实操经验，最坏结果也不过相当于损失学费。此外，受传统大型私募股权机构退出受阻、收益分配机制固化等因素影响，许多MBA毕业生和初中级私募从业者选择出走成立搜索基金。斯坦福大学2025年的一份研究显示，当年新成立的搜索基金达77家，通过独立操盘收购中小型企业以规避大机构体制束缚，正在成为年轻专业人士的新选择。 本期节目描绘了一幅高科技军备竞赛与实体商业实用主义并存的图景：前沿科技巨头在算力集群与智能体生态上展开数千亿美元级的资本押注，并引发了关于商业闭环、算力利用效率及联邦监管断路器的博弈；而另一批创业者则主动避开传统学术体系与华尔街大机构框架，转向传统现金流企业的收购整合。如何在高资本消耗与实际商业回报之间取得平衡，仍是贯穿两类叙事的核心挑战。

## Background 本期内容由科技播客频道 TBPN 于 2026 年 8 月 3 日发布在 YouTube 平台。作为其精简版节目 Diet TBPN，官方简介介绍该节目由 John Coogan 与 Jordi Hays 共同主持，旨在用约 30 分钟提炼科技圈最新动态，曾获《纽约时报》盛赞为“硅谷的新宠”。 ## TL;DR 节目现场直击旧金山 AMD Advancing AI 大会，重点解析了 AMD 发布的 Helios 机架系统及 Anthropic 价值 50 亿美元的算力采购协议，并拆解了谷歌 Q2 财报中 TPU 收入与搜索广告的争议。随后，两位主持人深入讨论了 OpenAI 计划将云基础设施预算上调至 7500 亿美元并在乔治亚州建设超大型数据中心的传闻，以及某中国前沿大模型创始人长达 4 小时的硬核融资路演。此外，节目还涉及 Cognition 收购 Poke 团队、年轻人利用搜索基金（Search Fund）收购实体小企业的致富新路径、扎克伯格主导的“AI 乐观主义”公关叙事，以及美国国会推进“AI 停机开关”法案等最新行业动向。 ## Atmosphere & style 整期节目充满了典型的硅谷极客科技播客氛围，两位主持人在大会现场边喝无糖可乐边侃侃而谈，节奏紧凑明快、互动默契且充满即兴幽默。他们既能熟练切入底层算力、财务报表与代码微架构的专业细节，又能无缝切换到调侃亚特兰大拥堵交通、恶搞大模型创始人马拉松式路演等生动闲聊中。整体听感轻松随性却不失敏锐深刻的行业洞察，将重大科技产业叙事包裹在充满烟火气的硅谷日常对话之中。 ## Bulletin - [00:01](https://www.youtube.com/watch?v=K_4FZywME6U&t=1s) AMD 在 Advancing AI 大会上发布 Helios 机架级 AI 系统与 MI450 加速器，并与 Anthropic 达成 50 亿美元部署 2 GW 算力集群的里程碑式深度合作。 - [02:33](https://www.youtube.com/watch?v=K_4FZywME6U&t=153s) 谷歌 Q2 财报表现强劲，总营收同比增长 24%，云业务增速高达 82%，但狂暴的 AI 数据中心资本开支也让其单季度自由现金流罕见转负。 - [07:37](https://www.youtube.com/watch?v=K_4FZywME6U&t=457s) OpenAI 将到 2030 年的基础设施预算大幅提升至 7,500 亿美元，并投资 200 亿美元在乔治亚州启动“卡米拉项目”超大型数据中心。 - [11:58](https://www.youtube.com/watch?v=K_4FZywME6U&t=718s) AI 程序员 Devin 母公司 Cognition 宣布收购爆款交互产品 Poke 的开发团队 Interaction，将前沿智能体能力与消费级产品交互设计深度结合。 - [13:00](https://www.youtube.com/watch?v=K_4FZywME6U&t=780s) 美国年轻人群体中掀起通过搜索基金（Search Fund）与 SBA 贷款拼凑资金收购暖通空调、管道维修等传统蓝领实体企业的创业新热潮。 ## Key takeaways - [00:01](https://www.youtube.com/watch?v=K_4FZywME6U&t=1s) AMD 发布 Helios 机架系统以正面抗衡英伟达 NVL72，并通过与 Anthropic 价值 50 亿美元的算力采购协议加速底层驱动与 ROCm 软件生态闭环。 - [00:01](https://www.youtube.com/watch?v=K_4FZywME6U&t=1s) AMD 扩大与 Cerebras 的战略合作，共同推出专为智能体 AI 设计的解耦/分离式推理方案以支撑超大规模推理吞吐。 - [02:33](https://www.youtube.com/watch?v=K_4FZywME6U&t=153s) 谷歌云由于 TPU 系统硬件出货量大增而引发外界讨论，其模型 API 吞吐量在 Flash 模型驱动下已突破每分钟 220 亿 Token。 - [02:33](https://www.youtube.com/watch?v=K_4FZywME6U&t=153s) 面对广告主关于 AI 对话蚕食搜索流量及劣质点击的质疑，数据显示谷歌搜索查询量依然处于历史高位，AI Overviews 广告变现率已基本拉平传统搜索。 - [07:37](https://www.youtube.com/watch?v=K_4FZywME6U&t=457s) OpenAI 凭借对推理与代码智能体中“规模定律”的坚定信心，将 2030 年算力资本预算推升至 7,500 亿美元，并落地乔治亚州 200 亿美元数据中心。 - [09:38](https://www.youtube.com/watch?v=K_4FZywME6U&t=578s) DeepSeek 创始人梁文锋在长达 4 小时的海外投资人路演中披露，其团队推理毛利率高达约 85%，GPU 投资回收期仅 10 个月，并坚决不碰视频生成与消费级独立 App。 - [11:58](https://www.youtube.com/watch?v=K_4FZywME6U&t=718s) Devin 开发商 Cognition 出人意料地击败苹果、亚马逊等潜在买家，收购了以精巧人机交互闻名的 Poke 团队 Interaction。 - [13:00](https://www.youtube.com/watch?v=K_4FZywME6U&t=780s) 传统 PE 晋升天梯日渐狭窄，年轻人正通过拼凑 SBA 贷款与微型资金采用 Search Fund 模式收购高现金流传统企业，迅速跻身千万富豪行列。 - [16:09](https://www.youtube.com/watch?v=K_4FZywME6U&t=969s) 替代教育模式引发关注：萨尔·可汗提议由数名年轻人集中大学学费去杠杆收购一家盈利实体，以经营实战替代昂贵的象牙塔学位教育。 - [19:06](https://www.youtube.com/watch?v=K_4FZywME6U&t=1146s) 扎克伯格公开发起“AI 乐观主义”公关攻坚战，直指竞品的末日恐慌营销，但面临公众对社交平台“信息毒药”刻板印象的信任挑战。 - [20:42](https://www.youtube.com/watch?v=K_4FZywME6U&t=1242s) 派拉蒙试水 AI 真实犯罪微短剧暴露出当下技术局限：虽然音画生成合格，但因缺乏主观观察与人性温度而呈现“维基百科式拼凑”的空洞感。 - [22:34](https://www.youtube.com/watch?v=K_4FZywME6U&t=1354s) OpenAI 推出由 GPT Live 驱动的桌面端多智能体实时语音协同控制，并上线可整合苹果健康数据的 ChatGPT Health 服务。 - [22:34](https://www.youtube.com/watch?v=K_4FZywME6U&t=1354s) 美国国会跨党派议员正式提出要求设立“AI 紧急停机开关”的监管法案，折射出联邦政府在激进监管与国家竞争力之间的深层分歧。 ## Detailed walkthrough ### [00:01](https://www.youtube.com/watch?v=K_4FZywME6U&t=1s) 现场直击 AMD Advancing AI：Helios 机架系统与 Anthropic 巨额订单 节目在旧金山 AMD Advancing AI 大会现场拉开帷幕，两位主持人 John Coogan 与 Jordi Hays 难掩兴奋。作为 AMD 年度规格最高的 AI 盛会，大会集中展示了其最新芯片、软件生态合作以及正面抗衡英伟达的具体战略。开场时他们打趣提到，根据科技通讯记者刚刚曝出的消息，AMD 居然在为帕尔默·拉奇（Palmer Luckey）旗下 ModRetro 的复古掌机（ModRetro M64）提供芯片支持。主持人调侃称，稍后专访 AMD CEO 苏姿丰（Lisa Su）时，简直想把整整 30 分钟全用来聊掌机游戏业务，探究他们究竟要如何把各种硬件塞进复古掌机里。 抛开玩笑，大会现场真正引爆产业关注的核心焦点主要集中在三大看点上： 1. **Helios 机架级 AI 系统发布**：AMD 正式揭晓了搭载全新 MI450 加速器的 Helios 机架级 AI 系统，全面向英伟达的 NVL72 机架方案发起冲击。 2. **Anthropic 达成多千兆瓦级深度合作**：AMD 与前沿 AI 实验室 Anthropic 达成里程碑式协议，Anthropic 承诺部署高达 2 GW（吉瓦）的 AMD Instinct MI450 算力集群，交易规模高达 50 亿美元。双方将在底层算力和全栈软件层面展开深度协同。 3. **AI 反哺硬件与生态飞轮**：现场特别谈到，Anthropic 自身的 AI 模型正在显著加速 AMD 新硬件的底层适配与软硬件磨合进程。主持人回顾了此前乔治·霍茨（George Hotz）在调试底层驱动微小 Bug 时与 AMD 开发者团队的直接交锋，认为一旦通过高强度深度合作打通专有闭源包，AMD 的软件生态飞轮正在以前所未有的速度咬合收紧。 4. **携手 Cerebras 推出分离式推理**：AMD 宣布扩大与芯片巨头 Cerebras 的战略合作，共同推出专为智能体 AI（Agentic AI）打造的高性能解耦/分离式推理方案（Disaggregated Inference Solution），将推理流水线的不同阶段精准匹配给最适合的计算引擎，以支撑超大规模的极致推理吞吐。 ```mermaid flowchart TD accTitle: AMD 与 Anthropic 软硬件协同飞轮 accDescr: 展示 Anthropic 大模型如何反哺 AMD MI450 硬件适配与软件栈优化的闭环 subgraph AMD["AMD 算力与硬件层"] MI450["MI450 加速器 / Helios 机架"] ROCm["AMD 底层软件栈 & 驱动包"] end subgraph Anthropic["Anthropic 实验室"] Cluster["2 GW 算力集群部署 ($50亿)"] Claude["Anthropic 前沿模型 (代码与调试)"] end MI450 -->|"交付超大规模算力"| Cluster Cluster -->|"驱动模型训练与推理"| Claude Claude -->|"AI 辅助代码修复 & 硬件磨合"| ROCm ROCm -->|"加速新硬件投产迭代"| MI450 ``` --- ### [02:33](https://www.youtube.com/watch?v=K_4FZywME6U&t=153s) 谷歌 Q2 财报透视：云业务狂飙、TPU 收入疑云与搜索广告争论 视角转向科技巨头的二季报战况。《华尔街日报》以“谷歌拥有压倒资本支出担忧的硬实力”为题展开报道，谷歌云业务交出了极其亮眼的增长答卷。不过，知名半导体分析机构 SemiAnalysis 在社交平台上的一句调侃引发了轩然大波——他们宣称“谷歌云现在的收入主要来自销售 TPU 系统”。主持人澄清道，这源于谷歌在 SEC 监管文件中一行极具迷惑性的措辞变动：“谷歌云产生的‘产品收入’（Product revenues）主要来自 TPU 系统的销售”。大众容易将“产品”等同于整体业务，但实际上云厂商还有体量巨大的“服务收入”（Services revenues）。这并不意味着谷歌云彻底转向卖硬件，但不可否认的是，TPU 正在被越来越多顶级 AI 实验室采购，其作为独立硬件销售的体量已经大到具备实质性财务意义。 与此同时，谷歌 CEO 桑达尔·皮查伊（Sundar Pichai）宣布了极其强劲的整体业绩，称“第二季度是令人惊叹的季度，AI 投资正在重塑我们每一项业务的边界”。数据显示，Alphabet 整体营收同比增长 24%，谷歌云增速更是惊人地飙升至 82%。在基础模型调用方面，谷歌的模型 API 每分钟处理的 Token 吞吐量已突破 220 亿（22 billion tokens/minute），这主要由其高性价比的 Flash 模型驱动。 然而，强劲的数据也引来了外部广告主的强烈质疑。长期月度广告投放超 50 万美元的资深从业者 Max Anderson 发长文炮轰，断言“搜索业务的营收增长是人造且极端不健康的”： - 他认为传统关键词搜索量正在被免费的 LLM 对话查询无情蚕食。 - 谷歌为了掩盖流量流失，采取了“目光短浅且敌视客户的榨取手段”，向广告主收取更高费用，却分配了大量未经授权且转化极低的劣质点击。 针对这一恐慌舆论，广告科技专家（Souvenirator）随后出面进行全面事实反驳： - **搜索总量并未下滑**：谷歌二季报明确指出，受今年世界杯等事件带动，搜索使用量创下历史新高，一季报查询量同样处于历史峰值。 - **AI 摘要变现率拉齐**：AI Overviews（AI 搜索摘要）中的广告变现效率已经与传统搜索广告基本持平；谷歌搜索收入在 Q1 增长 19%、Q2 增长 17%，虽有小幅放缓但底盘坚固。 - **竞价机制未做隐秘篡改**：所谓的“谷歌悄然废除广义次高价拍卖（Generalized Second Price Auction）”属于不实谣言，搜索竞价体系依然维持既有机制。 在整体财务数字上，Alphabet 本季总收入达 1,198 亿美元，云业务贡献 248 亿美元，核心搜索业务贡献 630 亿美元，净利润高达 1,120 亿美元。不过，净利润的暴增很大一部分来自其持有的外部股权估值变动（例如其持有的 SpaceX 股份目前估值接近千亿美元）。而狂暴的 AI 数据中心基础设施开支，也直接将谷歌本季度的自由现金流拉入了负数区间（首次出现转负），但这早已在华尔街的资本开支预期之中。 --- ### [07:37](https://www.youtube.com/watch?v=K_4FZywME6U&t=457s) 基础设施军备竞赛：OpenAI 预算调至 7500 亿美元与乔治亚数据中心 《华尔街日报》披露的另一则重磅消息将 AI 基础设施的军备竞赛推向了全新高度：OpenAI 大幅上调了其算力开支预期，计划到 2030 年将算力基础设施预算大幅提升至约 7,500 亿美元，较此前约 6,000 亿美元的预估大幅跨越。近期代码智能体（Coding Agents）所展现出的飞跃式生产力，让管理层对“规模定律”（Scaling Laws）在推理与智能体时代的持续有效性深信不疑。 为了在这场算力争夺战中锁定海量资源，OpenAI 正在紧锣密鼓地与各大云服务商签订全新锁定协议，算力资本开支已经成为 CEO 山姆·奥特曼（Sam Altman）领导团队的核心战略焦点。作为具体落地动作之一，OpenAI 宣布将投资 200 亿美元在乔治亚州埃芬厄姆县（Effingham County, Georgia）启动名为“卡米拉项目”（Project Camila）的全新超大型数据中心。 聊到乔治亚州，两位主持人开启了轻松的闲聊日常，调侃起亚特兰大那堪比洛杉矶的拥堵交通：“不管你去哪儿，都得老老实实堵在车里半小时以上，但这座城市确实非常棒”。节目最后，话题一转切向大洋彼岸的中国 AI 力量——量化对冲基金幻方背景的 DeepSeek 创始人梁文锋刚刚主持了一场投资人电话会，这也标志着此前一直依靠自有资金闭门攻坚的 DeepSeek，如今正在罕见地开启外部资本融资。 ### [09:38](https://www.youtube.com/watch?v=K_4FZywME6U&t=578s) 中国大模型创始人长达4小时的投资人路演与财务内幕 节目开场谈及一家正在海外筹集外部资金的 AI 独角兽公司。为了完成巨额融资，其创始人直接在 Zoom 会议上展开了一场堪称马拉松式的“梦想兜售”。令人惊叹的是，这位创始人竟然一个人连续发表了整整 4 个小时的独白。主持人打趣道，这种表现简直就像是录了一整期《乔·罗根体验》（Joe Rogan Experience）播客——毕竟正如罗根播客的规律一样，任何人只要聊过一个小时，事先准备的讲稿和公关套路就会被消耗殆尽，真正展露出未经修饰的真实想法与深度。能够连续独白 4 个小时依然保持强大吸引力，这种精力本身就足以给投资人注入极大的牛市信心（"bullish"）。 研究机构 Citrini（原音误作 Catrini）的 Zephr 在社交媒体上披露了这场长达数小时路演的大量核心财务与算力数据： - **推理毛利率高企**：该团队的推理利润率达到了惊人的 85% 左右。 - **算力储备与折旧**：截至当年 5 月，他们仅拥有约 20,000 张英伟达 Hopper 等效显卡（不过主持人指出，他们可能在全球范围内以各种合规或灰色渠道获得了更多电力与算力支持）。硬件资产按 3 至 5 年进行折旧。 - **GPU 投资回收期与造血能力**：GPU 的资本回收周期仅为 10 个月；只要 API 收入达到 10 亿美元，公司即可实现正向自由现金流。 在发展理念上，这位创始人展现出了与西方主流 AI 实验室领袖高度趋同的话术哲学。他明确表示公司不会把“单纯追求利润”作为当前目标，而是坚定践行“AGI 优先”（path to AGI first）。更为激进的是，他划定了极为克制的业务红线：公司绝不涉足世界模型、图像生成、视频生成，甚至不会去做消费级独立聊天 App。他直言自己“不想成为下一个阿里巴巴”，而是要用最精简纯粹的路线直奔通用人工智能。主持人回顾道，在 Moonshot（月之暗面，凭借 Kimi/K3 模型引发轰动）以及智谱 GLM 纷纷崛起的激烈竞争背景下，这家曾略显低调的团队如何在这场全球瞩目的路演后出招，将非常值得观察。 --- ### [11:58](https://www.youtube.com/watch?v=K_4FZywME6U&t=718s) 轰动行业的消息：Cognition 收购 Poke 开发团队 Interaction 随着标志性的敲锣声（"Gong moment"），主持人宣布了一笔重磅并购：推出首个 AI 软件工程师 Devin 的独角兽公司 Cognition 正式收购了打造出爆款产品 Poke 的团队 Interaction。 在科技圈内部，Poke 很久以来就被普遍视作极具吸引力的潜在收购标的。他们以打造新颖、极具交互愉悦感（"delightful novel products"）的消费级产品而闻名。业内原先普遍预期 Poke 团队会被苹果这样的大型消费电子巨头招致麾下——如果能让 Poke 团队去重构备受诟病的 Siri，无疑会是令人兴奋的绝配；此外，OpenAI 或是亚马逊也曾被视作合理的买家。 出乎外界预料的是，最终拔得头筹的竟然是主打程序员智能体的 Cognition。主持人指出，Cognition 创始人兼 CEO 斯科特·吴（Scott Wu）显然看中了该团队在人机交互与产品构想上的罕见天赋。目前尚不清楚收购后 Poke 产品是会继续独立运营，还是其团队与技术将被彻底融入 Cognition 的主营生态中，但两家极具极客气质的团队合并依然让科技界对后续产出充满期待。 --- ### [13:00](https://www.youtube.com/watch?v=K_4FZywME6U&t=780s) 逃离传统PE：通过 Search Fund 收购传统行业成为年轻富豪 节目随后翻出了一篇早前刊发于《华尔街日报》的精彩特写，讲述年轻人如何仅凭听播客、自学金融逻辑，依靠搜索基金（Search Fund）在传统的暖通空调（HVAC）和管道等实体行业里白手起家成为年轻千万富翁的真实经历。 故事的主人公巴里·阿基尔（Bari Akil）在 2015 年还是一名无家可归的大学辍学生，曾流落于 WeWork、地铁车厢以及机场候机大厅过夜。身无分文的他唯一的执念就是如何赚大钱。在辗转于各种个人理财播客、视频与书籍的过程中，阿基尔偶然读到了一篇哈佛商学院（HBS）的经典案例研究，获悉顶尖 MBA 毕业生正通过一种名为“Search Fund”（搜索基金）的模式，募资去全美各地收购成熟的中小微企业。他感觉自己窥破了商业世界的终极秘密。 十年之后，现年 37 岁的阿基尔已经先后收购了两家价值数百万美元的实体公司，个人净资产跻身七位数，过去三年甚至过上了每月换一个国家居住的数字游民生活。像他这样的年轻人正在掀起一股热潮：他们抛弃了传统投行与私募基金中“从底薪做表分析师（spreadsheet-wielding associate）熬到董事总经理（MD）”的漫长晋升天梯，转而扑向利润丰厚但极其分散的管道维修、HVAC、精密机械加工等蓝领中小企业。 与传统私募动辄撬动机构投资者与主权财富基金的庞大并购不同，这些草根创业者采用的是“一单一审（deal-by-deal）”的拼凑式融资： ```mermaid flowchart TD accTitle: Search Fund Acquisition Flow accDescr: Diagram illustrating how search fund entrepreneurs cobble together financing to acquire traditional small businesses. subgraph FundRaising["资金拼凑与融资端"] A["美国小企业管理局 (SBA) 贷款"] B["专业搜索基金投资人"] C["亲友资金 / 家族办公室"] D["持牌小企业投资公司 (SBIC)"] end subgraph DealExecution["收购与落地运营"] E["目标标的：HVAC / 管道 / 专用制造"] F["说服老业主转让企业所有权"] G["实施运营改造与数字化升级"] end subgraph Payoff["现金流与回报"] H["企业稳定净现金流"] I["偿还杠杆债务 + 支付高管年薪 + 股权增值"] end A --> E B --> E C --> E D --> E E --> F F --> G G --> H H --> I ``` 知名律所 Gibson, Dunn & Crutcher 曾代理大型私募机构的前律师阿尔布雷希特（Albrecht）给出了极具画面感的评价： > “我把这些人视作蛮荒西部的拓荒先驱（homesteaders of the Wild West）。他们口袋里没有一分钱就孤身出发，但只要能活着回来，前面就有泼天的财富等着他们。” --- ### [16:09](https://www.youtube.com/watch?v=K_4FZywME6U&t=969s) 读大学不如买家公司：重构教育与草根并购的热潮 这一现象引发了主持人对替代教育方案的深度探讨。主持人联想到可汗学院（Khan Academy）创始人萨尔·可汗（Sal Khan）近期提出的前瞻性倡议：与其让孩子去念昂贵且充满泡沫的大学，家长倒不如直接为孩子买下一家实体企业，或者让几个同龄年轻人合伙买一家公司： - **大学学费的机会成本**：如今一所美国大学的开销动辄达到 20 万至 25 万美元。 - **4人团队的收购模型**：如果将 4 个原本打算念大学的学生的学费集中起来，就能凑出 100 万美元的硬核现金资产。通过金融杠杆，这笔钱足以收购一家每年产生数十万美元稳定净利润的真实小企业。 - **学以致用与现金流闭环**：4 个年轻人在 18 岁到 22 岁之间直接组成公司的执行管理团队，企业的经营现金流直接支付他们自身的工资。他们不再在象牙塔里纸上谈兵，而是在市场实战中边干边学销售、财务、人事与供应链管理。 - **风险底线对冲**：最好的结果是获得一家盈利强劲、具备长久复利价值的企业；即使最坏的结果是企业破产清算，他们损失的钱也仅仅等同于那笔横竖都要交掉的大学学费，却换来了无比扎实的实操经验。 尽管主持人承认顶级名校的光环与人脉短期内依然不可替代，但在顶尖商学院内部，通过收购进行创业（Entrepreneurship through Acquisition, ETA）的课程早已成为抢手标配。推动这股热潮的核心动力，还在于华尔街神话的破灭——头部老牌私募早已被元老合伙人牢牢把持，行业退出与企业脱手陷入困局，中层员工很难拿到业绩分红，被迫出走自立门户。 斯坦福大学在 2025 年的一项年度调研中指出，当年全美新募集的传统独立 Search Fund 共有 77 支。节目主持人们一致认为这个统计数字显得过于保守，根本无法反映 2025 年整个创投圈对该模式的追捧热度。他们推测，大量实际从事此类杠杆微型并购的年轻人，很可能并未在机构处将自己注册为标准的 Search Fund，而是直接以微型私募机构自居，实际在暗中攻城略地。 ### [19:06](https://www.youtube.com/watch?v=K_4FZywME6U&t=1146s) 扎克伯格押注“AI 乐观主义”：对抗末日恐慌叙事与公众对“信息毒药”的顾虑 节目老朋友马克·扎克伯格（Mark Zuckerberg）近期展开了一场聚焦于“AI 乐观主义”（AI Optimism）的全新公关宣传。根据 Axios 的报道，扎克伯格系统阐述了他对智能体（Agentic）未来的乐观愿景，强调 Meta“连接世界”的核心使命只会被全新 AI 工具与技术进一步强化。Axios 指出，扎克伯格此举意在与部分竞争对手形成鲜明对比——在随发言发布的视频广告中，Meta 直指某些竞品不断散播恐惧心理与反乌托邦式的未来灾难论调。 主持人对这一宣传片给出了积极评价，认为站在乐观主义视角显然比迎合末日恐慌明智得多，并庆幸扎克伯格没有在模因上盲目跟风走上以恐惧为导向的老路。但两人也敏锐地剖析了受众心理的潜在反弹：对于重度社交用户来说，Meta 确实拉近了彼此距离（比如私信分享搞笑 Reels 短视频并乐在其中）；然而在大量批评者眼中，现代社交平台已被视作消耗注意力的“脑残信息”（brain rot）。一旦大众产生“如果你们要拿做社交媒体的方式去改造下一代技术”的戒备心理，就很难全然买账。尽管如此，从传播与公关角度看，传递乐观信念依然是远优于散布恐慌的最佳策略。 ### [20:42](https://www.youtube.com/watch?v=K_4FZywME6U&t=1242s) 派拉蒙试水微短剧与 AI 内容落地：缺乏主观视角的“维基百科式拼凑” 传统影视巨头派拉蒙（Paramount）正大举进军微短剧（micro-dramas）领域，这一动向引发了关于当下 AI 生成短视频与竖屏剧生态的讨论。主持人分享了自己观看真实犯罪类（True Crime）AI 短剧的切身体验：从纯技术表现来看，AI 生成的视频画面与画外音旁白虽然仍能看出技术痕迹且略显僵硬，但基本及格；然而作品真正致命的缺失在于“叙事的人性温度”。 在经典的真人真实犯罪播客中，受众最着迷的是跟随着主持人的思考漫游——主持人会按照个人趣味，挖掘案发小镇鲜为人知的冷门细节与背景轶事，赋予故事独一无二的质感。相比之下，目前的 AI 短剧更像是把维基百科条目或新闻通稿粗暴压缩提炼后的产物，内容平淡空洞、缺乏新鲜事实与深度叙事切入点。主持人感叹，去实体电影院看横屏长片依然是一种对抗短平快“信息毒药”的不可替代体验，同时调侃称纽约布鲁克林甚至已经有人把横屏影院反向利用，专门包场组织大家一起在大银幕上看 Instagram Reels 和 TikTok 短视频展映。 ### [22:34](https://www.youtube.com/watch?v=K_4FZywME6U&t=1354s) OpenAI 密集更新与国会“AI 停机开关”法案推进 OpenAI 当天的动态引发了社区密集猜测。OpenAI 成员 Thibault 在社交平台发文称“为即将到来的一切感到无比兴奋，明天感觉非常 Codex”（"Tomorrow is feeling codexy"），随后引发了社区对极速推理芯片 Cerebras 或更高速 Turbo 模式的广泛猜想。随后释出的具体功能涵盖了桌面端的多模态升级与个人健康整合，展现出 OpenAI 进一步接管用户操作系统的野心： ```mermaid flowchart TD accTitle: OpenAI Desktop and Service Expansion accDescr: Diagram showing desktop voice controlling multi-agent workflows and ChatGPT Health integrating medical data. subgraph Voice_Agents["桌面端交互与多智能体协调"] A["GPT Live 实时语音引擎"] --> B["桌面端语音控制"] B --> C["系统级权限与操作控制"] C --> D["调度 ChatGPT 工作区与 Codex 协同"] end subgraph Data_Integration["个人健康数据生态"] E["ChatGPT Health"] --> F["连接 Apple Health"] E --> G["连接支持的电子病历"] F & G --> H["构建个人私域健康上下文"] end ``` 在监管与政策前沿，针对此前 OpenAI 遭受的网络安全事件，美国国会两党议员正式提出了一项要求设定“AI 紧急停机开关”（AI kill switch）的跨党派法案。该法案的递交折射出国会内部在联邦政府是否应当激进干预、严格监管人工智能系统上的严重分歧。 - **桌面端实时语音全面铺开**：由底层实时模型 GPT Live 驱动，支持用户边听、边说并同时在应用内协调复杂的电脑端操作，可纯语音直接指挥 ChatGPT 工作区与 Codex 内运行的多个智能体。 - **ChatGPT Health 扩展上线**：面向更广泛用户推出健康连接能力，允许安全接入苹果健康（Apple Health）以及受支持的第三方医疗档案系统，使模型能基于个人生理数据提供个性化上下文支持。 - **TBPN 现场收官与行程**：本期节目由 TBPN 团队前往 AMD 现场实地直播完成，感谢了 AMD 团队的硬件与场地支持。主播预告团队次日返程，将于下周一太平洋时间上午 11 点整照常开播。 ## Notable quotes > [02:33](https://www.youtube.com/watch?v=K_4FZywME6U&t=153s) 谷歌拥有压倒资本支出担忧的硬实力 — 《华尔街日报》 > [02:33](https://www.youtube.com/watch?v=K_4FZywME6U&t=153s) 谷歌云产生的‘产品收入’（Product revenues）主要来自 TPU 系统的销售 — 谷歌监管文件 > [02:33](https://www.youtube.com/watch?v=K_4FZywME6U&t=153s) 第二季度是令人惊叹的季度，AI 投资正在重塑我们每一项业务的边界 — Sundar Pichai > [02:33](https://www.youtube.com/watch?v=K_4FZywME6U&t=153s) 搜索业务的营收增长是人造且极端不健康的 — Max Anderson > [13:00](https://www.youtube.com/watch?v=K_4FZywME6U&t=780s) 我把这些人视作蛮荒西部的拓荒先驱（homesteaders of the Wild West）。他们口袋里没有一分钱就孤身出发，但只要能活着回来，前面就有泼天的财富等着他们。 — Albrecht > [22:34](https://www.youtube.com/watch?v=K_4FZywME6U&t=1354s) Tomorrow is feeling codexy — Thibault > [07:37](https://www.youtube.com/watch?v=K_4FZywME6U&t=457s) 不管你去哪儿，都得老老实实堵在车里半小时以上，但这座城市确实非常棒 — John Coogan / Jordi Hays > [09:38](https://www.youtube.com/watch?v=K_4FZywME6U&t=578s) 不想成为下一个阿里巴巴 — 梁文锋 ## Mentioned AMD, John Coogan, Jordi Hays, Palmer Luckey, ModRetro, ModRetro M64, Lisa Su, Nvidia, Helios, MI450, NVL72, Anthropic, Instinct MI450, George Hotz, ROCm, Cerebras, Google, Alphabet, The Wall Street Journal, SemiAnalysis, TPU, Sundar Pichai, Gemini Flash, Max Anderson, Souvenirator, AI Overviews, SpaceX, OpenAI, Sam Altman, Project Camila, DeepSeek, 梁文锋, Joe Rogan, Citrini, Moonshot AI, Kimi, GLM, Cognition, Devin, Interaction, Poke, Apple, Siri, Amazon, Scott Wu, Bari Akil, WeWork, Harvard Business School, Search Fund, SBA, SBIC, Gibson, Dunn & Crutcher, Albrecht, Sal Khan, Khan Academy, Stanford University, Mark Zuckerberg, Meta, Axios, Instagram Reels, Paramount, TikTok, Thibault, Codex, GPT Live, ChatGPT Health, Apple Health ## Video description Sign up for TBPN’s daily newsletter at TBPN.com 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. TBPN is made possible by: Ramp - https://ramp.com Public - https://public.com Cisco - https://www.cisco.com Console - https://www.console.com CrowdStrike - https://www.crowdstrike.com Figma - https://www.figma.com MongoDB - https://www.mongodb.com NYSE - https://www.nyse.com Railway - https://railway.com Shopify - https://www.shopify.com/ 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=K_4FZywME6U)

[2:33] 《华尔街日报》: 谷歌拥有压倒资本支出担忧的硬实力 [2:33] 谷歌监管文件: 谷歌云产生的‘产品收入’（Product revenues）主要来自 TPU 系统的销售 [2:33] Sundar Pichai: 第二季度是令人惊叹的季度，AI 投资正在重塑我们每一项业务的边界 [2:33] Max Anderson: 搜索业务的营收增长是人造且极端不健康的 [13:00] Albrecht: 我把这些人视作蛮荒西部的拓荒先驱（homesteaders of the Wild West）。他们口袋里没有一分钱就孤身出发，但只要能活着回来，前面就有泼天的财富等着他们。 [22:34] Thibault: Tomorrow is feeling codexy [7:37] John Coogan / Jordi Hays: 不管你去哪儿，都得老老实实堵在车里半小时以上，但这座城市确实非常棒 [9:38] 梁文锋: 不想成为下一个阿里巴巴

The New York Times: Silicon Valley’s newest obsession [2:33] Max Anderson: artificial and extremely unhealthy [17:35] Albertch: the homesteaders of the Wild West [22:34] Thibault: Tomorrow is feeling Codexy

[0:01] Thank you to the AMD team for having us [0:03] here at AMD Advancing AI. Uh we're on [0:06] the road in SF today. Uh but let's you [0:09] take you through a little bit of the [0:10] news of what AMD is announcing. What's [0:12] happening today? [0:13] >> Uh Advancing AI is AMD's biggest AI [0:16] event of the year where it reveals new [0:17] chips, software partnerships, and how it [0:19] plans to compete with Nvidia. According [0:21] to Jackson Force newsletter, [0:23] >> we just found this out this morning. AMD [0:26] is actually powering the new [0:27] >> mod retro M64. [0:29] >> Very cool. [0:30] >> It's very Yeah. [0:31] >> Crossover Palmer Lisa. [0:33] >> Yeah. So, we're going to be spending all [0:34] 30 minutes of our time with Lisa Sue on [0:37] gaming specifically. Exactly. How is how [0:39] are they going to bring suits and suits? [0:40] Yes. [0:42] >> Uh so the three key stories that people [0:44] are focused on out of AMD advancing AI [0:47] uh are AMD unveil unveils Helios the new [0:50] rack scale AI system powered by the [0:52] MI450 accelerator. Uh we'll talk to Lisa [0:55] about that. Also a new partnership with [0:58] a Frontier Lab. AMD and Anthropic are [1:00] partnering to deploy up to two gigawatts [1:02] of AMD Instinct Mi450s. [1:05] >> Sizegong [1:06] >> size five billion dollar deal uh and a [1:09] whole bunch of collaboration across [1:11] compute and the software stack. The big [1:13] question around the software ecosystem [1:15] around AMD is always a hot topic. I'm [1:18] sure there's a ton of developments [1:19] there. [1:20] stage just a little bit ago they were [1:22] talking about how um Anthropic's own [1:24] models are effectively speeding up their [1:26] ability to ramp up new hardware which is [1:29] something that uh has been talked about [1:31] for years now some questions around yeah [1:34] uh the coupe right uh if you assume that [1:38] making great software will just get [1:40] easier and easier uh that that's been a [1:43] big question uh and they're starting to [1:45] see uh real movement there [1:47] >> yeah but also a lot of that speed up [1:49] requires actual deep partnership like [1:51] this because there are some close source [1:54] packages that you might need to uh you [1:56] might need to work on. This is the old [1:57] George hot interaction y um talking [1:59] about little bugs that he was frustrated [2:02] with wound up working with AMD developer [2:04] team who we've talked to on the team on [2:06] the show. Um fantastic outcome there but [2:08] interesting uh to see that flywheel just [2:11] get tighter and tighter and tighter and [2:12] I expect it will. Uh also AMD is [2:14] expanding its partnership with Cerebrus [2:16] teaming up on AI inference. Uh we love [2:18] the team over at Cerebras. Uh they [2:20] introduced a powerful disagregated [2:22] inference solution pairing the right [2:23] engine to each phase of the inference [2:25] pipeline. Uh this is what agentic AI has [2:28] been waiting for. The fastest production [2:31] inference at massive scale. Google has [2:33] the muscle to overpower spending worries [2:35] says the Wall Street Journal. Google has [2:38] put up uh some very uh impressive cloud [2:42] growth numbers. There of course was some [2:44] back and forth on the nature of that [2:47] revenue uh with semi analysis sort of [2:49] jokingly maybe posting Google cloud [2:52] generates revenue primarily from the [2:54] sale of TPU systems. The actual line [2:56] that did change in the SEC filing that's [2:59] drawing attention is that uh there's a [3:02] line in there that says Google Cloud [3:04] generates product revenues primarily [3:07] from the sale of TPU systems which is a [3:09] little confusing because you think of a [3:11] company's product as their business but [3:12] of course they have services as well and [3:15] so the it's not that Google cloud is [3:17] fully pivoting to TPU sales exclusively [3:21] it is that uh they [3:22] >> it's becoming meaningful [3:23] >> yes yes yes They they've sold them to a [3:26] number of labs and they're seeing lots [3:28] of process. [3:28] >> Rewind a year though that was still a [3:30] question. [3:30] >> Yes. Yes. Um we can go through uh [3:36] there's a number of different uh posts [3:37] here. We can take you through uh Eric [3:40] Suffort talking back and forth about uh [3:43] souvenirators [3:44] uh putting a little bit of truth zone on [3:47] uh Google ads growth. What's going on [3:49] there? They saw a lot of uh growth in [3:52] monetized LLM queries. It all started [3:54] with Sundar Pachai's news. He said, [3:57] "Que2 was an amazing quarter with our AI [3:59] investments redefining what's possible [4:01] across every part of our business. [4:03] Alphabet revenue grew 24% year-over-year [4:05] and Google Cloud accelerated to 82% [4:07] growth, which is insane. That needs the [4:09] fog horn for sure. Uh we saw exciting [4:12] momentum across the board from search to [4:14] YouTube to the Gemini app. Our model [4:16] APIs are processing 22 billion tokens a [4:18] minute. And at this point, I'm like the [4:21] numbers have gotten so big on tokens per [4:23] minute, tokens per month that it's very [4:25] hard to keep uh keep in sync with what's [4:28] actually uh impressive until you [4:30] actually see a full chart. But that's [4:31] driven by their workhorse flash model, [4:33] which makes sense. They've been focused [4:34] on very efficient token generation [4:36] because they uh vend them into so many [4:38] Google systems for effectively free [4:40] because they're still, you know, at [4:41] monetizing via advertising. So, Max [4:44] Anderson said, "As someone who has [4:46] personally spent $500,000 a month plus [4:49] on Google ads for years, I can tell you [4:50] with certainty, [4:52] >> this revenue growth in search is [4:54] artificial and extremely unhealthy for [4:56] Google's business long term." [4:58] >> Little bit of fear, uncertainty, doubt. [5:00] Search volumes are declining as legacy [5:02] searches being increasingly cannibalized [5:04] by non-monetized LLM queries. Google's [5:06] response: manufacture revenue growth via [5:08] short-sighted, highly extractive [5:10] customer hostile tactics, i.e. charge [5:12] advertisers more for lower quality [5:14] clicks, including clicks they do not [5:16] want and explicitly do not approve [5:18] Google to charge them for. Uh he gives a [5:20] few examples. He's going back and forth [5:22] saying that uh this is uh you know [5:25] shortterm optimization, but the [5:27] souvenirator comes in. He says, [5:28] "Unfortunately, every claim made in this [5:30] viral thread is either false or mix [5:32] characterizes Google's policy. One, [5:34] search volumes are declining as legacy [5:37] search is being increasingly [5:38] cannibalized by non-monetized LM [5:40] queries." Uh, Google stated in its Q2 [5:42] earnings that search usage hit an [5:43] all-time high during the World Cup this [5:45] year. It also stated in Q1 earnings that [5:47] search queries reached an all-time high [5:49] that quarter. Ads and AI overviews [5:51] monetize in parody with those in legacy [5:53] search and Google's search revenue [5:54] increased by 19% uh and uh in Q1 and 17% [5:59] in Q2. So, a little deceleration but [6:01] still growing. He also debunks the idea [6:03] that Google silently deprecated the [6:06] second price auction. Google still uses [6:08] generalized second price auction for [6:10] search. He also debunks the idea that [6:12] Google previously had precise keyword [6:14] targeting settings that allowed [6:15] advertisers to pick individual search [6:18] phrases to bid on. Uh there's a [6:20] different thing, but Souvenator goes [6:22] back and forth and uh and adds some more [6:24] context to the Google earnings that [6:26] people are digging into. And we can dig [6:27] into more later this week. Over in the [6:30] Wall Street Journal, there's a couple [6:32] articles. Uh, Alphabet rides cloud [6:34] business revenue jumps 24%. But data [6:37] center outlays push cash flow into the [6:39] red. I think it was the first quarter [6:41] for negative cash flow, but this has [6:42] been uh telegraph for a long time. [6:44] They're investing in the AI boom, not a [6:47] surprise. And there's been talks from [6:49] all the hyperscalers around debt [6:50] issuances, equity issuances, all [6:52] different things to fund the AI [6:54] buildout. So, Google parent company [6:56] Alphabet posted 24% revenue growth [6:58] year-over-year in the second quarter, [6:59] fueled by its booming cloud business. [7:01] But concern over the company's heavy [7:03] spending on AI infrastructure damped [7:05] investors enthusiasm. Alphabet sales [7:08] came in at 119.8 [7:10] billion exceeding analyst expectations. [7:13] Its cloud business brought in 24.8 [7:15] billion and its search business [7:16] generated 63 billion. Uh net income was [7:21] 112 billion uh for the April June period [7:24] also ahead of analyst expectations. So [7:27] they beat uh primarily driven by its [7:29] gains in other companies stocks that [7:31] Alphabet owns. They own I think almost a [7:34] hundred billion dollars of SpaceX. I'm [7:36] not sure about the current valuation. [7:37] But in other AI news in the Wall Street [7:40] Journal, uh two big uh AI infrastructure [7:45] stories, you have Lisa Sue herself right [7:48] here. Uh AMD anthropic sign AI gear [7:51] deal. Uh we talked to AJ a little bit [7:53] about that. OpenAI boosts cloud spend to [7:56] 750 billion. New projects and Georgia [8:00] data center uh will be part of efforts [8:02] to increase capacity. OpenAI is scaling [8:04] up its data center ambitions and its [8:06] budget for spending on them. The [8:07] artificial intelligence company has [8:09] raised its projected spending uh on [8:12] computing power to around 750 billion [8:15] through 2030 up from a projection of [8:18] roughly 600 billion is a huge number. [8:20] Obviously uh a lot of discussion about [8:23] that but given the most recent leg up [8:25] based on coding agents uh a lot of [8:27] people are looking at that positively as [8:30] a sign of continued uh faith in the [8:33] progress and uh the scaling law of [8:35] course because we are in the scaling era [8:37] >> serious momentum. [8:38] >> Yeah the scaling uh the the increase [8:41] reflects new agreements with cloud [8:42] computing providers as open AI races to [8:44] lock up the enormous amounts of [8:46] computing capacity it needs to develop [8:47] and run its AI models. Open AAI spending [8:49] on cloud computing has become a central [8:51] focus of chief executive Sam Alman's [8:53] leadership team and has been a source of [8:55] tension between him. Um the company said [8:58] Wednesday would invest 20 billion in uh [9:01] to kick off a data center called Project [9:03] Camila in Effingham County, Georgia. So [9:07] that's new news. Also Nvidia has a [9:09] supplier that's over [9:10] >> been to Georgia. [9:11] >> Yes, I've been to Atlanta. It's a great [9:12] town. [9:13] >> Got to visit. [9:14] >> You've never been? [9:14] >> Maybe on a layover. It's like I don't [9:16] think I've been outside. [9:17] >> It's like Los Angeles. You're in the car [9:18] for half an hour driving in the traffic [9:20] to get everywhere everywhere. [9:21] >> But it's uh it's a really nice city. [9:23] >> Leang, when Bung over at Deepseek had an [9:26] investor call, I I don't have context on [9:29] how this [9:31] >> uh actually came up because they're [9:33] >> they're a hedge fund and they have [9:35] Deepseek still a private [9:37] >> They're raising outside capital now. [9:38] >> They're raising outside capital. Yeah. [9:40] >> So, this was probably part of that road [9:42] show. I think at a certain point he [9:43] needs to raise enough money that you got [9:45] to just get on Zoom and uh sell the [9:47] dream. [9:47] >> But he was grinding for 4 hours, just [9:49] monologuing for four. I don't know who's [9:51] monologuing for all four, but it is it [9:53] is sort of a long call. Uh but [9:55] >> that's bullish though. [9:56] >> It is [9:56] >> because if you can go and basically do a [9:58] Joe Rogan podcast episode just [10:00] monologuing. [10:01] >> Uh no, it's just like can somebody [10:03] actually carry a conversation for that [10:05] long? Because after, you know, this has [10:08] always been your take on Joe Rogan. It's [10:09] like it gets interesting on Joe Rogan [10:11] because by the the a after an hour [10:14] people go off script. Yeah. [10:16] >> And so if a CEO can carry a conversation [10:18] for four hours straight. Yeah. [10:20] >> And still be compelling, it's uh it's [10:22] bullish. But there are some details that [10:23] Zephr uh shared. He's over at Catrini. [10:26] Uh their inference margins are around [10:29] 85% [10:31] strong. They only had around 20,000 [10:33] hopper equivalents until May. um still [10:37] sort of I can see how they would say [10:39] that. [10:40] >> That doesn't mean that they don't have [10:41] access to a lot more power globally. [10:44] Yeah. Some of which may or may not be [10:46] above board, but who knows? Uh they have [10:48] 1 billion in API revenue is enough to [10:51] turn the company cash flow positive. [10:53] >> Yeah. [10:53] >> And they have a GPU payback period of 10 [10:55] months and their hardware uh is uh [10:58] depreciated over 3 to 5 years. So [11:00] looking uh looking quite good over the [11:04] pond. He also shared a lot of philosophy [11:06] that felt sort of derivative of western [11:10] AI lab leaders thought. Uh I mean he was [11:13] focused on uh you know some of the same [11:16] language around we don't want to moni we [11:18] don't want to optimize for pure profit. [11:20] We're on the path to AGI first. It's [11:22] something that you've heard from a lot [11:24] of lab leaders. Uh he did say that he's [11:26] very focused and he wouldn't be doing [11:28] world models or image generation or [11:30] video generation or even a chat app. He [11:33] said he doesn't want to be the next [11:34] Alibaba is purely focused on the path to [11:38] AGI sort of taking this focused approach [11:40] and so we'll see. Uh they've they've [11:42] sort of been out of the game with [11:44] Moonshot sort of coming from behind with [11:46] K3 and a lot of hype there and then GLM [11:50] of course as well. Uh so it'll be it'll [11:52] be interesting to see uh what he has up [11:54] his sleeve get uh on the back of this uh [11:56] investor call that everyone's paying [11:57] attention to. Gong moment. Cognition is [12:01] acquiring interaction, the makers of [12:03] Poke. [12:04] >> Huge. [12:04] >> Uh I think Poke has been [12:07] >> an acquisition target or been talked [12:09] about as an acquisition target for a [12:11] long time now. They've just made been [12:13] able to deliver delightful novel [12:17] products. Uh I wouldn't have expected [12:19] them to go to Cognition. I think a lot [12:21] of people were thinking that they'd get [12:22] picked up by a big uh consumer company, [12:25] maybe a you know an Apple. Weren't [12:27] people seeing Apple? [12:28] >> Apple would have been amazing. Having [12:30] the Poke team working on Siri would be [12:31] would be quite cool. Uh could have seen [12:33] them go into OpenAI or really any, you [12:35] know, even even an an Amazon, right? Uh [12:38] so a little bit unexpected here. Uh but [12:40] not surprised that Scott saw potential [12:43] in the team and I'm very interested to [12:45] see uh how these businesses end up [12:48] actually merging. Um will they keep Poke [12:52] around? Will it be um will it be just [12:55] sort of folded into cognition? But uh [12:58] either way uh very excited to see what [13:00] they built. [13:00] >> There's an older article in the Wall [13:02] Street Journal that we never got a [13:03] chance to read through, but it's fun [13:05] because it tells a story of a young [13:08] person who basically started literally [13:10] listened to a podcast and started a [13:12] search fund and made I guess millions of [13:14] dollars. So [13:15] >> did the meme. [13:16] >> Yeah. Did the meme. [laughter] Uh this [13:18] shortcut to private equity riches is [13:19] minting young millionaires. And I saw [13:22] Sal Khan from Khan Academy talking about [13:26] something sort of adjacent to this. Uh [13:28] because Khan Academy of course sort of [13:30] opensource education. You can basically [13:33] learn everything that you learn in [13:35] college from Khan Academy just watching [13:37] videos. Um and doing their their online [13:39] open coursework. I I believe it's [13:41] nonprofit. [13:42] >> It's still crazy that that Khan Academy [13:45] exists and and plenty of people still [13:47] turn it down. [13:48] >> Yes. Oh, they do. They do. Um, yeah, [13:52] this person didn't though. Basically, [13:55] uh, go-getters are backing to buy are [13:58] backing to buy up HVAC outlet outfits [14:01] and specialized manufacturers. If they [14:04] make it back alive, there's a lot of [14:05] money to be made. What a quote from the [14:07] Wall Street Journal. So, uh, in 2015, [14:11] Bari Akil was a homeless college dropout [14:15] with a single focus, figuring out how to [14:17] get rich. [14:19] >> Sleeping in Wei works [14:21] >> on the subway in an airport waiting [14:23] areas, Akill watched videos, listened to [14:26] podcasts, let's give it up for some [14:28] podcasts, and poured over personal [14:30] finance books. He needed to own a [14:32] business, he decided. But how to do it [14:34] with how do you do it with no money? [14:36] Then Akil read a Harvard Business School [14:38] case study that described how MBA [14:40] graduates were running around the [14:41] country trying to buy companies with [14:43] money from something called a search [14:44] fund. He felt like he had discovered a [14:47] secret. In the decades since Akil, now [14:49] 37, has bought two multi-million dollar [14:52] companies. He has spent each month of [14:54] the past three years living in a [14:55] different country. So he's doing the the [14:57] nomad [14:58] >> HVAC companies. [14:59] >> I guess we'll get into it. Uh he's now [15:01] worth seven figures. Akil is part of a [15:03] growing wave of people looking for a [15:05] shortcut to private equity riches. Armed [15:07] with grit and determination, they are [15:09] ditching the well-worn path from [15:10] spreadsheet wielding associate to [15:12] managing director and finding ways to [15:14] buy HVAC and plumbing outfits, [15:16] specialized manufacturers, and other [15:18] small and midsize businesses. Financing [15:20] from for deals, financing such deals is [15:23] far cry from traditional private equity [15:25] where firms use funds they raise from [15:27] institutional investors and the ultra [15:29] wealthy plus debt. Instead, these would [15:32] be business owners scrape together [15:33] financing on a deal-by-deal basis. They [15:35] use loans from the U. US Small Business [15:37] Administration, seek funds raised from [15:40] specialized investors or cobble together [15:41] money from friends, family offices, [15:43] private equity funds, and SBA licensed [15:45] small business investment companies. And [15:47] that's just the start. Convincing the [15:48] owner to sell and executing an [15:50] improvement plan is a daunting task, [15:51] even for the most experienced [15:52] dealmakers. While these buyers typically [15:55] >> you're overseas constantly. [15:58] Yeah, I mean I imagine that he's [16:00] overseas now, but during the heyday he [16:03] was like actually going and you know [16:06] knocking on doors. But uh the reason I [16:09] brought up Salon was that he was uh [16:12] advocating for instead of college buy [16:15] your child a business or or take a group [16:19] of of peers your ch your child's age and [16:24] they can buy a business together. [16:26] Because if you look at the cost of [16:27] college, [16:28] >> yeah, [16:28] >> you know, $200,000, $250,000, sometimes [16:32] if you get four students and they're [16:34] effectively the executive team and you [16:36] say, "Look, we're we have a million [16:38] dollars in cash. [16:39] >> We're going to lever something up. So, [16:40] we're [16:42] of of net income." [16:44] >> Yeah. [16:44] >> And we're going to lever it up. [16:45] >> Yeah. And the cash flow is going to pay [16:47] your salary. So, you're going to have a [16:48] job. Yeah. [16:49] >> From 18 to 22 when you graduate. and [16:53] your job is just to learn everything on [16:55] the fly. And best case, you wind up with [16:58] a fantastic business that makes a bunch [16:59] of money, but worst case, you're no [17:01] worse off than the money that you would [17:03] have spent on college. And so, you get [17:05] to learn all these skills on the on the [17:07] fly. I don't know. I don't know if it'll [17:09] take off. It feels like alpha school [17:11] adjacent. Uh [17:12] >> well, when the time comes, [17:14] >> yeah, [17:14] >> maybe we can get all the kids and say, [17:16] >> I mean, it's hard to predict what the [17:19] world would be like in 15 years or 20 [17:21] years if you're just having kids now, [17:23] but uh it doesn't seem that crazy to [17:26] imagine, but I think I I I think the [17:28] aura of uh of of certain schools will [17:31] endure and that will be the right path [17:32] for some people. Sure. Uh anyway, uh I [17:35] view these guys as home the homesteaders [17:37] of the Wild West, said Albertch, who [17:39] previously represented a big private [17:41] equity firm and a law firm Gibson and [17:42] Dunn. Uh they're heading out without a [17:44] dollar to their name, and if they make [17:46] it back alive, there's a lot of money to [17:48] be made. Many caught the bug at a top [17:50] business school where entrepreneurship [17:52] through acquisition courses are now a [17:54] standard part of the curriculum. They [17:56] are drawn by the promise of a more [17:58] flexible schedule, a conviction in their [18:00] own operational know-how, and desire to [18:02] avoid grunt work that ends up lining the [18:05] pockets of others. Also, fueling their [18:07] rise, a perception that it isn't as easy [18:09] as it once was to make big money on Wall [18:11] Street. It makes sense. A lot of the big [18:13] private equity firms were founded [18:14] decades ago, have founders that have and [18:16] partnerships that have accured a lot of [18:18] the value and and it's much harder to [18:20] just scale up something new. Uh the [18:22] private equity industry is struggling to [18:24] profitably unload companies causing [18:25] mid-level employees dubious of their [18:27] chances of receiving pay tied to deal [18:29] performance to jump ship. Some 77 search [18:32] funds, money from specialized investors [18:34] to back in individuals looking to buy [18:36] businesses were raised in 2025. 77 [18:39] search funds, new search funds listed by [18:41] the uh according to an annual study [18:43] conducted by researchers at Stanford in [18:45] 2025. That feels low based on how big [18:47] the meme was all throughout 2025. Maybe [18:50] there's no there's no way that that's [18:52] accurate. [18:53] >> Yeah. Maybe there's some that are that [18:55] aren't being caught in like the in the [18:57] in the definition or they they they have [19:00] defined their company as a as a true [19:02] private equity firm instead of a search [19:03] fund. So they said, "No, we're not a [19:05] search fund even though they might [19:06] actually be acting as one." [19:08] >> You know who's optimistic about AI, [19:10] John? [19:11] >> Who? [19:12] >> Friend of the show, Mark Zuckerberg. [19:14] >> Oh yeah. Uh Mark has launched a new [19:16] campaign that is focused on AI optimism. [19:22] >> Uh in Axio Medio Mark Zuckerberg on [19:24] Thursday laid out an optimistic view of [19:26] the agentic future arguing the company's [19:29] focus on connecting the world will only [19:30] be strengthened by new AI tools and [19:33] technologies. [19:34] Axio says why it matters. His position [19:36] is framed as a stark contrast to some of [19:38] Meta AI's competitors who, according to [19:41] a video ad posted with Zuckerberg's [19:43] comments, promote fear and a dystopian [19:46] vision of the future. [19:48] >> I thought it was a cool video. I liked [19:50] it. It's interesting to imagine watching [19:54] it not as a fan of social media though, [19:57] because it's leaning a lot on like we've [19:59] connected people, we bring people [20:00] together, and I certainly see it that [20:03] way. when I log on to Meta Platforms, [20:05] I'm sending you funny reels and we're [20:07] having a blast. But a lot of people see [20:09] it as brain rot, right? And they don't [20:10] and and they're like, "Oh, if you're [20:12] going to do what you just did to the [20:14] next thing, I'm not certain. I'm not [20:15] fully on board." So, I don't know how [20:17] it'll be production value. [20:20] It's definitely the correct angle. The [20:22] opposite is worse. Like clearly like you [20:24] want to be optimistic and there's a lot [20:26] of data to show. [20:26] >> I'm really glad that he's not like [20:27] mometically going down the fearbased [20:29] path. God. [laughter] Yes. That would be [20:32] very very rough. [20:33] >> Working for some players. [20:36] >> Um anyway, so we we'll see how it lands. [20:39] What else is going [20:42] >> on? Paramount is moving into micro [20:44] dramas. This is the moment you've been [20:46] waiting for. You love dramatic movies. [20:49] >> I don't know. [20:50] >> Can I interest you in some micro dramas? [20:52] Have you tried any of the [20:54] >> real the true shorts? [20:56] >> Yeah. Yeah. Yeah. Yeah. I watched a true [20:57] crime one and I I felt like I was [21:02] actually fine with the AI video and [21:06] voice over element of it. What I felt [21:08] was missing was when you're when you're [21:11] tracking a true crime story and you turn [21:13] on a podcast about that true crime, you [21:16] know, whatever it was. Uh I like being [21:20] uh led through a meandering idea maze of [21:24] what is interesting to that host. So [21:26] maybe they find the the town and the [21:29] details about the town where something [21:31] happened particularly interesting and so [21:33] they're just putting in random facts [21:36] that flesh out the story in a particular [21:38] way. I felt like the one that I watched [21:40] at least was very much like the [21:42] Wikipedia level summary or like the [21:44] headlines just condensed and it was too [21:46] generic and it wasn't giving me enough [21:48] uh novelty and like new facts. So I even [21:52] though like the video wasn't quite [21:53] there. It was obviously AI. It wasn't it [21:55] wasn't super visually striking. The [21:58] audio was sort of you know jilted. Uh [22:01] that all would have been fine if it had [22:02] been someone who actually went and found [22:04] some really deep interesting novel way [22:08] to tell the story or novel details that [22:10] otherwise you wouldn't have heard of [22:12] from the headlines. So I don't know. [22:14] Still early but I don't think they're [22:15] going to get me. There's something about [22:17] going to the theater. It's an [22:18] experience. It's like it's the [22:19] anti-brainer. you put your phone if [22:21] anywhere were to turn the screen [22:23] sideways. [22:23] >> They've done that. They've done that. [22:25] There's there somebody in Brooklyn I [22:27] think put together like a whole like [22:28] we're going to watch Instagram reels [22:30] together or like Tik Tok like short form [22:33] film festival or something. [22:34] >> Yeah. Speaking of conjecture, there's a [22:36] lot of conjecture about what OpenAI is [22:38] releasing today. Tibo said, [22:40] "Unbelievably excited for what's coming [22:42] together. Tomorrow is feeling codeexy." [22:44] Uh people are speculating that it might [22:46] be a a cerebrous a faster version of [22:48] soul turbo mode. It is okay. [22:51] >> Catch voice is now in the desktop app. [22:53] Control your computer and direct [22:54] multiple agents running in chatd work or [22:56] codecs using just your voice. [22:58] >> It's powered by GPT live. So it can [23:00] speak, listen and coordinate work in the [23:02] app at the same time. [23:04] >> And then also health is rolling out to [23:06] users everywhere. Health and chat GPT. [23:08] You can securely connect Apple Health [23:11] and supported medical records to [23:12] understand your information context. [23:16] >> Got to get it on there for sure. [23:18] >> Um there's also uh uh a new law [23:21] introduced, a bipartisan AI kill switch [23:24] bill following the OpenAI cyber [23:26] incident. The bill enters Congress as [23:28] lawmakers remain divided on how [23:30] aggressively the federal government [23:31] should regulate artificial intelligence. [23:33] I'm sure this will be a a longer [23:36] conversation on a future show. And [23:39] that's our show, folks. Thank you so [23:41] much for tuning in to TBPN Live from [23:44] AMD. Uh we are traveling tomorrow. We'll [23:47] be back on Monday, 11:00 a.m. Pacific. [23:50] Thank you to the team here who flew up [23:52] and thank you to everyone at AMD for [23:54] making this possible. We really [23:56] appreciate [23:56] >> dialing in these IRL shows. [23:58] >> Yeah, we're getting there. Uh one step [24:00] at a time. [24:01] >> The team makes it the team makes it look [24:02] easy. Give us five stars on Apple [24:04] Podcast and Spotify. Sign up for our [24:06] newsletter at tbpn.com. And we will see [24:08] you on Monday, 11:00 a.m. sharp for [24:11] another one. Goodbye.