Alger's Crawford Sees AI Compute Shortage Driving Hardware Rallies Through 2028

Alger旗下Crawford预计AI算力短缺将推动硬件涨势持续至2028年 全球人工智能算力需求的激增是一场指数级的结构性转变，半导体供应受限局面将延续至2028年，这使得传统的泡沫担忧为时尚早。 Ankur Crawford博士指出，AI变革是由复利式的指数级增长而非线性发展所驱动，这使全球市场面临长达数年的结构性算力短缺。硬件与存储供应链（包括Micron和Western Digital等DRAM龙头）不仅未处于投机性泡沫中，反而正受益于极高的行业进入壁垒以及超大规模云服务商（hyperscaler）以充沛现金流支持的巨额资本支出。与此同时，传统企业软件因内部AI开发以及SaaS定价模式的转变而面临严峻的结构性阻力；硬件巨头Apple也面临长期风险，无屏幕的环境感知类AI设备可能会颠覆传统智能手机的主导地位。 金融市场习惯于对AI创新进行线性预测，未能考虑到由现代算力扩展和自主智能体（autonomous agents）驱动的指数级复利增长。 (11:21) Ankur Crawford认为当前AI市场并非处于泡沫之中，因为算力需求受到结构性限制，且受到巨额超大盘公司现金流的有力支撑，而非依赖补贴性债务。 (35:01) 物理极限与高昂的资本进入壁垒已使存储和芯片供应链整合为极端的寡头垄断格局，从而推高了DRAM和GPU等各类硬件的价格。 (26:25) 随着超大规模云服务商将长达十年的算力投资压缩至当下实施，其估值逻辑正经历转变，从纯粹的自由现金流创造者演变为重资产基础设施建设者。 (44:11) Token定价结构将不可避免地根据计算复杂度出现分化，从而取代现行的统一定价模式。 (51:36)

Alger portfolio manager Dr. Ankur Crawford argues that concerns over an artificial intelligence bubble are misguided because exponential compute demand and concentrated supply chains are backed by immediate monetization and robust operational cash flows. While acknowledging valuation friction as hyperscalers reinvest cash flow, Crawford contends that acute hardware shortages will sustain the infrastructure expansion well into 2028. Structural Supply Constraints Defy Bubble Comparisons (35:01) Crawford rejects parallels between the current artificial intelligence buildout and historical busts like the 2000s solar crash or dot-com bubble. Unlike the solar boom, where negligible capital barriers to entry spurred rapid commoditization and bankruptcies, modern semiconductor and hardware infrastructure has consolidated over decades into high-barrier oligopolies. Only three global firms supply Dynamic Random-Access Memory (DRAM)—Micron, Samsung, and SK Hynix—which allowed DRAM prices to double year-over-year and helped drive Micron's stock from $85 to $450. Similarly, Western Digital and Seagate operate an effective duopoly in hard disk drives, where Western Digital's forward earnings estimates expanded from $10 to $25 per share within three months without adding speculative unit capacity. Because Big Tech infrastructure spending is funded out of massive corporate operating cash flows rather than fragile debt or government subsidies, Crawford asserts that compute supply will remain in a structural deficit through 2026, 2027, and into 2028. Immediate Monetization Counters Telecom-Era Infrastructure Fears (53:57) Addressing investor concerns over hyperscalers transitioning from asset-light models to capital-intensive infrastructure builders, Crawford acknowledges that shift creates valuation friction as markets transition from pricing free cash flow to enterprise value or earnings multiples. However, she dismisses comparisons to the 2000 telecom crash. During the dot-com era, billions of dollars of unlit fiber optic cables sat idle because consumer software and high-bandwidth applications did not yet exist. In contrast, today's AI data centers run at full utilization immediately upon activation, with hyperscalers realizing full capital payback periods on compute hardware in approximately 18 months via hourly compute rental. Crawford also rejects Dr. Michael Burry's claim that tech firms are using aggressive five- to six-year depreciation schedules for short-lived chips, noting that GPUs produced in 2021 are still actively running complex workloads. Furthermore, hyperscalers retain the flexibility to throttle future capital expenditures if underlying enterprise demand softens. Structural Margin Compression Across Enterprise Software (75:00) Crawford disputes Nvidia CEO Jensen Huang's premise that agentic AI will expand utilization for legacy enterprise software, forecasting severe headwinds across traditional SaaS providers. Legacy software economics rely on annual price increases and seat expansion; however, AI-driven workforce efficiencies and low utilization of bloated enterprise software features are eroding pricing power. Consequently, Crawford projects long-term operating margins across enterprise software will compress from historic 40% highs toward 20%. The primary displacement risk comes not from casual user coding, but from enterprise IT departments utilizing models like Anthropic's Claude to build custom in-house software to replace expensive third-party vendor contracts. Crawford differentiates commoditized point solutions from specialized platforms like Cadence Design Systems, unique ad-tech engines like AppLovin, and integration layers like Palantir, which she views as insulated from the broader software selloff. Hardware Moats Threatened by Screenless Interfaces (83:02) Examining Josh Brown's "HALO" framework (Heavy Assets, Low Obsolescence), Crawford challenges the notion that Apple's smartphone dominance represents an unassailable moat. She criticizes Apple's delayed execution in rolling out an AI-capable Siri, attributing the lag to internal tension between maintaining a privacy-first brand identity and the extensive data requirements needed to power autonomous personal agents. Crawford predicts that within three years, alternative ambient form factors—such as smart glasses, pins, and wearable audio devices—could begin displacing daily smartphone screen interactions. This shift would undermine Apple's hardware gatekeeper advantage and directly threaten ad-supported platforms like Meta and TikTok, whose revenues depend on monetizing visual screen time. While Crawford presents a high-conviction outlook for sustained compute demand and rapid hardware monetization through 2028, substantial uncertainties persist. These unresolved questions center on whether legacy software firms can adapt to collapsing seat-based pricing, whether Intel can overcome immense technical hurdles to rival TSMC in advanced foundry manufacturing, and how quickly ambient AI hardware might disrupt screen-based digital ecosystems.

阿尔杰公司（Alger）执行副总裁兼投资组合经理安库尔·克劳福德（Ankur Crawford）指出，当前的算力与硬件投资周期受到真实、爆发性的需求支撑，并非投机性资产泡沫；与此同时，传统企业软件（SaaS）模式正面临生成式工具与企业自研替代带来的结构性重估。 算力瓶颈与供应链整合支撑硬件估值 (35:01) 克劳福德认为市场习惯以线性思维预测指数级增长，从而严重低估了底层算力需求。神经网络多层节点的复合膨胀使得对智能（Token）的消耗速度急剧加快，例如优步（Uber）在2026年4月就耗尽了全年的算力预算。在此背景下，关键硬件供应链的严重紧缺支撑了业绩大幅上调：动态随机存取内存（DRAM）价格同比翻倍，美光科技（Micron）股价从85美元飙升至450美元；存储巨头西部数据（Western Digital）的远期每股收益预期在三个月内由10美元跳升至25美元，市值由150亿美元跃升至约1250亿美元。 针对泡沫论调，克劳福德明确反驳并指出当前周期与2000年电信泡沫存在本质区别。当年铺设的大量光纤因缺乏下游消费级应用而闲置数年，而如今整个技术架构（她比作已造好的F1赛车）早已就绪，唯独缺乏作为燃料的电力与算力输入。半导体行业经历数十年整合后门槛极高，DRAM仅剩美光、三星和SK海力士三家供应商，先进制程代工亦高度集中于台积电（TSMC）。由于超大规模厂商的硬件资本投入可在约18个月内通过算力租金收回，且2021年产的芯片至今仍在满负荷运行，克劳福德驳斥了迈克尔·伯里（Michael Burry）关于芯片折旧存在会计欺诈的指控。 超大规模厂商资本支出转型与自研芯片分流 (44:11) 以亚马逊、微软、Alphabet和Meta为代表的超大规模科技公司，正从轻资产、高自由现金流模式向重资产基础设施建设者过渡。这种将绝大部分经营现金流再投资于数据中心的策略，导致华尔街在估值体系切换（从自由现金流倍数转向市盈率或企业价值倍数）过程中产生分歧与股价波动。针对“循环收入”与供应商融资的质疑，克劳福德指出英伟达向CoreWeave投资20亿美元属于锁定分销渠道的合理商业拓展，在其数千亿美元的硬件收入体量中占比微小，与电信泡沫时期的坏账贷款截然不同。 在巨头竞争层面，亚马逊（AWS）虽然在估值上以十余倍的GAAP市盈率对沃尔玛等传统零售商呈现折价，但其核心优势在于自研芯片Trainium的布局。尽管Alphabet的TPU研发领先约五年，但自研ASIC芯片能系统性降低基础算力成本，使云服务商不必完全受制于昂贵的通用GPU，从而针对特定企业工作负载提供差异化低成本方案。克劳福德预计，随着市场成熟，用于简单天气查询与用于高精尖药物研发的Token单价未来将产生显著的层级分化。 “SaaS浩劫”与传统企业软件的边际收缩 (75:00) 克劳福德对英伟达首席执行官黄仁勋关于“智能体将增强现有企业软件”的乐观言论表示异议，认为传统SaaS正面临长期的结构性逆风。传统软件依赖于每座席收费（Seat-based pricing）与合同净留存率（DBNR）的逐年递增，但AI驱动的企业人员精简和功能冗余（多数客户仅使用15%至20%的内置功能）正在打破这一垄断定价权。企业内部IT团队正借助Claude等高级模型开展定向自研，若自研能为公司在五年内节省8000万美元软件支出，采购第三方独立应用（Point solutions）的动力将大幅下降。 受此影响，软件行业的终局营业利润率预期可能从历史峰值的40%被压缩至20%左右，诸如赛富时（Salesforce）等软件巨头已被市场以低估值倍数重新定价。克劳福德认为软件板块正经历理性分化：缺乏护城河的长尾工具聚合商遭受重创，而拥有庞大平台生态的微软与CrowdStrike，或作为数据中枢协助企业落地的帕兰提尔（Palantir），以及依托专有算法拓展物理电商广告的AppLovin，则展现出更强的生存与变现韧性。 HALO交易分歧与无屏幕交互对硬件护城河的冲击 (83:02) 主持人乔希·布朗提出“重资产、低淘汰率”（HALO）概念，主张在数字化创作成本归零的时代，不可替代的实体硬件将享有估值溢价，并将苹果（Apple）与亚马逊列为典型代表。克劳福德对此表示赞同其底层逻辑，但对苹果的长期护城河提出严厉警告。她指出苹果多年来在AI语音助手Siri上的实质性落地屡次推迟，陷入了“隐私至上”的品牌定位与AI代理需要高强度调用个人数据之间的战略摇摆。 克劳福德预测，随着自主AI代理技术的成熟，三年内智能眼镜、可穿戴微型设备（Pins）或音频终端等无屏幕环境计算形态，可能开始分流智能手机的使用频率。一旦用户的注意力从物理屏幕中抽离，不仅苹果依赖屏幕控制权的终端硬件溢价将面临动摇，高度依附于屏幕浏览时长的社交媒体广告变现机制（如Meta旗下Instagram和TikTok）也将遭遇根本性冲击。 本期讨论表明，AI基础设施端凭借紧缩的产能与极高的资本回收率暂时证伪了“泡沫破裂”说，但企业级软件的商业逻辑与智能终端的屏幕形态正经历深刻解构。未决的关键问题在于：在超大规模厂商将自由现金流无限期压制于数据中心建设的同时，公开市场能否长期承受估值体系的重构；以及苹果等硬件巨头能否在环境计算技术成型前，顺利跨越从屏幕中心化到AI代理化的交互鸿沟。

## Background 该视频由财经频道 The Compound 于 2026 年 4 月 17 日发布在 YouTube 平台。节目为《The Compound & Friends》第 238 期，主持人 Josh Brown 和 Michael Batnick 邀请了 Alger 执行副总裁兼资深投资组合经理 Dr. Ankur Crawford，共同探讨 AI 浪潮背后的真实驱动力、全球算力短缺的物理瓶颈，以及为何当前市场盛行的“AI 泡沫论”是错误的。 ## TL;DR 拥有材料科学与硬核半导体投研背景的基金经理 Dr. Ankur Crawford 在节目中深入剖析了 AI 产业的底层逻辑，断言全行业当前正处于严重的“算力短缺”之中，指数级扩张的算力需求与寡头垄断的物理供应链决定了当前的繁荣绝非泡沫。她详尽对比了 2000 年电信泡沫与光伏/EV 行业的本质差异，指出 AI 硬件从一开始就具备真实盈利、充沛现金流支撑以及低至 18 个月的极快回本周期。同时，讨论深入拆解了超大规模云厂商的资本开支与自由现金流阵痛、SaaS 软件商业模式面临的严峻颠覆，以及诸如 Nebius、QXO、AppLovin、英特尔与苹果等具体标的的投资与战略博弈。 ## Atmosphere & style 整场对谈兼具华尔街顶尖基金经理的硬核专业度与老友录音室的轻松诙谐。两位主持人频繁插科打诨、互相调侃甚至戴反耳机，而拥有科学家严谨逻辑的 Ankur Crawford 则在欢笑中游刃有余地用一针见血的技术洞见降维打击常见的市场偏见。听众既能感受到探讨“指数级狂飙未来”时的智识兴奋与紧迫感，也能体会到面对宏观波动与资本开支重估时专业投资人所展现出的冷静与敬畏。 ## Bulletin - [36:28](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=2188s) Alger 基金经理 Dr. Ankur Crawford 指出，全球算力在 2026 至 2028 年前将持续处于供不应求的失衡状态，所谓“AI 泡沫论”忽略了基本面盈利预期的倍数级上修。 - [55:12](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3312s) 不同于 2000 年电信泡沫中巨额资本沉淀在闲置光纤，当前旗舰 GPU 硬件投资的回本周期压缩至约 18 个月，且旧款芯片在推理端仍具备持续长效的现金流产出能力。 - [01:15:33](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=4533s) 传统 SaaS 模式正面临 AI 智能体的严重颠覆，企业利用 AI 自研定制系统和缩减人员编制打破了按席位提价的逻辑，其稳态营业利润率可能从预期的 40% 腰斩至约 20%。 - [42:30](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=2550s) 半导体产业历经二十年整合已形成极高寡头壁垒，DRAM 存储与机械硬盘等底层硬件厂商克制扩产并推升产品密度，避免了光伏式的盲目恶性产能过剩。 ## Key takeaways - [08:48](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=528s) 成长股投资的关键在于捕捉“高单位出货量增长”与“生命周期转变”两类变革，后者常因低估值被误认为价值股，实为被市场低估的成长拐点。 - [11:21](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=681s) 分析师习惯以线性外推思维看待 AI，但底层神经网络层级拓展与智能体部署遵循指数级复合曲线，导致算力消耗迅速击穿企业预算预算。 - [17:45](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=1065s) 前沿超级独角兽持续推迟 IPO，公募基金通过 40 法案最多配置 15% 非流动资产的规则提前介入一级市场（如 Anthropic），捕获指数无法覆盖的阿尔法。 - [24:25](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=1465s) 长期的战略性乐观不等于盲目满仓，面对地缘政治摩擦等风险需主动提升现金仓位，保留在市场下挫 30%–40% 时捡拾优质资产的流动性弹药。 - [27:17](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=1637s) Token 本质是“思考与算力的通用度量衡”，随着大模型推理复杂度提升与智能体落地，Token 消耗上限被不断打破。 - [28:35](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=1715s) 算力瓶颈正向 DRAM 存储及大容量机械硬盘等整条供应链外溢，全球寡头格局与极高进入壁垒为相关硬件厂商提供了极强的定价权与盈利支撑。 - [38:11](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=2291s) AI 硬件与昔日光伏泡沫存在本质差异：先进制程具备极高的技术封锁与研发资本门槛，且巨头的 CapEx 依赖充沛的自身现金流而非外部资本补贴。 - [44:11](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=2651s) 超大规模云厂商（Hyperscalers）由轻资产向重资产模式切换，市场对自由现金流被资本开支吞噬的担忧正在迫使估值体系经历痛苦的重估与震荡。 - [51:36](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3096s) AI Token 并不平等，未来将根据任务复杂度出现清晰的分层定价体系，通用高算力 GPU 将主攻复杂推理，而轻量模型与定制 ASIC 承接流水线任务。 - [53:57](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3237s) AI 资本开支并非重演当年的元宇宙烧钱，元宇宙属于超前尝试且缺乏落地底座，而 AI 则是各巨头关乎核心业务生死存亡的刚需之战。 - [57:23](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3443s) 市场所谓“循环交易”应区分看待：类似英伟达参投 CoreWeave 属于正常的生态培育，而附带巨额股权认股权证的硬件绑定交易才潜藏较高风险。 - [01:19:49](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=4789s) SaaS 板块将出现深刻分化：拥有平台生态的网络安全巨头和工业级 EDA 软件护城河稳固，但单点工具（Point Solutions）与低壁垒协同软件将面临毁灭性替代。 - [01:23:07](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=4987s) 长远来看，具备自主智能体的常伴无感硬件可能逐步摆脱长方形手机屏幕，苹果若无法及时推出变革性个人代理，其依赖屏幕交互的商业壁垒将受到严峻考验。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=0s) 台积电财报会轶事与财报收听工具之争 节目在轻松的幕后闲聊中开场。Michael Batnick 提到自己当天刚第一次听了台积电（TSMC）的财报电话会议，发现了一个极为奇特的现象：每当分析师提问后，首席财务官（CFO）都要把刚才的问题一字不差地重复一遍才作答，整场会议都是如此，显得格外诡异。对此，在半导体领域深耕多年的 Dr. Ankur Crawford 解释道，过去很多财报会是在现场举办的，分析师坐在台下观众席提问，管理层在台上，为了让全场听清才会由高管复述一遍，这演变成了一种特殊的“企业文化习惯”。Batnick 还顺带吐槽了台积电的投资者演示文稿（deck），直言那份 PPT 简陋得像是在 20 世纪 80 年代用 MS-DOS 制作的，完全看不出这是一家市值过万亿美元的科技巨头。 紧接着，大家聊到了基金经理日常收听财报电话会的工作流。Josh Brown 询问 Ankur 是如何听财报会的，并力荐了他们投资并深度使用的平台 Quarter——该工具能在语音播放的同时高亮对应的文字，让人能“边听边看”，信息吸收效率极高。Ankur 则透露他们公司内部使用的是名为 Era 的工具，界面同样很便捷。她还打趣地提到自己曾和 Quarter 的销售聊过，Era 给出的报价更为划算；Brown 则开玩笑称 Era 背后说不定也是贴牌使用了 Quarter 的底层数据。 - **财报收听偏好**：面对财报季排山倒海般的信息洪流，基金经理不可能听完组合里的每一家公司。Ankur 依靠研究团队覆盖全盘，但自己有几类电话会绝不错过：最具代表性的是各大超大规模云服务商（Hyperscalers）；此外每季度她会抽样收听沃尔玛（Walmart）或好市多（Costco）以把握消费端脉搏，并挑选代表性的工业股或医疗保健股来感知宏观环境。 - **信息提取与 AI 工具**：Batnick 提到自己经常阅读订阅服务 The Transcript 提炼的各行业高管发言精要，并指出这背后很可能已经融入了 AI。Ankur 也表示，现在完全可以将每日日程交给 Perplexity 等 AI 工具，快速提取核心语录与关键观点。 --- ### [04:06](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=246s) 录制前热身：戴反的耳机与不存在的“即兴说唱” 在正式进入访谈前，录音室里发生了一连串幽默插曲。Josh Brown 调侃这是 Ankur 的“播客处女秀”，Ankur 笑着纠正说这其实是第二次，她之前曾做客过 Morningstar 的节目。Brown 则打趣道：“行吧，但我们就当这是你的第一次，上周你总不可能去上了 Joe Rogan 吧？”随后，Brown 突然抱怨耳机戴着很不舒服，Batnick 仔细一看才发现 Brown 居然把耳机完全戴反了，惹得全场大笑，Brown 连忙把责任甩给录音师 Daniel。 当节目的标志性片头音乐响起时，Brown 开玩笑地对 Ankur 说：“通常嘉宾听到这首主题曲，都要跟着节奏来一段即兴饶舌（freestyle rap），你有准备吗？”Ankur 笑着认输：“我真没准备。”Brown 打趣道：“那行，下次再补上。” --- ### [05:55](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=355s) 隆重介绍 Dr. Ankur Crawford：从科学家到华尔街顶尖基金经理 随着正式片头与赞助商信息结束，Josh Brown 正式揭开第 238 期节目的序幕。他隆重介绍了本期重磅嘉宾——来自联博旗下老牌资产管理机构 Alger 的执行副总裁兼投资组合经理 Dr. Ankur Crawford。 ```mermaid flowchart TD accTitle: Ankur Crawford 职业演进图 accDescr: 展示 Ankur Crawford 从科研学术背景到半导体分析师再到资深基金经理的职业生涯演变 subgraph Academic["学术科研背景"] A["科研学者"] --> B["拥有多项专利"] B --> C["美国国家科学院院士奖学金<br/>(National Academy of Sciences Fellowship)"] end subgraph Finance["华尔街投研历练 (22年)"] D["加入 Alger 担任半导体分析师"] --> E["晋升为 Alger 科技研究主管"] E --> F["2012年正式挂帅基金经理"] end Academic --> Finance subgraph Funds["核心管理产品"] F --> G1["Alger Capital Appreciation"] F --> G2["Alger Focus Equity"] F --> G3["Alger Spectra Large Cap Growth Equity"] F --> G4["Alger Concentrated Equity ETF (代码: CNEQ)"] end ``` Ankur Crawford 拥有极强的技术硬核背景，她不仅握有多项技术专利，还曾荣获美国国家科学院（National Academy of Sciences）的奖学金。22 年前，她带着深厚的工科底子迈入华尔街，在 Alger 从最底层的半导体行业分析师起步，凭借对芯片和软硬件产业的深刻理解，一步步成长为 Alger 科技团队的主管，并在 2012 年正式出任投资组合经理。如今，她负责掌舵多只旗舰成长型基金，包括 Alger Capital Appreciation、Alger Focus Equity、Alger Spectra Large Cap Growth Equity，以及 2024 年全新挂牌上市的集中型股票 ETF——Alger Concentrated Equity ETF（股票代码：`CNEQ`）。 现场在名字发音上也引出了一段互动，主持人明确了她的名字读作“En-core”（安可）。Brown 表示，本期对话不仅要探讨全市场最关注的宏观大局与 AI 浪潮，更会深入 Ankur 极为擅长的半导体核心腹地，为投资者拨开技术与估值的迷雾。 --- ### [08:21](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=501s) Alger 的投资哲学：成长型投资的先驱源流 在深入具体股票和行业之前，Josh Brown 请 Ankur 阐述她的核心投资方法论以及身处 Alger 的角色认知。 Ankur 回顾道，Alger 自 20 世纪 60 年代创立以来，始终是一家纯粹的成长型投资管理机构（growth manager）。事实上，创始人 Fred Alger 被公认为华尔街历史上最早确立“成长型投资”概念的先驱之一。半个多世纪以来，该机构的基因就根植于发掘那些能够通过技术颠覆、产品周期与商业模式创新实现非线性增长的优秀公司，而这种长期沉淀的成长股框架，也正是她理解当前这一轮 AI 与半导体变革的底层基石。 ### [08:48](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=528s) 寻找“未被认知的成长”：高单量增长与生命周期拐点 在谈及 Alger 的成长股投资哲学时，Ankur Crawford 阐明了其核心逻辑始终围绕一个核心提问——“变革究竟在哪里？”（Where is the change?）。她解释道，但凡发生剧烈变革的地方，往往潜藏着尚未被市场充分定价的巨大机遇；如果投资者能够抢在市场形成共识之前洞察这种变化，就能创造极其可观的超额收益。在实际选股中，这种“变革”主要被划分为两大类别： 第一类是典型的“高单位出货量增长”（high unit volume growth）。这类公司代表了大众认知中最标准的成长股：主营业务收入高速扩张，正在以某种创新方式颠覆现有市场格局；虽然它们当前可能尚未产生充沛的自由现金流，但其业务规模正迅速向成熟盈利期迈进。 第二类则是让 Alger 的投资哲学显著区别于多数成长型基金经理的关键所在——“生命周期转变”（life cycle change）。这类企业往往已经经历过上一轮波澜壮阔的高速扩张期，原有主营市场逐渐走向饱和，企业不得不直面中年危机并重新确立自我定位。它们必须回答一个残酷的问题：“未来究竟还能否重获增长，接下来该何去何从？” ```mermaid flowchart TD accTitle: 成长股投资的两大路径 accDescr: 展示 Alger 关注的高单量增长与生命周期转变两种企业类型 A["寻找市场中的'变革' (Change)"] --> B["高单位出货量增长<br>(High Unit Volume Growth)"] A --> C["生命周期转变<br>(Life Cycle Change)"] B --> B1["典型成长型公司<br>颠覆市场格局<br>营收高速增长"] C --> C1["跨越成熟期与市场饱和<br>低估值倍数 (Low Multiple)<br>寻找正向增长拐点"] C1 --> D["本质是'未被认知的成长'<br>(Unrecognized Growth)"] D --> E["关键抓手：卓越的管理层<br>(如 Satya Nadella 执掌微软)"] ``` 在讨论生命周期转变时，Josh Brown 和 Michael Batnick 相继列举了苹果（Apple）、微软（Microsoft）、西部数据（Western Digital）、独立发电商（IPPs）以及英伟达（Nvidia），Ankur 对此高度认同，并补充了丹纳赫（Danaher）等通过业务重塑自我迭代的经典案例。这类处于业务再造期的公司往往呈现出极低的估值倍数，常被外界误以为是“价值股”（Value stock）。但 Ankur 坚决反驳这种标签：“这绝不是价值陷阱，而是‘尚未被认知的成长’（Unrecognized growth）。”不过，主持人也指出其中的巨大陷阱：无数企业试图自我革新，最终却折戟沉沙。Ankur 强调，破局的核心在于管理团队的执行力，最典型的例子莫过于萨提亚·纳德拉（Satya Nadella）执掌微软时期，投资者必须敏锐观察管理层所释放的信号并审视其推进动作，以判断其能否兑现下一个十年的增长曲线。 --- ### [11:21](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=681s) 指数级狂飙的 AI：为什么分析师的线性思维正在失效 当话题转向当下席卷全球的人工智能浪潮时，Josh 询问 Ankur，自三年前大语言模型（LLMs）横空出世、世人首次领教其威力至今，她如今是否依然保持着最初的狂热。Ankur 毫不犹豫地回答，在过去三年里，她对 AI 能力的兴奋感与敬畏感不仅没有衰退，反而呈倍数级增长。 她回忆道，三年前团队曾撰写过一篇题为《AI 与创作成本的下降》（AI and the Declining Cost to Create）的研究报告。为了完成该文，他们密集走访了大量 AI 原生从业者与整个硬件/软件供应链专家。回顾当时的预测，Ankur 感叹如今现实中真实发生的技术演进，原本在她看来需要两到三年后才会出现——整个技术成熟的时间线被大幅提前了，而其根本原因在于技术能力正在以完全超乎想象的“指数级”速度狂飙。 - **线性认知与指数现实的脱节**：华尔街分析师与金融市场通常习惯于用“线性思维”做预测，但在面对真正的指数级增长时，传统估值模型和认知框架很难准确把握其终局高度。 - **智能体（Agents）的技术早产**：原本行业普遍预期要到 2028 年左右才能具备实用价值的自主智能体能力，在当下就已经初露锋芒并进入落地阶段。 - **高强度的学习焦虑**：面对日新月异的技术迭代，Ankur 笑称自己正处于持续学习的状态，“这辈子从未睡得这么少、工作得这么多”，因为每天睁开眼都有全新的更新与突破等待消化。 --- ### [13:26](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=806s) 从 Agent 熬夜到零代码开发：60小时打造大学升学应用 谈到令人感到“人类已然生活在未来”的震撼时刻，Ankur 提到了类似于 OpenClaw 与 Perplexity 等具备自主电脑控制或代码执行能力（Agentic products）的突破性工具发布。她回忆起 2024 年 2 月的那个周末，当时正陪儿子参加滑雪队集训，借住在朋友的滑雪木屋里，她为了研读新技术通宵未眠。虽然最初尝试在本地部署时因为繁琐的配置和安全权限顾虑而稍显混乱，但一个月前上线的具备计算机自动化操作能力的工具（如 Perplexity 推进的端到端方案），已经展现出惊人的实用价值——比如在缺乏开发背景的情况下，该工具成功帮她解决了自建应用发布及底层支付接口接入的复杂障碍。 在这项技术的启发下，Ankur 完成了一次极具戏剧性的个人实践。因为大女儿即将升入大学，在经历了极其痛苦、混乱且昂贵的高中升学辅导后（前后雇用了三位不同的升学顾问），女儿提议：“很多普通家庭的孩子根本请不起升学顾问，为什么不用 AI 把这个流程平民化（Democratize）呢？”于是在一个大雪纷飞的周末，没有任何编程经验的 Ankur 开始借助 AI 亲自编写一款升学辅助应用： - **深度自我挖掘**：内置 12 个直击内心、极具启发性的深度问题，帮助学生系统梳理自我认知、过往经历与活动背景。 - **全流程组织管理**：无缝对接 Common App（美国大学通用申请系统）与意向高校名录，具备版本控制功能，支持学生一键将文书草稿发送给家长或导师进行批注。 - **“战略智能”（Strategic Intelligence）模块**：针对特定申请文书题目，结合学生的个人数据、独特背景与课外活动，指导学生如何确立最打动招生官的切入角度与表达策略。 当 Josh 听完这一整套流程后感叹，未来的大学甚至应该强制要求申请者通过与 AI 的交互来提交申请，以此证明他们已经具备适应全新技术时代的能力。而整件事对 Ankur 而言，最令人震惊的不是应用本身的商业前景，而是一个从未写过代码的人，仅仅利用三个周末、累计投入约 60 个小时，就完整构建了一个复杂的生产级应用。主持人打趣道：“如果没有 AI 编程辅助，这大概需要耗费你 60 年。”Ankur 坦言：“不，如果没有 AI，耗时将是无限长（Infinite）——因为我根本连第一行代码该从哪里写起都不知道。” ### [17:45](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=1065s) 估值迈向万亿美元的超级独角兽与一级市场投资布局 基于阿尔杰（Alger）聚焦高增长资产的选股哲学，这些处于前沿的 AI 公司正处于爆发式增长期，远未达到市场饱和阶段。对话由 Anthropic 和 OpenAI 令人瞠目的最新一轮融资估值切入。安库尔·克劳福德（Ankur Crawford）透露，阿尔杰几周前刚刚参与了 Anthropic 的最新一轮融资。她指出，资本市场正在经历前所未有的历史性演变——历史上从未有任何时期的公司会一直保持私有化状态，直到估值接近一万亿美元才考虑上市。 ```mermaid flowchart TD accTitle: 一级市场超级独角兽的价值沉淀与公募介入逻辑 accDescr: 展示科技巨头推迟上市导致价值滞留一级市场，公募基金通过法规配额提前布局以保护客户利益并为未来IPO做准备。 subgraph MarketShift ["市场范式转变"] A["超级AI独角兽持续保持私有化<br/>(估值迈向数千亿至万亿美元)"] B["二级市场公开投资者<br/>错失绝大部分早期爆发红利"] A --> B end subgraph FundAction ["阿尔杰基金布局策略"] C["利用40法案基金框架<br/>(最高配置15%非流动性资产)"] D["锁定具备2-3年明确IPO路径的<br/>顶级私有标的 (Anthropic等)"] B -.-> C C --> D end subgraph FuturePlaybook ["上市后应对考量"] E["密切跟踪季度财务与管理层沟通"] F["上市波动评估：若二级市场抛售<br/>则逢低加仓；锁定保护后择机退出"] D --> E E --> F end ``` 这种推迟上市的现象导致巨大的价值沉淀在非公开市场，普通二级市场投资者根本无法触碰。为了不让客户错失这部分红利，阿尔杰选择深入一级市场，寻找他们深度信任且具备清晰路径在未来两到三年内完成 IPO 的领军企业。迈克尔·巴特尼克（Michael Batnick）询问这是通过何种投资载体实现的，安库尔解释称，这是横跨他们所有旗下基金的动作：依据美国《1940年投资公司法》（40 Act），传统公募基金可以将最高 15% 的资产净值配置于非公开发行的私有资产。虽然这不是阿尔杰的核心常规操作，但此前早有先例——他们早在 2012 至 2013 年左右就已是一级市场阶段 Palantir 的早期投资者，后来也投资了数字银行 Chime。 乔什·布朗（Josh Brown）指出，如今许多科技巨头在递交 S-1 招股书前估值就已经膨胀到数千亿美元，这固然让普通公众投资者感到无奈，但从另一个角度来看，这恰恰凸显了主动管理型成长基金的核心价值与护城河——主动型基金能够打破一二级市场壁垒，为出资人捕获基准指数无法包含的超额阿尔法。谈及上市后的博弈，安库尔表示团队未来将面临新的抉择：虽然需要遵守锁定期规则，但如果上市后二级市场因恐慌抛售导致股价大跌（"if the public market vomits in"），他们不仅不会离场，反而可能进一步加仓。在投后信息披露方面，私有企业虽不召开公开电话会议，但阿尔杰每季度均能获得财务报表，并能保持与核心管理层的直接沟通。 - **40法案基金配额上限**：公募基金依据监管规定，最多可将 15% 的资产配置于一级市场私有股权。 - **历史一级市场先例**：阿尔杰曾于 2012-2013 年间在一级市场布局 Palantir，并在后期投资了金融科技公司 Chime。 - **潜在上市观察名单**：Anthropic、Databricks 以及 OpenAI 未来两到三年的上市进程与市场承接力是机构关注的核心风向标。 --- ### [20:59](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=1259s) 全球“算力荒”：人类的线性思维 vs 神经网络的指数级需求 乔什将话题转向宏观核心议题：当前市场流传着“全球极度缺乏算力”的论调，安库尔对此深信不疑，断言“每个人都缺算力”（"Everyone's short compute"）。然而，二级市场投资者的最大梦魇在于周期见顶——担心自己成为最后的接盘侠，在全行业过度建设、产能过剩以及科技巨头削减资本开支（CapEx）的前夕买入。这种谨慎情绪在过去三年里让许多资金望而却步。 安库尔一针见血地指出了认知分歧的根源：人类的大脑受制于线性的思维训练。面对两个离散的数据点，常人的本能反应是画一条直线进行测距和向外推导；然而，人工智能的基础设施扩展是彻头彻尾的“指数级增长”。由于大部分市场参与者从未在现实世界中经历过真正的指数级爆发，因此才会对眼下科技巨头惊人的资本支出规模与激增的财务数字感到困惑和怀疑。 ```mermaid flowchart LR accTitle: 神经网络与智能体驱动的算力指数级消耗飞轮 accDescr: 描述从神经网络的层级指数连接到智能体落地，再到全行业算力与Token预算耗尽的过程。 A["神经网络深度拓展<br/>(每增加一层连接呈几何级倍增)"] --> B["算力需求呈指数级爆炸<br/>(非线性推导)"] B --> C["AI智能体 (Agents) 落地<br/>(赋能生产力任务)"] C --> D["Token消耗无上限<br/>(Token Maxing)"] D --> E["企业计算预算提前击穿<br/>(全行业算力陷入绝对短缺)"] ``` 从底层技术逻辑来看，构建神经网络本身就是一个指数级难题。智能由错综复杂的神经元互联所构筑，每一层网络的向外延伸，其背后的节点连接数量都是巨大的天文数字彼此相乘。当人类试图在数字世界中复刻甚至超越通用智能时，对计算资源的需求将大到超乎想象。一年前市场还在焦虑这是否是一场虚无的泡沫，但随着 AI 智能体（Agents）的普及，现实逻辑已经彻底改变——无数用户和企业正在拼命压榨 Token 额度（"token maxing"），云端账单呈现失控式飙升。 背后的驱动力非常纯粹：一旦 AI 能够切实接管具有高生产力价值的复杂任务，企业就会不计成本地进行大规模自动化部署。乔什随即补充了一个极具戏剧性的商业现实：Uber 已经在 4 月份前后提前烧光了整整一个年度直至 2026 年的全部计算预算。安库尔指出，Uber 只是冰山一角，算力技术已经彻底民主化，不再需要专业程序员编写复杂代码，任何普通人都可以通过自然语言调用智能。只要企业有足够的 Token 额度，实际的使用量就将持续突破天花板。 - **思维模式陷阱**：投资者习惯于线性外推，但 AI 需求与算力基础设施遵循的是以神经网络层级为倍增器的指数级增长曲线。 - **智能体（Agents）的生产力转化**：促使企业预算激增的根本原因，在于只要智能体能代替人工产生实质生产力，企业就会立刻加大调用规模。 - **Uber 计算预算击穿案例**：企业在 2026 年尚未到来之前，由于全员加速落地 AI 应用，其原定到 2026 年底的算力资本预算在当下（4月份）就已完全耗尽。 --- ### [24:25](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=1465s) 极度看好与战术避险：2月份主动提升现金头寸的背后逻辑 迈克尔和乔什向安库尔发问：她对 AI 指数级扩张的坚定信仰，是否构成了她作为投资组合经理抵御市场剧烈回撤的心理防线？具备这种极致多头信念的人，按理说应当能够承受远超怀疑论者的市场波动。 安库尔对此给出了极具操作层面的回答：长期的战略性乐观并不等于机械式的盲目持仓。她以当年的亲身操作为例：2 月份时，她通过敏锐的市场直觉感受到潜藏的危险信号，特别是中东地缘政治摩擦升温，美军舰队驶向伊朗周边海域，整个宏观环境的风险收益比急剧恶化。面对不安的市场信号，她并未因为看好 AI 的长期前景而按兵不动，而是果断在其管理的高集中度股票基金（concentrated equity fund）中将现金仓位提升至 7%。 乔什开玩笑地追问：“你总没有卖掉 Anthropic 吧？”安库尔笑着回应当然没有卖掉私有股权，而是通过平抑整个二级市场组合的公开持仓来筹措现金缓冲，以防御即将到来的潜在波动。她强调，战术性提升现金头寸绝不意味着对 AI 长期主线的倒戈；事实上在此期间，她对 AI 赛道的坚定程度反而有增无减。在市场共识产生分歧、他人产生恐慌心理时，手里握有流动性弹药才能成为真正的赢家——如果你能在高位获利了结，并准备在资产下跌 30% 到 40% 的深坑中低价捡回优质筹码，这就是作为专业基金经理在任何时候都无法拒绝的交易选择。 --- ### [26:25](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=1585s) 拆解“算力短缺”：算力的本质度量与物理瓶颈 迈克尔提出了一个代表广泛大众疑惑的直白问题：“当所有专业人士都言之凿凿地说‘我们严重缺乏算力’时，这在字面和物理层面到底意味着什么？真正的瓶颈卡在哪里，对投资者究竟意味着什么？” 乔什抢先插话调侃道：“电力！电力绝对是一个巨大的瓶颈！”随后两人展开了激烈的互怼与捧哏，迈克尔无奈催促他让安库尔回答，乔什则进一步把问题引向核心技术度量标准： > “听着，我们要讨论的核心度量单位其实是 Token（词元）。迈克尔可能需要恶补，我们台下 90% 的观众要么是替别人管钱的专业顾问，要么是精明的独立投资者——剩下的 10% 纯粹是疯子。你必须向大家阐明，为什么算力短缺最终会映射在 Token 上？你们究竟是如何从底层量化并思考这场计算资源危机的？” 这场插科打诨的交锋，直接揭开了后续关于电力基建限制、数据中心物理承载力与 AI 模型运算“通用货币”Token 之间紧密交织的深度技术探讨。 ### [27:17](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=1637s) Token 的本质：算力与思考的度量单位 谈话切入了一个核心概念：在当前的 AI 浪潮中，Token 到底意味着什么？是否所有的 Token 都具有同等价值？Dr. Ankur Crawford 赞叹这是一个极具深度的好问题。从技术层面来说，一个 Token 大约相当于两到三个英文字母；如果你发送一个包含 100 个字符的提问，输入端大约消耗 30 个 Token。但从更高的视角来看，Token 实际上是“思考的单位”与“智能的单位”（units of thinking and intelligence），归根结底则是“算力的单位”（units of compute）。所需的智能水平越高、推理越复杂，消耗的 Token 就越多。这不仅包括用户输入的 Prompt，还包括大模型在后台推演并反馈给用户的海量输出。 为了更形象地解释 Token 的来源与运转机制，Crawford 引用了 Reid Hoffman 提出的“AI 工厂”（AI Factory）概念。如果把芯片比作飞轮或生产车间，模型就像是一座高速运转的智能工厂：输入若干 Token 作为原料，机器在后台进行极其密集复杂的计算与“研磨”（grinding），最终交付出高密度的智能成果。计算的过程就是计算机进行底层数学运算并给出响应的物理动作。 ```mermaid flowchart LR accTitle: AI Factory Token Process accDescr: 流程图展示从输入 Token 到芯片研磨再到智能输出的过程 subgraph InputStage["输入端"] A["用户 Prompt 输入<br/>(如 100 字符 ≈ 30 Tokens)"] end subgraph Factory["AI 工厂 (算力研磨)"] B["芯片集群 / 算力飞轮<br/>(底层矩阵运算与推理)"] end subgraph OutputStage["输出端"] C["高密度智能产出<br/>(复杂推理与解答 Tokens)"] end A --> B B --> C ``` --- ### [28:35](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=1715s) 算力瓶颈的外溢与 DRAM 存储器超级暴涨 当用户支付月费订阅这些大语言模型以获取算力使用权时，瓶颈究竟卡在哪里？是因为英伟达（Nvidia）芯片供不应求，还是台积电（TSMC）产能受限？Crawford 指出，当下最直接的问题依然是“芯片不够”，但短缺已经开始沿着整条供应链迅速蔓延，最典型的就是 DRAM（动态随机存取存储器）市场。DRAM 价格同比暴涨了约 100%，这对于一种传统周期性大宗商品来说是极其罕见的现象。 Crawford 回忆，早在两年前她就预见到 AI 对高带宽与海量内存的巨大拉动，但当时整个存储市场极其惨烈，价格毫无起色，以至于她一度怀疑自己的判断是不是错了，误以为市场早已把一切定价完毕。然而事实证明，是市场本身滞后了。存储市场直到去年 9 月才彻底爆发：美光（Micron）在去年 8 月的股价还徘徊在 85 美元左右，如今已经飙升至 450 美元，市值跨越半万亿美元（约 5,150 亿美元）大关。这种暴涨让投资者与市场如梦初醒，突然意识到行业正面临着巨大的 DRAM 供给缺口。 - 美光科技（Micron）股价从去年 8 月的约 85 美元暴涨至 450 美元。 - 美光市值从原先的 1,750 亿美元膨胀至超过 5,000 亿美元。 - DRAM 存储芯片价格同比增幅高达约 100%。 --- ### [30:17](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=1817s) 纽村网红餐厅 vs. 寡头垄断的芯片格局 为了向不了解金融和半导体周期的外行解释这种狂热，Josh Brown 分享了一个他常用的类比：纽约目前最炙手可热的网红餐厅“The Corner Store”。这家店火爆到令人发指，人们周四下午 3 点就去排队，仅仅是为了在吧台占个座，排队盛况堪比“泰勒·斯威夫特的主题快餐店”。但 Josh 直言，那里的食物其实就是非常不错的烤牛肉三明治，同个街区还有 500 家其他餐馆，唯独这家店门庭若市。许多科技股目前也处于类似状态：资金蜂拥而入挤破门槛，哪怕股价短线回调也会被瞬间拉起，因为所有人都坚信这家店是唯一的去处。但 Josh 也提醒，追涨买入不能只看过去一年的走势，这种行情很难无限复制。 Crawford 对这个比喻提出了尖锐的反驳。她指出餐馆可以有成百上千家，但半导体产业链有着极其残酷的演进历史。在她职业生涯刚起步时，全球有十几家存储芯片公司——台湾有三家政府支持的厂商、韩国有三家、美国还有美光等。然而，随着芯片制程越来越微缩、研发制造成本呈指数级上升，行业经历了一轮又一轮惨烈的洗牌与合并。当年内存过剩时，这些公司的市盈率跌到只有 8 到 10 倍，几乎白送都没人要。如今洗牌结束，全球 DRAM 供应只剩下了三家寡头厂商（三星、SK海力士、美光）。投资者和科技巨头排队的原因不是盲目跟风，而是在物理世界与产业链中，根本没有第四个替代选项。 --- ### [34:53](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=2093s) 数据爆炸与机械硬盘复兴：为什么说“泡沫论”极其愚蠢 Michael Batnick 随即提出疑问：西部数据（Western Digital）的市值在 2025 年 1 月时仅有 150 亿美元左右，如今已经飙升至近 1,250 亿美元，这难道还不是泡沫吗？Crawford 坚定地回答：“不，这不是泡沫。”她解释称，随着 AI 视频生成和各类 AI Agent（智能体）的爆发，生成的数据量正在呈爆炸式增长，所有这些庞大的数据都必须在物理设备中找到存储归宿。 这一逻辑与机械硬盘（HDD）的行业结构如出一辙。五年前 HDD 还是一个无人问津的冷门行业，但如今市场上基本上只剩下两家核心玩家：希捷（Seagate）和西部数据（Western Digital）。更重要的是，这些厂商并没有盲目扩建新的物理生产线，而是通过提高盘片存储密度来增加容量，产能扩张节奏非常克制，不会引发恶性的供需失衡。Josh 调侃 Michael 依然停留在线性思维中，而 AI 时代必须用指数级思维来审视。Crawford 明确表示，现在很多市场看客挂在嘴边的“AI 泡沫论”是“极其愚蠢的”（really stupid），随后二人就这一反共识的核心分歧展开了进一步探讨。 - 西部数据（Western Digital）市值从 2025 年 1 月的约 150 亿美元飙升至近 1,250 亿美元。 - 全球大容量机械硬盘（HDD）市场形成了希捷与西部数据的双寡头格局。 - 供给端通过提升存储密度而非盲目扩建物理厂房，避免了产能过剩。 ### [36:28](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=2188s) 算力供需失衡与被盈利推高的估值 讨论从对 AI 投资是否构成泡沫的争论切入。尽管有人担心对算力的大量投资最终可能无法转化为实际收入，但 Ankur Crawford 明确表示，至少到 2026、2027 甚至延续到 2028 年，基于已公布和预期的资本开支（CapEx），算力市场绝不可能达到供求平衡状态，卖方市场具备极强的抗风险韧性。至于 2030 年会如何，完全取决于未来几年实际落地的基础设施规模；这就像在 1996 年去预测 2000 年是否会出现互联网泡沫一样，谁也无法准确预知几年后大众会把资金投向何方。但就未来几年的基本面与资本投入而言，算力短缺的失衡格局依然牢固地利好供应方。 针对市场上充斥的“泡沫论”，Ankur 强调必须正视企业业绩的基本面增长。很多人只看到股价暴涨，却忽视了盈利预期的飙升速度更为惊人。以西部数据（Western Digital）为例，在短短三个月内，其 2026 或 2027 年的每股收益（EPS）预期就从 10 美元大幅上调到了 25 美元，这本质上已经是完全不同的公司了。对于英伟达（Nvidia）、美光（Micron）等公司也是如此——在指责它们处于泡沫之前，必须先回答合理的盈利数字究竟是多少。如果基本面在以倍数级扩张，单纯依靠股价涨幅来判定泡沫是站不住脚的。 - **算力供需展望**：预计在 2026 至 2028 年间，算力需求将持续超过供给，卖方定价权依然稳固。 - **盈利上修示例**：西部数据未来几年的每股盈利预期在短短一个季度内从 10 美元直接跃升至 25 美元。 - **生命周期转折**：当前底层硬件与算力产业正经历典型的生命周期重构，市场结构的变化推动了企业基本面的蜕变。 --- ### [38:11](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=2291s) 模式匹配的误区：为何 AI 不会重蹈光伏与电动车的覆辙 Josh Brown 提出，投资者往往习惯于机械地做“历史模式匹配”，比如看到近期光伏和电动车（EV）股票的崩盘，就认为 AI 概念股也会走向毁灭。2021 至 2022 年间有约 20 家电动车相关公司上市，如今大量跌至个位数甚至归零；光伏板块也曾经历过大起大落。面对这种“涨得过猛就必然崩盘”的定势思维，Ankur 直言，许多人谈论泡沫仅仅是因为“当你不理解某种事物时，最省力的解释就是说它是泡沫”。她接触过许多医疗或消费品背景的投资组合经理（PM）与首席投资官（CIO），面对底层技术架构的剧变，缺乏半导体与材料科学背景的人很难真正理清其运转逻辑，而这恰恰构成了她在科技投资上的认知优势。 为了剖析 AI 硬件与昔日光伏泡沫的本质差异，Ankur 回忆了早年跟踪光伏行业的经历。当时尚德、天合光能（Trina Solar）、晶科（Jinko Solar）、阿特斯（Canadian Solar）等企业蜂拥而至。作为加州大学伯克利分校材料科学专业的毕业生，当她听到有人“在自家车库里搓出一家光伏组件厂然后以 24 亿美元估值 IPO”时，便敏锐地意识到这一产业的技术门槛极低。任何懂得光伏电池（photovoltaic cell）制造的人都能切入，一旦有利润空间就会引来疯狂的产能复制与价格战，再加上产业过度依赖忽冷忽热的政府补贴，最终导致大量资本血本无归。而如今的 AI 硬件无论在利润率、技术门槛还是现金流支撑上，都与当年的光伏不可同日而语。 - **认知偏差的心理学**：在华尔街，“泡沫论”往往是“我错过了行情”或“我无法理解复杂技术”的掩饰代名词。 - **技术门槛的本质区别**：光伏组装的技术壁垒极低，属于拼成本的重制造与补贴驱动型周期品；而先进制程与 AI 算力具备极高的研发与技术封锁门槛。 - **真金白银的现金流支撑**：AI 算力产业链各环节从一开始就具备极高的利润率，巨头用于采购的巨额资本开支来自自身充沛的经营现金流，而非靠资本市场补贴烧钱。 --- ### [42:30](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=2550s) 半导体行业二十年大整合与极高的技术壁垒 Ankur 进一步指出，当下的 AI 繁荣建立在过去二十年（2000 年至 2020 年）整个半导体供应链深度整合的基石之上。随着物理极限逼近，半导体技术的复杂度和成本呈指数级膨胀：晶圆尺寸从 6 英寸向 8 英寸演进，芯片流片（tape-out）的研发成本在每一个先进制程节点都会翻上 3 至 5 倍。天价的晶体管密度测试与研发预算，迫使半导体设备商、服务器代工厂、芯片设计公司和网络硬件商只能通过并购整合来获取规模效应，以便用庞大的营收来摊销研发费用。 这种残酷的自然淘汰让行业格局变得极度寡头化。过去每周都有一家半导体公司上市、市场上有四五十家主要玩家的光景早已一去不复返；如今市场上只剩下寥寥数家芯片巨头，晶圆代工领域甚至实质上只有台积电一家最为关键（“and it really one that matters”）。即便是像 2024 年上市的 Astera Labs 这样的稀缺标的，也仅是近年来极少见的特例。整个底层供应链被高度垄断在少数掌握极端技术壁垒的巨头手中，根本不存在外部竞争者凭空入局打价格战的可能。 ```mermaid flowchart TD accTitle: 半导体产业二十年整合与壁垒形成 accDescr: 展示技术复杂度和成本飙升如何推动半导体产业链从高度分散走向寡头垄断 subgraph CostPush["研发成本与复杂度指数级飙升"] A["先进制程节点演进"] --> B["流片研发成本翻 3-5 倍"] B --> C["晶圆与晶体管测试门槛剧增"] end subgraph Consolidation["2000-2020 供应链深度大整合"] C --> D["半导体设备商整合"] C --> E["网络与服务器厂商兼并"] C --> F["芯片设计巨头寡头化"] end subgraph MarketMoat["极高的竞争壁垒格局"] D & E & F --> G["先进制程代工高度集中\n(台积电等极少数晶圆厂)"] G --> H["彻底隔绝新入局者竞争\n保障极高利润率与定价权"] end ``` --- ### [44:11](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=2651s) 超大规模云厂商的阵痛：从轻资产向重资产的估值切换 针对整个 AI 叙事中最核心的矛盾，Josh 提出了所有投资人最悬心的问题：超大规模云厂商（Hyperscalers，如亚马逊 AWS、微软、Meta、谷歌）正在从以往被市场极其推崇的“轻资产（Asset Light）”模式，急剧转向大肆采购服务器和建设数据中心的“重资产（Asset Heavy）”模式，它们倾注了全部现金流所换来的投资回报率（ROI），能否支撑这台机器无限运转下去？ Ankur 坦言，即使是作为成长股基金经理，她也无法打包票这些巨头最终能百分之百获得预期中的回报率。在 Meta 等巨头大幅调高 CapEx 指引之前，她最大的担忧正是超大规模厂商的自由现金流可能直接被巨额资本开支吞噬殆尽，降至“零现金流”水平。以往成长型投资人习惯于依据高额现金流给这些平台型科技股赋予高溢价；而如今，巨头们把几乎所有的经营现金流重新砸进资本开支，导致市场估值体系被迫在市盈率（P/E）、自由现金流倍数、市销率（EV/Sales）之间经历痛苦的逻辑切换与重估（derating）。这种估值范式的转换期往往充满震荡与阵痛，也正是近期大型科技股在资本开支膨胀后股价陷入胶着和盘整的深层原因。 ### [45:55](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=2755s) 资本开支与云计算周期的对比：时间维度的极致压缩 当讨论转向科技巨头们前所未见的资本支出周期时，主持人指出自由现金流几乎被全部吞噬，源源不断地回流进数据中心的建设之中。Dr. Ankur Crawford 将这种剧烈的资本消耗与 2011 至 2014 年间云计算基础设施的起步阶段做了类比。在那个时代，亚马逊（Amazon）早在公众意识到之前就已经在内部大量投资构建云平台，而微软（Microsoft）虽然开销巨大，但整个进程并没有引发如今这般触目惊心的集中花费。 ```mermaid flowchart LR accTitle: 云计算与 AI 资本开支周期演进对比 accDescr: 对比传统云计算在漫长周期内的逐步扩展与当前 AI 因能力爆发导致的时间窗口极限压缩与资本集中投入 subgraph CLOUD["传统云计算周期 (2011–2014)"] direction TB C1["基础设施逐步部署"] --> C2["网络与算力线性扩张"] C2 --> C3["资本开支平摊至漫长时间跨度"] end subgraph AI["当前生成式 AI 周期"] direction TB A1["技术能力呈指数级爆发"] --> A2["时间维度极度压缩"] A2 --> A3["海量资本集中涌入数据中心与芯片"] end CLOUD -. "演进为超高速节奏".-> AI ``` 之所以当前 AI 的资金支出显得如此庞大与激进，并非两者在本质逻辑上有何不同，而是因为技术的成熟曲线被极度压缩了。Dr. Crawford 指出，如果在云计算初期就同时赋予市场 5G 级别的高速网络和完全成型的云端能力，当年的资金消耗同样会是天文数字。如今的 AI 具备指数级增长的复合能力，巨头们必须在极短的时间窗口内完成全套基础设施的布局。 - 资本支出的具体流向：数据中心土建、先进制程芯片、液冷及配电系统等一整套底层设施。 - 核心矛盾点：过去十几年间逐步分散消化的云基建投入，在 AI 浪潮中被强行压缩到短短数年内集中释放。 --- ### [47:18](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=2838s) 亚马逊的估值、Anthropic 合作与自研芯片 Trainium 的现实考量 谈到亚马逊的价值支撑时，主持人询问了其与 Anthropic 的战略关系，以及早期 AWS 推崇的 Bedrock「模型自由接入（Bring Your Own Model）」理念。Dr. Crawford 强调，尽管对 Anthropic 的投资和合作直接推动了 AWS 收入实现 30% 中高段的强劲增长，但这并不是投资亚马逊唯一的底层逻辑。Anthropic 未来不可避免地会多云布局、服务其他竞争平台，亚马逊的真正护城河依然在于其被严重低估的整体资产与自研技术层。 从估值角度来看，亚马逊按照 GAAP 市盈率（P/E）计算已跌入十几倍的低估值区间，甚至比好市多（Costco）或沃尔玛（Walmart）等纯零售超市低了近 10 个点。面对主持人关于 Andy Jassy 声称自研芯片若对外销售可形成 500 亿美元业务的探讨，Dr. Crawford 给出了客观而冷静的审视： - **自研芯片定位**：亚马逊的 Trainium 芯片与谷歌研发多年的 TPU 相比仍有明显代差。谷歌 TPU 在性能上能真正叫板英伟达（Nvidia），源于其整整抢先了五年的研发沉淀；Trainium 目前表现尚可（"an okay chip"），亚马逊仍在努力追赶。 - **商业模式现实**：亚马逊真正的赢面并不在于转型成为向第三方大举售卖芯片的硬件供应商，而是在于通过自研芯片构建更廉价的自营算力层，使企业客户在处理特定任务负载时拥有比使用昂贵 GPU 更经济的选择。 --- ### [51:36](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3096s) 并非所有 Token 皆平等：算力分层与企业定价重塑 对话触及了一个关于 AI 经济学的根本问题——"所有 AI Token 是否生而平等？" Dr. Crawford 斩钉截铁地回答："并不平等。" 目前市场仍然采用粗放的统一定价模式，但这只是基础设施初期的过渡现象。用最高规格的 GPU 算力去回答日常闲聊或查询天气，就像是在家里放进一头狮子来抓老鼠一样大材小用。 未来算力市场必然走向精细化分层。用于药物研发（Drug Discovery）、复杂推理或工业设计的核心 Token，其经济价值与算力消耗绝不可能等同于普通日常对话或教育辅助。随着软件架构和专业垂直应用的演进，企业端客户将被迫重新思考技术栈的搭配，针对不同重要级别的任务调用不同能耗与成本的芯片及模型，最终使得 AI 应用的售价与其所调用的 Token 真实价值深度绑定。 - **个人用户 vs 企业客户**：前端普通用户感知有限，但企业级服务的计费模型与应用定价将迎来颠覆性重构。 - **任务分流趋势**：通用且昂贵的高性能 GPU 将专注于高价值复杂任务，定制化 ASIC 芯片及轻量化模型则承接基础流水线操作。 --- ### [53:57](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3237s) 市场倒逼还是战略刚需？AI 支出与元宇宙幻灭的本质区别 主持人提出疑问：在当前科技股表现出现分化、部分巨头股价下行的背景下，市场力量是否会像当年逼迫扎克伯格（Mark Zuckerberg）终止元宇宙（Metaverse）烧钱计划一样，迫使萨提亚·纳德拉（Satya Nadella）等科技掌门人收紧资本开支？当年 Meta 在股价经历 70% 的断崖式下跌后，即便扎克伯格拥有绝对投票权，也不得不对市场情绪妥协并大举缩减开支。 Dr. Crawford 对此给出了明确的反驳，指出 AI 的投入逻辑与当年的元宇宙存在本质区别： | 维度 | Meta 元宇宙开支阶段 (Metaverse) | 当前生成式 AI 资本周期 (AI Infrastructure) | |:--- |:--- |:--- | | **战略定性** | 缺乏实际落地支撑的远期愿景探索 | 关乎核心业务生死存亡的生存之战（Existential Threat） | | **市场需求** | 纯供给端驱动，真实企业端与消费者需求薄弱 | 算力严重供不应求（Short Compute），各行业争相部署 | | **技术路径** | 在底层 AI 基础设施未成熟前试图强推虚拟交互 | 作为底座技术支撑一切未来应用（包括元宇宙最终落地） | 巨头管理层一再强调"投入不足是生存级威胁"，这并非顺应牛市情绪的口头狂欢，而是源于背后的真实需求迫近。即便是元宇宙想要真正变为现实，其前提也必须建立在今天正在疯狂布局的高阶 AI 算力底座之上。因此，即便短期内二级市场出现回调阵痛，巨头们踩刹车的门槛也远比元宇宙时期要高得多。 ### [55:12](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3312s) 元宇宙的超前教训与 AI 的 18 个月回本周期 主持人与嘉宾打趣谈起元宇宙，调侃大家当年是否在 Snoop Dogg 虚拟豪宅隔壁买过数字海滩别墅，或者陪孩子在 Roblox 里玩耍。Ankur Crawford 指出，元宇宙概念其实并非空想，未来确实可能成为现实，但它致命的缺陷在于“为时过早”——它需要强大的 AI 能力作为基石才能成立。在 2020 到 2021 年间，当扎克伯格（Mark Zuckerberg）甚至为了元宇宙将 Facebook 改名为 Meta 时，Ankur Crawford 就曾在一份分析中直言“这是属于 2030 年的事件”。当时 Meta 几乎单打独斗，在老旧且算力不足的硬件基础设施上强行推进，导致资本开支无法带来即时回报。 然而当前的 AI 热潮与当年的元宇宙有着本质不同。针对市场普遍焦虑的“巨额资本支出（CapEx）究竟何时能见投资回报率（ROI）”，Ankur Crawford 给出了一个出人意料的判断：科技巨头当下甚至能在 18 个月内收回投资成本。造成这种现象的根源在于算力供给严重短缺，导致算力成本逆势上涨——像英伟达 H100 这样的旗舰芯片，单价非但没有随着时间推移下降，反而一直在攀升。数据中心内每个芯片的时租定价极高，按现行价格测算，硬件投资只需约 18 个月就能回本。因此，在资本以极快速度回收的当下，企业根本不需要绞尽脑汁去为 ROI 辩护。更重要的是，一旦未来需求放缓，巨头们随时有能力迅速关小或关闭资本开支的水龙头，而目前的产业态势距离见顶还差得极远。 --- ### [57:23](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3443s) 所谓“循环交易”是利益输送还是商业常识？ ```mermaid flowchart LR accTitle: AI Ecosystem Investment Dynamics accDescr: Comparison between standard ecosystem seeding by Nvidia and high dilution hardware deals by AMD. subgraph NVDA_Model["英伟达生态培育模式"] NV["英伟达 (Nvidia)"] CW["CoreWeave 等云服务商"] NV -- "早期参投/算力赋能" --> CW CW -- "长期回购高额 GPU" --> NV end subgraph AMD_Model["高稀释度硬件绑定模式"] AMD["AMD"] BigTech["Meta / OpenAI"] AMD -- "附带股权认股协议 (高达10%)" --> BigTech BigTech -- "承诺采购达到指定营收" --> AMD end ``` 市场中坚定看空 AI 泡沫的人士经常抛出另一个质疑：当前的交易充满“循环注资”（Circular Financing）——供应商向客户投资，客户再拿这笔钱去采购芯片、算力或租赁数据中心，宛如 1990 年代末互联网泡沫时期的造假互换营收。对于此类质疑，Ankur Crawford 仔细拆解了不同交易的本质差异： - **真正令人担忧的模式**：例如 AMD 与 Meta 以及 OpenAI 之间的深度捆绑协议。在这类交易中，Meta 只要购买 AMD 芯片达到特定营收门槛，就能获得 AMD 甚至高达 10% 的股权认股权证。这种做法相当于 AMD 在以 0% 毛利的代价送出公司部分所有权来推销硬件，带有极重的资本输送色彩。 - **良性培育市场的模式**：像英伟达（Nvidia）对 CoreWeave 的投资则完全不同。英伟达未来的 AI 相关年营收规模高达数千亿美元，对 CoreWeave 投资的区区 20 亿美元根本无法在财务层面上人为虚增英伟达的收入大盘。 Josh Brown 补充道，这种商业策略在传统行业中随处可见且合情合理：正如百事可乐（Pepsi）会给餐馆免费提供招牌、遮阳篷和装潢赞助，换取餐馆常年独家采购百事饮料；或者啤酒品牌 Miller Lite 送给酒吧一张印有品牌 Logo 的台球桌，从而绑定长期的啤酒批发业务。AI 领域的巨头扶持伙伴本质上是一模一样的商业逻辑，只是金额从千美元级别放大到了数十亿美元，规模宏大得令普通投资者感到眩晕。 --- ### [01:01:00](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3660s) 2000年电信泡沫 vs 现今 AI：光纤沉睡与赛车加油 讨论随即深入到与 2000 年互联网及电信泡沫的深层技术对比。Ankur Crawford 强调，2000 年的崩盘在于资本开支投入的全是“长寿命重资产”（Long-lived Assets），例如被深埋地下的巨量光纤电缆。当时耗资数百亿美元铺设了海量暗光纤，但在长达 5 年内根本没有人使用，随后需求彻底坍塌。最关键的原因在于，当年的技术根本无法利用这些底层设施——那时的互联网体验极其糟糕，没有网络视频，没有社交媒体，也没有 YouTube，在 Napster 上下载一首歌曲需要整整 10 分钟，到了 2004、2005 年下载一部电影甚至需要几天时间。2000 年的人们构想了宏大的梦想，却没有实现该梦想的软硬件环境。 Ankur Crawford 提出了一个生动的赛车比喻：2000 年就像是 F1 工程师凭空构想了一辆超级赛车，到了赛场却发现连底盘和轮子都没做好；而今天的 AI 革命中，赛车的所有机械构件、四个轮子和传动系统早已完备，**现在唯一欠缺的仅仅是燃料（算力本身）**。资本支出直接注入到能够即刻运转、产生实用价值的成熟体系中，绝非当年的空转。 对于知名空头投资者 Michael Burry（“大空头”原型）指责大型科技公司在服务器与 GPU 折旧年限上做财务手脚、通过拉长折旧年限粉饰利润的言论，Ankur Crawford 予以了坚决驳斥。Burry 宣称这些硬件只有 4 年寿命，巨头们却按照 5 年甚至 6 年来折旧。Crawford 直言这种批评纯属无稽之谈：即便芯片迭代加快、不再是最尖端型号，旧款 GPU 在下游推理和特定负载上依然长久高效地运转并产生现金流，其真实寿命和利用率完全支撑现行的会计准则。 ### [01:03:50](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3830s) 算力芯片寿命误区与 Nebius 的传奇背景 对于市场上关于“AI 芯片三年就会报废”的说法，Dr. Ankur Crawford 直言无法理解。她指出许多 2021 年生产的算力芯片至今仍在高度活跃地运行，芯片并不像一些悲观论者想象的那样短命。随后，主持人提议聊聊她投资组合中的具体持仓代码，首先抛出的是神秘的代码 NBIS（Nebius）。Crawford 笑着感叹这是一个漫长而精彩的故事。 Josh Brown 补充自己的认知：Nebius 的前身是被称为“俄罗斯版谷歌”的搜索引擎巨头 Yandex。Crawford 纠正并梳理了这段离奇的历史。Yandex 原本由阿卡迪·沃洛兹（Arkady Volozh）掌舵。俄乌战争爆发后，因与克里姆林宫在政治立场上产生分歧，Arkady 属于因战事受到制裁但实质上已与普京决裂的商界人物。他毅然带领约 1500 名顶尖工程师出境离开俄罗斯，并由自己出资承担所有人员迁徙成本。带着剩余的 20 亿美元资金，核心团队坐在一起复盘未来，最终一致判定：“未来是属于 AI 的，而我们精通构建超大规模数据中心，且能比其他人做得更好。” ```mermaid flowchart TD accTitle: Nebius的历史演变流程 accDescr: 展示从Yandex分拆到欧洲新成立Nebius并在美股上市的路径 subgraph Origin["Yandex 分裂与重构"] A["俄罗斯巨头 Yandex\n(创始人: Arkady Volozh)"] -->|"俄乌冲突爆发\n创始人出资带 1500 名工程师离境"| B["手握 20 亿美元重组资本\n总部迁往荷兰阿姆斯特丹"] end subgraph Transformation["算力转型与重新上市"] B -->|"押注 AI 数据中心与原生算力"| C["转型为 AI Native 超大规模云厂商"] C -->|"纽交所通知紧急恢复交易\n代码: NBIS"| D["前 Yandex 套牢盘清盘离场\n机构投资者深度建仓"] end ``` - **团队转移成本**：创始人 Arkady Volozh 个人出资承担了全部 1500 名离开俄罗斯的工程师团队迁徙费用。 - **资本储备**：重组剥离出境后，团队手握约 20 亿美元可用启动资金，决定全力转型建设 AI 数据中心。 ### [01:05:38](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3938s) NBIS 意外挂牌与定位于“下一代 AI 原生超大规模云” 原 Yandex 股票在战争后遭美股退市冻结，新主体 Nebius 将总部设立在荷兰阿姆斯特丹。Crawford 透露，某周四团队突然接到纽交所通知，告知股票将于下周一挂牌交易。当时他们几乎等同于一家私有化运营的公司，整个团队连夜在周末加班冲刺完成上市准备，美股随之凭空出现了一个神秘代码 NBIS。开盘初期，市场上绝大部分交易者都是此前被深套在 Yandex 股票中多年的旧股东，大家蜂拥借机斩仓套现出逃。 当时外界对其业务一头雾水，Nebius 随即发布了一份长达 110 页的业务演示 PPT。Crawford 看到后立刻要求与管理层会面，然而对方行程排得密不透风。为了彻底弄清这家公司的底细，Crawford 一路追到瑞士达沃斯世界经济论坛，终于坐下来与管理层深入接触。虽然 Nebius 同时涉足自动驾驶等边缘业务，但在 Crawford 眼中，它的核心价值在于基础设施：它不仅是一家欧洲注册、全球化运营的数据中心运营商，更有潜力成为下一代原生 AI 云巨头（AI-native hyperscaler）。 - **上市波折**：在接到纽交所复牌通告后，团队用短短一个周末的突击工作实现了代码以 NBIS 重新挂牌。 - **核心定位**：Crawford 将其定性为具备自研全栈架构能力的“下一代 AI 原生超大规模云厂商”（the next AI-native hyperscaler）。 ### [01:07:08](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=4028s) 传统产业整合者 QXO：Brad Jacobs 的 50 亿美元 EBITDA 蓝图 讨论转向另一只持仓 QXO。Crawford 指出，这绝非传统的科技股，而是由连续创业传奇布拉德·雅各布斯（Brad Jacobs）操刀的建材分销平台。Brad Jacobs 此前曾成功一手打造了联合租赁（United Rentals, URI）与 XPO Logistics，功成身退后于大约一年半前创办 QXO，首笔核心交易便收购了屋顶建筑材料巨头 Beacon Roofing，随后又迅速将 Kodiak 收归麾下，意在整合极度分散的建筑建材供应链。 面对 Josh Brown“成长型基金为何要买传统屋面建材”的疑问，Crawford 给出了极其激进的业绩增长模型：目前收购的 Beacon EBITDA 约为 10 亿美元，但 Jacobs 的目标是在 2029 至 2030 年间，通过导入现代数字化管理、科技赋能、精细定价和效率革新，将公司的 EBITDA 推升至 50 亿美元。Jacobs 在此前行业中多次兑现过这类整合神话。Crawford 对其个人品质也赞不绝口，称其为“最令人印象深刻的领导者之一”，兼具工作狂特质与对员工极高的责任心。 | 核心维度 | 关键事实 / 战略规划 | |:--- |:--- | | **掌舵人历史战绩** | 连续创立 United Rentals (URI)、XPO Logistics (XPO) 等千亿级产业整合巨头 | | **整合赛道与逻辑** | 瞄准极度分散的建筑建材（Building Products）市场，整合大量缺乏科技能力的夫妻店（Mom and Pops） | | **近期并购推进** | 先后拿下 Beacon Roofing（屋顶材料）与 Kodiak，横向拓展建材全品类 | | **核心财务目标** | 计划将当前 Beacon 约 10 亿美元的 EBITDA，到 2029-2030 年大幅拉升至 50 亿美元 | ### [01:09:34](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=4174s) AppLovin 从游戏广告引擎跨入电商变现 随后话题转向被纳入 EC App 等基金中的持仓 AppLovin。面对“未来 AI 自行编写 App 可能会消灭应用生态”的质疑，Crawford 明确澄清，AppLovin 的核心逻辑不是做 App 开发工具，而是精细化的移动广告投放引擎。尽管由于被指数纳入科技软件类 ETF（如 IGV），其股价容易在“软件股大溃败”（SaaSpocalypse）的市场恐慌中被误伤躺枪，但它的内核是一家纯粹的广告科技巨头。 在手机移动端游戏广告领域，AppLovin 占据了大约 50% 的市场份额。过去其经典闭环是在《糖果传奇》（Candy Crush）中向玩家推荐下载《字谜游戏》（Scrabble）；而如今，公司利用其强大的 AI 算法模型，开始向广告流中精准投放口红等真实的实体消费品，正式切入电商广告变现。公司创始人兼 CEO 亚当·福罗吉（Adam Foroughi）作风饥渴果断，在营收规模狂飙的同时，严格控制公司员工总人数不扩张，将公司打造成了一台极高自由现金流的印钞机。 - **市场份额压制**：在移动游戏广告展示市场掌握约 50% 的市占率，积累了庞大的第一手用户行为数据。 - **第二增长曲线**：利用底层 AI 推荐引擎打破游戏内推游戏的局限，全面进军电商实体商品（如美妆个护）广告变现。 - **极致人效比**：创始人 Adam 严格压制员工 headcount 规模，在营收高速扩张期保持团队规模扁平固定。 - **管理层接触原则**：针对 Josh Brown 提问是否必须能接触到管理层才肯建仓，Crawford 坚定回应：高管最终是投资人资本的托管人（stewards of your capital），深入理解管理层至关重要。 ### [01:12:05](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=4325s) 管理层评估与英特尔（Intel）的复兴叙事 在探讨投资哲学时，主持人和 Ankur Crawford 深入探讨了如何评估上市公司管理团队的诚信与领导力。管理层是任何投资逻辑的关键支柱，但投资者也需警惕过度亲近带来的盲区——并非与高管成为私人朋友，而是考察他们是否为合格的“资本看门人（stewards of capital）”以及是否保持谦逊。 随后，Josh Brown 抛出近期市场备受瞩目的重磅标的——英特尔（Intel）。他认为英特尔堪称近期的“年度股票”，其从昔日道琼斯成份股、市值巨头的“折翼天使（fallen angel）”沦落至股价接近个位数，历经多任 CEO 换届失败后，终于迎来陈立武（Lip-Bu Tan）掌舵，重新引起成长型投资者的共鸣。面对市场对英特尔“大象转身”的兴奋情绪，Ankur 坦承自己尚未建仓，保持相对审慎： - **陈立武的扭转战绩**：陈立武是一位极具远见卓识的领导者，此前曾带领 Cadence 走出濒临破产的绝境，将其重塑为行业巨头；如果有谁能扭转英特尔，他无疑是最佳人选。 - **晶圆代工（Foundry）的执行挑战**：尽管市场当前将英特尔的代工业务视作具有国家战略意义的核心资产，且全球正处于“算力短缺”期，但造芯与运营晶圆厂极度复杂。目前全球顶级代工几乎只有台积电（TSMC）一家独大（台积电自身也在亚利桑那州建厂推进）。 - **工程执行力存疑**：代工业务的成败根本上取决于工厂内工程师团队的执行力，而陈立武本人并非一线半导体制造工程师，英特尔能否成功转型为代工厂仍是一个“生死攸关的存亡问题（existential question）”。 --- ### [01:15:33](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=4533s) SaaS 大崩盘（SaaS-pocalypse）与 AI 代理的颠覆 关于企业级软件公司是否正在遭遇灭顶之灾（SaaS-pocalypse），市场存在巨大分歧。英伟达 CEO 黄仁勋（Jensen Huang）此前曾为软件股站台，称空头完全看错了逻辑——智能体（Agentic AI）并不会凭空自建底层系统，而是会全面调用现有 CRM 和企业软件接口，因此传统 SaaS 公司的护城河与生态依然坚固。然而 Ankur 明确表示“恕难苟同”，认为黄仁勋受制于商业立场“无法直言不讳”。如果 Anthropic 等大模型厂商以当前的速度持续进化，传统软件企业很难维持既往的高估值与护城河。 ```mermaid flowchart TD accTitle: SaaS商业模式面临的AI颠覆压力 accDescr: 展示企业自研、AI减少员工数以及软件使用率低下如何瓦解SaaS席位收费模式与利润率。 subgraph CostPressure["企业端诉求与AI替代"] A["现有软件利用率低<br/>(企业仅使用15-20%功能)"] --> B["IT部门利用AI工具<br/>自研定制系统替代第三方案"] C["AI替代初级劳动力<br/>企业缩减员工编制 (Headcount)"] --> D["按席位收费 (Per-Seat) 模式受阻"] end subgraph Impact["SaaS公司财务与估值冲击"] B --> E["营收净留存率 (NDR) 恶化<br/>失去年度提价能力"] D --> E E --> F["营业利润率长期承压<br/>(从 40% 目标跌至 ~20%)"] F --> G["估值重塑：沦为类似于<br/>'受损科技资产 (Broken Tech)'"] end ``` 以 Salesforce 为例，剔除股权激励（SBC）后其企业价值倍数（EV/FCF）约为 10 倍，计入 SBC 后则在 14.5 倍左右，交易价格宛如“带有微弱期权价值的受损科技资产（broken tech with a little option value）”。SaaS 商业模式的核心痛点在于： - **功能浪费与付费模式重构**：企业通常仅使用打包软件中 15% 到 20% 的功能。客户越来越不愿按人头（per-seat）每年接受价格上涨，转向按使用量（per-use）付费。 - **企业内部自研威胁**：尽管出于安全性考虑，企业高管不会采纳员工随意用 AI 随性编写（vibe coding）的代码，但企业正规 IT 部门利用 AI 工具，完全有能力在内部复刻外部软件，为公司在未来五年内节省数千万甚至高达 8,000 万美元的开支。 - **席位增长神话破灭**：随着 AI 对各行业雇佣人数构成潜在替代，企业净增员工规模放缓，SaaS 最关键的净营收留存率（NDR / DBNR）面临长期下行。 - **利润率预期腰斩**：过去支撑高估值的“40% 终局营业利润率（operating margin）”梦碎，Ankur 直言在此竞争格局下，这类企业的稳态营业利润率可能只有 20% 左右。 --- ### [01:19:49](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=4789s) 网络安全护城河与单点工具（Point Solutions）的消亡 在软件板块普遍承压的背景下，主持人和 Ankur 讨论了网络安全（Cybersecurity）是否具备免疫力。Ankur 认为网络安全领域将产生分化： - **平台型巨头具备韧性**：如 CrowdStrike、Palo Alto Networks 以及微软等具备平台属性的企业，能将 AI 能力原生集成到自身庞大的生态架构中，具有较强的抗周期与抗颠覆能力。 - **单点解决方案（Point Solutions）岌岌可危**：所有专注于单一细分功能的软件公司都将面临灭顶之灾，它们极易被企业内部开发的代码、AI 原生初创公司，或平台型巨头的捆绑功能所迅速吞噬替代。 - **垂直利基软件受冲击**：如 Constellation Software 旗下收购的大量细分垂直行业软件（例如专供全美干洗店使用的标准化管理系统），在 AI 降低定制开发门槛的浪潮下，其原先被视作高壁垒的利基市场正在重新受到市场审视。 ### [01:20:47](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=4847s) 软件股分化、可替代性与被错杀的标的 讨论聚焦于软件行业近期的大幅回撤以及不同标的之间的明显分化。Josh Brown 指出，市场并非在盲目恐慌，而是表现出了相当的理性：那些看似非常容易被 AI 替代的轻量级边缘软件跌幅极深，而像 CrowdStrike 凭借其 Falcon 平台所构筑的核心防御体系，受影响程度则完全不同。在整个软件板块下挫的背景下，Cadence 的抗跌表现远好于 Atlassian（TEAM），因为 Atlassian 被认为缺乏真正的护城河。 ```mermaid flowchart TD accTitle: 软件板块分化逻辑 accDescr: 展示在 AI 浪潮下不同类型软件公司的防御力与脆弱性差异 subgraph Vulnerable["高脆弱性 / 护城河薄弱"] A["轻量级单点工具 / 任务协同"] --> B["Atlassian (TEAM) 等缺乏护城河标的"] end subgraph Defensible["高护城河 / 核心生态壁垒"] C["EDA与工业设计核心软件"] --> D["Cadence"] E["底层安全防护平台"] --> F["CrowdStrike Falcon"] G["复杂企业级数据中枢"] --> H["Palantir (跨越鸿沟者)"] end ``` 当被问及当前是否有软件股被市场过度错杀、反而迎来绝佳买点时，Ankur Crawford 提到了 AppLovin。她认为 AppLovin 只是因为被编入相关软件指数而连带遭受抛售，加上空头报告发难以及市场担忧“智能体广告”（agentic advertising）对其模式的冲击；但实际上它被误伤了。谈及 Palantir 时，Crawford 将其形容为带领大家“跨越鸿沟”的牧羊人（"the shepherd that is basically having everyone cross the chasm"）。尽管 Palantir 在这一波 SaaS 惨跌潮之前曾从 210 美元回调到 160 美元，但其基本面表现出的韧性远超普通 SaaS 概念股，Alger 在其其他投资组合中也持有该股票。 ### [01:22:22](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=4942s) Adobe 的现状与新兴 AI 视频工具的探索 Josh 质疑 Adobe 是否已经走向衰退，市场上充斥着 Adobe 毫无特殊之处的言论；但与此同时，OpenAI 关停或调整了 Sora，外部涌现的大量新型工具又令人怀疑究竟有没有用户愿意为此付费。 Crawford 分享了自己对前沿工具的实操体验。她最近正在测试 Higgsfield AI 这一视频生成工具，其直观感受是“相当惊艳”（"pretty amazing"）。作为投资人，她习惯于亲手试玩各类工具以探寻其真实能力边界。她测试 Higgsfield 的目的是为了验证能否直接用它为自己的 App 制作社交媒体广告视频（比如生成一条 TikTok 推广短视频）。实际测试结果显示虽然尚未达到完美无缺的程度，但实用性已经非常可观。 ### [01:23:07](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=4987s) HALO 概念与“脱离屏幕”的未来硬件形态 Josh 谈及他提出的概念——“HALO”（重资产、低受创资产/重度物理资产属性），并打趣向 Crawford 索要评价。Crawford 笑称 Josh 的包装极为巧妙，但她指出在 AI 时代，随着内容创造边际成本的暴跌，任何具备稀缺性的有形资产（tangible asset）必然会享受估值倍数扩张（multiple expansion）。 在讨论美股七巨头（Mag 7）中谁最具 HALO 属性时，Josh 认为苹果和亚马逊排名最高。尽管苹果没有重资产数据中心，但其核心硬件——“每个人手中的长方形屏幕”构成了人机交互的绝对核心入口。然而，Crawford 提出了针锋相对的观点：如果苹果不迅速行动，其 HALO 属性将会彻底瓦解，因为长远来看，AI 最终将使用户从屏幕中解脱出来。 ```mermaid flowchart LR accTitle: 交互形态从屏幕到无感硬件的演进 accDescr: 从当前依赖手机屏幕向屏幕无感化硬件设备的迁移推演 A["当下交互模式<br>手机屏幕 (Rectangle)<br>每日开启150次"] --> B["过渡态与探索<br>智能眼镜 / 手环 / 吊坠 / 随身AI设备"] B --> C["终极形态<br>去屏幕化个人智能代理 (Screenless Agent)"] C -.-> D["潜在冲击<br>依赖滑屏时长的广告巨头 (Meta / TikTok)"] ``` - **屏幕替代的形态猜想**：未来的硬件不一定是失败的单品，可能是智能眼镜、手环、吊坠（amulet）或常伴身旁的个人智能体硬件。 - **苹果的战略迟缓**：Crawford 与 Josh 均对苹果至今未能将 Siri 打造成全能个人助手感到费解（"Why doesn't Siri book my plane tickets?"）。两年前宣传的全新 AI 表现至今仍未兑现。 - **隐私安全的包袱**：苹果迟缓的症结或许在于其根深蒂固的“保护隐私与安全”标签，如何在提供深度智能体服务的同时守住这一人设，构成了其内部战略困境。 - **对社交媒体广告模式的降维打击**：内部讨论曾推演，一旦人类不再频繁低头滑动长方形手机屏幕，Meta 旗下 Instagram 以及 TikTok 等依赖信息流滑动的商业模式将面临致命打击。 ### [01:27:51](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=5271s) 苹果资本开支克制与巨头盈利能力剖析 面对新型硬件何时替代手机的争论，Josh 态度相对保守，举例谷歌眼镜（Google Glass）早在 12 年前便遭遇惨败，Snapchat Spectacles 毫无声量，苹果的 Vision Pro 更是销量惨淡（调侃 Michael 是唯一买过且后来退货的人）。他坚信短时间内没有任何设备能割裂人类与手中的长方形手机。Crawford 认同眼前现状，但警告在 3 年周期的投资视野中，如果苹果持续交不出变革性的解法，其商业模式风险将急剧攀升。 对于苹果资本开支（capex）同比下降的现象，Crawford 认为这一策略其实异常精明：苹果让其他科技巨头去疯狂砸钱承担重资本开支，自己则安心充当最上层的平台分配者，坐收 App Store 抽成并调用他人的算力资产。 在争论“苹果是否仍是全球最好的商业模式”时，讨论转向了利润率与体量的对比： - 苹果凭借庞大体量拥有出色的盈利质量，但其毛利率并非全行业顶尖，大约维持在 38% 到 40% 的区间。 - 相比之下，英伟达（Nvidia）与谷歌（Google）的毛利率均明显高于苹果，其中英伟达在当前算力供应链中占据着最为强势的利润制高点。 ### [01:29:13](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=5353s) 节目收尾、未来邀约与致谢 在对人工智能与半导体生态展开了极为深入的探讨后，节目进入尾声阶段。主持人 Josh Brown 与 Michael Batnick 对嘉宾——来自 Alger 的基金投资经理 Ankur Crawford 博士——给予了极高的赞誉。Josh 感慨地表示，本期对谈内容极其发人深省（illuminating），自己在整场交流中收获颇丰，盛赞 Ankur 的表现无与伦比。现场响起了掌声，录播间气氛轻松愉悦。当被问及未来是否愿意再次作客节目时，Ankur 毫不犹豫地表达了喜爱之情：“噢，我非常乐意。”Josh 随即半开玩笑地顺水推舟追问“那你下周有空吗”，引得现场一片欢笑，而搭档已在一旁连连摇头表示档期显然不会排得这么急促。 谈及投资者如何获取更多专业研究与基金资讯时，Josh 指引大家前往 Alger 的官方网站（alger.com）。他随后好奇地询问 Ankur 平时是否活跃于社交媒体平台，在得到“完全不玩”的否定回答后，Josh 赞许地表示这反而是他最欣赏的一点。Ankur 则幽默地调侃道：“或许之后我会写个机器人程序（bot）帮我发动态。”此外，Josh 敏锐地捕捉到了 Ankur 对机器人技术（Robotics）话题流露出的极大热情与兴奋表情，Ankur 更是直言关于机器人领域“我们还有太多内容可以聊”。双方当即达成默契，敲定在今年晚些时候再度合作录制一期专门探讨机器人产业的特辑。 在最后向听众告别的环节，Josh 再次向 Ankur 致以诚挚感谢，并热情呼吁各大音视频平台上的所有观众与听众点击点赞按钮支持本期内容。制作人 John 也在一旁连声附和“百分之百赞成”，为这场聚焦 AI 算力短缺、泡沫争议与科技前沿演进的高密度对谈画下了圆满句号。 - **机构资讯**：关注 Ankur Crawford 及其投资团队的听众可通过其资产管理公司官网（alger.com）查阅旗下基金与研究动态。 - **个人社交态度**：Ankur 本人几乎不参与社交网络运营，并打趣称未来或可编写 AI 机器人代劳。 - **后续专题预告**：双方约定将于今年晚些时候重聚，专门就“机器人技术（Robotics）”展开更深入的专题对谈。 ## Notable quotes > [08:48](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=528s) 这绝不是价值陷阱，而是‘尚未被认知的成长’。 — Ankur Crawford > [13:26](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=806s) 不，如果没有 AI，耗时将是无限长——因为我根本连第一行代码该从哪里写起都不知道。 — Ankur Crawford > [20:59](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=1259s) 每个人都缺算力（Everyone's short compute）。 — Ankur Crawford > [26:25](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=1585s) 听着，我们要讨论的核心度量单位其实是 Token（词元）。迈克尔可能需要恶补，我们台下 90% 的观众要么是替别人管钱的专业顾问，要么是精明的独立投资者——剩下的 10% 纯粹是疯子。你必须向大家阐明，为什么算力短缺最终会映射在 Token 上？你们究竟是如何从底层量化并思考这场计算资源危机的？ — Josh Brown > [34:53](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=2093s) 不，这不是泡沫。现在很多市场看客挂在嘴边的‘AI 泡沫论’是极其愚蠢的。 — Ankur Crawford > [38:11](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=2291s) 当你不理解某种事物时，最省力的解释就是说它是泡沫。 — Ankur Crawford > [51:36](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3096s) 所有 AI Token 是否生而平等？并不平等。 — Ankur Crawford > [55:12](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3312s) 这是属于 2030 年的事件。 — Ankur Crawford > [01:01:00](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3660s) 2000 年就像是 F1 工程师凭空构想了一辆超级赛车，到了赛场却发现连底盘和轮子都没做好；而今天的 AI 革命中，赛车的所有机械构件、四个轮子和传动系统早已完备，现在唯一欠缺的仅仅是燃料。 — Ankur Crawford > [01:03:50](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=3830s) 未来是属于 AI 的，而我们精通构建超大规模数据中心，且能比其他人做得更好。 — Arkady Volozh > [01:15:33](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=4533s) 带有微弱期权价值的受损科技资产（broken tech with a little option value）。 — Ankur Crawford > [01:20:47](https://www.youtube.com/watch?v=w5KlIB4JmKY&t=4847s) Palantir 就像带领大家跨越鸿沟的牧羊人（the shepherd that is basically having everyone cross the chasm）。 — Ankur Crawford ## Mentioned Josh Brown, Michael Batnick, Dr. Ankur Crawford, Fred Alger, Satya Nadella, Reid Hoffman, Andy Jassy, Mark Zuckerberg, Michael Burry, Arkady Volozh, Brad Jacobs, Adam Foroughi, 陈立武 (Lip-Bu Tan), 黄仁勋 (Jensen Huang), Snoop Dogg, Alger, TSMC (台积电), Quarter, Era, Walmart, Costco, Danaher, Perplexity, OpenClaw, Anthropic, OpenAI, Databricks, Palantir, Chime, Uber, Nvidia, Micron (美光), Samsung (三星), SK Hynix (SK海力士), Western Digital (西部数据), Seagate (希捷), Suntech (尚德), Trina Solar (天合光能), Jinko Solar (晶科能源), Canadian Solar (阿特斯), Astera Labs, Amazon (AWS), Meta, Google, AMD, CoreWeave, Nebius (NBIS), Yandex, QXO, Beacon Roofing, Kodiak, United Rentals (URI), XPO Logistics, AppLovin, Intel (英特尔), Salesforce, CrowdStrike, Palo Alto Networks, Constellation Software, Cadence, Atlassian, Adobe, Higgsfield AI, Apple, TikTok, Instagram, CNEQ (Alger Concentrated Equity ETF), Common App, DRAM, H100, Trainium, TPU, Bedrock, HALO, Falcon, Sora, Vision Pro ## Video description What if the AI boom isn’t a bubble, but just getting started? On Episode 238 of The Compound & Friends, Josh Brown and Michael Batnick sit down with Dr. Ankur Crawford (Portfolio Manager, Alger) for a deep dive into what’s really happening behind the AI explosion, and what most investors are still missing. They cover: • Why the world is “short compute” (and what that actually means) • The truth about AI stocks, Nvidia, and the semiconductor trade • Why Amazon might be one of the most misunderstood AI winners • The real reason AI demand could be exponential, not cyclical • The biggest risk to Big Tech (and it’s not what you think) • Whether this is a bubble… or the early innings of something massive Plus: inside insights on chips, data centers, tokens, hyperscalers, and how AI is already changing the way businesses, and investors, operate. This episode is sponsored by Invesco. To learn more, visit ⁠https://www.invesco.com/us/en/solutions/invesco-etfs/income-advantage.html?audienceRole=FinancialProfessional ►00:00 - Cold Open ►05:55 - Intro ►07:43 - Ankur Crawford's Investment Philosophy ►11:21 - Where are we in the AI story ►20:57 - Market Imbalances and Compute Shortage ►34:57 - Is This a Bubble? ►44:06 - Are All AI Tokens Created Equal? ►01:04:04 - Talking Stocks: NBIS, QXO, INTC ►01:15:29 - SaaSpocalypse ►01:23:02 - Have you heard about HALO? 👕 Check out The Compound shop: https://www.idontshop.com 🎙️Listen to our podcasts: TCAF: https://podcasts.thecompoundnews.com/show/TCAF/ Animal Spirits: https://podcasts.thecompoundnews.com/show/animalspirits/ Ask The Compound: https://podcasts.thecompoundnews.com/show/ATC/ The Compound Newsletter: https://www.thecompoundnews.com/subscribe Linkedin: https://www.linkedin.com/company/the-compound-media/ Instagram: https://instagram.com/thecompoundnews Twitter: https://twitter.com/thecompoundnews Tik Tok: https://www.tiktok.com/@thecompoundnews Investing involves the risk of loss. This podcast is for informational purposes only and should not be or regarded as personalized investment advice or relied upon for investment decisions. Josh Brown and Michael Batnick are employees of Ritholtz Wealth Management and may maintain positions in the securities discussed in this video. All opinions expressed by them are solely their own opinion and do not reflect the opinion of Ritholtz Wealth Management. See our disclosures here: https://ritholtzwealth.com/podcast-youtube-disclosures/ The Compound Media, Incorporated, an affiliate of Ritholtz Wealth Management, receives payment from various entities for advertisements in affiliated podcasts, blogs and emails. Inclusion of such advertisements does not constitute or imply endorsement, sponsorship or recommendation thereof, or any affiliation therewith, by the Content Creator or by Ritholtz Wealth Management or any of its employees. For additional advertisement disclaimers see here https://ritholtzwealth.com/advertising-disclaimers. #thecompound #ai #investing #stocks #techstocks #artificialintelligence #aistocks [Source](https://www.youtube.com/watch?v=w5KlIB4JmKY)

[8:48] Ankur Crawford: 这绝不是价值陷阱，而是‘尚未被认知的成长’。 [13:26] Ankur Crawford: 不，如果没有 AI，耗时将是无限长——因为我根本连第一行代码该从哪里写起都不知道。 [20:59] Ankur Crawford: 每个人都缺算力（Everyone's short compute）。 [26:25] Josh Brown: 听着，我们要讨论的核心度量单位其实是 Token（词元）。迈克尔可能需要恶补，我们台下 90% 的观众要么是替别人管钱的专业顾问，要么是精明的独立投资者——剩下的 10% 纯粹是疯子。你必须向大家阐明，为什么算力短缺最终会映射在 Token 上？你们究竟是如何从底层量化并思考这场计算资源危机的？ [34:53] Ankur Crawford: 不，这不是泡沫。现在很多市场看客挂在嘴边的‘AI 泡沫论’是极其愚蠢的。 [38:11] Ankur Crawford: 当你不理解某种事物时，最省力的解释就是说它是泡沫。 [51:36] Ankur Crawford: 所有 AI Token 是否生而平等？并不平等。 [55:12] Ankur Crawford: 这是属于 2030 年的事件。 [61:00] Ankur Crawford: 2000 年就像是 F1 工程师凭空构想了一辆超级赛车，到了赛场却发现连底盘和轮子都没做好；而今天的 AI 革命中，赛车的所有机械构件、四个轮子和传动系统早已完备，现在唯一欠缺的仅仅是燃料。 [63:50] Arkady Volozh: 未来是属于 AI 的，而我们精通构建超大规模数据中心，且能比其他人做得更好。 [75:33] Ankur Crawford: 带有微弱期权价值的受损科技资产（broken tech with a little option value）。 [80:47] Ankur Crawford: Palantir 就像带领大家跨越鸿沟的牧羊人（the shepherd that is basically having everyone cross the chasm）。

[7:43] Ankur Crawford: When he incepted what growth is, it really came in two different forms... the essence of it was where is the change? Because oftentimes where there's change, there's unrecognized opportunity. [11:21] Ankur Crawford: As analysts we think linearly, and the market thinks linearly, but we are in a time of exponential growth. And so it's really hard to get your arms around what exponential actually means... [35:01] Ankur Crawford: Tell me what the right numbers are for Nvidia. Tell me what the right numbers are for Micron and Western Dig, and I'll tell you if we're in a bubble or not. [44:11] Ankur Crawford: I'm really worried that the hyperscalers are going to zero cash flow... Whenever you have a valuation regime handoff to going from PE to cash flow or cash flow to PE... there's always a period of trickiness. [51:36] Ankur Crawford: I fundamentally believe that this market is going to... there's going to be tokens that you pay a lot for, and there's going to be tokens that... when I ask GPT what the weather is going to be... why should we be paying the same dollar value for each of those tokens? [83:02] Dr. Ankur Crawford: If Apple doesn't act soon, that HALO designation is going away. Because over time, I do believe that AI will allow us to extract ourselves from our screens.

[0:00] I listened to my first time when semi [0:01] Ernie's call today. [0:02] >> Oh, you did? [0:03] >> It was very strange, I thought. I'm sure [0:05] you listened to a million of them, that [0:07] the analyst ask questions and then the [0:09] CFO repeats the question that has been [0:12] asked [0:13] for every single question. It's bizarre. [0:16] >> I think um [0:18] >> know what's happening there because he [0:18] he literally is like, all right, so the [0:20] question was I'm like [0:21] Huh, that's weird. And then he just did [0:22] it the entire call. [0:23] >> So it used to be that there was like [0:25] actually an audience and so people would [0:27] ask in the audience and they were on [0:29] stage. [0:29] >> Oh, that makes sense. [0:31] >> was how it used to be and um [0:34] >> But you heard the call. Like I heard the [0:35] question. And then they just I don't [0:37] know, but that's that's a possible [0:38] explanation. [0:38] >> it's a cultural, you know, thing. [0:40] >> I thought it was odd that their their [0:41] deck looks like it was made in the '90s. [0:43] Like their investment [clears throat] [0:44] deck. What what is that? [0:45] >> [laughter] [0:46] >> Well, I mean it's it's only like a [0:48] trillion-dollar company, so you got to [0:50] give them some time. [0:51] >> looks like it was made on MS-DOS. Like [0:53] honestly. Like I I pulled one chart, but [0:55] it just looks like it hasn't been [0:56] updated since 1984. [0:57] >> you How do you listen to conference [0:59] calls? Do you like you dial in or do you [1:01] have you discovered quarter yet? [1:03] >> Yeah, I have I I've have quarter, but we [1:05] use something called Era. [1:06] >> Oh, okay. What's that? [1:07] >> And so it's like it's similar to [1:08] quarter. You basically go through and [1:10] you just click click through into a [1:12] >> Okay. [1:12] >> into a call. Makes it really easy. [1:14] >> Well, so we're investors in quarter, but [1:16] also like power users. I'm obsessed with [1:19] it. I I don't know how I ever used it [1:21] because I like seeing the words get [1:22] illuminated as they're being spoken. [1:24] >> Yeah. [1:25] >> reading it and hearing it at the same [1:26] time and I feel like it sinks in better. [1:28] >> So it's interesting. I talked to one of [1:29] your sales guys [1:30] >> Yeah. [1:30] >> from from quarter [1:31] >> Okay. [1:32] >> and told them that I think I forget [1:34] exactly what the pricing was, but Era [1:36] was giving us better pricing. [1:37] >> Okay. I know some people there, so [1:39] >> Yeah. [1:40] >> we'll we'll [1:40] >> they actually use quarter on the back [1:42] end. [1:42] >> Oh, really? [1:43] >> Yeah, you should [1:44] >> Yeah, white white labeling it. [1:45] [laughter] [1:46] Okay. All right. How many [1:48] >> And I'm supposed to [1:49] >> Yeah, please. Yeah, yes. [1:51] How many calls are you listening to [1:53] uh each quarter? Like do you have [1:56] do you have like people listening for [1:58] you on some or do you get to everything [2:00] yourself? [2:01] >> Um no, I have we have a team. Right? Um [2:05] and [2:07] yeah, so we have a team that um that [2:09] listens to everything and then [2:11] selectively will you know, I'll I'll go [2:13] through and and listen to certain calls. [2:16] >> Okay. What are the ones that you [2:18] personally won't miss? [2:20] >> Oh, I I mean definitively the [2:22] hyperscalers. Um I TS I did miss TSM [2:25] today because I had a missed a personal [2:27] um appointment. Um [2:30] >> Don't worry the AI stories and talked. [2:32] >> That's what they said. [2:33] >> Great. [laughter] [2:34] I read your notes. [2:37] Um [2:38] Yeah, so I I don't know. It depends on [2:40] which quarter it is. I'll I'll try to [2:42] listen to like Walmart or Costco to [2:44] understand what's happening on that side [2:45] of the consumer. I'll an industrial [2:47] company, a healthcare company. So just a [2:49] smattering because there's no way that [2:51] as a portfolio manager is where [2:52] [clears throat] we're going to be able [2:54] to get through the entire portfolio. [2:55] >> It's too Yeah, it's it's too many. It's [2:57] and it and they're all together and it's [2:59] like a fire hose. [3:00] >> Yeah. [3:00] >> You know what's a great service? There's [3:01] a company There's a subset called The [3:02] Transcript that I subscribe to where [3:05] they pull from like the most important [3:07] quotes from all different companies that [3:09] of course you can't listen to all of [3:10] them. So that gives you a good sense of [3:12] like I don't obviously don't listen to [3:14] >> They're doing that with AI. [3:16] >> Um [3:17] I don't know, man. This guy listens to a [3:18] lot of calls. But yeah, I'm sure. I'm [3:20] sure. You imagine. [3:21] >> alternatively you can just go into [3:22] Perplexity and I'll I'll give them the [3:25] the lineup for the day and [3:28] um [3:29] >> It works. But no, quarter can do that. [3:30] They can all do that. Like give me the [3:31] give me the best quotes. What did they [3:32] say about this? It's it's great. [3:35] >> Um all right. This is going to be fun. [3:36] So this is your first my understanding [3:38] your first podcast. [3:40] >> No. [3:40] >> Yes. [3:41] >> Um [3:42] it's my No. [3:44] >> He says no. [3:45] >> second. [3:45] >> What was What did you do Joe Rogan last [3:48] week or something? [laughter] What was [3:49] What was your [3:50] >> I don't know. What was it, Scott? [3:51] >> What What else is [3:52] >> Morningstar. [3:53] >> Oh, Morningstar. [3:54] >> All right, fine. But, we're going to [3:55] pretend this is your first one. [3:56] >> [laughter] [3:57] >> I'm sure it was great. Shout-out to [3:58] Morningstar. I'm sure it was terrific. [4:00] All right. Uh th- this this one's going [4:02] to This is going to be a lot of fun. [4:04] And, uh I'm so appreciative of you [4:06] coming and doing this. [4:07] >> Are you okay? [4:07] >> Yeah, these are kind of uncomfortable. [4:09] >> you have them on backwards. [4:10] >> Oh, I do? [4:11] >> [laughter] [4:11] >> Why didn't Why didn't you tell me that [4:13] before I told you I had them on? [4:14] >> I just noticed now. So, this should be [4:16] on the left your left side. [4:18] >> Is that better? [4:19] >> We got it now. [4:20] >> it's better. [4:20] >> All right. [4:20] >> Is my hair okay? [4:21] >> You look great. [4:22] >> Okay, great. Thanks. [4:22] >> Daniel, I blame you. [4:24] Yeah, I feel like you're not You're not [4:25] on You're not on top of these things. [4:27] So. [4:29] >> [laughter] [4:30] >> All right, guys. How How How we looking [4:31] at Our guest has nine more conference [4:33] calls to listen to today. [4:34] >> We're looking pretty good. Let me come [4:35] in and search you up. [4:36] >> All right. [4:39] >> All righty. [4:40] >> That's our theme song. [4:42] >> Oh. [4:43] >> So, usually the guest will deliver a [4:44] freestyle rap over this. Are you Do you [4:47] have something prepared? [4:48] >> I do not. [4:49] >> No? All right. All right. We'll do it [4:50] next We'll do it next time. [4:51] >> Sorry. [4:52] >> Episode 238 [4:54] >> Um [4:59] >> [music] [5:05] >> Adding income has its advantages. [5:08] Invesco's Income Advantage ETFs are [5:10] designed [music] to provide consistent [5:12] monthly income and maintain growth [5:14] potential, all with less volatility [5:16] [music] [5:17] and downside risk mitigation. Access [5:19] some of the world's best-known stock [5:21] indexes now with added income with [5:24] [music] Invesco's Income Advantage ETFs. [5:27] Invesco, let's rethink possibility. [5:30] There are risks when investing [music] [5:31] in ETFs, including possible loss of [5:33] money. ETFs risks are similar to those [5:35] of stocks. Investments in the tech [5:37] sector [music] are subject to greater [5:38] risk and more volatility than more [5:40] diversified investments. Before [5:42] investing, consider the fund's [5:43] investment objectives, [music] [5:45] risks, charges, and expenses. Visit [5:47] invesco.com for a prospectus containing [5:50] this information. Read it carefully [5:52] before investing. [5:53] >> [music] [5:53] >> Invesco Distributors, Inc. [5:55] >> My God, 238? [5:57] >> [applause] [5:57] >> That's your lucky number. [5:59] >> It really is. [6:00] Ladies and gentlemen, welcome to the [6:02] 238th edition of The Compound and [6:05] Friends, your favorite investing [6:06] podcast, [6:08] arguably, and many would say the best [6:10] podcast in all of investing. We [6:12] appreciate you guys listening each week. [6:15] We are doing John, we're doing record [6:17] numbers this year. Record numbers. [6:19] [music] [6:20] All right. Uh our our guest today has [6:22] never ever before been on a podcast. I I [6:26] have it on high authority. Um she is Dr. [6:30] Ankur Crawford, the executive vice [6:33] president and portfolio manager of the [6:36] Alger Capital Appreciation, Alger Focus [6:39] Equity, and Alger Spectra Large Cap [6:43] Growth Equity funds. And more, one more. [6:47] The Alger Concentrated Equity ETF. I [6:50] just counted four. You doing four? [6:51] [music] [6:51] >> Yeah. [6:52] >> Okay. That's amazing. Congratulations. [6:55] The ticker on that last one is CNEQ, [6:58] and that was [music] launched in 2024. [6:59] Ankur holds several patents and was [7:01] awarded a fellowship by the National [7:03] Academy of Sciences. She came to Wall [7:05] Street 22 years ago and trained at Alger [7:09] as a semiconductor analyst before [7:11] becoming head of the technology team and [7:14] began her career as a portfolio manager [7:17] in 2012. [7:20] >> [applause] [7:20] >> In high school, she was No, I'm just [7:22] kidding. All right. Well, welcome to the [7:23] show, Ankur. You excited? [7:25] >> Yeah, I'm excited. I totally am. [7:27] >> All right. So, we're going to have some [7:28] fun. We're going to talk big picture. [7:30] >> core. I'm I'm we're told. [7:32] >> is pronounced on core. Like [7:33] >> Like on core? [7:34] >> Yeah. [7:35] >> All right. I I apologize. Let's do it [7:36] that way instead. Are you sure though? [7:39] Cuz [laughter] my I like the way that I [7:40] did it. All right. [7:42] On core, I want to I want to start with [7:44] like the big picture stuff that um [7:47] everyone is talking about today, but [7:48] then we're going to go much more in [7:50] depth on semiconductors and your like [7:54] sort of your expertise. Michael and I [7:57] have so many questions that we want to [7:58] ask you and some of them will be like a [8:01] Wall Street person trying to understand [8:02] the chip business better, but some of [8:04] them might even be insightful questions [8:06] and I I'm certain by the end of this [8:08] uh recording all of our audience will be [8:11] much more knowledgeable about that area [8:13] of the market. Sounds good? [8:14] >> Yeah, I hope I can answer all your [8:15] questions. [8:16] >> Okay, question one. What is the fastest [8:19] land mammal? [8:20] >> [laughter] [8:21] >> No, I'm just kidding. [8:21] >> A cheetah. [8:22] >> All right. Can we talk about your [8:23] approach to investing just generally [8:25] speaking? How do you manage your funds [8:27] and how do you think about um your role [8:30] as portfolio manager? [8:31] >> Yeah, I mean, look, Alger is a growth [8:34] manager. We've been a growth manager [8:35] since the '60s. [8:37] Um the idea of growth was really [8:39] incepted by Fred Alger. And [8:42] >> Is that right? [8:42] >> Yeah, like he was the one of the first [8:44] growth managers, I think ever. And um so [8:48] when when he incepted what growth is, it [8:51] it really came in two different forms. [8:52] Like uh it it the essence of it was [8:56] where is the change? [8:57] >> Mhm. [8:57] >> Because often times where there's [8:59] change, there's unrecognized [9:00] opportunity. And um if you can recognize [9:03] that change before the market, then you [9:06] make a lot of money. [9:07] >> Right. [9:07] >> So, you know, that is the core of what [9:09] we look for is an incredible amount of [9:12] change or just change in general. And um [9:15] that comes in two buckets. One, what we [9:17] call high unit volume growth, which is a [9:18] typical growth company that is having [9:21] high revenue growth and they're, you [9:23] know, disrupting the market in some way. [9:25] Typical, you know, they may or may not [9:28] have a lot of cash flow right now, but [9:29] they're growing into it. [9:30] >> This is what will attract your attention [9:32] to a stock. [9:33] >> Right. So, so high top line growth um [9:36] and the other side of equation makes us [9:39] a little bit different in our philosophy [9:41] relative to most growth managers because [9:43] it's what we call life cycle change. And [9:45] these are companies that have already [9:47] been through their growth um initiative. [9:50] They've probably saturated out their [9:52] markets and then they have to figure out [9:54] who they are. [9:55] >> Mhm. [9:55] >> Right? They've grown [clears throat] up [9:56] and they're like, do I have more growth [9:58] left in me or do I like what do I do [10:01] from here? [10:02] >> There are some great examples of that. [10:04] >> Great examples. [10:05] >> Apple is an example of that? [10:06] >> Apple's an example of that. Microsoft is [10:08] an example of that. You know, Western [10:10] Dig right now is an example of that. [10:11] >> Oh, Western Dig and the IPPs are and [10:14] Nvidia. [10:15] >> Nvidia is an example of that. That's [10:17] right. [10:17] >> Right. Okay. [10:18] >> So it's it's as these companies try and [10:20] reinvent themselves. Danaher was an [10:22] example of this. Right? So it's almost [10:24] as if, you know, they they have to [10:26] figure out how they're going to grow [10:28] again or not grow. [10:30] >> Okay. [10:31] >> Um but we're looking for the ones who [10:32] are inflecting positively on growth. And [10:34] what's interesting about that leg of the [10:37] the philosophy is that often times those [10:39] are incredibly low multiple stocks. And [10:41] people will say, why are you buying [10:42] value? [10:43] >> Mhm. [10:44] >> And what I say in response is, no, this [10:46] isn't value. This is unrecognized [10:47] growth. [10:48] >> Okay. I like that. The pitfall there [10:51] a lot of companies try to reinvent [10:53] themselves and can't. [10:54] >> That's right. And so part of that thesis [10:56] is, you know, what kind of management [10:58] team do you have? And are they capable [11:01] of actually turning around the ship? I [11:02] mean, so think about Microsoft. When [11:04] Satya came on, um you kind of have to [11:07] hear what he's saying and then see what [11:09] he's doing [11:10] >> Mhm. [11:11] >> in order [clears throat] to execute on [11:12] the next decade of growth. So that's [11:15] really what we're looking at. The [11:16] management team is crucial [11:19] for executing that kind of strategy. [11:21] >> Okay. And then away from the company [11:23] itself, every once in a while in the [11:26] markets we have a moment where an [11:29] innovative new technology emerges and [11:31] the entire world changes right before [11:33] our eyes. So, obviously we we're talking [11:36] about AI today and um I'd just love to [11:39] hear your take on [11:41] where you think we are in the AI story [11:44] and are you still as excited about it [11:46] now as you might have been 3 years ago [11:49] when that change first came along and we [11:51] got you know, the original LLMs and [11:53] people started talking about it for the [11:55] first time. [11:55] >> Yeah, I think my excitement and my awe [11:59] of the capabilities has only increased [12:02] >> Increased? [12:03] >> over the last 3 years. I mean [12:05] you know, I remember 3 years ago we [12:07] wrote a paper called AI and the [12:09] declining cost to create and in order to [12:11] to write that paper, there was a lot of [12:13] research that that needed to be done. [12:16] And I remember talking to a lot of kind [12:18] of AI natives, you know, people who are [12:20] in the AI supply chain, talking to them [12:22] about the capability of of what was [12:24] possible. [12:25] What is happening today is happening [12:27] what I thought would happen 2 or 3 years [12:29] from now [12:30] but you know, the timelines have been [12:32] pulled in and in part because the [12:34] capability is just um exponential. [12:37] >> Something specific that's happening [12:39] today or you're just saying like the [12:40] pace of it? [12:41] >> I think they like agents. Agents today [12:44] and what they're capable of today, I [12:45] thought were going to happen you know, [12:47] in '28. [12:48] >> Okay. [12:48] >> And that capability would come in '28, [12:51] not necessarily today. And so I would I [12:54] would you know, oftentimes as as [12:56] analysts we think linearly. [12:58] Um and the market thinks linearly but we [13:01] are in a time of exponential growth. [13:03] And so it's really hard to get your arms [13:06] around what exponential actually means [13:09] and what the end point is. [13:10] >> Okay. [13:11] >> Um and so for for me it's been a I mean, [13:14] it's a continuous learning process. I [13:15] have never slept less and [13:19] worked more because there's just every [13:21] day there's there's new updates that are [13:23] happening and there's just something new [13:25] to learn. [13:26] >> What's like the most jaw-drop dropping [13:28] example of an agentic product or [13:31] something that you've gotten a glimpse [13:32] at where you just said, "Oh my god, we [13:34] really are living in the future." [13:35] >> Um I'll tell you the weekend that [13:37] Openclaw came out, [13:39] >> Okay. [13:39] >> I didn't sleep. [13:40] >> Not long Not Not that far Not that um [13:42] long ago. [13:43] >> ago. It was February. [13:44] >> Yeah. [13:44] >> Um and I had taken my my my son's on the [13:47] ski team and like I had to wake up and I [13:50] get him ready and like I stayed up the [13:52] entire night. [13:53] >> Doing what Doing what? Playing with it [13:55] or reading about it? [13:56] >> I was just reading about it. And at the [13:57] time I was at I was, you know, I had a [13:59] phone I I was at someone's ski house and [14:01] so um [14:02] I didn't really have access to to the [14:04] capability of of what it could do. But [14:08] and I and I tried to launch one at home [14:10] to be completely honest and, you know, [14:12] it was kind of a mess. I was a little [14:14] bit afraid of the security aspect of it [14:16] and but and then then Perplexity [14:19] computer came. [14:20] And that was probably a month ago. [14:23] >> Oh, Perplexity computer came. [14:25] >> What is What is that? [14:26] >> Perplexity computer is almost like [14:28] Openclaw, but it can't It doesn't have [14:30] all of the functionality because it [14:32] won't go log in and do things for you [14:34] after having logged in. Right. So, um [14:38] you know, it will again go do things [14:40] like I built an app. And in that app I I [14:44] was having a really hard time publishing [14:46] it and kind of putting or putting [14:48] payments into it. And Perplexity [14:50] computer completed the task for me. [14:53] >> Okay. So, you have this thing paying [14:54] bills for you on the internet? [14:56] >> Well, this thing I mean as as I was [14:58] telling Mike before it um you know, I [15:00] have a daughter who's going to college [15:01] and the entire college process was so [15:03] horrible. [15:04] >> Oh. [15:05] >> I thought that no one should ever [15:07] >> No No one should ever have to go through [15:09] that. [15:10] >> And so we we we hired these college [15:12] counselors so I just recreated it using [15:14] an AI bot. [15:15] >> They what? The Common App Filling out [15:17] the Common App? [15:17] >> It's the Common App, the colleges and it [15:19] basically just guides your child through [15:21] the entire application process. [15:23] >> And you feel comfortable enough that [15:25] it's doing exactly what a traditional [15:28] college advisor would do? [15:30] >> I do. [15:30] >> Really? [15:31] >> I mean, we worked with three different [15:32] college advisors. I would say, you know, [15:35] >> Is your first one going? [15:36] >> Yes. [15:37] >> Okay. It's a lot of guts that you have. [15:39] Uh [15:40] that cuz this cuz you haven't been [15:41] through the process before. Yeah. Okay. [15:44] >> But now that I saw it, [15:45] >> Yeah. [15:46] >> it was like and and it was actually my [15:47] daughter's idea. She's like, you know, [15:48] there's a lot of kids that can't get a [15:50] counselor. [15:51] >> That's true. [15:51] >> You know, why don't we democratize this [15:53] process? [15:53] >> That's true. [15:54] >> And so, you know, on one of the snowy [15:55] weekends, um we couldn't do anything, so [15:58] I started building this app. [15:59] >> Okay. So, it's filling it's filling out [16:01] the common app with your kid and she's [16:04] interacting with it and it's [16:06] not telling her what schools to apply [16:08] to, but like [16:09] >> [clears throat] [16:09] >> but helping her organize all the things [16:11] she has to do for each school. [16:13] >> Well, it goes through and it it asks you [16:15] questions like uh it asks you 12 pretty [16:18] deep questions. I think all of us should [16:19] answer these questions cuz it tells you [16:21] a lot about yourself. [16:22] >> Okay. [16:22] >> Um and then it uses those questions. You [16:25] tell it you put on it put in all your [16:27] stats. You tell it what colleges you [16:28] want to apply to. It takes the questions [16:30] and just organizes it for you. So, it [16:32] has version control. [16:33] >> Yeah. [16:33] >> You can send you can send it to your [16:36] parents or your counselor for comments. [16:38] >> Okay. [16:38] >> Um but it has something called strategic [16:40] intelligence that allows you to allows [16:44] the kid to say like, how am I supposed [16:46] to answer this question? [16:47] >> Yeah. [16:48] >> You know, um given my background, given [16:50] my stats, given what I've been involved [16:52] with, how is the best way like what [16:55] should I really be saying? [16:56] >> I I started explaining this. I'm [16:58] starting to think like the colleges [16:59] should almost force [17:01] the kids applying to apply via AI [17:04] interaction to prove that they're going [17:06] to be ready for the world that's coming [17:09] or that's already here. Quite frankly, [17:11] it like it it almost seems like this [17:13] should be part of the process. [17:15] >> Yeah, I mean, I think it should be. I [17:16] built it. So, um you know, I think the [17:19] point is like not necessarily the app um [17:22] and and how it will, you know, can [17:24] really change the experience of applying [17:26] to college, but the fact that like I've [17:28] never coded before. [17:29] >> Yeah. [17:30] >> And I built this in probably it took me [17:31] like 60 hours to build it. [17:33] >> That's a That's a long time, actually. [17:35] >> I know. It was three weekends. It was [17:36] three weekends. [17:37] >> before. So, without it, it would have [17:39] taken you 60 years. [17:40] >> That's I would It would have been It [17:41] would have been infinite because I would [17:43] never have done it. [17:43] >> wouldn't know where to start. [17:44] >> Right. [17:45] >> Based on what you said earlier about the [17:47] type of companies that you buy, these [17:49] companies are experiencing exponential [17:51] growth. There's nowhere near saturation. [17:53] I mean, they just started. [17:55] Based on I guess let's use like the most [17:57] recent round. Anthropic raised at $800 [17:59] billion. [18:00] Would this be interesting to you or is [18:02] it like too early in their growth phase? [18:04] >> So, interestingly enough, we recently [18:07] participated in the last round. And that [18:10] last round 3 weeks ago was at 380 [18:12] billion. [18:13] >> I I might be wrong. I might be wrong. [18:15] Open AI was 800. [18:16] >> And Open AI was 850, I think. [18:19] >> So, okay, so say more. [18:20] >> we did actually participate in that [18:23] round. Um and in part what's happening [18:25] is I mean, we have never before in our [18:27] in [18:28] in the course of the market seen [18:30] companies that are staying private till [18:33] they're almost a trillion dollars. [18:35] >> Yeah. [18:35] >> And so, there's a lot of value that our [18:38] clients aren't really capable of [18:40] touching. Um and so, we're actually [18:43] going in and finding really great [18:45] companies that we believe in that we [18:48] think will grow in the private markets [18:50] that have a road to IPO over the next [18:54] two to three years. [18:55] >> Which vehicle is this in? [18:56] >> Um it's actually across all our funds. [18:59] And so, as part of the the I guess the [19:02] 40 Act funds, we're we're able to take [19:05] 15% of the fund and put it in private. [19:08] >> into private. [19:08] >> this the first time in the history of [19:10] Alger that you guys have done this? [19:11] >> No, we've [19:11] >> done it before. Like we were early [19:13] investors in Palantir um back in I think [19:16] 2012, 13, something like that. [19:20] Um and we were investors in Chime. So [19:22] it's not something that we've done [19:24] often. [19:25] But we are seeing a really interesting [19:28] pipeline of companies now. [19:29] >> If you have if you have companies [19:32] going to $500 billion before they even [19:36] file an S-1 and you're a growth investor [19:39] on the one hand it's sad that they're [19:40] not public and more people can't get [19:41] access, but part of me feels like this [19:44] helps make the case for active [19:46] management, at least for growth funds. [19:48] Like growth funds should be doing this [19:50] as a differentiator versus, you know, [19:52] their their whatever their benchmark is. [19:54] >> Right. Um well, I mean we should be [19:56] doing it in part because there's so much [19:58] value for our clients that is residual [20:01] and and we're not able to to capture in [20:04] the public markets. [20:04] >> And then and then these things come [20:06] public and you have to another decision [20:08] to make. [20:09] >> That's right. [20:09] >> up a lot. [20:11] Do we stick around? I mean, you probably [20:13] have a lockup or whatever it is, but [20:15] >> Yeah, and look, we haven't had to [20:17] contend with that quite yet. [20:18] >> Okay. [20:19] >> Um and we'll see. It'll be really [20:22] interesting as Anthropic and Databricks [20:25] and OpenAI come public what happens. [20:28] >> cuz you don't know how the public market [20:29] will react to it. Like if the public [20:31] market vomits in, you might buy more. [20:33] >> That's right. [20:33] >> So what sort of access [20:36] uh financial information do you have as [20:37] investors in these companies? [20:39] >> Um we get depending depending on which [20:41] company it is, it's it's quarterly we [20:43] get quarterly statements. Um now, [20:47] you know, not all of them are like have [20:50] full they won't have conference calls [20:51] and but we have access to the management [20:53] teams or someone at the company. [20:55] >> Need need it as Berkshire Hathaway, so [20:57] don't don't worry. Um I want to I want [20:59] to talk big picture. You you say that [21:02] we are short compute globally. Short [21:04] every everyone is short compute? [21:07] >> Everyone's short [21:08] >> Okay. For how long do you think that [21:10] persists? Because one of the things if [21:11] you're invested in a lot of the stocks [21:13] in the space, one of the things is like [21:17] am I going to be like the last buyer? Am [21:19] I like am am I going to buy right before [21:23] all of a sudden there's like a glut and [21:25] everyone over built and you know, the [21:27] CapEx plans start to moderate and that's [21:29] everyone's biggest fear. Arguably that's [21:31] kept people out of this for 3 years, [21:33] um some people. But how do you think [21:35] about that compute shortage and how much [21:38] more runway there might be? [21:39] >> So I think of it in in many different [21:41] aspects. First of all, I'm going to go [21:43] back to that first thing I said, that we [21:45] think linearly. We've been trained to [21:47] think linearly. You give someone two [21:48] data points, what do they do? [21:50] >> They connect them. How long how far is [21:51] the distance between here and there? [21:53] >> That's right. [21:53] >> Yeah. [21:53] >> And and it's also they extrapolate it, [21:55] right? So [21:56] um but what happens if in fact [21:59] that growth is exponential? [22:01] And again, like and and that's this is [22:02] why people are struggling with the CapEx [22:05] numbers and the growth numbers because [22:07] we really haven't experienced [22:08] exponential growth and [22:10] um [22:12] so [22:13] you know, that that's one aspect. I I do [22:15] believe that this is exponential kind of [22:18] growth in in the amount of compute that [22:21] we need. [22:22] Um and that is predicated on the fact [22:25] that intelligence [22:28] is what when you start to build a a [22:30] neural network, [22:32] it is an exponential problem. [22:34] >> Especially my intelligence. [22:35] >> Like yes. [22:36] >> It's hugely exponential problem. I have [22:37] always said this. Right. In other words, [22:40] the [22:40] because when you talk about neural and [22:42] you talk about what it's the amount of [22:44] connections. [22:46] >> That's right. [22:46] >> And the amount of connections is [22:47] infinite. It's a big number multiplied [22:49] by another big number multiplied [22:50] >> And for every layer beyond, it's an [22:54] exponential problem, right? So for every [22:56] every connection here has, you know, an [22:59] exponential kind of more connections [23:01] beyond it. So, [23:03] um [23:05] I I think that when you think about how [23:08] how we can use intelligence and as we [23:10] try to recreate intelligence, [23:13] that demand is [23:16] it's it's difficult to to fathom, [23:19] but it is very real. So, so think about [23:22] agents. You know, a year ago we were [23:24] worried about it, you know, oh, are we [23:26] are we in a bubble? [23:28] >> Yeah. [23:28] >> Then we got agents. [23:30] And right now we're all stopped out on [23:32] on the agents that we can use and the [23:33] tokens we can use and, you know, people [23:35] are trying to token max and [23:37] um their their bills are getting out of [23:39] control. Why is that? [23:41] >> Yeah. [23:41] >> Because once we can ask someone to do or [23:44] something to do something for us that is [23:47] productive, we will deploy it. [23:49] >> at Uber said they already blew their [23:51] entire 2026 budget on compute. It's and [23:54] it's like April. Did you see Did you see [23:56] that news? [23:57] >> see that. [23:57] >> What What do you make of that? Is that [23:59] Do you think that they just have the [24:00] guts to say it and there were a lot of [24:01] companies in that same position because [24:03] of the cost of all this compute? [24:05] >> Yes. I think I think there are a lot lot [24:07] of companies that, you know, want to use [24:10] more of the compute and now that it's [24:11] becoming accessible, [24:12] >> Yeah. [24:13] >> um and democratized because you don't [24:15] have to be a coder to use it and anyone [24:17] can use it. I mean, usage I I do think [24:20] will go through the roof if they have [24:22] the tokens. [24:23] >> We're going to do some market stuff, um [24:25] but to segue there, [24:27] does having the level of conviction that [24:29] you do in the exponential opportunity [24:33] enable you to live through the periodic [24:36] drawdowns not just for tech and AI, but [24:38] just like market wide? Is that a big [24:40] part of your [24:41] portfolio manager persona? [24:44] >> Um that's I I I I don't know if I have a [24:46] portfolio manager persona per se, but [24:48] >> seems like you're very It seems like you [24:50] have this like this innate bullishness [24:52] about the the possibilities. And so [24:55] maybe that would enable you to endure [24:57] more volatility than somebody who is [25:00] either skeptical or doesn't have that [25:02] same level of conviction. [25:03] >> Well, let let me give you an example. [25:06] Um, you know, in February, I got pretty [25:08] worried about the market. And it was [25:10] kind of like a guttural instinct of, you [25:12] know, something doesn't feel right. [25:14] We're we're starting to put ships into [25:16] into the seas near Iran. Like [25:19] And instead of just sitting there, [25:21] despite my bullishness, you know, I [25:24] raised 7% cash in our in our [25:25] concentrated equity fund. [25:27] >> Yeah. [25:27] >> Right? [25:27] >> Not by selling Anthropic. [25:29] >> Not by selling Anthropic. [laughter] [25:31] Um, and so, it was kind of like more of [25:33] like a a a take you take everything [25:35] across the portfolio just to raise the [25:37] cash to [25:38] um to buffer yourself for volatility. [25:41] Um, that doesn't mean necessarily that [25:45] you know, you're you're incredibly [25:46] negative on the AI trade. In fact, [25:49] through the course of that period of [25:50] time, I've only become more bullish. So, [25:53] um [25:54] that doesn't mean there won't be [25:55] volatility with that trade because not [25:57] everyone believes the same thing I do. [26:00] >> Well, you can raise cash because you [26:02] have the impression that others are [26:03] going to sell and that that's going to [26:05] open up opportunities. And that doesn't [26:07] make you bearish, it just makes you I [26:10] think cognizant of the fact that [26:13] other people are going to create those [26:14] opportunities for you and you have to [26:16] have cash to take advantage of that. [26:17] >> That's right. And you know what? If you [26:18] can sell something for a profit and buy [26:21] it back 30, 40% cheaper, [26:23] you you do that any day. [26:24] >> I have a dumb question that I know Josh [26:26] is thinking, but he's afraid to ask you, [26:27] so I'll do it. When when you and [26:30] everybody says we're short compute, [26:33] cuz everybody's saying the same thing. [26:34] What literally does that mean? Like, [26:35] where are the bottlenecks and what does [26:38] that mean for investors? [26:39] >> I'll handle this. Electricity is a very [26:41] big bottleneck. [snorts] [26:42] >> Stop. Stop. Stop. That's enough for you. [26:43] >> No, go ahead. [26:44] >> Uh, well, I mean, the unit of measure [26:47] that we're talking about is tokens, [26:48] right? Okay. So, for Michael's benefit [26:51] and some of the people in our audience. [26:52] >> Could you please answer the question? [26:55] >> No, no, I'm [26:56] I'm asking you. [26:57] >> Josh really wants to answer it. [26:58] >> Why Why do we Why do we measure this in [27:01] tokens? What does Our audience is like [27:05] 90% people that are either investing for [27:08] other people or for themselves and just [27:10] 10% absolute lunatics. [27:12] What does the audience need to [27:14] understand about the way that you're [27:15] thinking about the compute shortage and [27:17] why tokens matter to this conversation? [27:19] >> And are all tokens created equal? [27:21] >> Ooh. [27:22] >> That's a really great question. [27:23] >> That's one of the best questions I've [27:24] ever heard. [27:25] >> Um so, tokens are a unit of [27:28] intelligence. And think of it as I mean, [27:31] technically, a token is two or three [27:34] letters. [27:35] So, you send out a question [27:37] um and it's 100 letters long, that's 30 [27:40] tokens. [27:41] Right? Um that's technically what like [27:43] your input token would be. [27:46] But [clears throat] really, you should [27:47] think of tokens as being units of [27:49] compute. And the more the more thinking [27:52] are the units of thinking and units of [27:54] intelligence. And the more intelligence [27:56] you need, the more tokens you use. It's [27:58] as simple as that. [27:58] >> Cuz not just your your prompt, but then [28:00] the response coming back to you. [28:02] >> And the tokens come from where? You buy [28:04] them from these companies? [28:05] >> The tokens is just Think about it as [28:07] like if your chip is a flywheel or if [28:10] it's a a factory. And you know, this is [28:13] why Reid Hoffman calls it an AI factory. [28:16] So, it's a factory where you're saying, [28:17] "I'm putting in three tokens, I'm [28:19] getting out a lot of intelligence or I'm [28:22] putting in this many tokens and I need a [28:23] lot of intelligence out." It has to [28:25] grind through it. And that grinding is [28:28] compute. [28:29] >> Right. The computer literally has to [28:31] carry out its calculations [28:33] and then reply back to you. [28:35] >> guess the question is it's like, all [28:36] right, so I pay a monthly subscription [28:38] to these companies to these LLMs for the [28:40] privilege of using the compute. And the [28:42] bottleneck is where like they can't get [28:44] enough because Nvidia can't get enough [28:46] chips because Taiwan Semi can't [28:48] manufacture enough of them. Like where [28:49] where is it getting caught? [28:51] >> I I right now it's you don't have enough [28:53] chips. Right? [28:55] You You still don't have enough chips [28:56] although there's other parts of the [28:57] supply chain now that are get that are [28:59] becoming short. So, DRAM is you see [29:01] what's happening in the DRAM market. [29:03] DRAM pricing's, you know, grown I think [29:06] 100% [29:07] year-over-year. And what that is unusual [29:11] for a commodity. [29:12] >> Was that foreseeable for you? Did you [29:14] know that was going to happen? [29:15] >> You know what? Early on I I like two two [29:18] years ago I was like, "God, this is [29:20] going to really pull on DRAM." But the [29:22] memory market was horrible. So, quite [29:24] honestly I thought I was wrong. I was [29:26] like, "My gosh, I'm not seeing a [29:28] response in the memory market." [29:29] >> You were so ahead of your time that [29:31] nobody else had gotten to that [29:32] conclusion. [29:33] >> interesting. The market was giving you a [29:34] wrong signal and you thought it was so [29:35] obvious that that it would have priced [29:37] it in. You're So, you thought you were [29:39] wrong when the market was wrong. [29:39] >> Yeah, and I thought I thought I was [29:41] wrong. And then, you know, it basically [29:43] just took off at the September of last [29:45] year. So, Micron was at $85 in August or [29:49] August of last year. [29:50] >> Where is it now? [29:52] >> 450. [29:54] Right? So so it wasn't until just now [29:56] that the market was like, "Oh my god, [29:57] you know, we we we have this massive [29:59] DRAM shortage." Because the the pricing [30:02] had just started to take off. [30:05] >> Micron was like a $175 billion market [30:07] cap. I think it's like 400 now or [30:09] something. Maybe even more. [30:10] >> really? I I only know the share price. I [30:11] don't know what the the market cap is [30:12] anymore, but And these things move so [30:14] fast. Like it's hard to keep track of [30:16] the market caps. [30:17] >> You know what You know what's the thing [30:18] with the memory stocks? I try to explain [30:20] this to I try to explain this to [30:22] somebody that doesn't know [30:24] uh finance or stocks at all. [30:26] >> 515. Sorry to interrupt. [30:28] >> And [30:29] >> uh Market cap? [30:30] >> Is that nuts? It's a half a trillion. No [30:32] big deal. No big deal. Um [30:35] I try to explain this to somebody who is [30:37] like not in our industry cuz they're [30:39] like, "Should I buy SanDisk?" Or [30:41] whatever. I'm like, "Yeah, sure. I don't [30:42] care what you do. [30:44] He's like, no, no, no, like explain it. [30:45] Explain it, explain it, explain it. So, [30:47] rather than try to like go into like [30:49] memory chips and shortages, and it's [30:51] cyclical, but maybe this time it's not [30:53] cyclical. [30:54] Um, I explained it I explained it to him [30:57] in a language that he would understand. [30:58] I said, think about like restaurants in [31:00] New York, right? So, you have a [31:02] restaurant called The Corner Store. It's [31:04] the hottest restaurant in the world. You [31:05] literally cannot get in. I've been there [31:07] twice, not to brag. [31:08] Um, [clears throat] [31:09] people line up at at 3:00 in the [31:11] afternoon on a Thursday, and like hold [31:14] places for people. And that's just to [31:16] put their name down, so they could sit [31:18] at the bar. Okay? You like cannot go in [31:21] this place. It's like Taylor Swift's [31:23] diner. Okay? [31:24] >> Well, was it good? [31:25] >> No, it's great. It's it's unbelievable. [31:27] But, the the point is, I mean, it's not [31:29] it's not like It's [31:30] >> It's very good. [31:31] >> It's roast beef sandwiches. Like, it's [31:33] not like you're not going to fall out of [31:34] your chair. And that's the point. There [31:36] are [31:38] uh 500 other restaurants in the same [31:41] neighborhood, [31:42] but people aren't lined up out the door [31:44] of any of these other ones, or maybe one [31:46] or two of these other ones. [31:47] And so, my my point was like, right now, [31:50] so many people recognize the opportunity [31:53] in this small handful of stocks, there's [31:56] like a line to buy them. And [32:00] they sell off, and then they recover [32:02] just as fast, because [32:04] just everyone wants to be there. It's [32:06] not that you can't buy another stock. [32:07] It's not that you can't go to another [32:09] restaurant. There's only one Corner [32:11] Store. And if you open five Corner [32:14] Stores on the same street, all next door [32:17] to each other, each one of them would [32:19] have a thousand people waiting online. [32:21] Because right now, everyone just [32:23] believes these stocks are going to keep [32:26] doing what they've been doing, or that [32:28] restaurant is like the only place that [32:29] they want to be. But, it it changes over [32:32] time. So, if you're about to buy it [32:34] today, [32:36] don't look at what the chart has done [32:37] over the last year because it's not [32:39] going to repeat or if it does repeat it [32:41] would be extremely anomalous. [32:43] >> So [32:44] the the one contention I would have with [32:47] with that analogy is that the corner [32:49] store [32:50] >> bastard. [32:51] >> [laughter] [32:52] >> No, please give it to me. [32:53] >> Is is that you know, you're saying there [32:55] could be many other corner stores [32:57] around? Um but in reality because these [33:01] are technology plays, there's actually [33:03] for each technology segment, there's [33:05] only a handful of [33:07] >> But you and I are old enough to remember [33:09] when [33:11] when memory was the ultimate uh [33:14] semiconductor related commodity. [33:16] >> Okay, but let's talk through that life [33:18] cycle. [33:18] >> Yeah. [33:19] >> So when I started um [33:21] you know, my career there was like 12 [33:24] different memory companies. Three were [33:26] in Taiwan, they were funded by the [33:28] Taiwanese government. [33:30] >> Yeah. [33:30] >> There was like three in Korea, funded. [33:32] There was Micron, [33:34] um [33:35] I've I've even forgotten all of the all [33:37] of the memory companies. [33:38] >> Now, over time because of the nature of [33:41] how difficult it is to and this is [33:43] actually a story of the semiconductor [33:44] industry. [33:46] Because it was so hard to actually um [33:49] produce things at a smaller and smaller [33:52] node [33:53] they were all forced to consolidate. [33:55] >> Yeah. [33:56] >> And so what happened? We basically came [33:59] to this point in time as a consolidated [34:01] industry where you basically have three [34:03] players providing [34:05] >> DRAM. [34:05] >> SanDisk got bought by somebody and then [34:07] came back out. [34:08] And then spun back out. [34:10] That's right. [34:11] These stocks were eight, nine, 10 times [34:13] earnings. When there was a chip when [34:15] there was a glut in memory, you could [34:17] not give these stocks away. [34:19] >> I mean there's eight times earnings in [34:21] August, eight, nine times earnings in [34:22] August. And so [34:24] um [34:25] so the the the thing is like there [34:27] there's only three companies you can buy [34:29] to play memory. [34:31] Right? To play DRAM. [34:32] >> Right. [34:33] >> And so there is no alternative. So [34:35] >> And only three suppliers to the [34:36] industry. Forget about the stocks. Like [34:38] the commodity the the product itself. [34:40] >> That's right. There's only three of [34:41] them. [34:41] >> Yeah. [34:41] >> So there is actually that there the [34:43] reason there's a line out the door is [34:45] because if you want to play memory, [34:47] >> Mhm. [34:48] >> there is actually no alternative. [34:49] >> When you say play, you mean from the [34:50] investors point of view? [34:51] >> From an investors point of view. [34:53] >> So Western Dig in January 25 [34:56] had a market cap of 15 billion and it's [34:58] now 125 almost. Is this a bubble? [35:01] >> No. And in part what's happened is like [35:04] think about the data that is produced. [35:06] So this this is a very interesting um [35:08] >> The problem is that Michael can only [35:10] think linearly. [35:11] >> [laughter] [35:12] >> You have to think exponentially, [35:13] Michael. [35:14] >> I love him, but [35:16] >> Um [35:17] So so think about all the data that's [35:19] being produced. There's video, there's [35:21] you know, these agents that are [35:22] producing things for you that need to [35:24] now be stored and um so all of that data [35:27] has to go somewhere. [35:29] Now, the hard disk drive is very similar [35:31] to the memory industry. There's two [35:33] players. [35:34] >> Mhm. [35:34] >> Seagate and Western Dig. [35:36] And they're not adding capacity. [35:39] Right, they're not actually adding [35:40] physical units. They're increasing the [35:43] density of their hard disk drives, which [35:45] is how they add capacity, but they're [35:47] not doing it at a rate that is going to [35:49] cause a supply demand imbalance. [35:51] >> Right. [35:52] >> And [35:53] >> business nobody wanted to be in 5 years [35:55] ago. [35:55] >> That's right. [35:56] >> Yeah. You know, I love hearing you [35:57] hearing you say it's not a bubble cuz [35:59] these aren't your biggest positions. [36:01] So it would be very easy for you to say [36:02] like, yeah, that is really stupid. [36:05] >> Yeah. Um well, I do think the the bubble [36:07] talk is really stupid. [36:09] >> Go. [36:10] >> But um but we can we can get to that [36:12] >> Name names. No, I'm just kidding. Why is [36:15] it stupid though? Because to take the [36:17] the contra side, [36:19] at a certain point, like these stocks [36:21] will not be rising at the same rate that [36:24] they have been. So it could be that it's [36:26] not a bubble for the products and the [36:28] chips, but it might be a a bubble for [36:31] the investments people are making. [36:33] >> Or and or it could be a bubble in demand [36:36] for compute because there won't be [36:37] revenue at the end of the [36:38] >> But that's the part that she doesn't [36:39] believe. [36:40] >> don't believe, right? At least At least [36:42] through 2026, 2027, and into 2028, based [36:45] on the CapEx that we expect, I don't [36:47] think that we'll be in supply-demand [36:49] balance for compute. We have that kind [36:50] of resiliency. I I don't know if if 2030 [36:53] is going to be that way. [36:54] >> Right. [36:54] >> Depending on how much CapEx we put in [36:56] the ground. It's almost like asking me [36:57] in 1996 [36:59] whether or not we're going to be in a [37:01] bubble in 2000. Well, I mean, I don't [37:03] know. You tell me where the You tell me [37:05] where the valuation [37:06] >> predict what people will do. [37:07] >> Right. [37:08] >> You can't predict what humans will do [37:09] with their money. It's impossible to [37:11] know what they'll do 3 years from now, [37:13] but based on the fundamentals of the [37:15] announced CapEx over the next couple of [37:17] years, you think the market is still in [37:20] an imbalance benefiting the sellers. [37:22] >> That's right. [37:23] >> Okay. [37:23] >> That's right. And so, but but the the [37:25] reason that we have to reconsider the [37:27] all this bubble talk is [37:29] yes, the stocks have gone up. Yes, it's [37:31] gone from 15 to 125 billion, but look [37:34] what the numbers have done. [37:36] I mean, just in the last 3 months, [37:37] Western Dig's numbers have gone from 10 [37:39] to 25. [37:40] >> Wow. [37:41] >> Yeah. [37:41] >> Right? So, [37:42] >> Which numbers? [37:43] >> The earnings numbers for 26 or 27, one [37:45] of those. Like, 10 to $25. [37:48] >> Yeah. [37:49] >> It's like a different company. [37:50] >> It's a completely different company, and [37:51] this is this is classic life cycle [37:52] change for for us, right? Where where [37:54] the industry is changing because of, you [37:57] know, the market changing. So, So, when [38:00] like Nvidia, for a long time, people [38:01] were like, "Isn't this a bubble? Isn't [38:03] it over for Nvidia?" Well, tell me what [38:05] the right numbers are for Nvidia. Tell [38:07] me what the right numbers are for Micron [38:08] and Western Dig, and I'll tell you if [38:10] we're in a bubble or not. [38:11] >> And core, one of the things people have [38:12] trouble with though is pattern matching [38:15] and their over-eagerness to say this [38:17] thing looks like that thing, and I won't [38:19] even go into like dot-com stuff cuz it's [38:21] played out, but just in the last few [38:23] years [38:24] um the solar stocks came and went. Um [38:27] the electric vehicle stocks came and [38:29] went. We had 20 electric vehicle related [38:31] IPOs in 2021, 2022. Um and a lot of them [38:37] literally went to zero. Most of them are [38:39] single digits even today. And that's [38:42] despite their fundamentals having having [38:44] improved. So people say [38:46] here we go again. They're bidding these [38:48] stocks up like crazy. They think [38:52] ABC. [38:54] If A, B, and C don't come through, these [38:56] stocks are all going to crash. [38:57] >> Okay, so so so let me let me comment on [38:59] that in that the first thing [39:01] >> you differentiate is is the the [39:02] question. [39:03] >> Um so so let me just about the bubble [39:05] and like people who often talk about [39:07] this being in a bubble right now. [39:09] >> Mhm. [39:09] >> Um what I have found is that sometimes [39:11] it's almost easier [39:13] if you don't understand something be [39:14] like, oh, it must be a bubble. [39:16] >> I agree with that 100%. [39:17] >> there's a lot of like portfolio managers [39:19] that that I have talked to and met with [39:21] and CIOs that [39:23] and it's interesting because the the PMs [39:26] are often like they came from health [39:28] care or consumer. And getting your arms [39:31] around what is happening with that kind [39:33] of background is really hard. [39:36] >> Is that part of your advantage as a [39:37] growth manager is that this you're [39:39] native to this um sector? [39:41] >> Personally, I think so because it seems [39:44] so obvious to me like what is about I've [39:46] had I've never had this much clarity in [39:48] my career. And in part because it's it's [39:51] pretty clear what's about to happen. [39:52] >> Okay, so you think you think that it's [39:54] people that cuz we say [39:56] bubble talk is code for I missed out or [40:01] I'm so smart and these people just made [40:04] all this money in XYZ. [40:07] It's impossible, therefore it must be a [40:09] bubble. [40:10] Like that's how people think. [40:12] >> I know, but that's kind of crazy. That's [40:13] >> but [40:14] >> Right, so so so so that's on the bubble [40:16] part um, that I thought was important to [40:18] highlight because I do think when I hear [40:21] people talk about a bubble, I always [40:22] ask, like, tell me tell me where you [40:24] think it's a bubble. [40:24] >> Yeah. [40:25] >> Is it just the CapEx numbers? Okay, let [40:26] me justify Why don't you justify or why [40:28] don't I justify to you that CapEx [40:30] number? And And I can. [40:31] >> Yeah. [40:32] >> Um, so wait [40:33] >> of all [40:33] >> Can you? [40:34] >> I I can in a second. Let me get to a [40:35] solar question because the solar [40:38] question is a really important one. [40:40] Solar was a massive market. [40:42] And in 2000, when was it? 2000, when [40:46] when when we had this company called [40:47] Trina Solar and JK Solar, there was all [40:50] these [40:50] >> the Chinese panel makers, the Canadian [40:52] panel, right? [40:53] >> Right. And all of those [clears throat] [40:55] guys, CSIQ, Canadian Solar, they came [40:58] They came to market. [40:59] And, you know, I was particularly like I [41:02] graduated from Berkeley. I You know, I I [41:04] loved alternative energy. I thought like [41:06] oil was evil. And, um [41:10] you know, I would go to these meetings [41:12] and the guy was like, yeah, we we [41:13] basically built the solar company in my [41:15] garage. [41:17] And I was like, that's so weird. You You [41:19] built [laughter] a You built a solar [41:20] company in your garage and you're IPOing [41:22] for 2.4 billion dollars, which now seems [41:24] like nothing. [41:26] But, um [41:27] I'm like, how does that work? And I was [41:29] just a young analyst. I was like, [41:30] there's something wrong with this. Like, [41:32] if I can go I'm a material scientist, I [41:34] can just [41:35] go build a 2.4 [41:36] >> a photovoltaic cell right now. [41:38] >> That's right. And so, like, why why do [41:39] these guys get to win? [41:41] And so, what happened there was a [41:43] technology tech technological hurdle was [41:46] so low. [41:47] >> Yeah. [41:48] >> That as if as long as there was any [41:50] profit to be had, it would basically get [41:52] depreciated. [41:53] >> a manufacturing challenge to just make [41:56] them the cheapest, and of course China [41:58] will always win. But, then there was [42:00] also this big government subsidy [42:01] component to it, which, you know, would [42:03] come and go. [42:05] Um, but I just [42:06] >> But, but that's why I sold it [42:07] differently. [42:07] >> People lost a lot of money. So, all [42:09] right. So, in this case, the dollars [42:11] being spent on this are real and [42:13] growing. The profitability of it is high [42:16] from the outset. It's not like one of [42:19] these things where everybody has to lose [42:21] money in order to subsidize it. A few [42:24] companies are willing to do that, but a [42:26] lot of these companies are coming out of [42:27] cash flow in their capex. And that makes [42:29] it very different. [42:30] >> and and I think what you have to [42:32] remember is if you look across the [42:33] entire supply chain, [42:35] because from 2000 to 2020, [42:40] we have had the the entire supply chain [42:43] has been consolidated. [42:44] >> Mhm. [42:45] >> So, think about that, right? You had, [42:47] you know, semicap equipment consolidated [42:50] over the last 20 years. [42:53] >> The hard the box makers [42:54] >> The box makers consolidated. [42:56] Um, you know, the chip makers [42:58] consolidated. The I mean, even like the [43:01] networking guys consolidated. And so, [43:04] you're coming at this problem from And [43:07] And why did they consolidate? Not [43:09] because they had to. It's just like [43:10] technology became so complicated [43:13] that in order to go from like a 6-in [43:16] wafer to an 8-in wafer, you almost had [43:19] to join forces because the R&D budgets [43:21] were out of control. [43:22] >> Okay. [43:23] >> Right? And And that happened for the [43:24] chip guys, too. For For to do a tapeout, [43:27] which is basically spin up a new chip, [43:29] the the the cost of doing a tapeout was [43:32] going up, you know, three, four, five X [43:34] for every single node. [43:35] >> To fit all the transistors, do all the [43:37] testing. [43:38] >> Right. So, the way to do it is you [43:39] consolidated the market, so you just had [43:41] more scale, so the R&D budgets were [43:43] amortized over more revenue. [43:45] >> As you said, I can't remember the last [43:47] like chip IPO. Like we [43:49] >> Astera Labs. [43:50] >> What And when was that? [43:51] >> Um, 2024. [43:53] >> But that's the point. Like I remember an [43:55] era where there was a semiconductor IPO [43:57] every week. [43:57] >> Right. [43:58] >> And it's just we don't have that. [44:00] >> there are like [44:01] uh just a handful of semiconductor [44:02] companies, you know? Um, there used to [44:04] be 45 50 [44:05] >> the same three foundries. [44:07] >> That's right. And it really [44:08] [clears throat] one that matters. [44:09] >> So, can I ask you this? [44:11] Um I think the biggest question [44:13] for investors, if we could see into the [44:14] future, that would put this debate to [44:17] bed, is the transition from the [44:20] hyperscalers from asset light [44:23] to asset heavy, [44:24] spending all of this money, [44:26] and I see you're getting ready to cook [44:27] and I can't wait for for for you to jump [44:29] in. Like Amazon, which which is one of [44:31] your biggest positions. All right, so [44:32] they're spending whatever they're [44:33] spending on CapEx. Are they going to get [44:36] a return on their investment that will [44:38] allow for this to keep going? [44:40] >> So, [44:42] what I would say is [44:45] I can't 100% say that they will get the [44:48] same return that they expect. And and [44:51] going into this year, this was my number [44:53] one, like back in before Meta put out [44:57] their big CapEx number, I remember [44:58] talking to our team and saying I'm [45:00] really worried uh [45:02] that uh you know, the hyperscalers are [45:05] going to zero cash flow. [45:07] And that handoff of we we [45:10] we value things on cash flow. [45:12] You know, as growth managers, we we want [45:14] to see the cash flow and [45:16] >> all of their cash flow now. [45:17] >> They're reinvesting all all of them are [45:19] investing all of their cash flow. And [45:21] that handoff can be a tricky one. [45:24] Because whenever you have a valuation [45:27] regime handoff [45:29] um to going from PE to cash flow or cash [45:32] flow to PE from EV to sales to to [45:35] earnings, um there's always a a period [45:39] of trickiness. [45:40] >> And we're living through it right now, [45:41] which is why they haven't done anything [45:43] for a while. [45:43] >> they derated. We used to prize how asset [45:46] light they were and how high their cash [45:48] flow was. [45:49] And now the story's different. Now, [45:52] they're in the most massive investments [45:54] uh [45:55] uh life cycle we've ever seen, and there [45:58] is no there is no more free cash because [46:00] they're plowing it right back into data [46:03] centers. [46:03] >> That's right. And And what I would liken [46:05] it to is [46:06] when they started building out the [46:08] cloud. Right? So, in 2011, [46:11] 12, 13, 14, right? So, So, Amazon [46:13] started building their cloud probably a [46:15] little bit earlier than that because [46:16] they're doing it for internal use. [46:18] Um [46:20] they were spending a lot. We just didn't [46:21] know because they It was kind of inside [46:24] of their their customer [46:26] >> Amazon. They weren't even telling people [46:28] like right? It's different companies. [46:30] >> And even Microsoft, if you look at how [46:31] they spent, they were able to kind of [46:34] build this cloud, although they spent a [46:36] lot of money. [46:37] >> But it wasn't like this sort of spend, [46:38] was it? [46:39] >> It wasn't this kind of spend, but only [46:40] because the cloud didn't come as fast as [46:43] AI has come. [46:44] So, if we basically had said, "Oh, let [46:47] Let me give you the capability of the [46:48] cloud with the you know, with the 5G [46:50] network and you know, give it to you [46:52] immediately," the spend would have been [46:55] significant. And that's what's happening [46:57] with AI. Like that the time The time [46:58] horizons have have compressed so much, [47:00] and the capability is so [47:03] um it's [47:04] ex- exponentially growing. [47:05] >> So, what is Amazon spending the money [47:07] on, and where is the return going to [47:08] come from? [47:09] >> Um well, they're spending the data [47:11] centers, chips, cooling, you know, the [47:13] entire shebang [47:15] of [47:16] you know, how you have to build up a [47:18] data center. [47:18] >> the Amazon What's the Amazon [47:20] relationship with Anthropic in your [47:22] mind? Is it an investor? Like how key is [47:24] that? It's like because I remember 3 [47:27] years ago [47:28] Amazon was talking about Bedrock AI and [47:30] saying, "We're basically going to be a [47:32] bring your own model. Like anybody can [47:34] use any AI model they want at AWS." And [47:38] the street like that. And then Anthropic [47:40] just took off for the enterprise. [47:43] And all of a sudden, it started being [47:44] more like, "Well, you know, if you want [47:46] to play Anthropic, the way to do it is [47:48] via Amazon because of that [47:51] relationship." So, is that Is that part [47:53] of why you're so bullish on Amazon or [47:54] not necessarily? [47:55] >> No, look, I think I think Amazon, if you [47:58] look at the valuation for Amazon today [48:00] on APE on a GAAP PE basis. [48:04] And I [48:05] I forget exactly what it is, but it is [48:07] absolutely not egregious. [48:09] >> Yeah. [48:09] >> Um [48:10] you know, [48:12] it's at like teens. [48:13] >> Yeah. [48:14] >> For on on an earnings basis. [48:16] >> maybe the lowest valuation you've ever [48:18] been able to buy the stock for since [48:19] it's been public. [48:20] >> That's right. I mean, compare it to [48:21] Costco or a Walmart, it's I think 10 [48:24] points lower [48:26] >> Yeah. [48:26] >> than either of those. So, is that is [48:29] that really kind of the right valuation [48:31] for a company like Amazon? [48:32] >> It is a grocery store. So [48:34] >> a grocery store, that's right. But it's [48:35] a it's a very efficient grocery store. [48:37] >> The partnership with Anthropic and the [48:39] investment, like how important is that [48:40] to your thesis for why AWS could could [48:44] be one of the biggest beneficiaries of [48:46] AI? Or is it not that important? [48:48] >> I mean, it's important because it drives [48:50] the it drives the top line for it for [48:51] AWS, right? So, I mean, if AWS is going [48:54] to grow in the mid-to-high 30s this [48:56] year, and part of that is because of [48:58] Anthropic. [48:59] >> Okay. [48:59] >> So, it is important in that aspect, but [49:02] really for the duration, because [49:04] Anthropic is going to work with other [49:06] companies as well. [49:08] >> Okay, they do. [49:08] >> And they will. All right, all right. And [49:10] so, um it can't be the only crux. So, [49:13] Amazon, look, they have Trainium. Like [49:16] just like Google has their TPU, Amazon [49:19] has Trainium. And so, it is important to [49:21] own the stock. [49:22] >> Okay. [49:23] >> And because it does lower your cost of [49:25] compute. [49:25] >> Andy Jassy was talking about Trainium [49:27] and some of the other chips that Amazon [49:29] is making and saying, [49:31] "If we were to be valued or if people [49:34] were to think about them as though we [49:36] were selling these chips to third [49:37] parties, it would be like a $50 billion [49:40] business." Do you think that [49:43] Do you think that the street is starting [49:44] to I don't want to say re-rate cuz it [49:46] hasn't really, but do you think the [49:48] street is starting to feel a little bit [49:50] more bullish about Amazon's [49:53] uh A ability to produce chips, but B [49:56] more importantly, potential willingness [49:59] to become a big chip seller given the [50:02] shortage of compute. Like, could that be [50:04] a whole new leg to why the stock should [50:06] go higher? [50:06] >> Um [50:07] >> And I own the stock, so I want you to [50:09] consider my question through that lens. [50:11] >> That's an interesting [50:12] Yeah, I [50:12] >> You [laughter] definitely own more of [50:13] that than I do. [50:14] >> Um so, what I would say to that is that [50:17] look, it's an important aspect of the [50:19] technology. [50:20] >> They've been working on training and for [50:21] a long time. I think they are behind the [50:24] curve, actually, and they are catching [50:26] up. So, training and for is still [50:28] >> Delete all of that. [50:29] >> It's still an okay [50:31] >> Okay. [50:31] >> chip. It's not a fantastic chip. I think [50:33] TPU is a fantastic chip [50:35] >> Yeah. [50:35] >> that competes with Nvidia. [50:37] >> processing unit from Alphabet. [50:39] >> That's right. Um and mind you, they they [50:42] they've been working on that for 5 years [50:43] longer than Amazon's been working on [50:45] their training and products. [50:46] >> Okay. [50:47] >> So, um Amazon will catch up. I think [50:50] what is exciting about Amazon um on that [50:54] aspect, not that they're going to sell [50:55] their chips, is that they can [50:58] effectively make a compute layer that [51:00] is, you know, cheaper. [51:03] So, you can do compute on a training and [51:06] that will, you know, be cheaper for you [51:09] to run, potentially. Um [51:11] >> So, more supply for a a supply [51:14] constrained market. [51:15] >> So, um well, I think it's just more more [51:18] so to think about the stack. [51:20] >> Okay. [51:20] >> Right. So, if if you go and say, I'm [51:22] going to I'm going to use AWS with a [51:24] training and chip to do a certain [51:25] workload, you know, might it be cheaper [51:28] if you do it on a training and it may [51:29] be. [51:30] >> versus a very broad, powerful GPU that [51:33] maybe is not necessary overkill [51:35] >> That's right. [51:35] >> one task. [51:36] >> And it goes to your question about [51:37] tokens, actually. Like, are all tokens [51:40] created equal? [51:42] We didn't answer that question, did we? [51:44] >> And the answer is? [51:45] >> No. Okay, so um all all all tokens are [51:48] not created equal, right? Um, I [51:52] I fundamentally believe that this market [51:54] is going to There's going to be tokens [51:56] that you pay a lot for. [51:58] And there's going to be tokens that, you [52:00] know, when I ask GPT what the weather is [52:03] going to be, um, [52:05] >> Right. First, help me create an atomic [52:07] bomb. I understand. [52:08] >> They're like a very different. And so, [52:09] why should we be paying the same dollar [52:12] value for each of those tokens? And not [52:15] like the atomic bomb. I don't ask it to [52:16] do that. You might, [52:17] >> [laughter] [52:18] >> All right, get my kid into Princeton. [52:20] I got it. Same idea. No, I I think [52:21] that's a really important part of the [52:23] conversation is that, [52:25] you know, the initial build is like, all [52:27] right, it's GPUs. These GPUs are [52:30] insanely powerful and probably over It's [52:33] like sending putting a lion in your [52:35] house catch a mouse. If your daily use [52:37] of Claude is like, what should I have [52:39] for lunch today? So, so the the market [52:43] will figure out what to use for water [52:45] maybe already has, you're saying. [52:47] >> Right. It It hasn't figured it out quite [52:49] yet because all tokens are still kind of [52:51] the same token value. [52:53] >> Yeah. [52:53] >> Um, and so, [52:55] you know, I do think drug discovery [52:57] tokens will have a completely different [52:59] value in the market than, you know, your [53:03] educational tokens. [53:04] >> So, that won't impact me as a user of [53:08] these LLMs, but that will impact the way [53:10] the enterprise version is being priced [53:13] for a corporate customers. [53:14] >> Right. And And I also think it might be [53:17] I it might impact what model you use, [53:19] right? So, [53:20] >> Well, so people are already figuring [53:22] that out. People are already saying, [53:24] I use this for this. I use this for the [53:26] regular people, not professionals. I [53:28] would imagine professionals [53:30] they go way more in depth on what which [53:33] of the company's operations are we [53:35] carrying out here versus there. [53:37] >> Right. Or Or an app that you build, [53:39] right? If you're going to build an app [53:40] for drug discovery, you might, you know, [53:43] you may you may price the app at a level [53:45] that it allows you to use the most [53:46] advanced tokens or or the most kind of [53:49] valuable tokens. [53:50] >> Okay. So, the products will be in part [53:53] priced on how which tokens are being [53:56] utilized. [53:57] >> That's right. [53:57] >> So, these companies keep saying that not [54:00] spending enough is an existential [54:02] threat. They have to spend and they're [54:04] able to say that because we're in a bull [54:07] market and even though some of these [54:08] stocks haven't done amazing in the very [54:10] short term, like investors are still [54:12] giving them the benefit of the doubt. [54:14] Um is there It sounds like there It [54:16] sounds like there's You don't think [54:17] there's a risk that they might say on a [54:18] qual, "Hey, you know what? We're [54:20] actually going to take our foot off the [54:21] gas pedal." But like is there a chance [54:23] that the market and I know we're [54:24] speculating here that maybe the market [54:26] could force their hand? Like if [54:27] Microsoft didn't stop going down and the [54:29] stock is down 45%. At what point does [54:31] Satya say, "All right, guys. Like maybe [54:33] we should chill out with all this." [54:34] >> example. The like the stock market [54:37] enforced the end of metaverse spending. [54:40] I think we all would agree on that. Like [54:42] the He probably would have kept going [54:44] two more quarters had the stock not been [54:46] in a 70% free fall. Now, Zuckerberg [54:50] controls all the votes. So, it really [54:52] had to just be his acknowledgement that [54:54] all of these investors who are selling [54:56] the stock can't all be wrong. [54:58] >> Right. [54:58] >> That's the That's the concern there. [54:59] >> Um I think I think again it's very [55:01] different than the metaverse because the [55:03] metaverse the promise of the metaverse [55:05] >> I don't believe in it, by the way, [55:06] credit to me. [55:07] >> I actually I actually believe in the [55:09] metaverse still, but you needed AI to [55:10] get there. [55:11] >> you buy [55:12] in the meta? Did you buy a beach house [55:14] next to Snoop Dogg's virtual beach [55:15] house? [55:16] >> Roblox with my kids. [55:17] So. [laughter] [55:18] >> Um but you know, the metaverse is [55:21] actually I do I do think it can be a [55:22] reality. It just needed the capability [55:25] of AI to get there. [55:26] >> It's so early. [55:27] >> In 2020, like I remember [55:30] doing doing like a little thing for for [55:32] Al Gore on on the metaverse and I said, [55:35] "This is a 2030 event." [55:37] >> Mhm. [55:37] >> Right? It's 10 years out. So, he was [55:40] trying to invest in 2020 [55:43] or 20 When was that? 21? [55:45] >> 2020 2020 2021. [55:47] >> Yeah, so [55:48] >> Change the name of the company. [55:49] >> For right. [55:50] >> Yeah. [55:51] >> To basically get to 2030. That was a [55:53] long period of time. [55:56] >> all by himself also all by himself. [55:58] >> All by himself with kind of deprecated [56:00] compute. [56:01] >> Yeah. [56:01] >> Right? So, he didn't have GPUs then that [56:04] were going to drive the same allocation. [56:05] >> is different. This is not that. [56:07] >> So, this is different and and in part [56:09] because [56:10] you know, everyone's so worried about [56:11] the ROI. Like, will will we get the ROI [56:13] back? And what I would say right now at [56:15] this moment in time, [56:18] they're basically making their money [56:19] back in 18 months. [56:21] >> Wait. [56:22] How? What? [56:24] >> So, the cost of compute is actually [56:26] rising right now. If you look at the [56:27] cost of an H100, it has been rising. It [56:30] hasn't been going down. [56:32] And so, [56:34] you look at the compute stack and and [56:36] how the the per hour pricing of each [56:38] chip inside of one of these data [56:40] centers, the payback is 18 months. [56:43] >> Meaning the Meaning because the price of [56:46] the chips is going up, [56:47] >> That's right. [56:48] >> an investment that you might have made [56:49] 18 months ago, now you're in the money. [56:53] >> But wait, is that the right way to [56:53] measure it? [56:54] >> For now, it is. But, the point is what [56:56] is the ROI? [56:58] Right? So, if you get paid back your [56:59] capital in 18 months, [57:02] then do you really have to justify your [57:04] ROI? Now, the other thing I think that [57:06] people like miss in this is that they [57:09] can turn As soon as they see the demand [57:12] start to decline or get curbed, guess [57:15] what? They'll they'll turn off their [57:16] CapEx. [57:16] >> And we're just we're nowhere we're [57:17] nowhere near. [57:18] >> We're nowhere near that. But, the point [57:19] is it's not like a it's not like the [57:21] telecom, you know, bubble of 2000. [57:23] >> go next. So, this is the other thing the [57:25] bubble people will say. [57:28] Look at these deal announcements. Every [57:30] one of them is circular. [57:32] Every one of them involves a company [57:34] vendor financing, some sort of an [57:36] investment, or paying a customer to pay [57:40] them back to either buy compute or chips [57:43] or use of a data center or lease [57:45] something um and [57:49] I've heard Jensen I've heard Jensen [57:51] deflect that away it's like of course [57:52] it's circular. We're all in the same [57:54] ecosystem doing business with each [57:55] other. How could it not be so you know [57:58] and I I'm sympathetic to that but like [58:00] what would you say to people that that [58:02] say oh I remember the late 1990s when [58:05] one company needs to make needed to make [58:07] earnings this quarter so they would do a [58:10] deal and then the next quarter the other [58:11] guy had to uh make earnings so they [58:13] would do the a reverse deal. I [58:15] understand this is not the same thing [58:17] but how do you answer that? [58:18] >> Um I answer that by I look every deal [58:21] has been different. I think that Amazon [58:23] Meta deal I worries me more because [58:25] you're giving away part of your company [58:26] to say use my chips [58:28] and I'm going to give you 10% of my [58:29] company. [58:30] >> Meta deal? [58:31] >> Um not Amazon AMD Meta deal. [58:33] >> Oh okay. I see. [58:34] >> um Meta [58:36] Meta gets 10% of AMD if they get to a [58:38] certain amount of revenue and AMD chips. [58:40] Right? So effectively you're at they're [58:41] giving it for 0% gross margin. [58:43] >> Yeah. [58:44] >> Um you know that to me is a little bit [58:46] more quote unquote circular [58:50] >> OpenAI and Meta [58:51] >> OpenAI and Meta both are both had deals [58:53] with with AMD that that looked like [58:55] this. [58:56] Nvidia basically seating the market to [58:59] enable the market is less worrisome [59:01] because I mean Nvidia's revenues on [59:05] on AI like the AI based revenues is [59:08] going to be hundreds of billions of [59:09] dollars. [59:09] >> Yeah. [59:09] >> So the 2 billion that they gave to [59:11] CoreWeave is like are they really [59:13] enabling their own revenue? Not really. [59:15] >> Well they're seating a partner who if [59:17] that partner is successful it means [59:20] years and years and years of future chip [59:22] sales. [59:23] >> That's right. [59:23] >> is that bad? [59:24] >> Why is that bad? don't [59:25] >> it's bad. [59:25] >> It's not bad. I think it's common [59:27] It's actually common sense. [59:29] >> It's It's not Well, [59:31] we see examples of this in every single [59:32] industry is another thing to point out. [59:35] So, you see Pepsi [59:37] gives a restaurant all kinds of free [59:39] advertising and banners and things to [59:42] decorate their bar uh and therefore [59:44] sells more Pepsi into that bar for years [59:47] and years to come. This is just a [59:48] high-tech version of it with billions of [59:50] dollars, not thousands of dollars, but [59:53] it's like a fairly common thing. [59:54] >> Yeah, I I That's a great example. I um [59:57] You know, I I do think you probably see [59:59] it in a lot of different industries. [60:00] >> Light says, "Here's a billiards table. [60:02] Put it right in the middle of your bar [60:03] and the lamp hanging over it is going to [60:05] say Miller Light." And the rest the the [60:07] tavern owner is like, "Okay, cool. We [60:09] have a pool table." Um they're buying [60:12] Miller Light forever and the bar does [60:14] better and like this is obvious stuff. [60:16] >> Yeah. [60:16] >> We're just talking about it on a much [60:17] greater scale here. [60:18] >> Yeah, and and that's what scares people. [60:20] I think again, if you if you think about [60:22] the magnitude of what's happening [60:23] >> metaphors have to do with eating and [60:25] drinking. I don't know if you picked me [60:26] up on this. So, the audience definitely [60:28] is. [60:28] >> Um So, so the um [60:32] you know, the the the magnitude of [60:34] what's happening here is just it's hard [60:35] to fathom. Again, I we go back to that [60:38] like it's really hard to think how big [60:40] and how massive this change is. [60:42] >> Right. [60:43] >> And that's why I think people are just [60:44] struggling with it. [60:45] >> Okay, so Give Give the audience Give the [60:48] audience three stocks that will double [60:49] before the end of the year. [60:50] >> Oh, no. [60:51] >> I'm just kidding. Can [laughter] we talk [60:52] about like some of the actual [60:54] investments that you're making in the [60:55] public markets? [60:56] >> Sure. [60:57] Um [60:58] >> All right, so we're looking at some of [61:00] your holdings. [61:00] >> Wait, can can we go back to one thing? [61:02] >> Yeah, shut up, Josh. [61:03] >> So, so one of the things I want to go [61:04] back to is that this fact like in in [61:07] 2000, [61:09] one of the issues with what happened in [61:11] 2000 and that 2000 bubble is the capex [61:13] that was spent spent was long-lived [61:16] assets. [61:17] So, you put it into the ground. You [61:18] think about like fiber optic cable. You [61:20] put it in the ground, you spend billions [61:22] of dollars, [61:23] and they were effectively long-lived [61:25] assets. [61:25] >> Like nobody needed them for 5 years. [61:27] >> No one needed them for 5 years, and then [61:29] the demand fell apart. But, you also [61:31] didn't have the technology to take to [61:33] take to make use of those cables. [61:35] >> web video. [61:36] >> There was no I mean, do you remember [61:38] using the internet in 2000? [61:40] >> Yeah. [61:40] >> It was a horrible experience. [61:42] >> Yeah, it was chat It was like chat rooms [61:43] where you're typing. It was There's no [61:45] social media yet. There was no YouTube. [61:47] >> There was no There I mean, you could [61:49] barely like download a webpage without [61:51] it, you know, glitching or having a lag. [61:53] >> 10 minutes to download a song from [61:55] Napster. [61:55] >> Right. Like that was the internet. [61:56] >> Right. I mean, 2004, 2005, like [61:59] downloading a movie would take 2 hours. [62:02] >> Days. [62:02] >> Yeah. It would be It would be a horrible [62:04] experience. And so, we had dreamed the [62:06] dream [62:07] in 2000, but we actually didn't have the [62:10] technology to support it. [62:11] >> a a timing mismatch before the demand [62:14] for all that stuff they invested in [62:16] would show up. [62:16] >> And what's really different now is [62:18] actually all the technology is in place, [62:21] and it it's like I I've likened this to [62:23] when I I've spoken to advisors just to [62:25] make it a little bit more tangible. I've [62:27] likened it to [62:29] um [62:29] you know, if you were an F1 car [62:32] engineer. [62:33] >> Thought you're going to say Burger King. [62:34] She's not using food analogies. [62:35] >> Oh, yeah, sorry. I I can't think of a [62:37] good food analogy. But, if you if you [62:39] were like, you know, an F1 engineer, and [62:43] you dreamt up this amazing car, [62:45] and you're like, this is the way it's [62:46] going to work, and I think I'm going to [62:48] win because you know, I've I've dreamt [62:51] this amazing car up. And you get to the [62:53] race, [62:54] and you have four wheels. [62:55] >> Yeah. [62:56] >> Right? This time, you've actually built [62:58] the car, and the only thing that you're [63:00] waiting for is gas. [63:02] >> Okay. [63:03] >> Right? Or, you know, electricity to [63:05] charge your battery. But, like the point [63:06] is that [63:08] now, the only thing you need is the [63:10] compute. [63:11] You have all the other parts in place. [63:13] >> is here, so therefore the capex is going [63:16] directly into something that is [63:17] currently working. [63:18] >> That's right. [63:19] >> But speaking of the CapEx, there's [63:20] another doctor investor who's on the [63:21] other side. Dr. Barry says that there's [63:23] accounting shenanigans with with the [63:25] child there amortizing his expenses. [63:28] >> Yeah, I I don't think that's right. So, [63:29] he he's basically saying, "Oh, they're [63:31] using 5-year depreciation, but they're [63:33] 6-year depreciation, and they're [63:34] actually 4-year." Which is absolutely [63:36] just incorrect. [63:38] >> is he why is he keep saying that? [63:39] >> I don't know. [63:40] >> He's saying that [63:41] >> [laughter] [63:41] >> He's saying that the So, you're saying [63:43] these chips last a really long time. [63:45] They might not be the newest They might [63:46] not be the newest versions, but [63:49] >> are still working. [63:50] >> And those are from 8 years 7 years ago? [63:52] >> 20 what? 26? They're from 2021. [63:55] Right? So, [63:56] >> chips are still very actively being [63:59] used. [63:59] >> mean, the idea that these computes are [64:01] just or that the the the chips are going [64:04] to die in 3 years? [64:06] >> Yeah. [64:06] >> I don't I don't really understand it. [64:08] >> All right, let's do some tickers. So, [64:11] um some of your holdings and [64:13] I sort of know a little bit about some [64:15] of these, um but why are you bullish on [64:17] Let's we'll just go like a few. [64:19] Uh tell us about Nebius. [64:22] >> Oh, wow. [64:23] >> What? [64:23] >> That's like a big long story, which is [64:26] really fascinating. I will try to [64:27] shorten it for you guys. [64:28] >> Yeah, keep it to like 90 minutes. [64:31] Uh I understand that Nebius My my [64:33] understanding is this was [64:35] originally part of Yandex, [64:37] >> Yep. [64:38] >> which was the Russian Google. [64:39] >> It wasn't part of Yandex. [64:41] >> Oh, it's what Yandex became. [64:42] >> No. So, so Yandex um was the the Russian [64:46] Google. [64:47] >> Yeah. [64:47] >> Um headed by a man named Arkady. [64:51] >> Yeah. [64:52] >> And when the Ukraine [64:54] >> In Russia, Google searches you. [64:56] >> That's right. [64:57] >> All right, go ahead. [64:58] >> When when the Ukraine war happened, you [65:01] know, he had it just a disagreement with [65:03] the the politics. [65:04] >> Okay. [65:05] >> And [65:06] you know, he was one of the [65:08] quote-unquote oligarchs that um [65:11] you know, was A sanctioned, but he was [65:14] also on on the outs with Putin. [65:16] >> [clears throat] [65:17] >> He ends up leaving. 1,500 of his [65:19] engineers end up leaving. [65:21] I think Arkady paid for all of them to [65:22] leave. [65:23] And then he sat there, he's like, "Okay, [65:25] I have $2 billion. What are we going to [65:26] do?" [65:27] >> Mhm. [65:28] >> And they said the future is really AI. [65:31] We know how to build data centers. Let's [65:33] build AI data centers because we think [65:35] because we think we can do it better [65:36] than everyone else. [65:37] >> Okay. [65:38] >> And so, you know, remember they were [65:41] Yandex, they got delisted. [65:42] >> So, it's in Europe now, it's in uh [65:44] Belgium or [65:45] >> It's in the Netherlands. They're based [65:47] in in the Netherlands. But one day they [65:49] get this call from the New York Stock [65:51] Exchange being like, "You're going to [65:52] get listed on Monday." [65:55] Right? And they're like, "What do you [65:56] mean? Like, we're like a private company [65:57] right now." And it was a Thursday, so [65:59] they worked all weekend, they get [66:00] listed, and all of a sudden there was [66:01] this like weird ticker on the [66:03] on the exchange, N B I S. [66:05] >> Yeah. [66:06] >> And [66:06] >> I didn't know what it was, either. [66:07] >> Yeah, and so like I was like, "Is it a [66:09] word? Like, what is this thing?" [66:10] >> People were trading it, though. [66:11] >> People may have been And but you know [66:13] what? Initially, when people traded it, [66:15] it was it was all the people that were [66:16] stuck in Yandex. [66:18] >> Right. [66:18] >> Right? And that were like, "Thank gosh, [66:20] I can get out now." [66:22] >> Okay. [66:22] >> And and so they're like, "Get me out. [66:24] I've been stuck here for for years." [66:26] >> And then and but they didn't have it at [66:29] No one knew what the story was. So, they [66:30] put out this 110-page deck [66:32] on online, and um [66:36] I don't know. I remember hounding them [66:37] being like, "I need a meeting with you." [66:39] And they had no time. [66:40] >> Yeah. [66:40] >> So, I I hunted them down in Davos. [66:43] >> Okay. As one does. [66:44] >> Um as one does. And it was [66:47] And I just got to we got to know the [66:48] company really really very well. [66:50] >> Okay. So, you you look at this as cuz I [66:52] know they're in autonomous vehicles, and [66:54] they have a lot of stuff going on, but [66:56] like you look at this as a data center [66:57] play [66:58] that's domiciled in Europe but doing [67:00] business all over the world. [67:01] >> What I would say is that this is [67:02] actually potentially the next hyper [67:05] scalar. The next AI native hyperscaler. [67:08] >> Really? [67:08] >> Yes. [67:08] >> Okay. Uh Um tell us about tell us about [67:12] QXO. [67:13] >> Oh, QXO. QXO is a fantastic company. [67:15] It's not It's not a typical tech [67:17] company, right? Um but we all know Brad [67:20] Jacobs. Brad Jacobs was the um serial [67:24] entrepreneur. Um he put together URI, uh [67:27] XPO. Um left XPO and began QXO. QXO is a [67:33] building products company. He started it [67:36] uh probably about a year and a half ago. [67:39] Bought Beacon Roofing. [67:40] And he's basically going to consolidate [67:42] the industry. [67:43] >> Okay. [67:44] >> What's really exciting about this is [67:45] that A, it's Brad Jacobs. Um we really [67:48] like you talked about the management [67:49] team. Um we know what he is. [67:52] >> That guy's amazing. [67:53] >> He is [67:54] Brad Jacobs might be one of the most [67:55] impressive people. [67:57] >> is he in roofing right now? [67:59] >> Um because he's consolidating the home [68:02] the the building products market because [68:04] it's incredibly fragmented. [68:06] >> Yeah. [68:06] >> So what he's seeing is that well, [68:08] there's all these mom and pops that are [68:09] in building products. They don't know [68:11] how to [68:12] >> and like like all kinds of materials. [68:14] >> He will. He just bought Kodiak. [68:16] >> Yeah. [68:16] >> And he's starting to get into the rest [68:18] of it. So he started with roofing. Then [68:20] he's going to get into everything else. [68:22] He's going to take pricing at [68:24] efficiency, at technology. [68:26] >> a growth investor. Like how much [68:27] potential do you need there to be in a [68:29] story like this to include in your [68:30] portfolio? Do you think this is like if [68:32] it all I know we don't know that what's [68:34] going to work, but if it really works, [68:36] you're not looking for like a 20% [68:38] return. That's not how you invest. [68:40] [clears throat] [68:40] >> No. [68:40] >> Okay. [68:40] >> So let me give you just the like the [68:42] framework for how he thinks. [68:44] Um Beacon is about a billion dollars in [68:46] EBITDA. [68:47] He thinks that he's going to increase [68:49] the EBITDA of this company to 5 billion [68:51] in 2029 to 2030. [68:52] >> big deal. [68:53] >> That's a big deal. [68:53] >> In a quiet, sleepy industry. [68:55] >> That's right. [68:56] >> And he's done this over and over. This [68:57] guy [68:57] >> And he's done it over and over again. [68:58] >> I never heard of this guy. I think maybe [69:00] 4 years ago he was on Patrick [69:01] O'Shaughnessy's podcast. He's been on I [69:02] think I think again. [69:04] Um what an incredible story. [69:06] >> Oh, and cred- You know what? He's also [69:07] an incredible human being. If you If you [69:10] see how he takes care of the people that [69:12] work for him, um A, he's he's a [69:15] workaholic. Um B, he truly takes care of [69:19] them. Like he is like a good boss. [69:20] People He's a great leader. So, he has [69:23] all the everything you want to see in a [69:25] fantastic CEO. [69:26] >> I just listened to his book. I think it [69:27] was called How to Make a Billion [69:28] Dollars. [69:29] >> Yeah, the first one or the second one? [69:31] There's two of them. [69:32] How to make a [69:32] >> make a few billion dollars. [69:34] >> What's What's the investment case for [69:35] AppLovin in a world where we think the [69:38] apps are going to write themselves? [69:40] Um tell us about that. You You own that [69:42] in uh your EC app and some of your [69:44] funds. [69:45] >> yeah. So, AppLovin is um it's not really [69:48] an app creating tool. It's an [69:50] advertising tool. [69:51] >> It's the Lovin. [69:52] >> It's the It's the Lovin. [69:53] >> It's the Lovin part that matters. [69:55] >> Um and what's exciting about AppLovin It [69:56] used to be [69:56] >> stock did get caught up in the software [69:59] >> It did. And And part because it's part [70:01] of the IGV. And this is This is part of [70:03] We have to talk about software at some [70:04] point, but [70:05] >> Yeah. [70:05] >> Um [70:06] >> We'll go have dinner. [70:07] >> getting started. [70:08] >> I know. Like I need another hour with [70:10] you guys. [70:10] >> Okay. [70:11] >> Um so, with AppLovin, [70:13] you know, they're they're basically an [70:14] advertising company. [70:16] And what they realized is that they have [70:18] a lot of data. They have probably 50% [70:20] share of, you know, served ads in mobile [70:22] games. [70:24] And what they realized is, you know, [70:25] what? We're pretty good at this. We're [70:26] using an AI engine to serve up these [70:28] ads, and we're improving kind of the the [70:31] return for our clients. Why don't we [70:33] start putting in lipstick or [70:36] um you know, actual goods? [70:38] And monetize our ads differently. [70:42] >> Okay. [70:42] >> And so, that whole story of we're not [70:45] only going to show you games to be [70:46] downloaded inside of games. [70:48] >> So, you're So, you're playing Candy [70:49] Crush and an ad pops up for Scrabble. [70:53] Like digital. Like an app. Okay. So, [70:54] they've got very good at that. If people [70:57] play this game, they'll probably play [70:58] that game. Okay, fine. They know how to [71:00] price it. The The game app wants wants [71:03] more exposure. Great business. You're [71:05] saying they're now taking this business [71:07] in another direction. [71:08] >> That's right. So, all of a sudden [71:09] they're entering the e-commerce market [71:10] to sell to sell goods. [71:12] And it helps them monetize their [71:14] inventory significantly. [71:15] >> Okay. [71:16] >> And so, really interesting. I mean, [71:18] Adam, another Look, one of the the keys [71:21] for us is the management team. Adam is a [71:24] founder CEO, um hungry. He's very [71:28] involved with with who gets hired, how [71:30] they get hired, the headcount. Like, [71:32] he's keeping his headcount flat. [71:33] >> Okay. [71:34] >> Regardless of the the top-line growth. [71:36] So, it's just a cash generation, like [71:38] >> You seem to be on a first-name basis [71:40] with a lot of the founders of the [71:42] businesses that we're talking about. How [71:43] important is access to management for [71:45] you as a portfolio manager? If there's a [71:47] company like a Nebius that says, "We're [71:49] not going to talk to you. I don't care [71:51] if you come to Davos. We don't do like [71:54] or does that cool you off in terms of um [71:57] wanting to be an investor? Or is it not [71:59] that important? [72:00] >> No, it's important because they end up [72:02] being the stewards of your capital. [72:05] Of our clients' capital, right? We're [72:07] betting I mean, it is an important leg [72:09] of of any story is like who is leading a [72:12] company. [72:13] >> Yeah. [72:13] >> You know, what is their integrity? [72:15] >> Right. [72:15] >> You know, it's a it's really important [72:17] to know to know the management team. [72:19] Obviously, like that that access [72:21] >> to them though, where you like them [72:22] personally and they they screw over [72:24] shareholders or they're more inept than [72:26] you thought? Like, is that the the [72:28] downside to that? [72:29] >> Um well, it's not like I'm friends with [72:32] them, per se, right? [72:33] >> need to be able to ask them questions. [72:35] >> That's right. And I need to I need to [72:36] understand, like, are they are they good [72:39] Are are they are they good stewards of [72:41] capital? I mean, it's important to me [72:43] for me to understand that Brad is a [72:45] great leader. [72:46] >> Okay. [72:46] >> But that that is really like he is he [72:48] has incredible humility. That is [72:50] important for me to understand. [72:52] >> I've two more things we have to get to [72:53] or the audience will kill me. [72:55] One of them is software, but before we [72:57] go there [72:59] I said two days ago I think Intel [73:01] is like the stock of the year so far. [73:04] >> Yeah. [73:04] >> The comeback in Intel is extraordinary [73:07] for a lot of reasons. Number one [73:09] it's not every day that a former Dow [73:11] component fallen angel, formerly one of [73:14] the biggest market caps in the country [73:17] drops almost to single digits [73:20] has a new CEO come in after multiple [73:22] failed new CEOs and actually strikes a [73:26] chord with growth investors again and [73:28] reinvents itself. So that thing that we [73:30] talked about earlier in the show about [73:32] that reinvention [73:34] this looks like that on steroids, but I [73:36] don't know as much about semis as as you [73:38] do obviously. So I'd love to get your [73:40] take on like [73:42] is this a buy here or did we miss the [73:44] whole thing or investors getting too [73:45] excited? [73:46] >> I'll be honest, I don't know. [73:48] >> Okay. [73:48] >> What I do know is Lip-Bu is [73:51] >> Is the new CEO. Yeah. [73:52] >> Lip-Bu Lip-Bu Yeah, Lip-Bu is a new CEO [73:54] of Intel. Um he is again an incredible [73:59] >> Is he a visionary? [73:59] >> leader. He is a visionary. And I mean [74:02] his story [clears throat] with Cadence I [74:03] mean someone should do a deep dive on on [74:05] Cadence and um see how he turned Cadence [74:08] around. [74:09] >> What? [74:09] >> He took it from almost going bankrupt to [74:12] the the company that it is today. [74:14] >> Okay. [74:15] >> And that was under Lip-Bu's direction. [74:17] >> Okay. [74:17] >> So if there's anyone that could take [74:18] this asset and turn it around it's [74:21] Lip-Bu. [74:21] >> The foundry is now being seen as a [74:24] strategically important asset whereas [74:27] they were being ridiculed as recently as [74:29] three years ago these idiots thought [74:31] they could build Taiwan semi in the [74:33] United States. [74:34] >> Yeah, and and what I would say that's [74:36] where I'm skeptical. [74:37] >> You don't So you don't believe [74:38] necessarily that that's a great asset. [74:41] And Taiwan semi's doing it too in [74:43] Arizona, aren't they? [74:43] >> Taiwan semi's building it building their [74:45] fabs in Arizona. Um [74:46] >> But if we short compute and Intel is a [74:49] company that can help bring more compute [74:51] to the market, [74:52] >> Right, but you have to it's not easy, [74:54] right? So so if you if you've been [74:55] inside of a fab and built a chip, [74:58] it's it's difficult to build a chip. [75:00] >> I I wouldn't know, but I wouldn't doubt [75:01] it. [75:02] >> you only have one, right? You only have [75:03] one foundry now. You have TSM. [75:05] >> So you're not you're not convinced that [75:07] this foundry is going to be as [75:08] successful as the bulls in Intel. [75:11] >> right? Because it is it is a bet on the [75:13] execution of the engineers inside the [75:15] foundry. Lip-Bu is not an engineer. You [75:17] know, so hopefully he will hire the [75:19] right people to make this happen. [75:21] >> Okay. [75:21] >> But I don't know. [75:23] >> Okay. So you're not in the you're not [75:24] currently in Intel. [75:25] >> No. [75:26] >> Have you looked at it recently? [75:27] >> We have looked at it recently. I just I [75:29] it's it's an existential question, like [75:31] can you be a foundry or can you not be a [75:33] foundry? [75:33] >> Okay. SaaS-pocalypse. [75:35] Um Jensen has said the bears have the [75:39] story completely wrong. [75:41] A gen- agents and agentic AI [75:45] is actually going to utilize all of the [75:47] these CRMs and and corporate software to [75:52] help people. They're not going to spin [75:54] up their own versions. Therefore, these [75:57] companies do still have moats and they [75:59] are relevant for the age of agentic AI. [76:02] How do you feel about that? [76:03] >> Um I don't agree, respectfully. [76:05] >> Jensen's being too nice? [76:06] >> I think I think Jensen can't really say [76:09] what what is happening because [76:11] >> what I think. [76:12] >> And so why do you think that? [76:15] >> I [snorts] think what you think. Like [76:18] no way. Like there's no there's no way [76:20] all of these companies can remain [76:22] relevant if [76:24] um if Anthropic is going to continue to [76:27] come on at the rate that it's been [76:28] coming on. So that doesn't mean I think [76:30] they're all sells, [76:32] but I you know, I think it's really hard [76:35] to know, obviously, but right now [76:37] they're all being treated like sells. [76:39] >> Well, cuz the growth story at Salesforce [76:40] cannot do what they've been doing for [76:43] the last 15 years, where you re-up and [76:46] the contract goes up. [76:46] >> Right. [76:47] >> And that's a huge part of the story. [76:48] >> raises prices every year and no one's [76:50] going to push back. How could that be [76:51] possible? [76:51] >> But and of course we can't know, is the [76:53] market overdoing it? [76:55] With the selling. [76:56] >> so if you look at let's take Salesforce. [76:59] Salesforce as of yesterday it was [77:00] trading at 10 times EV to free cash [77:02] flow. [77:03] Um if you include stock-based [77:04] compensation, right? If you Right, oh [77:06] no, if you if you exclude stock-based [77:08] compensation. Including stock-based [77:10] compensation it's like 14 and a half. [77:11] >> That's not that compelling. [77:13] >> It's um it's kind of [77:15] trading like broken tech [77:18] with a little option value. [77:21] That said, I think there's other [77:23] interesting things out there right now [77:24] that you know, I I'm definitive will [77:27] cross this chasm. [77:29] So, you know, the issue for me on the [77:31] entire software space is that [77:35] we only use 15 to 20% of the software [77:38] of any of these package software. [77:40] >> Yeah, they're selling it per head and [77:42] people might say, you know what? I'll [77:44] pay you per use. I'm done with this [77:47] game. [77:47] >> That's right. And so so it it's a more [77:50] competitive market. [77:52] Right? People can write their own [77:53] software and maybe, you know, maybe and [77:55] there are a lot of people that will have [77:57] will [77:58] will say that's not right because you [77:59] have the security issues or or [78:01] or whatever. [78:01] >> Valid. [78:02] >> Right? And it's valid, but again, am I [78:04] am I paying you more if I don't use the [78:06] software? [78:07] >> If one of my employees comes to me and [78:08] says I wrote this sick program, [78:10] I run a wealth management firm. I'm not [78:12] going to be like, oh cool, plug that [78:13] into our system. Let's see what happens. [78:15] I don't care how much money I'm saving. [78:18] I will destroy potentially destroy my [78:19] business. [78:21] So, [78:22] that's the vibe coding thing. But the [78:24] the there's a middle ground [78:26] where the IT department at a company [78:30] >> can do it themselves. [78:31] >> has a makes a concerted effort to [78:33] actually replicate something that [78:35] they're paying a third party for and [78:37] they show it to the board of directors [78:39] and they say [78:40] like guys we really can do this and we [78:43] can save the company $80 million over [78:45] the next five years. And the company [78:47] says [78:48] all right, that's what we're going to [78:49] do. Like that's the real threat. [78:51] >> Top line is coming down for this [78:52] company. [78:53] >> That's right. And so I think the best [78:55] case scenario is they don't get you [78:57] know, uplift as you were saying on that [78:59] DBNR, right? Which is or is that the [79:01] basically the expansion the expansion [79:03] revenue. [79:04] >> Yeah. [79:04] >> So they no longer can grow on a seat by [79:07] seat basis how much they're selling. [79:09] >> And that's big part of the growth story. [79:10] What's the net retention rate? [79:11] >> You know that if we think AI [79:13] >> seats [79:13] >> if we think that AI is threatening like [79:16] employment in general [79:19] then higher head counts are going to be [79:21] harder to come by period. Even if you [79:23] can get a company to sign another [79:25] package, doesn't mean they're going to [79:26] have the same amount of employees. [79:28] >> That's right. And so that that dream of [79:30] that 40% operating margin that you know, [79:33] was why these SaaS businesses traded at [79:35] the multiple that they did. [79:37] I think that's a different terminal [79:39] value. I don't know if it's [79:41] 30, I don't know if it's 20. My my gut [79:44] is that given the landscape it should [79:46] these should be 20% operating margin [79:48] businesses. [79:49] >> Is cybersecurity immune or non-immune? [79:51] >> I think parts of cyber will be immune. [79:53] Parts of it will not be. [79:55] Right? So um you know, Apollo or Alteryx [79:58] or CrowdStrike more likely to be immune [80:00] because they're platforms and they can [80:02] build on that platform just like [80:03] Microsoft. [80:04] >> Yeah. [80:05] >> But point solutions [80:07] across all of software [80:09] >> or okay. [80:10] >> Point solutions across all of software [80:12] are going to be at risk cuz they'll just [80:14] be usurped by either, you know, new [80:16] players or people who will like you said [80:18] write code inside of the company and um [80:22] or by bigger players that basically just [80:25] add that on. [80:26] >> I think the market has this right. When [80:28] I look at Constellation so Constellation [80:31] is like a collection of a thousand tiny [80:34] niche software businesses where it's [80:37] like this is dry cleaning software. [80:40] So I don't know. This is [80:42] 20,000 dry cleaners in America and they [80:44] all standardize on this one [80:47] tiny piece of software. That seems super [80:50] susceptible in a way that CrowdStrike's [80:53] Falcon platform is not susceptible. [80:56] And if you look at the dispersion even [80:59] within software, let's say they're all [81:01] down. The ones that are [81:04] um the ones that seem more replaceable, [81:06] those are down more. Like I it doesn't [81:08] look like just a wholesale panic. It [81:10] looks fairly reasoned. [81:12] >> Yeah, I mean like Cadence isn't down as [81:14] much as [81:16] um [81:17] Team. Atlassian. [81:18] >> Right. [81:19] >> Right. Atlassian has zero moat. [81:21] >> Yeah. [81:21] >> And that's why it's trading down to I [81:24] don't know where it is anymore, but [81:25] >> Bad. Are there any software stocks that [81:28] you think it's already over done and the [81:31] market really has it wrong and are buys? [81:33] >> Um [81:35] you know, I do think AppLovin was put [81:36] into that bucket. [81:38] >> Okay. [81:38] >> And [81:39] >> Just cuz it's in the index and they [81:40] treated it like it was [81:41] >> in the index and there were some short [81:42] reports on it and you know, people are [81:45] worried about advertising and agentic [81:47] advertising and how does how does [81:48] AppLovin get involved with that? Um [81:52] you know, I think companies like [81:53] Palantir [81:55] they're they're like the shepherd that [81:56] is basically having everyone cross the [81:58] chasm with them. [81:59] >> Do you own Palantir? [82:00] >> We do own Palantir. Not in the the ETF, [82:03] but across the other ones. [82:04] >> That one didn't come down until the end. [82:06] >> Right. [82:06] >> Like that one held up better than the [82:08] rest of SaaS. [82:09] >> Well, it it's kind of sold it actually I [82:11] don't know, went from 210 to 160 [82:14] >> Yeah. [82:15] >> before like the SaaS-pocalypse. And then [82:17] it went, you know, later. [82:19] >> Okay. [82:19] >> But um so it has it has pulled in quite [82:22] a bit. [82:22] >> Is Adobe dead? [82:25] They're acting like there's nothing [82:27] special about Adobe and meanwhile I see [82:30] open AI shut down Sora [82:32] and I don't know who's using all these [82:34] other tools, but is anyone paying for [82:35] anything? [82:36] >> I don't know. I don't [82:36] >> are paying for Adobe. [82:38] >> I I just started to use Higgs Field, [82:40] Higgs Field AI, which is basically a [82:42] video generator. It's pretty amazing. [82:44] >> What are you using it for? What sort of [82:45] videos are you generating? [82:46] >> Well, I'm just honestly I just play with [82:47] all these tools to figure out what what [82:49] they can do. And with Higgs Field I was [82:52] trying to figure out if I can make an [82:53] advertisement for my app. [82:55] Right? And to see if like if I wanted to [82:58] to have a Tik Tok video, how might I [83:01] make it? [83:02] Right? Um [83:03] and it was pretty good. I mean it wasn't [83:05] great, but it was pretty good. [83:07] >> I don't know if you I don't know if [83:08] anybody told you this, but I am the [83:11] creator of something called Halo. Have [83:13] you heard this term? [83:14] >> I I I was told you were. [83:16] >> Okay. On [laughter] a scale of 1 to 10, [83:18] how how clever is that would you say? [83:20] >> It's clever. I would say I [83:22] >> No, that's good right there. [83:23] >> No, no, no. Wait, wait. [83:24] >> [laughter] [83:26] >> When [gasps] when we did when I when [83:28] when we first thought about this like AI [83:30] and declining cost to create, the [83:32] corollary to this [83:34] was not I mean Halo is like a really [83:36] awesome way to [83:36] >> She's trying to say she came up with [83:38] Halo. Listen to this. [83:39] >> But was that any tangible asset would [83:42] get multiple expansion. [83:43] >> And the and and here we have it. I mean [83:46] basically [83:47] >> you you're packaging of it is beautiful. [83:49] >> Yeah, gorgeous. Well, all right. So the [83:51] reason I'm asking you this question [83:53] though, [83:54] um [83:55] could we overdo it? [83:57] >> Could we overdo which part? [83:58] >> In the stock [83:59] >> Oh, I'm sick of it. I'm sick of hearing [84:00] about Halo. Michael's going to throw [84:01] himself out the window if he hears it [84:02] again. [84:03] >> Um all right. Looking at the Mag 7. [84:07] If you had to rank all of them from most [84:09] Halo to least Halo, [84:11] and I said [84:13] I think so my personal opinion Apple and [84:15] Amazon are the two most Halo [84:17] stocks in the Mag 7, meaning [84:20] the physicality of what they have and [84:22] do. Okay, they don't own data centers, [84:25] Apple. But like [84:28] the my rectangle is my rectangle and [84:30] it's not going anywhere and it's central [84:32] to how the user is interacting with all [84:34] of this AI. I think it's very Halo. What [84:37] do you think about that? [84:38] >> Um [84:39] Okay. [84:40] If Apple doesn't act soon, [84:42] that impact that that Halo designation [84:46] is going away. [84:47] Because over time, I do believe that AI [84:50] will allow us to extract ourselves from [84:53] our screens. [84:54] >> Okay, and into what though, glasses? [84:56] >> It could be glasses, it could be I mean [84:58] you've seen you've seen the open AI [85:00] you've seen the open AI button. It [85:02] doesn't matter it does maybe not that [85:05] >> lens? [85:05] >> but some form [85:07] >> Something will come along. It could be [85:09] it could be a bracelet, it could be [85:10] spectacles, it could be something. [85:12] >> it could be something that that hangs [85:14] with you that acts as acts [85:17] What Apple should be doing is becoming [85:20] that personal agent. [85:22] >> I can't believe they're not doing it [85:23] already. [85:23] >> I don't know I don't understand why [85:24] they're not [85:24] >> gigantic Siri? [85:25] >> I don't know. [85:26] >> Why doesn't why doesn't Siri book my [85:28] plane tickets? [85:29] >> know. [85:29] >> Okay, I agree with you. [85:30] >> so but they are not doing it and maybe [85:33] >> they can't they can't do I think they [85:34] can't do it. [85:35] >> Well, I think they can do it but they [85:38] are so [85:39] >> the AI guy recently. I think they don't [85:41] have the horses. [85:42] >> They they didn't I don't I don't know if [85:44] they fired him or if he left. [85:45] >> Sure. [85:45] >> But um [85:47] I think the issue is that they're so [85:48] concerned with security as their their [85:51] whole being has been we are going to [85:53] keep your data we're the safe tech and [85:56] so how do you do this [85:58] and still maintain the story that you're [86:01] basically the safest tech. [86:03] So, you know, I think there's a there's [86:06] probably a bit of a [86:07] >> You think that's cool? [86:08] >> conundrum? [86:09] >> You think that's why they announced they [86:11] announced [86:12] um new and improved Siri and AI [86:15] capabilities two full years ago have not [86:19] delivered. Multiple iPhone models have [86:21] come and gone [86:23] and they don't have it. It's I've never [86:25] seen anything like it. [86:26] >> I'll tell you it's incredibly [86:27] frustrating because you see the [86:29] capability. [86:30] >> Are they Are they nervous? [86:31] >> so. I certainly hope so. Because I don't [86:34] know if they're nervous. I think that to [86:36] some extent they've, you know, milked [86:38] this device enough. I personally hate [86:40] carrying around an iPhone. [86:41] >> Okay. [86:42] >> I wish I didn't have to. [86:43] >> I love it. I take it in the shower. [86:44] >> Okay. [86:45] >> Now, if there were a new form factor [86:47] that everyone was doing, I would If it's [86:49] an amulet that you wear around your neck [86:51] and you talk into it and people just got [86:53] used to that, I would roll with it, but [86:55] there just isn't. [86:56] >> But there will be. [86:57] >> So So you're sure? [86:58] >> So our job is to figure out what what [87:01] the possibilities are in 3 years. [87:03] >> Okay. [87:03] >> Right? Is it possible [87:06] for there to be a different device that [87:08] is screenless? Is there Is it possible [87:10] for there to be a different device that [87:12] you interact with differently so you're [87:13] not, you know, opening this 150 times a [87:16] day? [87:16] >> Okay. [87:17] >> Right? And I think there is. [87:18] >> But you know, but if that goes away, [87:19] what happens to meta? [87:21] >> Yeah, so so there This is a discussion [87:24] that we've had internally of as you go [87:26] to a more screenless world, what happens [87:29] to other business models? [87:30] >> That's the Instagram. [87:31] >> Right. [87:31] >> [clears throat] [87:32] >> If I don't If I don't have my rectangle [87:35] and I'm not doing this [87:37] then where does TikTok and Instagram get [87:39] revenue from? [87:40] >> Zero. [87:40] >> Wouldn't it be wonderful to not have [87:42] TikTok and Instagram? You know how bad [87:43] it is for our brains? [87:45] >> But [87:45] >> It's like really good for my [87:47] ego, though. [87:48] At least Instagram. I don't All right, I [87:50] get it. I mean [87:51] >> But [87:51] >> But they said, "Listen, I remember [87:54] Google Glass was already like 12 years [87:56] ago. [87:57] It was a bust. The Snapchat Spectacles [88:00] didn't go anywhere. Apple's own stupid [88:02] helmet that only Michael bought, no one [88:04] else on Earth. He sold one. [88:06] Um He returned it to his credit. [88:08] Like [88:09] I'm waiting, but until [88:12] I think the rectangle is the rectangle [88:14] and I don't think that anything can get [88:15] in between us and our rectangles. [88:16] >> Yeah, I think for now you're right. [88:18] >> Yeah. [88:19] >> But the longer they don't have a [88:20] solution, the more at risk their model [88:23] becomes. [88:23] >> They their capex went down [88:25] year-over-year. They seem super [88:26] comfortable with [88:27] >> laurels. [88:28] >> Well [88:29] >> Yeah. [88:29] >> I I don't know if it's rest. I actually [88:30] think this part is brilliant. They're [88:32] like, you guys spend the capex, we're [88:34] just going to be the layer on top that [88:36] >> 30% [88:38] in the App Store. [88:39] >> We're just going to we're just going to [88:40] sit on top and use your compute. [88:42] >> It's up [88:43] >> go spend all the capex. [88:44] >> it's it's either Apple or Amazon are the [88:46] best of the Mag 7 year-to-date. Um for [88:49] different reasons. [88:50] >> Well, is Apple still the best business [88:51] in the world? [88:53] >> Uh I think the biggest size and the [88:56] highest gross margins, probably. [88:58] >> No, Apple's gross margins are not the [89:00] highest gross margins. Nvidia's gross [89:01] margins are higher. [89:02] >> Oh, Nvidia Yeah, Nvidia would be the [89:03] best. [89:04] >> Even Google. Google's gross margins are [89:05] higher. [89:06] >> Is that right? Okay. [89:07] >> Yeah, Apple's [89:09] Yeah. [89:10] 40% 30 38 40% is Apple's. [89:13] >> That's [clears throat and snorts] it? [89:13] >> Yeah. [89:14] >> What a piece of [89:15] And Courtney, Jeff on the show today? [89:17] >> It was great. [applause] [89:18] >> All right. So, we're going to take a [89:18] brief intermission and then we'll I'm [89:20] kidding. Um I have to say, this has been [89:22] so uh illuminating for me. I've learned [89:25] so much just talking to you today and uh [89:28] I think you were amazing on the show and [89:29] thank you so much for doing this. [89:31] >> Thanks for having me. You guys are fun. [89:32] >> Would you come back? [89:33] >> Oh, I would love to. [89:34] >> Okay. What are you doing next week? [89:36] >> [laughter] [89:36] >> Not next week. [89:37] >> This guy This guy is already shaking his [89:39] head no. All right. Um I want to tell [89:41] people how they can learn more about [89:42] Alger and your funds. So, uh is it [89:44] alger.com? Is it that simple? Okay. [89:47] Um and do you do you you're not doing [89:50] social media really, are you? Little [89:52] bit? [89:52] >> No. [89:52] >> No. Good for you. I So, I respect that [89:55] the most, I think. [89:56] >> Maybe I'll have Maybe I'll write a bot [89:58] to do it for me. [89:59] >> Um we want to have you back to talk [90:00] about robotics and also I know look at [90:02] your face. Look how excited you are. All [90:04] right. Well, so we're going to [90:05] >> Oh, we have a lot to talk about. [90:06] >> All right. So, we'll do maybe we'll do [90:07] that later this year. [90:09] Um, I just want to say thank you so much [90:10] for joining us and guys, thank you for [90:12] listening. [90:13] Compound appreciates you. If you have [90:16] not yet hit the like button on this [90:17] episode on whatever platform [music] you [90:20] are uh watching or listening to us, now [90:22] would be a pretty good time to do that. [90:23] What do you think, John? [90:25] Right now? [90:25] >> 100%. [90:26] >> 100% he says. [90:28] All right. Guys, thanks again. We'll see [90:29] you soon. [90:30] >> [music] [90:33] >> All right. Do you like how that came [90:34] out? Yeah. [90:36] All right.