视频库 / NO.071ASK THE BEST MINDS THE BIG QUESTIONS一人,一实验室
视频库 / NO.071
字幕 字幕位置
--:--
点击播放,这里会跟随视频显示当前句的中英字幕。
EDITED TRANSCRIPT · 依据现场录音编译整理,可划线生成便签
编者按:本文整理自播客《Invest Like the Best》的一期对谈,主题为《AI 热潮才刚刚开始》。受访者是 Whale Rock Capital Management 创始人兼首席投资官 Alex Sacerdote,他于 2006 年在波士顿创办这只科技主题基金;提问者是该节目主持人、Positive Sum 创始人 Patrick O'Shaughnessy。两人从 Anthropic 这笔投资谈起,一路谈到 S 曲线框架、竞争优势的形态、应用软件的困境、数据中心硬件的重估,以及一位在高盛工作四十一年的父亲。全文依据现场录音编译整理,仅删去口语枝节、寒暄与重复,论证与细节均予保留。

开场:曲线笔直向上

萨瑟多特: 当你踩中 S 曲线的正确位置,你拿到的是指数级的销量增长。如果这门生意的商业模式足够强,利润就不是线性增长,而是指数级增长。问题在于,这个世界不习惯用指数去思考。相信自己能准确预测两年、三年、四年之后的人,非常少。但只要你跟得住并且真正理解 S 曲线,懂得护城河,会做模型,你确实可以把这些了不起的事情提前算出来。企业级 AI 应用这个市场,渗透率还不到 1%。我们从没见过这样的曲线,平时我们谈 S 曲线,这一条我们干脆叫它 L 曲线,笔直向上。

发令枪响之后

主持人: 你说你眼下确信度最高的持仓是 Anthropic。能不能讲讲你是怎么发现它、怎么投进去的?我想借这个案例,把你我此刻共同关心的所有话题都串起来:像你这样的投资人如何进入一级市场、Anthropic 这门生意、AI 本身。为什么它的确信度最高?故事从哪里开始?

萨瑟多特: 2022 年 11 月,OpenAI 发布 ChatGPT,那一枪响了之后,我们立刻把整个公司投进去,带着十个人的团队做了一次彻底的深挖。每当出现一种新的计算范式,就会出现一套新的技术栈,而这套新栈会在旧栈上制造出新的赢家和输家。这套栈的结构,现在黄仁勋讲得很多了:最底下是电力,再往上是芯片,然后是云,再上面是基础模型,最上层是应用。

那是 2023 年初,我们的判断是:先站在芯片和基础设施这一层。理由有两条。第一,需求最先落到他们头上。第二,我们知道谁会赢。上面几层谁会胜出,当时我们并不确定,但无论谁赢,都需要海量的算力。我们围绕这一层做了非常深的功课,这部分稍后可以细说。

接下来两三年,我们对基础模型层怎么演化才慢慢看清楚。两三年前,冲这个方向的公司有六十家,OpenAI 大致领先。2023 年 4 月我们做过一场线上分享,当时给出的是几种可能:也许是赢家通吃;也许彻底商品化,因为有开源玩家;也许是一场向零的竞赛;也许最终变成三四家领先者的寡头格局。

后来三年发生的事情是,几乎所有创业公司都掉队、死掉了。然后是世界上最大的几家公司下场,包括亚马逊和 Meta。亚马逊其实一直没真正出现过。Meta 会怎样还要看,他们进场时来势很猛,但努力半途卡住,只能彻底重启。与此同时,Anthropic 这匹黑马跑了出来,它是一家创业公司,路线非常纯粹,就是死磕企业市场;而 OpenAI 基本拿下了消费端;Gemini 则永远不能低估,我们同样喜欢谷歌,它是我们最大的持仓之一。

于是格局越来越像一场三强赛,某种程度上的寡头。这跟云市场当年的演化非常像:三家公司托起了整个 SaaS 云世界,而且它们的生意都极其出色。

开源到底能追多近

萨瑟多特: 我们当然也意识到开源的风险,尤其是来自中国的开源。但我们逐渐能安心的一点是,最前沿模型产出的 token 质量确实更高。原因在于,如果你已经做到榜单顶端的 80% 水平,从 80 走到 85 是一次巨大的解锁。而开源阵营没有那么多算力,他们可以逼近前沿,却没法实现越级超车,然后就会掉下来。

与此同时,规模定律以及其他改进模型的手段,比如各种反馈回路,都还有很长的跑道。我们接触的每一个行业内部的人都认为规模定律会继续奏效。所以我们形成了三强赛的判断。

代码,真正的解锁

萨瑟多特: 然后最大的那记重锤来了,是代码。这才是 AI 真正的解锁点。

头几年我们知道 AI 会很大,但我们是有疑虑的。我们下了重注,因为我们知道训练需求一定在,可是能带来多少收入、能不能真正替代人力,我们心里没底。你回想早期那几版模型,是不错,但企业侧有不少负面反馈,它们究竟能不能真正做到智能体化,是个问号。

到 2025 年,答案出来了。Claude Code 和一批编程工具开始爆发。第一代产品是微软 Copilot 那种,一个月二十美元,能帮你把代码的「语法」修得漂亮点,也许找出一个 bug,也许写出一段类似一个自然段体量的代码块。然后年中前后 Anthropic 拿出的东西,能力远不止于此,它开始能以智能体的方式跑起来。我们眼看着这件事发生,编程这个市场直接炸开了。

后来我们听到的说法是,那些可以不受限制地使用工具的人,包括 Anthropic 内部的人,当时一天在 token 上要花掉一百美元。你算一算,一年就是两三万美元。再想想全世界有多少程序员,两千万。光是编程这一项,市场规模就有五千亿美元。而且请注意,这是建立在只有七八九个月历史的技术之上的。仅凭编程这一个市场,我们就能看到 Anthropic 面前摆着一个巨大的机会。

有意思的是,我们当时在给投资人的信里写了这么一句。我们是在 1800 亿美元估值那一轮投进去的,而当时他们希望做到九十亿美元的收入。

主持人: 从一亿到九十亿。

萨瑟多特: 对。那些数字是我们从没见过的:从一亿冲到十亿,然后奔着九十亿去。而我们做这笔投资是在 2025 年 8 月,那时候还没有人知道 2026 年会是什么样子。

最近的第二次大解锁是,Claude Code 几乎完全智能体化了。去年安德烈·卡帕西和林纳斯·托瓦兹这两位公认最懂代码的人,说法还是相当保留的,后来他们彻底改口。卡帕西说,去年的工具能写 20%,剩下 80% 得手写;最新模型出来之后,这个比例反过来了,现在他除了用英语,一行代码都不写了。更别提那些原本根本不会编程的人,他们身上要释放的能量还没算进去。所以单是编程这一块就已经完全起飞,而 Anthropic 一直守在前面。

模型层为何没有商品化

萨瑟多特: 云和 AI 公司之间有一个关键差别。云在本质上是商品,卖的是服务器和存储,上面当然叠了很多软件,也有黏性,但底子是商品。大家原以为 AI 模型也会是纯商品,结果里面存在巨大的差异化:训练方法不同,各自擅长的技能不同。很多人做了路由器在几家之间切换,这看上去像是在说它们可以互换,可实际上,Anthropic 在跟私募股权和金融沾边的任务上非常强,谷歌在读取 PDF 上非常强。所以这里面有大量差异化,有关键的知识产权,这本身就是极好的竞争优势。冲着编程这块地盘来的公司很多,Anthropic 一直守在前头。

基础模型公司,特别是 Anthropic,还有一点好:他们卖的不只是 API 或者模型本身,而是围绕 API 建起一整套产品生态。SDK、面向协作的 Claude、编排层,以及所有配套工具,他们把 API 外面这层软件叫做「脚手架」(harness),作用是把模型的能力压榨到极限。

这件事我们在 2013 年的 AWS 身上很早就看到过。当时人们觉得那不就是仓库里的商品化服务器吗,有什么大不了的。可他们看到的是一种全新的计算方式,于是发明出一堆别人还看不到的产品,一点一点把锁定效应建起来。

从十个基点开始

萨瑟多特: 我们的另一个视角是:我们在 S 曲线的什么位置。基础设施层这条曲线,我们判断渗透率大约 10%。顺带说一句,我们认为它仍然是参与 AI 最好的方式之一,稍后可以讲这一层怎么把收益反哺回来。

你想想,虽然已经有两亿、八亿,我记不清具体多少人在用 AI,但他们用的是 AI 1.0,相当于打了兴奋剂的搜索引擎。现在有了新的原语:Claude 装进你的电脑,接上各种系统,然后你构建技能。接下来,人和公司会开始构建技能,再往后构建真正的 AI 机器人,然后大企业会构建规模大得多的东西。可是真正在做这件事的人有多少?桑达尔说过,只有全世界知识工作者的十个基点。Anthropic 大概有一千四五百万日活,其中真正把 AI 用出应有样子的,恐怕只是一小部分。

这十个基点,是典型 S 曲线的起点:先是折腾者,然后是早期采用者,再然后是早期主流人群。未来四年,这个数字会从十个基点走到 1% 到 2%,再到 3% 到 5%,再到 15%。而今年在企业侧,开关「啪」地一声被打开了,所有人突然意识到必须马上做,而且要快。

主持人: 有点像互联网 1.0 的那种感觉。1998 年你知道自己必须有个网站,但做那个网站很难。

萨瑟多特: 是的,但这一次东西凑齐得非常快。所以我们认为,企业级 AI 应用这个市场的渗透率还不到 1%。我们从没见过这样的东西,平时我们谈 S 曲线,这一条我们叫 L 曲线,笔直向上。

再把这个结论带回基础设施层:现在只有十个基点的人在真正使用 AI,算力就已经卖光了,全世界的算力都不够用。Anthropic 现在拿到的算力只有它需要量的一半,而这还是在那一轮巨大的采用潮到来之前。马克·安德森说过,未来四年他唯一确定的一件事,就是算力不会够。

公开市场基金怎么挤进一级市场

主持人: 我很好奇,像你这样出身公开市场的投资人,过去想买什么按一下买入就行,现在却要在许多最重要的非上市公司里持仓,比如 Stripe、Databricks、OpenAI、Anthropic。你怎么拿到你想要的仓位规模?有多少是靠直接和公司打交道的创造力?如果是直接谈,那就是双向选择,人家也得同意让你进来。关于怎么拿到理想的额度,你学到了什么?毕竟这不是你原来的看家本领。

萨瑟多特: 就 Anthropic 这一笔来说,我们是先认识了这家公司。我们有位分析师认识他们财务团队的人,其实六百亿美元那一轮我们看过,但没投。那时我们对这家公司了解得不够,毛利率是负的,而且坦率讲,我们还没有看到编程业务爆发成后来那个样子。

公开市场有个好处,你可以用很长的时间去了解一家公司,可以按自己的节奏出手。后来我有机会跟达里奥相处了一段时间,加上播客里听到的,我开始意识到这个管理团队非常优秀:专注、投入、几乎没有人员流失,代码质量很高,而且商业计划真的开始跑通了。从一亿做到十亿是一回事,做到九十亿是另一回事。

于是我们尽一切可能去接触公司,他们愿意见我们。我们做了一份九十页的演示文稿,用 Claude Code 把互联网上关于编程市场的反馈全部扒了一遍,包括他们的产品强在哪里、可能需要在哪里改进,同时把我们对编程市场规模的完整判断也放了进去。他们欢迎我们进入这一轮,之后我们和他们的首席财务官一直保持密切联系。能和他们建立这样的关系很好,我认为在额度上我们拿到的超出了自己的体量。这一笔是彻底的全垒打。

至于其他情况,我们现在所处的时代是,独角兽市场比欧洲多数国家的股票市场都大,甚至可能比它们加起来还大,肯定比德国大,肯定比英国大。所以在我们投一级市场之前,我们就必须了解这些公司,第一笔是 2020 年。现在更是必须了解,因为它们有时就是某个赛道里最大的公司,影响力巨大。

我们每年和管理层做两三千场面对面会议,其中大约 10% 到 15% 是非上市公司。然后我们会聚焦到真正想深入了解的那几家,想办法见到他们,参与他们的融资。

Stripe 那一笔

萨瑟多特: 我们的第一笔是 Stripe。当时我们在 Adyen 上有很重的仓位,大概是 2017 到 2020 年。Adyen 是一家出色的支付公司,属于新一代云支付,从 WorldPay 手里抢生意。当时云原生的现代支付只占八十万亿美元总盘子的 5%。但是,你如果不像了解自己的手背那样了解 Stripe,你就没资格投 Adyen。所以我们做了大量尽职调查,为 Adyen 访谈了两百个客户。可是当我们问起 Adyen,人家同时会说起 Stripe,我们意识到这是可口可乐和百事可乐的关系。于是我们说,必须找到办法投进去。2019 年我终于见到了科里森兄弟。

那时我们在一级市场上没什么名气。我有个朋友在一家持有大量份额的创投机构,我跟他说过,如果哪天你们想卖一点,记得告诉我。2020 年 4 月,疫情期间,他给我打了电话。我们对 Stripe 了解很多,虽然没有完整的财务数据,但在那个估值上,我们知道的已经足够了,我记得是三百五十亿美元。他们披露过总支付额超过五千亿美元。我们知道 Adyen 的费率是二十五到三十个基点,Stripe 是四十到五十个基点,我们也知道他们有多少员工,所以大致能推出盈利能力。

结果是,实际费率更高;他们对总支付额的说法很谦虚,真实数字远高于五千五百亿,接近一万亿。我们按自己的假设完成了承保,而现实比假设好得多。后来我们还把这笔从卖方手上加码到了一个亿美元的大单。

有时候卖方也乐见这样的买家:创投基金持有一段时间之后大多会卖出,而他们喜欢我们不但会持有,还会在公司上市后继续在公开市场持有。我们在 Nubank 上就是这么做的,上市后又在公开市场长期持有。

三根支柱

主持人: 也许现在正是时候,请你把你关于 S 曲线的全部心得摊开讲。你的公司几乎就是建立在技术采用生命周期这个理念之上的,在某次平台变迁或者 S 曲线变化中,在正确的时点投进正确的公司。S 曲线的基本概念大家都知道,你提到的折腾者、早期采用者、早期大众这些阶段也不陌生。但既然这是你长期以来观察市场和个股的透镜,我想请你讲得极其具体,讲清楚为什么 S 曲线对投资如此有用,讲到那些细枝末节的分寸。

萨瑟多特: 我们有一套投资框架,三条。第一是 S 曲线,第二是竞争优势,第三是被低估的盈利能力,我们一条条来说。

当你踩中 S 曲线的正确位置,你拿到的是指数级的销量增长。如果商业模式很强,而科技行业里靠各种不同护城河支撑的强模式非常多,那么利润就不是线性增长,而是指数级增长。这就引出第三条:在长期盈利能力被低估的时候买入。每股收益常常能从 1 美元长到 10 美元,长到 20 美元,这种情况发生的频率比你以为的高得多。正因如此,你能用极低的市盈率买到世界上最好的一批公司。

2023 年我们买英伟达的时候,付的是四倍市盈率。2019 年我们为汽车这条 S 曲线买特斯拉的时候,付的是五倍。持有苹果的那些年,我们付的是四倍。我们为 AWS 买亚马逊的时候,AWS 相当于白送。

这个世界不习惯用指数思考,大家的注意力全在明年、下个季度。相信自己能准确预测两年、三年、四年后的人非常少。但只要你跟得住 S 曲线,懂护城河,会建模,你确实能把这些了不起的事提前算出来。

拐点是怎么来的

萨瑟多特: 先说 S 曲线本身。它之所以关键,是因为每一项技术都遵循同一个形态。东西早就出来了:智能手机在 iPhone 之前已经存在了十年,互联网在网景之前已经存在了二十年,AI 一直藏在这些公司内部,直到 ChatGPT 把它推到公众面前,把这团火点着。电动车也一样,特斯拉在 2019 年那次垂直起飞之前,已经上市十五年,因为采用的障碍实在太多。

早期的智能手机笨重、没有触摸屏,那不是苹果做的,也没有无线数据网络,而且太贵,五六百美元。乔布斯把价格压到两百美元,AT&T 有了 3G 网络,屏幕能触摸,简单到你奶奶都会用。苹果搭起生态,把事情变简单。所有采用障碍被一一清除,之后就是火箭升空。那就是需求的龙卷风,全世界都知道自己马上就需要这个东西。这就是那一次翻转。

电动车也是同一回事。价格太高,马斯克把价格压到四万美元;有里程焦虑,他把续航做到三百英里;供应链终于就位,他可以一年产出上百万台。这些条件凑齐,拐点就被触发了。

曲线有多高

萨瑟多特: 还有一个更细的层次:不是「哦,它起飞了」就完事,你还得知道这条 S 曲线有多高、有多大。因为我们承保的是两三年后的样子,你必须知道那之后的增长长什么样,这样才知道该什么时候卖、该拿多久。

而且 S 曲线是会变化的。当年 AWS 还只是亚马逊财报里一个隐藏的科目,跟着它的是研究零售互联网的分析师,不是研究硬件芯片的人,何况那还是一种全新的商业模式。我们意识到它的潜在市场是企业 IT 史上最大的一块,因为在此之前,市场是路由器、内存、存储、戴尔、EMC 这些,而 AWS 把这些活儿全干了。

于是我们去算这条 S 曲线有多高,算出来他们直接对标的 IT 系统市场是六千亿美元。然后我们假设这件事会带来 50% 的通缩效应,据此推出当时的渗透率只有百分之一二。但后来我们发现,它其实并没有通缩。你现在去问任何人,他们都会说自建的成本和用云差不多。这意味着潜在市场比我们原来算的还要大得多。

所以既有巨型 S 曲线,也有嵌套其中的子曲线。我们算是幸运,一路赶上了互联网 1.0、移动、云、电商,现在是 AI,而 AI 可以肯定是其中最大的一条,而且这些浪潮层层叠加。

不过也要留神。电动车那条曲线,我们当时判断也许 40% 到 50% 的汽车会电动化,结果它在 10% 到 15% 就撞上了一堵墙。通常 S 曲线会一路走到底,但这一次由于种种原因没有。所以你必须调整,必须持续盯住。

一般来说,当渗透率到了三四成,指数增长就结束了,这意味着卖方分析师的预期追上来了,不再有大幅超预期。

主持人: 那通常就是你卖出的时候?

萨瑟多特: 我们偏好高增长。苹果那次算是个失误。头五六年苹果太棒了,是我们最大的持仓,除了 2008 年,每年上涨百分之五十到七十。我们在 2012 年卖掉了,当时美国大约一半人有了智能手机。而苹果后来保住了领先地位,经历了两年跑输,估值压到很低,又添了几块附加业务,还能从应用商店抽三成,参与到应用这一层。所以它依然复合得相当漂亮,比如年化 20%。但真正的大年份,是在渗透率从 0 走到 50% 的那一段。

什么时候开始动心

主持人: 我很着迷于这些曲线开头那段有时长达十年以上的平坦期。这让我想问,关于什么时候该买、甚至什么时候该开始留意,你学到了什么?怎么衡量?每次都不一样吗?你踩过哪些坑?卖出我们谈过了,那什么时候该开始考虑买入?

萨瑟多特: 安迪·格鲁夫说过,遇到战略拐点的时候,你不能相信数据。战略拐点靠的是直觉,是零散的现场证据。我很喜欢一本讲全脑投资的书,叫《道琼斯平均律》(The Tao Jones Averages),讲右脑和左脑。最好的投资人都有创造性的那一半,那是视觉化的,是把点连起来。

我们在手机游戏这条 S 曲线上待了很久。早年手机游戏之所以只能是休闲小游戏,是因为屏幕太小、算力不够。后来我在中国,看到一个十二岁的小男孩拿着一台巨大的手机,玩着一款相当出色的游戏,我当时就想,天啊,它真的搬到手机上来了。这就是视觉线索。

企业市场难就难在你看不见。所以我们去 Gartner 的 IT 峰会,三万名美国的首席信息官会到场。Splunk 当年就是这样被我们看到的,它曾经是一家极棒的数据库公司,讲解它的那个会场是站着都没地方。VMware 也是,那得是三十年前了,他们把服务器虚拟化,会场同样挤到只能站着,你能直接看见企业需求开始萌动。AWS 那次我们去,大宴会厅九点钟坐得满满当当,十点钟还是满的,十一点钟依然是满的。你真的能在需求爆发之前把它看见。

所以我们到处找线索,这里面有一整套模式识别。顺带说一句,晚一点是可以的。很多情况下错过头一两年、三年都没关系,因为如果这条 S 曲线的顶部是五千亿美元,增长可以持续很久。你不必非在第一时间进场,错过第一个翻倍是可以接受的。我在富达起步的时候,彼得·林奇很喜欢带年轻人,我跟他有过一些相处时间。他说:把走势图涂白,一切都关乎未来。

收音机与洗碗机

萨瑟多特: 所以晚一点没关系,而 S 曲线帮你判断的,是这件事还能走多久。除此之外还有曲线的斜率,这一点很重要。很多人以为现代世界什么都快,其实决定采用速度的因素相当多。我们请霍拉斯·德迪乌做过一项研究,他以前和克莱顿·克里斯坦森共事,我们让他回到历史里去看。我们墙上挂着过去一百年里那些重要的 S 曲线。

收音机那条是史上最快的之一,七年左右就接近百分之百渗透。洗碗机那条则很平,因为它必须接到后端的管线上。

主持人: 你还发现了什么?这太有意思了。

萨瑟多特: 比如 B2B 的东西通常需要很长时间,因为它必须接进现有系统。就像洗碗机,得先装进房子里。而面向消费者的东西一般快得多。

主持人: 我很喜欢这个说法,收音机和洗碗机,两种采用模型。

萨瑟多特: 我在富达是看互联网的,我的第一只股票是亚马逊,那是另一个很好玩的故事。但我也看 B2B 互联网,当年围绕它有一整套宏大的多头论述,可底层基础设施根本没就位,所以 B2B 那件事没发生,一直到二十年后 SaaS 出现才真正实现。

这也是 AI 的一个风险。大公司非常在意安全,行动可能很慢,围绕 AI 还有很多文化上的阻力:你需要几个真正的布道者去推,需要高层去推,可 IT 部门会说这有风险。云当年也是这样,最大的顾虑之一就是把数据放在云上不安全。后来我们看到中情局用了,看到第一资本用了,我们跟第一资本的首席信息官聊过,结论是放在云上更安全,之后才真正起飞。

但正因为 SaaS 像洗碗机,云也像洗碗机,必须接管线,所以它们虽然在长,增速大概是 30% 到 40%,也许 50%。而 AI 惊人的地方在于,不管是消费者还是企业,你只要打开浏览器,它就在那里。

主持人: 打开浏览器就在那儿。

萨瑟多特: 所以我们才会看到笔直向上的曲线。而且至少在中短期,从十个基点的真实使用率走到百分之二到五,跑道足够,足以让它继续直上。所以我们把它叫做倒过来的 L 曲线,真的挺激动人心。

谁能从队伍里跑出来

主持人: 关于赢家群体什么时候从竞争队伍里分化出来,你学到了什么?你刚才讲的主要是 S 曲线整体的需求和增长,而抢这块蛋糕的玩家总是很多。看起来你的做法是等某家已经从队伍里跑出来之后再投,而不是在混战中挑赢家。大致对吗?

萨瑟多特: 我们的顺序是:先找到 S 曲线,然后对所有在这个领域有暴露的公司做穷尽式研究,从中找出具备强大竞争优势的那一家。

很多人不喜欢科技股。沃伦·巴菲特不喜欢科技,因为他没法预测未来,变化太快。而 S 曲线就是我们看向未来的地图。另外,很多人对科技的担忧在于颠覆太多,你没法相信一家公司能成为长寿资产。但这些年我们发现,数字世界里的某些竞争优势,其强度不亚于线下世界,甚至更强。

第一是网络效应,领英、脸书、阿里巴巴都靠它,例子举不完。第二是成为行业标准。甲骨文和彭博就是行业标准。甲骨文收费很高,市面上有免费版本,也有开源的替代,可数据库管理员全在他们那儿,所有软件都为他们调优过,所以他们对关系型数据库市场的锁喉持续了很久很久。第三是快速做到规模。这些曲线增长太快,Anthropic 忽然就做到了九十亿、三百亿美元的销售额;亚马逊同样拿到了巨大的规模,而且拿得很快,沃尔玛用四十年建立的规模优势,他们五年就有了。第四是成为平台,让别人在你之上构建。第五是关键知识产权,高通就是例子,不给他们付钱你造不出手机;ASML 也是,没有他们的光刻机你造不出芯片。第六是品牌,品牌非常重要,谷歌和亚马逊要增长,却从来不需要打广告,马斯克做任何事都不需要打广告,这背后是获客成本对终身价值的关系,也就是整个商业模式。

我刚才提到的那些公司,比如苹果,几乎是把这些优势全都集于一身。有时候我们能比世界上其他人更早注意到这些东西。我们的高光时刻之一,是 2013 年在罗宾汉投资者大会上讲亚马逊,讲的是 AWS。我们说,连多头都不知道自己手里坐着什么。亚马逊在战争开始之前就已经赢了。当时我们说,这里只有可口可乐,没有百事可乐。后来证明还是有一个百事的,但市场大到足够容下。我们能看到他们领先七年,所以先发很重要。然后他们变成了完整的生态和平台,然后拿到规模,体量是所有人的十倍,没人能砸出足够的研发去追。

但你说得对,如果没有竞争优势,你哪怕站在史上最好的 S 曲线上,一样会输。

主持人: 一样会出局。

萨瑟多特: 如果你的名字叫 RIM、Palm、诺基亚、HTC、LG、摩托罗拉,我可以一直数下去,结果就是零、负、负、负、负。我们在基础模型层看到的正是这一幕:五十家公司在做,全都掉队了,两三家浮到顶上,而且有很多理由相信他们能守住位置。

模型公司的护城河

主持人: 谷歌稍微复杂一点,因为 Gemini 后面挂着另一门庞大复杂的生意。但如果把 Anthropic 和 OpenAI 当成纯粹标的,把它们的竞争优势一条条推演一遍,为什么在足够长的时间里它们不会被侵蚀?

萨瑟多特: 我们做过的所有 S 曲线里,AI 是最复杂、变化最快的一条。所以我们必须记住风险确实存在,但回报也是最高的,因为我们谈论的是以万亿计的市场。刚才说云也许是八千亿美元,AI 我们现在的判断是三万亿到五万亿。高风险,高回报。

就 Anthropic 来说,第一,他们看起来握有关键知识产权,总体上守住了在代码上的高市场份额。第二,他们在企业端建立了很强的品牌,你去问任何一位首席信息官,他们脱口而出的第一个名字就是 Claude。第三,他们会拿到逃逸速度和规模。OpenAI 和 Anthropic 要跟谷歌这类拥有巨大现金牛的公司缠斗,本来是件吓人的事。但两家管理团队都值得称赞,他们在极度资本密集的行业里找到了融资的办法。就 Anthropic 而言,十倍的收入增长加上融资能力,看起来他们已经达到了逃逸速度,于是有了规模。

还有一点是 Anthropic 和 OpenAI 都可能拥有的:Anthropic 现在在代码上领先,他们把代码能力再喂回自己的模型,这就是递归改进。看他们的创新节奏,是在加速的。所以他们也许能进入某种腾空阶段。

OpenAI 之前的精力分散在很多不同赛道,但他们在企业端开始变好,编程工具不错,那一侧的增长也开始加速。另外要看到,消费端有海量的眼球。不过就目前而言,企业端的生意明显更好,因为你我都愿意付很多钱,它替代的是人。消费端也许能靠广告,也许如果你能把一个 Claude 式的助理做到足够好,用户愿意付费。

你说得对,格局是会变的,但通常来说,我们几乎每次路演都会拿出的那张图显示:在互联网上,领先者会变得更大、更快,并且赢下去。多数时候就是领先者赢。Shopify 成为领先者,就一路走下去;亚马逊是领先者,就一路走下去;某家 SaaS 公司拿到领先,就在互联网上自我复利。另外你必须够大,必须有规模,必须有算力,还得付得起算力的钱,能做到的人就那么几家。这些就是我们认为正在成形的护城河。

当然也有例外,通常发生在范式切换的时候。美国在线没能从拨号跨到宽带。网景出得很早,但商业模式不够强。不过你去问硅谷任何一个人、任何一家创业公司,他们都会告诉你,他们在这三家之上构建。世界很大,经济很大,他们能在这三家的基础上做出差异化。

为什么几乎清空了应用软件

主持人: 那我很好奇,这一切对软件意味着什么。翻你的组合,我看不到多少大型软件公司、企业软件公司。不知道你以前有没有持有过又卖掉了,或者你原本就是这么看的。但只要有过用 AI 搭一些真正好用的小工具的体验,哪怕它们还只是玩具,你都很难不冒出一个念头:如果我花足够多的时间,哪怕我不是技术出身,也许我能给公司做一个替代 ERP 的东西。似乎没有什么根本性的理由说这不可能,而如果可能,那些公司就有大麻烦了。这件事上每个人都有很强的立场。你是怎么看这类公司的?

萨瑟多特: 大概五年前,我们组合里可能有 40% 到 50% 是软件。而且在 2023 年 4 月那场研讨里,我们说得很明确:一定要先投芯片。但在应用层,我们最初的想法是,这些公司体量大、销售队伍庞大,可以拿这些 AI 的 API 去做产品,而且他们有数据,这对软件来说会是天大的好事。

结果我们很快发现,他们的 AI 产品并不好,拉不动业绩,也没人收得上钱。于是我们基本上把应用软件全部卖光了,现在只剩一两个很小的仓位。今年年初我们在这个板块实际上是净空头,这在第一季度帮了我们大忙。

这里面有很多层。老式软件像用纸和笔,或者说像马车;新式软件像喷气发动机,坦白说更像《星际迷航》里的传送装置。它的革命性大到让你觉得,它现在就必须具有颠覆性,哪怕它此刻还没有,哪怕不是马上。

软件公司还有另一个问题:在任何首席信息官的待办清单和优先级排序里,他们的位置大幅下滑了。所以就算 AI 短期内不构成颠覆,预算也在往 Anthropic 的 token 上走,因为那边的投资回报更快。第二,钱花在那边,预算就被挤压,这对他们是伤害。第三,很多软件公司过去每年都能提价,现在他们大概不太敢了。第四,还要看就业会怎么演变,这件事两边都有聪明人在争论,但我们确实看到有些公司在大幅削减岗位,或者冻结招聘,这会伤到席位数。

至于他们自己做应用,如果你想乐观一点,可以说他们只是花了点时间。我们前面讲过 AI 的原语还非常早期,所以也许他们只是还没走到可以商业化的那一步。但也可能他们没有合适的人,也可能他们不明白,卖一套固定系统和卖一个替人干活的东西,销售动作完全不同。如果你部署的是替代人力工作的系统,你必须蹲在客户旁边,确认活儿真的干成了,你需要前置部署工程师,而这些人他们内部未必有。

当然还有「自己造」的风险。多头会说,他们永远不会自己造一套 ERP,这大概是对的。老技术确实很有黏性,这也是事实:手机游戏没有伤到主机游戏,平板没有伤到 PC,智能手机也没有伤到 PC。软件里有大量集成和积累的工作量,而且企业确实更愿意买,不那么愿意自己造。这些都对。

但你没法想象不出这样一个世界:一到五年之内,针对每一家强大的在位者,都出现一家 AI 原生的新公司来抢,它们的数据优势可能被抹平,借助 AI,把旧的拆下来换上新的也许并不难。

对做空的人来说,好处是这些公司估值很高,而且所有人都知道它们承压。有些人会忍不住去抄底,可 AI 编程工具在越变越好。所以只能继续看。我们非常密切地盯着这些软件公司,看它们能不能拿到足以改变轨迹的收入。但这很难,因为如果你是 Salesforce 这种体量,四百亿美元的销售额,AI 相关的年度经常性收入可能只有五亿到七亿,基数太大了。也许这件事最终会跑通,但需要时间。

软件行业有个「40 法则」,增长率加上营业利润率。20% 的增长加 20% 的利润率,就算不错。对 AI,我们有一个新版的 40 法则,其实主要是用在芯片投资上:你的销售额里 AI 占比多少,比如 30%;你在这个品类里的市场份额多少,比如 30%;加起来是 60。这是个很好的观察起点,因为你既有暴露度,又有强势的市场地位。软件的问题是,它们的 AI 占比现在只有百分之一二,路还长得很。

不过最近我们也在捕捉另一条线索,这个想法还不成熟:AI 也许反而会让某些软件平台变得更重要。你拿到 Claude 之后第一件事做什么?把它接进 Slack。如果 Slack 能因此成为关键的数据仓库,那它在组织内部就变成了永久性的固定装置。所以也许这些智能体,也许 AI 的下一波,就是会使用工具的智能体,它们会在现有在位软件工具内部运行,像人一样去使用这些工具。

主持人: 顺着这条线索说,它们会用到的、黏性最强的工具,共同点似乎是基于网络的工具。Slack 就是个好例子,Slack 本身的软件我说不好,总归有让人不满意的地方。特别的地方不在软件,而在于所有人都在上面。

萨瑟多特: 对。

主持人: 我很好奇你会想要什么样的东西。是不是只有网络效应这一条真正重要?

萨瑟多特: 我们在这块的思考还很早期。但我觉得未必只有网络效应。比如 Workday 或者人力资源系统,或者那些大型的记录系统,智能体可能就跑在它们之上。CRM 正在无头化,或者说他们在做无头版本,这也是空头论述的一部分:你被降级成一个数据库;本来有人机界面,现在需要做 AI 界面,而 AI 界面就是没有界面,它直接钻进数据里去。这样你就失去了和客户的交互。但反过来说,如果智能体直接进入 CRM,在 CRM 内部完成工作,那反而会把 CRM 固化下来,你就不用担心它会消失。

数据中心硬件的去商品化

主持人: 我们能聊聊芯片吗?你提到过好几次。基础设施芯片,或者说数据中心周边的一切,我不确定你是怎么划分的。为什么这一块对你这么有意思?我很喜欢那个改良版的 40 法则,AI 收入占比加品类市场份额,这个指标很有意思。今天哪些公司在这个指标上闪光?哪些落后者出人意料?

萨瑟多特: 过去四十年,数据中心里什么都没变。哪怕到了云时代,基本上还是英特尔 x86,它在九十年代某个时点成了数据中心的芯片。云时代算力在增长,计算负载每年增长 25% 到 40%,但摩尔定律的改进速度也差不多是这个数,所以并不需要多大的创新。硬件这个行业多年多年地几乎没有增长,整个产业链的每一个部分都被商品化了:每一颗芯片、服务器的每一个部件、印刷电路板、内存、机箱、网络,都没有创新。从一G 到十G 要花七年。刚切换的头一年确实需要一些创新才能做到十G,会带来一个小周期,然后又商品化。

现在到了 AI,负载每年增长十倍,把硬件的每一个方面都推到物理极限。所以你不仅拿到了巨大的销量增长,整个行业还出现了我们所说的硬件产业的去商品化。

大概三年前我见过肖恩·马奎尔,他说他真希望自己能回去做一只硬件对冲基金,因为这些公司全是上市的,而且个个握有强大的知识产权。红杉最好的一批投资,正是在硬件年代做的,苹果、思科等等。我们现在处在芯片的文艺复兴里。所以不只有巨大的销量增长,服务器的每一个环节都需要巨大的创新。

比如内存,过去是纯粹的商品,而现在的高带宽内存是十颗芯片叠在一起,输入输出是过去的十倍,三星花了好几年才做出来,而且这是极其关键的一环,还在不停迭代升级,所以他们必须提前三四代和英伟达一起工作。

我们在 Celestica 上就经历了这样一段。Celestica 是代工厂,这个行业从 1999 年之后就是场灾难,全部转移到了中国,彻底商品化。但他们撑住了,Celestica 的血统是 IBM 的超级计算业务,他们把那些人才和技艺留了下来。后来我们注意到,他们是谷歌 TPU 服务器的独家供应商。我们当时就傻了,那大概是三年前,股票交易在八倍市盈率。他们同时还有一整块生意,是把以太网白盒交换机卖给云厂商,白盒就是商品化的代称。

结果发现,这些其实是极好的生意。不只是增长惊人,而是要做一台 AI 服务器,它是液冷的,运行温度高得多,一台机器值二三十万美元,而过去一台服务器只要五千美元。旧服务器坏了扔掉就是,这台东西一旦坏了,整套系统全部停摆。所以你就变成了关键基础设施,像给飞机供应关键部件一样,永远不会被替换掉。而且他们在液冷上做得相当好,很多人试过,失败了,所以他们守住了这个位置。

以太网市场也一样。过去从一百G 到四百G 再到八百G,是七年一个升级周期,现在每年升级一次,这非常难做。上面还有一整层软件,就是开源的 SONiC。Celestica 的一些工程师正是这套开源软件的作者,他们和博通合作得非常紧密。所以我们原以为只是一个很好的增长驱动,结果发现那其实是很强的竞争优势,他们在云以太网交换机市场拿到了 50% 到 60% 的份额,而这个市场对 AI 至关重要,因为 AI 极度依赖网络。

再比如印刷电路板。普通服务器需要十层,这些 AI 服务器需要四十层,能做出来的 PCB 供应商非常少,里面还有各种各样的复杂工艺。我们也持有台光电子,他们做的是最关键的原材料铜箔基板,用在这些板子上。PCB 的出货量在涨,层数在增加,单是出货量的复合增速就有 50% 到 60%,同时单价在涨,毛利在涨。而可见度呢,过去是「下周需要你我们再打电话」,现在是「未来四年我们都需要你,一起来设计路线图」。所以一家过去 5% 增长的低利润公司,变成了未来四年营收复合增速 35% 到 50% 且利润率上行的公司。在这之上,什么都缺货。所以哪怕它确实是商品,这也会是一轮很棒的周期。

这种情形在供应链上下游到处都是。比如康宁,他们做光纤,市场份额高得离谱。我读到微软刚建成的一座数据中心,光是那一处,里面的光纤长度足够绕地球四圈半。他们的光纤更细、更能弯折,而且可以按精确规格定制,利润率更高,是他们增长最快的业务。

网络这块还要分层。有横向扩展,就是把所有服务器机架连起来;有跨域扩展,就是把数据中心之间连起来。当你要建这种巨型集群,而一个地方拿不到足够的电力去做训练时,你就得把它们用线连起来。可这些线要粗得多,用量是十倍级别的差异,这又制造了巨大的增长。真正的重头戏是纵向扩展,就是把机架里的每一颗 GPU 都互连起来,现在走的是铜,最终会走光纤。那件事一旦发生,康宁的机会会翻两到三倍。所以在机架的每一层上都有故事。

主持人: 让人应接不暇。

萨瑟多特: 确实应接不暇。再比如电源,英伟达的每一代芯片或机架,用电量要多出 50% 到 125%,这直接推高了台达和先进能源的产品单价。我听到这些故事的时候真的难以置信:等一下,你是说你们的单价接下来连续四年每年上涨 40%,而且利润率更高?

更大的图景是,如果我们对这条 L 曲线的判断没错,AI 的需求会一直往上。而以现在的状况,DRAM、NAND、PCB,我们已经短缺了 30%。

绝对值还是变化率

主持人: AI 收入占比和市场份额这两个指标,你更看重绝对值还是变化率?

萨瑟多特: 这是个好问题。2024 年我们做过一份材料,把所有人的份额和占比都列了出来,我让 Claude 把它画成图,结果它没做对,因为它没抓住变化率。变化率非常重要,而且这件事本身很惊人:你从 10% 走到 30%,增长率会加速,利润率也会加速。所以变化率极其重要。

为什么别人复制不了

主持人: 那为什么在公开市场上,把这件事做对的人不多?如果你整套框架就是 S 曲线、竞争优势、被低估的盈利能力,过去二十五到三十年这部电影已经放过很多遍了。

萨瑟多特: 我妈也问,你为什么要把自己的秘密告诉所有人。这就好比赌场为什么要教人怎么玩二十一点。因为难,真的很难做到。你必须有足够深的功底,必须能安心地投这个领域。我在 Whale Rock 做科技已经二十年,我们的团队一直在做这件事,经历过很多轮周期,我们知道其中的差别。

而且,几乎没有人真正关注过硬件和芯片,所以这个领域涌进来的全是新手。

主持人: 你和加文,就这两家。加文做得也很好。

萨瑟多特: 大家对这块不适应,而且它比看上去难。这些公司的走势图都在往上走,所以很吓人,你会犹豫还能不能买。另外你必须有整体视角,否则你没有确信度。英伟达过去四年,每一次都是这样:哦,他们有了很棒的一年,天啊这肯定是泡沫;然后又是很棒的一年,横盘六个月,这肯定是泡沫,太失控了,太吓人了。而且那些看空的理由并不是毫无道理。但如果你能看到整幅画面,理解这些事情是怎么展开的,并由此建立确信,你就守得住。坦白讲,如果你只是个半导体分析师,很多人正是因为看不到基础模型层真正发生了什么而错过了这一轮。所以拥有大局观很有帮助,手里握着几十条跨越几十年的 S 曲线,知道它们在不同情境里如何演绎,也很有帮助。

什么让他担心

主持人: 我会把你这一小时的立场描述为对 AI 的影响和由此可得的回报都极度看多。那在这整幅画面里,什么让你最担心、最不确定?还是说只是这一切变化的速度?在如此强烈的乐观之中,什么让你放不下心?

萨瑟多特: 有一件事让我不舒服,就是普通民众里对 AI 的负面情绪很多,政府某些方面对 AI 的负面情绪也很多。缅因州好像刚刚禁止了数据中心,只有 20% 的人对 AI 持乐观态度,还有出现负面监管的可能。不过我确实认为,精灵已经出瓶了。

另一个风险是 AI 的能力提升放缓。我认为即使模型不再进步,仍然有大量的采用有待发生。但黄仁勋很多年前讲显卡的时候说过一句话:如果「够用就行」真的够用了,我这门生意就没了。他每年都把图形做好一点点,而人们永远想要最好的。在 AI 里,如果 Anthropic 或者 OpenAI 撞上一堵墙,不再进步,开源模型就会追上来,那可能变成一场向下的竞赛,对股票大概不是好事。但对芯片公司可能是好事,芯片公司不在乎谁赢。

主持人: 不在乎谁赢下 token,对吧。

萨瑟多特: 不在乎谁赢。所以这算是另一个正面因素,如果开源起来了,他们照样受益。黄仁勋是真心希望开源能起飞,上一次 GTC 上他反复提的就是这件事。

还有一个风险是,如果一两个玩家掉队,守不住位置、竞争不下去,那就意味着未来有一大批算力用不上了。当然,如果 AI 足够大,别人会把它吸收掉。我们已经见过一次:甲骨文有个大单被取消,Meta 直接顶了上去。但假设 Meta 决定不再参与,说我们追不上了,这只是浪费资源。这种情况我们会非常小心地盯着。总体上,我们看到的是越来越多公司真心实意扑进来,连微软都在试着自己造。这些就是我认为的几个主要风险。

应用层为何还要等

主持人: 看起来你在 AI 的应用层几乎没做什么。历史上,市值最终大部分会落在应用上,而不是基础设施,而且过去也不存在模型层这一层,硬要说的话可能就是云。为什么把重心放在黄仁勋那个五层蛋糕的底部几层,而不是放在真正被消费者使用的应用层?

萨瑟多特: 我们也有一些。OpenAI 的一部分价值就是 ChatGPT,那是一个应用。但关于应用层,第一,它总是来得更晚。iPhone 的头三四年,应用真正起来是需要时间的。所以也许现在才刚刚开始。

第二,到今天为止,我们觉得这块相当有风险,因为基础模型的边界到哪里结束、应用从哪里开始,并不清楚;这些应用能不能建起足够的护城河去抵御,并把生意做成,也不清楚。我们原以为会在一些在位者身上看到,比如 CRM,他们确实开始做了,也许只是时间问题,但在企业世界里我们真的还没看到。市面上有一些非常好的应用型创业公司,可整个生态并不清晰。我们做芯片的时候,生态是清晰的;我们进基础模型的时候,生态不清晰,现在对我们来说清晰多了;而应用层,眼下还是不清晰,甚至有点危险。

但一定会诞生伟大的应用公司。我们一直在关注布雷特·泰勒和他的 Sierra。布雷特当过 Salesforce 的 CEO,写过谷歌地图,做过脸书的技术负责人,现在在建这家很出色的公司叫 Sierra,我们没有参与。真正的考验就在那里:他能不能把它变成一家巨大的公司。目前他做得相当不错,再看吧。

这类东西什么时候真正成气候、什么时候证明自己可持续,是个时机问题。通常不会发生在头三四年,而是稍微再往后一点。

研究机器与那面奖墙

主持人: 你办公室里有一面巨大的奖墙,颁给年度最佳研究项目的分析师,具体名称我记不清了,我记得你自己也拿过,一个人干活的时候自己发给自己。总之现在已经积累了二十年历史,每年有一位或几位分析师因为当年最出色的研究项目把名字留在墙上。我很好奇这类研究的性质,以及它在这一切之下正在如何改变。比如今年会拿奖的那个人,他做的那种需要人来完成的工作是什么?毕竟 2009 年那种能让你拿奖的活儿,今天可能用 Claude Code 一小时就自动做完了。研究的性质,以及什么样的工作才配上那面 Whale Rock 奖墙,正在实时发生什么变化?

萨瑟多特: 我很想说我们的 AI 体系已经先进到带来了翻天覆地的变化,但到目前为止并没有。它帮我们更快地进入状态,我们也有几个很好用的应用,但它还没有取代分析师的工作。

我们做的事情里,很大一部分是尽人力所能地去见更多公司,和我们覆盖的管理团队建立关系,去跟竞争对手聊。我们用的这套方法直接来自菲利普·费雪 1950 年代写的《怎样选择成长股》,就是「闲聊法」,是成长股投资的路子:走出去,跟供应商、客户、竞争对手聊,寻找那些领先公司身上的关键特征,从而真正建立起确信。

现在如果碰上一个新的复杂领域,比如 ABF 载板或者 PCB,我们能很快把功课补上。但它没法以任何方式替你选股。

我的看法是,如果你是一个基本功扎实的分析师,这个角色仍然有价值,但你必须在上面加上洞见。我们现在用 AI 写纪要、复盘季报,写出来的东西比以前好得多。可最上面那一段必须写得非常好,那一段是智慧:这件事意味着什么?它跟我们的投资论点怎么对上?什么变了?不要只做一个记者。AI 可以做一个很好的记者,但它还不太能看进未来。

就像两年前几位同事在 AppLovin 上做的那件事。我认为我们有两位最好的广告技术分析师,是他们说服我买的。我懂广告技术,做完投行之后我就在纽约附近的一家互联网广告创业公司起步,所以我了解互联网广告和广告技术,而这个行业历史上是个糟糕的行业。但迈克尔和山姆比所有人都更早把 AppLovin 的故事想明白了,公司还没上市的时候他们就在跟。他们了解所有竞争对手,了解所有门道,这里面有大量术语。山姆跑去拉斯维加斯的应用广告大会,我们也去了展会,跟一批又一批的人聊。他们还把模型做扎实了,并且和亚当·福鲁吉建立了很好的关系,那是最出色的管理者之一。这样的活儿,我看不出 AI 能做。

和同行交换想法

主持人: 和公司之外的其他投资人交流,在你的生活里扮演什么角色?

萨瑟多特: 最好的收获之一,就是和这么多聪明的投资人建立起来的友谊。坦白说,菲利普·费雪也讲过,他的流程之一就是在全国范围内结识十到十五位志同道合的人,彼此分享想法。他们都是很好的朋友,其中不少人上过你的播客。你会建立起真正的友谊,交换和讨论想法,重要的是这必须是双向的。

我把这个叫做「三脚架」:当我喜欢一个东西,我的分析师也喜欢它,然后一位我非常尊重的人同样喜欢它,这三条腿撑起来,确信度会大大增强。

产品线是怎么长出来的

主持人: 关于给投资人提供什么样的产品,这些年你学到了什么?现在已经不是单一结构了,如果我是投资人想把钱交给你,有好几种方式。你是怎么走到这一步的?这段经验能给那些想为自己的出资人提供合适选项的投资人什么建议?

萨瑟多特: 头十五年我们只有一只多空基金。你需要专注,一旦分散注意力就会很难。我们把它做起来,做到了我们想要的规模。

公司成立二十年,大概到第十年的时候,开始有人要求做纯多头产品,于是 2020 年我们推出了纯多头基金,到现在六年,规模已经超过了多空基金。绝大部分资产在这两只上。

大约 2015 年,我们正式确立了可能会投一级市场,于是给投资人选择权,可以选择加入或退出,比例可以是 15% 或 25%。但真正破冰要到 2020 年。2021 年我们推出了一只混合基金,最多可以有 80% 投向一级市场,思路类似,只是给想要更多一级暴露的人。

最近我们推出了 Whale Rock 大市值科技基金。我们认为,市场对全世界最大的那批科技公司存在巨大的结构性低配。原因之一是,我们自己回头看,这些年很大一部分业绩正是来自最大的那几家公司,无论是苹果、亚马逊还是特斯拉,而人们很难把它们超配到应有的比重。

我们很多最大的资金池,比如捐赠基金,他们意识到自己在过去这些年里对全球最大的科技公司严重低配。因为他们有大量一级投资,公开市场部分本来就不多,公开市场里可能又有一半是国际的;而在公开市场这一格里,他们相信大盘股没有超额收益,所以低配大盘,转而配置很多做选股的中小盘管理人,毕竟「大盘不可能有阿尔法」在直觉上很顺。再加上对冲基金那一格,即使是偏多头的,也不会有 15% 的英伟达和其他这些东西。

我们意识到,人们担心这些巨头,但这其实是数字经济的必然产物:在科技里,领先者通常越长越大并且赢下去,很快就拿到很高的市场份额,拥有极强的竞争优势,而且面向全球销售。这会带来巨大的利润池和巨大的市值,而且未来还会继续这样。所以大多数捐赠基金实际上是在跟这件事对赌,因为他们完全低配。

后来有人来问我们该怎么办、该用哪个指数。我在汉密尔顿学院的董事会里,他们的投资委员会也在琢磨同一件事。这类声音听得多了,最后有一位客户提出来,我们就说,那我们给你做一只,因为这里面有大量阿尔法可挖。所谓「七巨头」也好,「FANG」也好,未来会变成别的组合,而且 2022 年它们一起涨,去年分化极大,今年又在跌。

于是我们做了 Whale Rock 大市值科技基金,投资范围是全球市值前三十的公司,我们从中挑出最好的十二到十三家。我认为最大市值这块的阿尔法非常可观。你想想,一只小盘股,只要一个人看明白它好,就能把它推上去;可要让市场承认谷歌不是输家而是赢家,需要一百个人,需要一百位持仓分散的组合经理都想通。那我们能不能比其中 95% 的通才组合经理更早想明白?我们过去是做到了。

主持人: 这种赔率我们愿意押你。

萨瑟多特: 所以那里确实有阿尔法可拿。而且作为一类资产,这也非常好,因为这些公司按定义就拥有极好的护城河,也许它们不在超级 S 曲线上,但有时候恰恰就在。英伟达当然在,台积电对它有极高的杠杆,海力士的杠杆更是极端,ASML 也有杠杆。这只基金我们才做了四个月。

主持人: 那也许可以这样理解:你真正建起来的是一台研究机器,用公司这个透镜去理解世界,你不断想改进的就是这台机器,而产品只是把它的输出乘以不同的方式表达出来。如果我要理解 Whale Rock,首先应该去研究这台机器。

萨瑟多特: 我们管它叫 Whale Rock 学习机器,是一个十人的资深团队。沃伦·巴菲特读书,我们也读书、读博客,但在科技行业你还必须走出去跟人谈。所以我们每年做两千五百到三千场与管理层的面对面会议。芒格和巴菲特讲知识的复利,我们已经复利了二十年。团队当然有人员变化,但整体上非常稳定。安德鲁和迈克尔跟了我十九年和十八年,团队的平均从业年限在十年上下,这还包含了一些较新的成员。

这台研究引擎可以支撑所有这些产品,公开市场和一级市场是同一批人在做。我们不会去把全世界的石头都翻一遍,但只要看到符合我们体系的东西,我们就有能力出手。

父亲

主持人: 和你聊天太愉快了。每次结尾我都会问同一个传统问题:别人为你做过的最善良的一件事是什么?

萨瑟多特: 那一定是我父亲。我实在太幸运了。我父亲在康奈尔读的是电气工程双学位,后来转到华尔街,在高盛有一段很棒的职业生涯。八十年代他主管公司金融,九十年代做私募股权部门的主席。他极其聪明,但为人非常谦逊,是个真正的绅士。

我创办 Whale Rock 的时候,找亲友募资,他是我打的第一个电话。他说,我在高盛待了四十一年,不如我来加入你吧。我来做那个头发花白的人,做监督,做董事长。你做你擅长的,你在波士顿把公司搭起来,把团队带起来,把钱管好,我帮你募一些钱。

我们一起工作了六年,直到 2011 年他去世。能和他共事,我觉得自己非常幸运。管一只基金并不容易,可我们从来没有对彼此提高过嗓门。他也是很多人的好导师。他去世后,我收到许多来信,人们说,你父亲对我影响太大了,他是位真正的绅士,是我极好的引路人。所以我真的觉得自己很幸运。如果我能成为他那样的人的一半,我就算彻底赢了。

主持人: 他是怎么做到的?他的方法是什么?为什么那么多人这样说?

萨瑟多特: 我说不好。他谦逊,极其聪明,有智慧。他在投资上也很有名,而这在很多投资银行里并不常见。他还在承诺委员会任职,让公司避开了不少棘手的处境。他非常温暖,人们可以带着问题走进他的办公室,无论是个人的麻烦还是工作上的事,他都处理得很有分寸。他有一种柔和的方式,而且很有幽默感。

主持人: 真幸运。

萨瑟多特: 是的,我太幸运了。

主持人: 阿历克斯,非常感谢你的时间。

萨瑟多特: 谢谢你。

排版 + 横图 + 来源,粘贴即成稿
章节 · 点击跳转视频
0:00 开场:不到 1% 渗透率与「L 曲线」 ▶ 正在看
0:53 从 ChatGPT 发令枪到押注 Anthropic ▶ 正在看
4:26 代码:AI 真正的解锁点 ▶ 正在看
8:45 模型层的差异化、脚手架与渗透率 ▶ 正在看
14:09 公开市场基金如何挤进一级市场 ▶ 正在看
19:26 S 曲线框架:拐点、高度与斜率 ▶ 正在看
26:14 何时买入:直觉、现场线索与模式识别 ▶ 正在看
32:24 竞争优势的六种形态与反面清单 ▶ 正在看
39:54 为什么几乎清仓了应用软件 ▶ 正在看
48:07 数据中心硬件的「去商品化」 ▶ 正在看
58:11 框架为何难复制,以及风险清单 ▶ 正在看
65:09 研究机器、产品线与父亲 ▶ 正在看
本期小问 · 档案清单
19:26 技术采用的 S 曲线,真能预测几年后的赢家吗? ▶ 正在看
8:45 这轮 AI 的价值,最终会沉淀在哪一层? ▶ 正在看
26:14 好的判断从哪里来? ▶ 正在看
65:09 研究是怎样做成的? ▶ 正在看
本期讲者
亚历克斯·萨塞尔多特Whale Rock Capital Management 创始人、首席投资官,2006 年在波士顿创立该科技主题基金,产品线覆盖多空、纯多头、一级市场混合基金与大型科技股基金。职业生涯始于富达(Fidelity),研究互联网与科技行业逾二十年。
Patrick O'Shaughnessy播客《Invest Like the Best》主持人,Positive Sum 创始人、原 O'Shaughnessy Asset Management 首席执行官。该节目以长时段深访投资人与创业者著称。
彼得·萨塞尔多特讲者的父亲,康奈尔大学电气工程毕业后转入华尔街,在高盛任职 41 年,1980 年代主管公司金融、1990 年代任私募股权部门董事长,后出任 Whale Rock 董事长,2011 年去世。
01开场:不到 1% 渗透率与「L 曲线」
0:00
When you get the right part of the S-curve, you get exponential unit growth. If you have a very strong business model, your earnings don't grow linearly, they grow exponentially. You know, the world doesn't think exponentially. Very few people believe you can accurately predict 2 3 4 years out. But if you follow and understand the Scurve and you you know the moes and you know how to model, you really can uh predict these these great things. the enterprise AI or enterprise application AI market is less than 1% penetrated and we've never seen, you know, we talk about S-curves, we call this an L curve, just [music] straight up.
当你处在S曲线的正确位置上,你就能获得指数级的单位增长。如果你有一个非常强的商业模式,你的盈利不会是线性增长,而是指数增长。你知道,这个世界不会用指数思维思考。很少有人相信你能准确预测两三年、四年以后的事。但如果你跟踪并理解S曲线,你了解护城河,你懂得如何建模,你真的能够预测这些了不起的事情。企业级AI,或者说企业应用AI市场的渗透率还不到1%,而我们从来没见过,你知道,我们常说S曲线,但这个我们叫它L曲线,就是一路笔直向上。
便签笔记
02从 ChatGPT 发令枪到押注 Anthropic
0:53
Alex, you were saying that your highest conviction position is anthropic right now. Can you tell the story of discovering it, making the investment, using this anecdote as an excuse to talk about all the things that I think you and I are mutually interested right now, investors like you, investing in private markets, anthropic, the business, AI, everything. It's a great great way to zoom in. Why is it your highest conviction? And how did you get started? >> Yeah. Well, when when the gun went off with OpenAI chat GPT in November 2022, we immediately took the firm and did a massive deep dive with our 10 person team. And we anytime you have a new compute paradigm, there's a new stack and on the and and that creates new winners and losers on the old stack. And in this stack, you know, it's now Jensen talks a lot about it, but it's power at the bottom, chips at the bottom, the clouds, and then the foundational models, and then the applications on top. And at [snorts] that time, this was 2023 early, we said, we want to be in
Alex,你刚才说你目前信念最强的持仓是Anthropic。你能讲讲发现它、做出这笔投资的故事吗,就借这个例子聊聊我觉得你我现在都很感兴趣的那些话题,像你这样的投资人、投资私募市场、Anthropic这家公司、AI,所有这些。这是一个非常好的切入点。为什么它是你信念最强的持仓?你又是怎么开始的?>> 是的。嗯,2022年11月,当OpenAI的ChatGPT打响发令枪的时候,我们立刻带着公司10个人的团队做了一次大规模的深度研究。而且每当出现一种新的计算范式,就会有一个新的技术栈,这会在旧的技术栈上造就新的赢家和输家。而在这个技术栈里,你知道,现在黄仁勋经常谈这个,最底层是电力,然后是芯片,再往上是云,然后是基础模型,最上面是应用。而在那个时候,也就是2023年初,我们说,我们想先投芯片和基础设施。
便签笔记
1:55
the chips and the infrastructure first. And not only do they get the uh demand first, but we know who the winners are. And no matter who wins above, which we weren't sure at the time, we know we're going to need tremendous amounts of compute. And we did a deep dive into that, which we can talk about later, but over the next 2 or 3 years, we started to get more clarity on how the foundational model, the layer would evolve. And at the time, two or three years ago, there were 60 different companies going after it. OpenAI was kind of in the lead. And we did a webinar in April 2023. We said, look, this might be a winner take all. It might be a total commodity because there's open-source players. It might be a race to zero or it might be an oligopoly where there's three or four leading players. And what we saw over the following, you know, 3 years was that almost all the startups fell away and died. And then some of the largest companies in the world including Amazon and and Meta. Amazon really never
它们不仅最先拿到需求,而且我们知道赢家是谁。而且不管上面谁赢——当时我们并不确定——我们知道都会需要海量的算力。我们对此做了一次深度研究,这个我们待会儿可以聊,不过在接下来的两三年里,我们开始越来越看清基础模型这一层会怎么演化。而在当时,也就是两三年前,有60家不同的公司在争这块市场。OpenAI算是领跑的。我们在2023年4月做了一场网络研讨会。我们说,你看,这可能是赢家通吃。也可能彻底商品化,因为有开源玩家。可能会变成一场向零竞赛,也可能形成寡头格局,有三四个领先玩家。而接下来这三年我们看到的是,几乎所有创业公司都掉队、消亡了。然后是世界上一些最大的公司,包括亚马逊和Meta。亚马逊其实从来没有
便签笔记
3:05
really showed up. We'll see what happens with Meta, but they were they came in strong and then basically their effort faltered and they had to do a total reboot. In the meantime, Anthropic kind of was this dark horse candidate, the startup and um they focused uh really purely on the enterprise and OpenAI had kind of won the consumer and then Gemini can never be counted out. We we love Google as well. It's one of our largest positions. So it really started to look like a three-horse race and somewhat of an oligopoly very similar to how the uh cloud market evolved where three companies underpin the entire SAS cloud world and and have really excellent businesses and then we also were aware of the open- source risk um from China and we started to get comfortable that the quality of the tokens from the leading edge were superior because if you're 80% close to the top of the benchmarks going from 80 to 85 is a huge unlock and the um the open- source guys they don't have as much compute so they can come close
真正登场。Meta会怎样我们再看,但他们来势很猛,然后基本上他们的努力受挫了,不得不彻底重启。与此同时,Anthropic算是那匹黑马,那家创业公司,而且他们非常纯粹地专注于企业市场,而OpenAI算是赢下了消费端,然后Gemini永远不能被排除在外。我们同样很看好谷歌。它是我们最大的持仓之一。所以局面真的开始看起来像是一场三马赛跑,有点寡头格局,非常类似于云市场当年的演化路径——三家公司支撑起整个SaaS云世界,而且都有非常出色的生意,然后我们当时也意识到了开源的风险嗯,来自中国的(模型),我们开始逐渐确信最前沿模型产出的 token 质量更优,因为如果你处在 80%接近榜单顶部的位置,从 80 分提升到 85 分就是一个巨大的突破,而开源阵营的那些玩家并没有那么多算力,所以他们可以逼近最前沿,但没法实现反超,然后就有点后劲不足。与此同时,Scaling Law 以及
便签笔记
03代码:AI 真正的解锁点
4:26
to the leading edge but they can't leapfrog it and then they kind of falter. Meanwhile, the scaling laws and other means of improving the models, the feedback loops, etc. Uh we saw that there was a very strong runway and everyone we talked to close to the industry saw that the scaling laws would continue. So we developed this thesis that it would be a three-horse race. And then the big kicker was code. And this is the true unlock of AI. In the first few years, we knew AI would would be big, but we were skeptical. Also, we made large investments because we knew the training was would be there, but we weren't sure how much revenue might come and if it could truly replace labor because if you remember the early versions of the models were good, but it there was a lot of uh some negative feedback from corporates and could they be truly agentic? We realized in 2025, the first cloud code and and the coding tools really began to explode and you saw the first gen was like Microsoft C-Pilot which is like $20 a month and then and
其他改进模型的手段、反馈闭环等等。呃,我们看到这里还有非常长的跑道,而且我们接触到的所有业内人士都认为 Scaling Law 会继续成立。于是我们形成了这样一个判断:这会是一场三强争霸。接下来最关键的推动力就是代码。这才是 AI 真正的突破口。在最初的几年里,我们知道 AI 会很大,但我们是有怀疑的。同时我们也做了大额投资,因为我们知道训练这块肯定会有需求,但我们不确定能带来多少收入,也不确定它是否真能替代劳动力,因为你要记得,早期版本的模型是不错,但企业客户那边有不少负面反馈,它们究竟能不能真正具备 Agent 能力?到了 2025 年我们才意识到,最早的 Claude Code 和那些编程工具真的开始爆发了,你看第一代产品大概是微软的 Copilot,一个月 20 美元左右,然后它开始起步,大致能
便签笔记
5:42
then it started and that could sort of improve your grammar of coding, maybe find a bug, maybe make a block of code like a paragraph and then Anthropic came out sometime in in the middle of the year and it could do so much more. Um, and it started to get to this point where it could run agentically and we kind of saw that happening and the coding market just exploded and then we started hearing that people who could use it unfettered. We heard we heard that you know even within Anthropic at that time people were spending $100 a day on tokens which if you do the math comes out to 20 or $30,000 a year. And if you think about how many coders there are in the world, 20 million, you've got a half a trillion dollar market just from coding alone. And mind you, that was on 7 8 9 month old technology. We could see just on the coding market alone that Anthropic had a tremendous opportunity ahead of it. So I think at the time, this is pretty funny, we wrote in our letter, you know, we made the investment um at the 180 valuation. And
改进你写代码的“语法”,也许帮你找个 bug,也许写出一段像一个自然段那样的代码块,然后 Anthropic 在年中某个时候推出了产品,它能做的事情多得多。嗯,而且它开始发展到可以以 Agent 方式自主运行的程度,我们也算是看着这一切发生的,编程市场就此爆发,然后我们开始听说,那些能不受限制使用它的人。我们听说,就连当时 Anthropic 内部也有人每天在 token 上花掉 100 美元,你算一下,一年就是两三万美元。而如果你想想全世界有多少程序员,2000 万,那光是编程这一块就有半万亿美元的市场。而且请注意,这还是基于七八九个月前的技术。我们能看到,仅从编程市场这一项来看,Anthropic 面前的机会就极其巨大。所以我记得当时,这事挺有意思的,我们在致投资人的信里写道,我们是在 1800 亿估值那轮投的。我们当时说,我记得他们希望做到 90 亿
便签笔记
6:50
we said, and I think they were hoping to get to a nine billion >> one to nine. Yeah. And and then the numbers were like nothing we'd ever seen before. 100 to a billion on the way to 9. But when we did it in August of 2025, we nobody had any idea what 2026 could be. the the the second big unlock lately which is that you know claude code has gone to almost completely agentic um where you had Andre Carpathy and Lionus Torvalds last year saying two of the smartest people in coding and they completely flip-fpped and Karpathy said you know last year's code tools could write 20% and 80% would be handwritten that flipped when the the latest model came out and now he hasn't written a line of code not except in English and not to mention the pure unlocked that we're going to get for the people that never knew how to code. So just coding alone has completely taken off. Anthropic has been able to stay ahead in coding. And so one difference between the cloud, GCP, AWS, and the AI companies is the cloud's generally it's commodity.
(对方)从 1 亿到 90 亿。是的。而且那些数字是我们从没见过的。1 亿到 10 亿,然后往90 亿走。但我们在 2025 年 8 月投的时候,没有人知道 2026 年会是什么样子。最近的第二个重大突破,就是 Claude Code 已经变得几乎完全 Agent 化了,去年还有安德烈·卡帕西和林纳斯·托瓦兹——这两位是编程圈里最聪明的人——他们的看法彻底反转了,卡帕西说,去年的代码工具能写 20%,剩下 80% 还得手写,等最新的模型出来之后这个比例就反过来了,而现在他除了用英语之外,一行代码都没写过,更别提那些从来不会写代码的人将获得的巨大解放。所以光是编程这一块就已经彻底腾飞了。Anthropic 也一直在编程领域保持领先。所以云厂商——GCP、AWS——和 AI 公司之间的一个区别是,云本质上大体是商品化的。
便签笔记
8:15
They're they're selling you servers and storage. You know, they have a lot of software on top and there is stickiness to it. But in the AI models, everyone thought it would be pure commodity. But there's tremendous differentiation with within. There's different training methods and different skills that they're good at. And a lot of people have routers that switch in between, which sort of makes it sound like they're commodity, but anthropic, they're very good for anything that has to do with private equity and finance.
他们卖给你的是服务器和存储。当然他们上面有很多软件,也确实有黏性,但在 AI 模型这块,大家原本以为会是纯粹的商品化。可实际上内部存在巨大的差异化。训练方法不同,各自擅长的能力也不同。很多人用路由器在不同模型之间切换,这让人听起来觉得它们是同质化的,但 Anthropic,它在任何和私募股权、金融相关的事情上都非常出色。
便签笔记
04模型层的差异化、脚手架与渗透率
8:45
Google's very good for ingesting PDF. And so there's a lot of like differentiation critical IP, which is a great competitive advantage. and companies many companies have come after the coding franchise and Anthropic has been able to keep ahead. The other thing that's good about the foundational models and anthropic is it's not just the API or the model. They're building a whole monopoly or whole ecosystem of products around the API. So we've got the SDK claude for co-work uh orchestration layer and and all the tools and and they call it sort of a harness which is the software around the API that gets the most out of the model. This was one of the things we saw with AWS really early on in 2013 was oh people thought it was a commodity server up in a warehouse big deal and what they they saw this was a new way of do doing computing. So they had they invented all these products that they could see before everybody else that slowly built lock in. The other way we think about this is where are we on this
Google 在读取和处理 PDF 方面非常强。所以这里存在大量差异化的关键知识产权,这是很强的竞争优势。而且很多公司都来抢这块编程业务,Anthropic 一直能保持领先。基础模型公司和 Anthropic 另一个好的地方在于,它不只是 API 或模型本身。他们在 API 周围构建的是一整套垄断性的、完整的产品生态。所以我们有了 SDK、Claude forCowork,呃,编排层,以及各种工具,他们把这套东西叫做 harness(脚手架),也就是围绕 API 的软件层,用来把模型的能力发挥到极致。这也是我们 2013 年很早就在 AWS 身上看到的一点:当时大家以为那不过是仓库里的一堆商品化服务器,有什么了不起,可他们看到的是一种全新的计算方式。所以他们比所有人都更早地发明出这一系列产品,慢慢建立起了锁定效应。我们思考这件事的另一个角度是:我们现在处在 S 曲线的什么位置?我们认为基础设施
便签笔记
9:58
scurve and we have this infrastructure layer scurve which [snorts] we think is somewhat like 10% penetrated. And by the way we think it's still uh one of the best ways to play AI and we'll talk about how that feeds back through. Um but if you think about it, um even though you know 200 or I don't know how many 800 million people are using AI, they're just using AI 1.0 which is like a a search engine on steroids. But now with these new primitives where you have claw on your computer linking it in, then you build skills. Companies are going to build people and companies are going to start building skills and then they're going to build true AI bots and then big corporations are going to build much larger but where are we in terms of the amount of people doing that? I mean Sunder said it's 10 bips of the uh knowledge workers the world. So Anthropic has something like 14 or 15 million DAUs. Probably a small portion of those are truly doing AI the way you can do it. So that 10 bips, it's classic
这一层的 S 曲线,[吸鼻子] 大概只渗透了 10% 左右。顺便说一句,我们认为这仍然是押注 AI 最好的方式之一,我们待会儿会讲这如何反馈回来。嗯,但你想想看,虽然现在有两亿,或者我也说不准是八亿人在用 AI,他们用的其实只是 AI 1.0,相当于一个加强版的搜索引擎。但现在有了这些新的基础组件,你可以把 Claude 装在自己电脑上、接进各种系统,然后你就能构建 skills。公司会开始构建——人和公司都会开始构建 skills,然后他们会构建真正的 AI 机器人,接着大公司会构建规模大得多的东西。但就做这些事的人数而言,我们又处在什么阶段?我是说,桑达尔说过这只占全球知识工作者的 10 个基点。Anthropic 现在大概有一千四五百万日活用户。其中真正在用 AI 发挥出全部潜力的可能只占很小一部分。所以那 10 个基点,就是典型的
便签笔记
11:05
Scurve where these are the tinkerers and then [snorts] it's going to go to the early adopters, then it's going to go to the early mainstream. But you're going to go from 10 bips to one to two or 3% to 5% to 15% in the next four years. And kind of a light switch this year went off in the enterprise where everybody realizes they need to do this now and do it fast. It's still >> like internet 1.0 I know when it's like you knew you needed a website in 1998 but it's like hard to build that website but this is coming together fast and so you know we think the I don't the enterprise AI or enterprise application AI market is is like less than 1% penetrated and we've never seen you know we talk about S-curves we call this an L curve just straight up and then we'll take this to the infrastructure which is even we're at 10 basis points of people really using AI and we're already sold out of all the there's not enough compute in the world. So Anthropic has half of what they need right now and that's before this huge
S 曲线,现在这批人是折腾派、极客,[吸鼻子] 接下来会扩散到早期采用者,然后是早期主流人群。未来四年里,你会从 10 个基点走到 1%、2%、3%,再到 5%、15%。而且今年企业端就像开关被打开了一样,所有人都意识到必须现在就做,而且要快。现在还是——(对方)就像互联网 1.0,我知道那种感觉,1998 年你知道自己必须有个网站,但要真做出那个网站很难,但这一次的东西成型得非常快,所以我们认为,企业 AI 或者说企业应用层的AI 市场渗透率还不到 1%。我们从没见过这种情况,我们平常谈 S 曲线,而这个我们叫 L曲线——直上直下地往上走。然后我们再说回基础设施,就连在只有 10 个基点的人真正在用 AI 的情况下,我们就已经卖光了所有产能,全世界的算力根本不够。Anthropic 现在拿到的只有他们所需的一半,而这还是在这一轮巨大
便签笔记
12:17
takeup. So Mark Andre said in the next four years one thing he's sure of is there's not going to be enough compute. >> Most software companies try to maximize your time on their app to juice engagement. Ramp does the exact opposite. RAMP understands that [music] no one wants to spend hours filing expense reports, reviewing expense reports, and checking for policy violations. So, they built their tools to give that time back using AI to automate 85% of expense reviews [music] with 99% accuracy. And since RAMP saves companies 5%, it's no wonder that Shopify runs on RAM, Stripe runs on RAM, and my [music] business does too. To see what happens when you eliminate the busy work, check out ramp.com/invest.
需求爆发之前。所以马克·安德森说,未来四年他唯一确定的一件事就是:算力肯定不够用。(广告)大多数软件公司都想最大化你在它们 App 上的停留时间来拉高互动数据。Ramp 恰恰相反。Ramp 明白 [音乐]没有人愿意花好几个小时提交报销单、审报销单、检查有没有违反公司政策。所以他们做产品的思路是把这些时间还给你,用 AI 自动完成 85% 的报销审核 [音乐],准确率达到 99%。而且既然 Ramp 能为企业节省 5% 的开支,那么 Shopify 用 Ramp、Stripe 用 Ramp 也就不奇怪了,我自己的 [音乐] 公司也在用。想看看消灭这些琐碎工作之后会发生什么,去 ramp.com/invest 看看。
便签笔记
12:56
OpenAI, Cursor, Enthropic, Perplexity, and Verscell all have something in common. They all use work OS. And here's why. [music] To achieve enterprise adoption at scale, you have to deliver on core capabilities like SSO, skim, arbback, and audit logs. That's where work OS comes in. Instead [music] of spending months building these mission critical capabilities yourself, you can just use work OS APIs to gain all of them on day zero. That's why so many of the top AI teams you hear about already run on work OS. Work OS is the fastest way to become enterprise ready and stay focused on what matters most, your product. Visit works.com to get started.
OpenAI、Cursor、Anthropic、Perplexity 和 Vercel 有一个共同点:它们都在用 WorkOS。原因如下。[音乐] 要实现大规模的企业级采用,你必须交付一些核心能力,比如 SSO、SCIM、RBAC 和审计日志。这正是 WorkOS 的用武之地。[音乐] 与其花好几个月自己去搭这些关键基础能力,你可以直接用 WorkOS 的 API,第零天就全部拥有。这就是为什么你听说过的那么多顶级 AI 团队都已经跑在 WorkOS 上。WorkOS 是让你最快具备企业级能力、并且能继续专注于最重要的事情——你的产品——的方式。访问 workos.com 即可开始。
便签笔记
13:31
Every investor should know about Rogo because Rogo Aai's platform is not just another generic chatbot. Instead, it was designed to support how Wall Street bankers and investors actually work. From sourcing, diligence, and modeling to turning analysis into deliverables. For me, three key things differentiate Robo. First, it connects directly to your systems so it can work with your actual data. Second, it understands your workflows, how work really happens across a deal or an investment. And third, it runs end to end and produces real outputs the [music] way the best people do. Auditable spreadsheets, investment memos, diligence materials, and slide decks that match your standards. This all comes [music] from the fact that Rogo is built by finance professionals for finance professionals.
每一位投资者都该了解 Rogo,因为 Rogo AI 的平台不是又一个通用聊天机器人。相反,它是围绕华尔街银行家和投资人真实的工作方式设计的。从项目寻源、尽职调查、建模,一直到把分析变成可交付成果。对我来说,Rogo 有三点关键差异。第一,它直接连接你的系统,因此能处理你真实的数据。第二,它理解你的工作流,理解一笔交易或一项投资中工作到底是怎么推进的。第三,它能端到端地跑完,并像最优秀的人那样产出真正的交付物 [音乐]:可审计的表格模型、投资备忘录、尽调材料,以及符合你们标准的演示文稿。这一切都 [音乐] 源于 Rogo 是由金融专业人士为金融专业人士打造的。
便签笔记
05公开市场基金如何挤进一级市场
14:09
And it's already being adopted by some of the most [music] demanding institutions in the world. To learn more, visit rogo.ai/invest. I'm so curious when an investor like you who historically was a public markets investor, you could hit buy and buy whatever you want, is now operating in lots of the most important private market companies. We can talk about Stripe or Data Bricks or OpenAI or Anthropic. How do you get the positions at the size that you want coming from the legacy of been being able to just buy? How much of it is um creativity just directly with the company? If it is directly with the company, so they have it's a double opt-in. they have to decide to let you in too. How do you do that? Like what what have you learned about getting the allocation you want or the amount of equity you want in a private company given that you know that wasn't your original background?
而且它已经被全球一些最 [音乐] 挑剔的机构采用。想了解更多,请访问 rogo.ai/invest。我很好奇,像你这样一位历来做公开市场的投资人,以前想买就买、想买什么买什么,现在却在很多最重要的私募市场公司里持仓。我们可以聊 Stripe、Databricks、OpenAI 或者Anthropic。在你原本那种“随时可以买入”的背景下,你是怎么拿到自己想要的持仓规模的?这里面有多少是靠创意、靠直接跟公司打交道?如果是直接跟公司谈,那这就是双向选择,他们也得决定要不要让你进来。你是怎么做到的?在拿到你想要的配额、或者说在一家私营公司里拿到你想要的股权比例这件事上,你学到了什么,毕竟这并不是你最初的老本行?
便签笔记
14:57
>> In that case, you know, we we got to know the company. One of our analysts knew people in in the finance group there and we actually we had a look at the at the $60 billion round and we we didn't do it. we didn't know the company as well and [snorts] we um the gross margins were negative and and and frankly we hadn't seen coding explode the way it had and and one thing about public markets is you get to know companies over a long period of time and you can kind of invest on your own schedule. I got a chance to spend some time with Daario. I obviously listen to on podcasts and it I started to realize these guys their management team is excellent. the focus, the dedication, they had almost no turnover, the quality of code, and then the business plan was really starting to play out. And uh it's one thing to grow from, you know, 100 to a billion, but it's another to do nine.
(受访者)那一次,我们是先去了解这家公司。我们有一位分析师认识他们财务团队里的人,其实我们看过 600 亿美元那一轮,但我们没投。当时我们对这家公司了解得没那么深,而且 [吸鼻子] 呃,他们的毛利率是负的,坦白讲我们那时也还没看到编程业务像后来那样爆发。公开市场的一个好处是,你可以在很长的时间跨度里慢慢了解一家公司,可以按自己的节奏去投。我有机会和 Dario 相处了一段时间。我当然也听他的播客,我开始意识到,这些人的管理团队非常出色。那种专注、投入,几乎没有人员流失,代码质量,再加上商业计划真的开始兑现了。而且,从 1 亿做到 10 亿是一回事,要做到 90 亿又是另一回事。
便签笔记
15:54
And then so we reached out to the company as much as we could. They took a meeting with us. We did a 90page PowerPoint deck where we used Claude Code to scour the internet for all the feedback we could about the coding market. and their and what their products were good at, where they might need to improve and we also did our whole overview of what the coding market would be. They welcomed us into this round and then we stayed close with the CFO and uh it's it's been great to build a relationship with them and I think we p punched above our weight in terms of the allocation. So that one was a total home run. In the rest of the the world, we are in this period where the unicorn market is bigger than most stock markets in Europe, maybe even combined. It's definitely bigger than Germany. It's definitely bigger than the UK. And we even before we invested in privates, the first one was 2020. We meet with these, we have to know these companies and you really have to know them now because sometimes they're the biggest companies
于是我们尽一切可能去联系这家公司。他们同意跟我们见面。我们做了一份 90 页的 PPT,用 ClaudeCode 把互联网上所有能找到的关于编程市场的反馈都扒了一遍,包括他们的产品强在哪里、哪些地方可能需要改进,我们还完整梳理了对整个编程市场的判断。他们欢迎我们进入这一轮,之后我们和 CFO 保持着密切联系,能和他们建立这样的关系真的很棒,我觉得我们在配额上拿到的份额超出了我们的“体量”。所以那一笔完全是本垒打。至于其他情况,我们正处在这样一个时期:独角兽市场比欧洲大多数国家的股市都大,甚至可能比它们加起来还大。它肯定比德国大,肯定比英国大。而且我们,甚至在我们投私募之前——我们第一笔是 2020 年。我们要见这些公司,必须了解这些公司,而现在你尤其必须了解它们,因为有时候它们就是这个领域里最大的公司,影响力巨大。
便签笔记
16:57
in the space and and have huge impact. So we, you know, we do two to 3,000 face-to-face meetings with management teams a year and about 10 or 15% of those are with privates and then we kind of focus in on the companies that we really want to learn about and find ways to meet with them and uh get involved in their rounds. And our first one was Stripe. And we had a large investment at the time. This is 20 2018 2017 1819 we and 2020 we own Audion which is a fantastic payments company and they're a next-gen cloud payments company taking from world pay and you know the cloud the cloud modern payments was 5% of total you know $80 trillion market or what have you but you can't invest in aud unless you know Stripe like the back of your hand so we did tremendous amounts of due diligence talked to 200 customers in Audon but when we asked about audio and we asked about Stripe and we realized this is Coke and Pepsi and um we said we got to find a way to invest and I finally got to meet the Coulson brothers in 2019 and
所以我们每年会和管理层做两三千场面对面会议,其中大约 10% 到 15% 是私营公司,然后我们会聚焦在那些真正想深入了解的公司上,想办法去见他们,并且参与他们的融资轮。我们的第一笔是 Stripe。当时我们有一笔很大的投资。那是 2017、2018、2019 年,还有 2020 年,我们持有 Adyen,一家非常棒的支付公司,是新一代的云支付公司,从 Worldpay 手里抢份额,你知道,云端的、现代化的支付当时只占整个大约 80 万亿美元市场的 5% 之类,但你没法在不像了解自己的手背一样了解 Stripe 的情况下去投 Adyen,所以我们做了大量的尽职调查,为了 Adyen 访谈了 200 家客户,可每当我们问起 Adyen,我们也会问起 Stripe,然后意识到这就是可口可乐和百事可乐的关系,我们说我们必须想办法投进去。我在 2019 年终于见到了 Collison 兄弟,那就是我们的第一笔。当时我们
便签笔记
18:06
so that was our first one. We weren't really known for privates. I've got a friend um who's who has a involved with a a venture firm that has tremendous amounts and I talked to him about it and I said let me know if you ever want to sell some and then I get a call from him during co in April of 2020. We knew a lot about Stripe. We didn't have the full financials, but we knew enough that at that valuation, I think it was 35 billion. We knew they had they disclosed we had over half a trillion of TPV. And we knew that Audience's take rate was 25 or 30 bips and we knew Stripes was 40 or 50. And we knew how many employees they had. So, we could kind of get at the profitability.
在私募领域并不出名。我有个朋友,他和一家规模非常大的风投机构有关系,我跟他聊过这事,我说如果你哪天想卖一些份额就告诉我,然后 2020 年 4 月疫情期间我接到了他的电话。我们对 Stripe 了解很多。我们没有完整的财务数据,但我们知道的已经足够判断,在那个估值下——我记得是 350 亿美元。我们知道他们披露过 TPV 超过 5000 亿美元。我们也知道 Adyen 的费率是 25 到 30 个基点,而 Stripe 是 40 到 50 个基点。我们还知道他们有多少员工。所以大致能推算出盈利能力。
便签笔记
18:52
It turned out the take rate was higher. It turned out they were being modest about their TPV. It was much higher than the 550. It was closer to the one 1 trillion. And you know, we underwrote the thing under our assumptions and it was much better. And then we were able to upsize that from the seller to a $100 million block. Sometime they like it that you know the VCs are going to own and then most of them are going to sell. They like it that we'll own and own in the public market which we did with new bank uh as well. all owned it for a long period of time in the public market as well.
结果发现他们的费率更高。结果发现他们在 TPV 上说得很保守。实际数字远高于 5500 亿,更接近 1 万亿。万亿。你知道,我们是按照自己的假设来做的承销测算,结果比预想的好得多。然后我们就得以跟卖方把这笔交易加码到 1 亿美元的大宗。有时候他们会喜欢这样,你知道,VC 会持股,然后大部分人到时候都会卖掉。他们喜欢我们会持有,而且是在公开市场里持有,我们在 Nubank 上就是这么做的。也在公开市场长期持有过很久,在公开市场上也持有了很长一段时间。
便签笔记
06S 曲线框架:拐点、高度与斜率
19:26
>> Maybe now's the right time to lay out everything you've ever learned about S-curves. Obviously, your firm is sort of predicated on this idea of technology adoption life cycles >> and investing in companies at the right time amidst a certain platform change or S-curve change. >> And I think everyone knows the basic idea of an S-curve and and the sort of uh the stages you mentioned, tinkerers and and early adopters and early majority. But I'd love you to go into the the super deep detail of what you've learned since this is the lens through which you've viewed markets and stocks for a long time. Bring us into like the nitty-gritty fine grading nuance detail of why S-curves can be so useful for investing. We have an investment framework. It's >> S-curve and we'll dive into each one.
>> 也许现在正是把你关于 S 曲线所学到的一切都摊开来讲的好时机。显然,你们公司的整个基础就建立在技术采用生命周期这个理念上 >> 以及在某个平台变革或者 S 曲线变化当中,在正确的时点去投资那些公司。S 曲线的变化。>> 我想大家都知道 S 曲线的基本概念,还有你提到的那些阶段——修补者、早期采用者、早期大众。但我很想请你讲讲更深层的细节,因为这是你长期以来观察市场和股票的透镜。带我们进入那些最细致入微的层面,讲讲为什么 S 曲线对投资这么有用。我们有一套投资框架。就是>> S 曲线,我们会逐个深入讲。
便签笔记
20:13
Competitive advantage and then underappreciated earnings power. And when you get the right part of the S-curve, you get exponential unit growth. If you have a very strong business model, which in tech there's so many of those for so many different types of modes, uh your earnings don't grow linearly, they grow exponentially. And that's the last piece. Invest when there's underappreciated long-term earnings power. And very often the earnings can grow from $1 to $10.50 to 20. And it happens way more than you think. And it allows you to buy some of the best companies in the world for extremely low pees. When we were buying Nvidia in 2023, we were paying four times earnings. When we bought Tesla in 2019 for the car scurb, we were paying five times earnings. When we were owning Apple, we were paying four times earnings. When we bought Amazon for AWS, we we were getting it for free. And you know, the world doesn't think exponentially.
竞争优势,然后是被低估的盈利能力。当你抓住 S 曲线上正确的那一段,你会得到指数级的销量增长。如果你有一个非常强的商业模式——科技里这样的太多了,各种类型的护城河都有——你的盈利就不是线性增长,而是指数级增长。这就是最后一块。在长期盈利能力被低估的时候投资。而且盈利常常能从 1 美元涨到 10 美元、15 美元、20 美元。这种情况发生的频率比你以为的高得多。它让你能以极低的市盈率买到世界上最好的一些公司。我们 2023 年买英伟达的时候,付的是四倍市盈率。2019 年我们为汽车这条S 曲线买特斯拉的时候,付的是五倍市盈率。我们持有苹果的时候,付的是四倍市盈率。我们为了 AWS 买亚马逊的时候,等于是白送给我们的。你知道,这个世界不会用指数的方式思考。
便签笔记
21:14
and they're so focused on the next year, the next quarter. Very few people believe you can accurately predict two, three, four years out. But if you follow and understand the S-curve and you you know the modes and you know how to model, you really can uh predict these these great things. So let's go to the scurve. So the S-curve is crucial because every technology follows this pattern where it comes out, you know, the I the smartphones were out 10 years before the iPhone. The internet was out 20 years before Netscape. AI has been out hidden inside of these companies, but it wasn't until Chachi PT took it public uh and ignited what it was. So electric vehicles, Tesla went public 15 years before 2019 when it went vertical it because there were so many barriers to adoption. The first smartphones, you know, they were clunky, they didn't have touchscreen, not Apple, there wasn't a wireless data system. And then uh and they were too expensive. They were $500 or $600. Steve Jobs got the price to
他们太专注于明年、下个季度了。很少有人相信你能准确预测两年、三年、四年之后的事。但如果你跟踪并理解 S 曲线,你懂护城河,你懂怎么建模,你真的能预测到这些很了不起的事情。那我们就来讲 S 曲线。S 曲线之所以关键,是因为每一项技术都遵循这样的模式:它先问世,你知道,智能手机在 iPhone 之前十年就有了。互联网在网景之前二十年就有了。AI 一直藏在这些公司内部,但直到 ChatGPT 把它推到公众面前,才点燃了这一切。再说电动车,特斯拉在 2019 年垂直起飞之前 15 年就已经上市了,因为当时采用的障碍太多了。最早的智能手机,你知道,很笨重,没有触摸屏,不是苹果,也没有无线数据网络。而且它们太贵了,要五六百美元。乔布斯把价格做到了 200 美元。
便签笔记
22:24
200. There was AT&T had a 3G network. It was touchscreen. It was so easy your grandmother could do it. So Annie built an ecosystem and made it simple. So all the barriers to adoption were eliminated and then you rocket when those barriers are removed. That's the tornado of demand that everybody in the world knows they need this right away. And so that's the flip that happens. It happened with electric vehicles. The price was too high. Elon got the price to 40,000. Range anxiety was there. he got the the range to 300 miles. The supply chain was was finally in place so he could churn out millions of these things. So that triggers the inflection. Now the other nuance, it's not just, oh, it's taken off now. It's how tall, how big is this S-curve, how tall it is, so you know when to sell, how long to hold on, cuz we're underwriting out 2 or 3 years. We have to know what the growth looks like thereafter. [snorts] And these S-curves can be dynamic. So when Amazon had AWS and it [snorts] was a hidden line item
AT&T 有了 3G 网络。有触摸屏。简单到你祖母都会用。所以他构建了一个生态系统,并且把它变简单了。于是所有采用的障碍都被清除了,障碍一旦被移除,你就一飞冲天。那就是需求的龙卷风,全世界的人都知道自己马上就需要这个东西。这就是那个翻转时刻。电动车也是这样。价格太高,马斯克把价格做到了 4 万美元。有里程焦虑,他把续航做到了 300 英里。供应链终于就位,于是他能量产几百万辆车。这就触发了拐点。还有另一个微妙之处,不只是「哦,它起飞了」。而是这条 S 曲线有多高、有多大,这样你才知道什么时候卖、要持有多久,因为我们的测算是往后看两三年。我们必须知道之后的增长是什么样子。而且这些 S 曲线是会动态变化的。比如亚马逊有了 AWS,当时它只是亚马逊内部一个隐藏的报表科目,覆盖它的是零售
便签笔记
23:32
inside of Amazon covered by retail internet analysts, not hardware chip, it was a new business model, what have you. But we realized the TAM for this, it was the largest TAM in enterprise IT ever because previously the TAM was routers, memory, storage, Dell, EMC, but they were doing it all. And so we figured out, you want to know how tall the S-curve is. So we figured out they were addressing 600 billion of IT systems directly addressing that. And then we said it's probably going to be 50% deflationary. Therefore, we're 1 or 2% penetrated. But then over time, we realized it it actually wasn't deflationary. If you talk to anybody now, they say if you build it yourself, it's about the same price. So that means the TAM was so much bigger. So there's mega S-curves and there's subass curves.
互联网分析师,不是硬件芯片分析师,它是个新的商业模式,诸如此类。但我们意识到它的 TAM——那是企业 IT 史上最大的 TAM,因为在此之前 TAM 是路由器、内存、存储,是戴尔、EMC,而它们把这些全干了。所以我们算了一下,你想知道这条 S 曲线有多高。我们算出他们直接对应的是 6000 亿美元的 IT 系统市场。然后我们说,这大概会有 50% 的通缩效应。因此我们当时的渗透率只有 1% 或 2%。但随着时间推移,我们发现其实并没有通缩。你现在去问任何人,他们都会说如果你自建,价格差不多是一样的。所以这意味着 TAM 大得多。所以有超级大的 S 曲线,也有子 S 曲线。
便签笔记
24:28
You know, we've been lucky that we've had, you know, internet 1.0, uh, mobile, cloud, e-commerce, and now AI, which we can confidently say is the biggest and all these things build upon one another. So, you know, with with the electric vehicle S-curve, you you you have to pay attention too because, you know, at the time we we thought probably maybe 40 to 50% of the cars would go electric, but it did hit a big wall at 10 or 15%. Usually the S-curves go kind of all the way. Um, but in this case, for a variety of reasons, it didn't. So, you have to adjust and you have to stay on top of it. And generally you want to um when something gets to sort of 30 40% penetrated then you stop having exponential growth which means the sell side catches up and there's no longer big beats >> and is that when you sell typically >> generally we like we like the high growth and it was a mistake with Apple because in the first five or six years of Apple um it was awesome. I mean it was our largest position. and it would
你知道,我们挺幸运的,经历了互联网 1.0、移动、云、电商,现在是 AI——我们可以有信心地说这是最大的一个,而且这些东西都是一层叠一层的。所以,就电动车这条 S 曲线来说,你也必须保持关注,因为当时我们以为大概会有 40% 到 50% 的汽车转向电动,但它在 10% 或 15% 的时候撞上了一堵大墙。通常 S 曲线是会一路走到底的。嗯,但这次因为各种原因,它没有。所以你得调整,你得持续盯紧。一般来说,当某个东西渗透率到了 30%、40% 左右,就不再有指数级增长了,这意味着卖方分析师跟上来了,也不再有大幅超预期 >> 那通常就是你卖出的时候吗>> 一般来说我们喜欢高增长。苹果那次是个错误,因为在头五六年里苹果表现太棒了。它是我们最大的仓位。每年能涨 50%、70%,除了 08 年。
便签笔记
25:35
go up 50 70% a year um except for '08 and then we sold in 20 2012 when it got to sort of 50% of the US had a smartphone and with Apple you know they maintained their leadership position it had a couple years of underperformance and then the multiple got low and they added several ancillary things and then they al also got to play in the uh the application because they get 30% % of the app. So they were able to compound very nicely, say 20%, but the the big years were in the 50, you know, the the 0 to 50% part of the curve.
然后我们在 2012 年卖了,当时美国大约 50% 的人有了智能手机。而苹果呢,你知道,它保住了领先地位,有几年表现落后,然后估值倍数变低了,他们又加了好几项周边业务,然后他们还能在应用这块分一杯羹,因为他们抽应用 30% 的分成。所以他们能以不错的速度复利增长,比如 20%,但真正的大年份是在 50% 那段,你知道,就是曲线上 0 到 50% 的那部分。
便签笔记
07何时买入:直觉、现场线索与模式识别
26:14
>> I'm so fascinated by this uh you know, sometimes decade plus long flatline at the beginning of one of these curves, which makes me wonder what you've learned about the right moment to buy or even start paying attention before you buy. >> How do you measure that? Is it always different? What are the pitfalls that you've fallen into? How do you know when when to we talked about when to sell, but how do you know kind of when to start thinking about buying in one of these things? >> Yeah. And you know, Andy Grove says sort of when you have strategic inflection points, you can't trust the data. And and strategic inflection points are about intuition, anecdotal evidence. I love this book called The Towel Jones Averages, a guide to whole investing, which is rightrain and leftrain. And the best investors have the right the creative side where they it's visual.
>> 我特别着迷于这一点,就是这些曲线开头那段有时长达十几年的横盘期,这让我想问,关于什么时候买入、甚至在买入之前什么时候该开始关注,你学到了什么。>> 你怎么衡量?是不是每次都不一样?你踩过哪些坑?你怎么知道什么时候——我们聊了什么时候卖,但你怎么判断什么时候该开始考虑买入这类东西?>> 是的。你知道,安迪·格鲁夫说过,当出现战略拐点的时候,你不能相信数据。战略拐点靠的是直觉和轶事性的证据。我很喜欢一本书叫《The Tao JonesAverages》,一本关于全脑投资的指南,讲右脑和左脑。最好的投资者拥有右脑,那种创造性的一面,是视觉化的。
便签笔记
27:03
It's connecting the dots. Um you know we invested in the mobile video game S-curve for so long. Mobile video games were just the screens were small on the phones and the processing power wasn't good. So you had all these uh casual games. But then I was in China and I saw this little 12-year-old boy with a huge phone and he was like playing a awesome video game. I'm like oh my god it's now coming to the phone. So, it's visual. Um, enterprise is hard cuz you can't see it. We go to the Gartner IT Symposium.
是把点连起来。你知道,我们在移动游戏这条 S 曲线上投了很久。移动游戏以前手机屏幕小,处理能力也不行。所以只有那些休闲游戏。但后来我在中国,我看到一个 12 岁的小男孩拿着一部大屏手机,在玩一款很棒的游戏。我当时就想,天哪,这东西现在真的来到手机上了。所以,这是视觉化的。嗯,企业级的就难了,因为你看不见。我们会去 Gartner IT 研讨会。
便签笔记
27:35
30,000 American CIOS go there and like we saw this happen with Splunk where that used to be an amazing database company and like their their room where they were explaining was like standing room only or we saw that with VMware you know I'm talking like 30 years ago where they virtualized the server and like there was standing room only and you could just see the corporate demand just beginning and with AWS We went there and the grand ballroom was completely packed and that was at nine o'clock and at 10 o'clock the grand ballroom was completely packed 11:00. So you could you could actually see the demand exploding before it happened. So um we look for for all kinds of clues and there's a whole pattern recognition that happens. And by the way it's okay to be late. It's okay to miss the first one, two, three years in a lot of cases because if the top of the S-curve is half a trillion, um the growth can go on for a long time. So, you don't always have to be right there. It's okay to miss the first 100%. Peter Lynch, I
三万个美国 CIO 会去那里。我们在 Splunk 身上就看到过这种情况,它当年是家很厉害的数据库公司,他们讲解的那个会场挤得连站的地方都没有。VMware 也是,你知道我说的是大概三十年前,他们把服务器虚拟化,会场也是挤得站都站不下,你能直接看到企业需求刚刚开始起来。AWS 也是,我们去了,大宴会厅完全爆满,那是九点,十点的时候大宴会厅还是完全爆满,十一点也是。所以你真的能在需求爆发之前就看见它。所以我们会寻找各种各样的线索,这里面有一整套模式识别在起作用。顺便说一句,晚一点也没关系。很多情况下,错过头一年、两年、三年也没关系,因为如果这条 S 曲线的顶部是五千亿,那增长可以持续很久。所以你并不总是非得踩在起点上。错过第一个 100% 是没关系的。彼得·林奇——我职业生涯是从富达开始的,他很喜欢带年轻人。所以我跟他
便签笔记
28:49
started at Fidelity and he loved to mentor the young kids. So, I got some time with him. He said, "Wite out the chart. It's all about the future." Um, and so it's okay to miss, but but what helps about the S-curve is is sort of how long it goes for. Then there's sort of the the slope of the Scurve, which is important. And a lot of people think cuz we're in a modern world, everything's so fast, but there's a lot of factors that determine the pace of the adoption. And we we um commissioned this gentleman, Horus Du used to work with Clayton Christensen, to go look in history. And we have the big S-curves on our wall over the last 100 years. And the radio Scurve is one of the fastest ever. It took 7 years to reach like 100% penetration. But the dishwasher Scurve is like that because it needs to be plugged into the back end.
有过一些相处的时间。他说:「把图表涂掉。重要的全是未来。」所以错过是没关系的,而 S 曲线的帮助在于它能告诉你这个过程会持续多久。然后还有 S 曲线的斜率,这也很重要。很多人以为,因为我们身处现代社会,一切都很快,但其实有很多因素决定采用的速度。我们委托了一位叫 Horace Dediu 的先生——他以前和克莱顿·克里斯滕森共事——去研究历史。我们墙上挂着过去一百年的那些大 S 曲线。收音机的 S 曲线是有史以来最快的之一,大概七年就达到了接近 100% 的渗透率。但洗碗机的 S 曲线就是那样,因为它需要接到后端管线上。
便签笔记
29:46
>> What are some Yeah. What else did you learn? That's fascinating. What else did you learn? >> So like the B2B stuff can take a long time because it needs to be plugged into the existing systems. It's like it's got to be put >> the dishwasher >> inside the house and then um and consumers generally tend to go a lot faster. Um >> I love that the radio and the dishwasher, the two models for adoption. >> Yeah. And and I I covered internet of fidelity. I c, you know, my first stock was Amazon. That's a whole other story which is a lot of fun. But I also did B2B internet and you know there was a whole huge bullcase on that. But the basically the underlying infrastructure wasn't in place for B2B to happen.
>> 还有哪些?是啊,你还学到了什么?这太有意思了。你还学到了什么?>> 比如 B2B 的东西可能会很慢,因为它必须接入现有系统。就像它必须被安装 >> 洗碗机 >> 装进房子里。然后消费端一般来说会快得多。嗯,>> 我太喜欢这个了,收音机和洗碗机,两种采用模型。>> 是啊。我在富达是看互联网的。你知道,我的第一只股票是亚马逊,那是另一个很有意思的故事。但我也做过 B2B 互联网,你知道当时有一整套宏大的看多逻辑。但基本上底层基础设施还没就位,B2B 根本发生不了。
便签笔记
30:28
Ultimately happened 20 years later with SAS. And so that is a risk with AI in that you know these big companies are very security conscious. Uh they're can be slow to move. There's a lot of cultural issues with a with AI where you know you really need a few evangelists to push it through and the top management needs to push it through but then the IT's saying this is this is risky and that happened with cloud too that was one of the big things with cloud where it was too it was everybody was afraid it's unsecure to have your data in the cloud and then we saw the CIA do it and we saw Capital One and we talked to the Capital One CIO we that it's more secure in the cloud and then it really started to take off. But but those takeoffs maybe because SAS is like the dishwasher and because cloud is like the dish it's got to be plugged in it it meant that yeah it was growing but it was sort of a 30 to 40 maybe a 50% growth rate but what's amazing about AI is you just at least with consumers or even business you just open up
最终是二十年后靠 SaaS 实现的。所以 AI 也有这个风险,因为你知道这些大公司非常在意安全。他们行动可能很慢。AI 还牵扯很多文化上的问题,你知道,你真的需要几个布道者去推动,高层管理也得推,但 IT 部门会说这个有风险。云当年也是这样,那是云面临的一大问题,当时所有人都害怕把数据放在云上不安全。后来我们看到 CIA 这么干了,看到第一资本这么干了,我们跟第一资本的 CIO 聊过,说放在云上其实更安全,然后它才真正开始起飞。但这些起飞——也许因为 SaaS 就像洗碗机,因为云就像洗碗机得接管线一样——这意味着,是的,它在增长,但大概是 30% 到 40%、也许 50% 的增长率。而 AI 令人惊叹的地方在于,至少对消费者、甚至对企业来说,你只要打开 >> 浏览器它就在那儿 >> 所以这就是为什么我们看到的是
便签笔记
31:40
>> the browser and it's there >> and so that's why we're getting this straight up >> and I think there's enough runway in the me in the near term going from 10 bits of people really using it to two to five or whatever which is going to cause it to keep on going straight up. So this we this we call it a backwards L curve. Um so it's really pretty exciting. >> What have you learned about uh when the group that ends up being the leaders separates itself from one of these competitive packs? So you're talking there mostly about overall growth of the S-curve and demand. There's always multiple players fighting for it. You know, you've invested, it seems like you kind of invest after someone has separated themselves from the pack, not try to pick the winners from the pack.
直线上冲 >> 而且我觉得短期内还有足够的空间,从只有 10% 的人真正在用,到两倍五倍什么的,这会让它继续直线上冲。所以我们把这个叫做反过来的 L 形曲线。嗯,所以真的挺让人兴奋的。>> 关于最终成为领跑者的那一批公司是在什么时候从竞争群里脱颖而出的,你学到了什么?你刚才讲的主要是 S 曲线整体的增长和需求。但总是有多个玩家在争夺。你知道,看起来你的投资方式是等到有人已经从群体中脱颖而出之后才投,而不是试图从一堆人里挑出赢家。
便签笔记
08竞争优势的六种形态与反面清单
32:24
Is that is that like roughly? >> Well, >> correct? >> Well, we're definitely So, you look for the S-curve, then we do an exhaustive study of everybody with exposure in that area and try and find the one with a very powerful competitive advantage. And a lot of people didn't like tech. Warren Buffett didn't like tech because he couldn't predict the future too fast. Yeah. And so the Scurve is our map for looking in the future. Now a lot of people were worried about tech because they thought there was so much disruption you could never trust a company to be a longived asset. And [snorts] what we've found over the years is some of the competitive advantages within the digital world are more powerful, if not equally or more powerful than than in the offline world.
大致是这样吗?>> 嗯, >> 对吗?>> 嗯,我们确实是——你先找到 S 曲线,然后我们会对这个领域里所有有敞口的公司做一次穷尽式的研究,试着找出那个拥有非常强大竞争优势的。很多人不喜欢科技股。沃伦·巴菲特不喜欢科技,因为变化太快,他没法预测未来。对。所以 S 曲线就是我们看向未来的地图。很多人对科技心存疑虑,因为他们觉得颠覆太多了,你永远不能指望一家公司成为长寿资产。而这些年我们发现的是,数字世界里的某些竞争优势比线下世界的更强大,至少也是同等或者更强。
便签笔记
33:13
You've got the network effect that was so powerful for LinkedIn, Facebook, Alibaba, you name it. Then you can become an industry standard. Oracle and Bloomberg are the industry standard. Oracle, you know, they charge a lot and, you know, there's free versions, there's open- source Oracle, but they had all the database administrators. They they had all the software that was tuned to work with them. So, they they basically had a chokeold on the relational database market forever. um you can get to scale very quickly because these scurves grow and all of a sudden Anthropic is doing 90 30 billion in sales or Amazon you know had so much scale and they got it quickly. So they got a Walmart size scale advantage in 5 years versus 40 years for Walmart. So you can have network effects scale you can become industry standard. You can be a platform that people build on top of.
你就拥有了当年让 LinkedIn、Facebook、阿里巴巴等等公司如此强大的网络效应。然后你就能成为行业标准。Oracle 和 Bloomberg 就是行业标准。Oracle,你知道,他们收费很高,而且市面上有免费版本,有开源的 Oracle 替代品,但所有的数据库管理员都在他们那边。所有的软件都是针对他们调优的。所以他们基本上永远地扼住了关系型数据库市场的咽喉。嗯,你可以非常快地做到规模化,因为这些 S 曲线增长很快,突然之间 Anthropic 的销售额就做到了 90 亿、300 亿,或者亚马逊,你知道,他们有那么大的规模,而且很快就达到了。所以他们用 5 年就获得了沃尔玛级别的规模优势,而沃尔玛用了 40 年。所以你可以有网络效应、规模效应,你可以成为行业标准。你可以成为一个别人在上面构建东西的平台。
便签笔记
34:13
You can have critical intellectual property, which was what Qualcomm had. You couldn't make a phone without paying them, or ASML has critical intellectual property. You can't make a chip without their lithography. And I think what's interesting is maybe these AI foundational companies, you know, they've got scale. Oh, you can also have brand. And brand's very important because Google, Amazon, they got to grow. They never had to advertise. Elon's never had to advertise for anything. and cost to acquire versus lifetime. It's the whole business model.
你可以拥有关键的知识产权,这就是高通所拥有的。不付钱给他们你就造不出手机;或者 ASML 拥有关键知识产权,没有他们的光刻机你就造不出芯片。我觉得有意思的是,也许这些 AI 基础模型公司,你知道,他们有规模。哦,你还可以有品牌。品牌非常重要,因为谷歌、亚马逊,他们必须增长,但他们从来不需要打广告。马斯克从来不需要为任何东西打广告。还有获客成本对比生命周期价值。这就是整个商业模式。
便签笔记
34:46
And so almost all the companies I mentioned have Apple, they have all of these rolled into one. Um, so we can sometimes we can notice these things before the rest of the world. And one of our high points was we pitched Amazon for AWS at 2013 at the Robin Hood investors conference and we said the bulls have no idea what they're sitting on. Amazon's won the war but before it even started and at that time we said there's Coke and there's no Pepsi. Did turn out there was Pepsi but it was big enough to last. And we could see they had a seven-year lead. So first mover is important. Then they became a whole ecosystem and a platform. Then they got scale. So they were 10 times the size of everybody else. Nobody could invest in the R&D to to catch them. So um but you're right that if you don't have a competitive advantage, you can be in the best S-curve of all time >> and still lose out.
所以我提到的几乎所有公司,还有苹果,都把这些优势集于一身。嗯,所以我们有时候能比世界上其他人更早注意到这些事情。我们的高光时刻之一是,2013 年我们在罗宾汉投资者大会上推荐了亚马逊的 AWS,我们说,那些看多的人根本不知道自己手里握着什么。亚马逊在战争还没打响之前就已经赢了。当时我们说,这里只有可口可乐,没有百事可乐。后来确实出现了百事,但市场足够大,容得下。而且我们能看出他们有七年的领先优势。所以先发者很重要。然后他们变成了一整个生态系统和平台。然后他们获得了规模。他们的体量是其他所有人的 10 倍。没人能投入那么多研发去追赶他们。所以嗯,但你说得对,如果你没有竞争优势���你可能处在史上最好的 S 曲线上,>> 但仍然出局。
便签笔记
35:44
>> But if your name was Rim, Palm, Nokia, HC, LG, Motorola, I can go on forever. 0 negative negative negative negative. And that's what we saw at the foundational model layer where there's like 50 companies trying to do that and they all have fallen away and two or three have emerged at the top and there's a lot of reasons to think they will continue to hold their position. >> So to take Google's a little trickier because they have this other huge massive complex business attached to the Gemini business. But if you take anthropic and open AI as pure plays and you dig through those and you reason through their competitive advantages, why aren't they susceptible to erosion of those things in the fullness of time?
>> 但如果你的名字是 RIM、Palm、诺基亚、HTC、LG、摩托罗拉,我可以一直数下去。全都是负的、负的、负的、负的。而这正是我们在基础模型层看到的情况,当时有大概 50 家公司在做这件事,他们全都掉队了,只有两三家脱颖而出站到了顶端,而且有很多理由相信他们会继续保住自己的位置。>> 那么谷歌就要稍微复杂一点,因为他们还有另外一个庞大复杂的业务捆绑在Gemini 业务上。但如果你把 Anthropic 和 OpenAI 当作纯粹的标的来看,你深入研究它们、推演它们的竞争优势,为什么假以时日,它们不会面临这些优势被侵蚀的风险呢?
便签笔记
36:28
>> Yeah. Of all the S-curves we've we've done, [snorts] AI is by far the most complex and the fastest changing. So it can be we have to keep in mind that there are risks but also the rewards are the highest cuz we're talking about a market in the trillions. You know we we just said cloud you know maybe cloud's 800 billion. This might be you know we now think 3 to five but there's higher risk higher reward. But let's just say with anthropic now they have it looks like they have critical intellectual property generally they've been able to maintain their their high market share and code.
>> 是啊。在我们研究过的所有 S 曲线里,[吸鼻子] AI 是迄今为止最复杂、变化最快的。所以我们必须记住这里面是有风险的,但回报也是最高的,因为我们谈论的是一个数以万亿计的市场。你知道,我们刚才说云计算,也许云计算是 8000 亿。而这个,我们现在觉得可能是 3 万亿到 5 万亿,但风险更高,回报也更高。不过就说 Anthropic 吧,现在看起来他们拥有关键的知识产权,总体上他们一直能够保持在代码领域的高市场份额。
便签笔记
37:10
Number two is uh they've built a a strong brand for enterprise to where go talk to any CIO and they'll just the first thing they'll say is claude. They're going to have escape velocity and scale. And what was scary for OpenAI and Anthropic fighting these big companies like Google was they had these huge cash cows. And to both of the management teams credited Open and Anthropic, they were able to work in these super capital intensive industries and find ways to raise capital. And certainly with Anthropic, with their 10x sales growth, it looks like and their fundraising ability, it looks like they've reached escape velocity. So now they have scale.
第二点是,呃,他们在企业市场建立了很强的品牌,你去问任何一位 CIO,他们第一个会说出口的就是 Claude。他们将会达到逃逸速度和规模。对 OpenAI 和 Anthropic 来说,跟谷歌这样的大公司竞争之所以可怕,是因为那些公司有巨大的现金牛业务。而 OpenAI 和 Anthropic 两家的管理层都值得称赞,他们能够在这种超级资本密集的行业里运作,并找到融资的办法。而对 Anthropic 来说,以他们 10 倍的营收增长,以及他们的融资能力来看,他们似乎已经达到了逃逸速度。所以现在他们有了规模。
便签笔记
37:52
And the other thing that Anthropic and OpenAI could have is Anthropic now that they're leading in code, they set that code back onto their model and it's this concept of the recursive improvement. And if you look at the pace of their innovation, it's accelerating. >> Um, and so maybe they can have this liftoff stage. you know, Open AI has, you know, they they were focused on so many different other sectors, but they're starting to do better in enterprise and their coding tools good and they're starting to see accelerating growth on that side. And then look, the consumer franchise, it it looks like enterprise right now is much better because you're, you know, you and I, we're willing to pay a lot because it's replacing human beings. you know, consumer, maybe you can get advertising, but maybe they would pay for a a clawbot type assistant if you could make that perfectly well for them.
Anthropic 和 OpenAI 还可能拥有另一样东西:Anthropic 现在在代码领域领先,他们可以把那些代码反哺回自己的模型,这就是所谓的递归式自我改进。而如果你看他们的创新节奏,是在加速的。>> 嗯,所以也许他们能进入这种起飞阶段。你知道,OpenAI,你知道,他们之前的注意力分散在很多其他领域,但他们在企业市场开始做得更好了,他们的编程工具不错,而且他们在那一块也开始看到加速增长。然后你看,消费者这块的基本盘,目前看起来企业市场要好得多,因为,你知道,你和我,我们愿意付很多钱,因为它替代的是人力。你知道,消费者这边,也许你能靠广告赚钱,但如果你能给他们做出一个完美的 Claude 机器人式助手,他们或许也愿意付费。
便签笔记
38:54
Um, but they have gazillion eyeballs there. But you're right, things do shift, but it usually on the we have this charts that we almost do for all of our pitches. On the internet, the leader goes bigger, faster, and wins. And it's it's it's happened you know most of the time the leader gets it. Shopify becomes the leader. It just keeps on going. Amazon the leader keeps on going. SAS company XYZ just you get the lead. It compounds on internet company compounds on itself. And and another thing is you need to be big. Another is scale. You need the compute and you got to pay for the compute because so there's only so many people that can do that. So that those are some of the modes that we think are now showing up. Now there are some exceptions to that rule. usually with the paradigm shifts AOL and then dialup went to broadband and and they didn't make make the change. You know, Netscape came out early and it wasn't as strong of a business model. But I think if you talk to anyone in the valley or
嗯,但他们在那边有海量的用户眼球。不过你说得对,情况确实会变,但通常来说——我们有这么一张图表,几乎每次推介都会用到。在互联网上,领先者会变得更大、更快,然后赢。而且这种情况大多数时候都会发生,领先者拿下市场。Shopify 成为领先者,就一直这么走下去。亚马逊是领先者,就一直走下去。某某 SaaS 公司,你一旦拿到领先,它就会复利累积。互联网公司会自我复利。还有一点是,你需要体量大。另一点是规模。你需要算力,而且你得付得起算力的钱,所以能做这件事的人就那么几个。这些就是我们认为现在正在显现出来的一些护城河。当然这个规律也有例外。通常是在范式转移的时候,比如 AOL,然后拨号上网转向宽带,他们没能完成这个转变。你知道,网景很早就出来了,但它的商业模式没那么强。不过我觉得,如果你去问硅谷的任何人,或者
便签笔记
09为什么几乎清仓了应用软件
39:54
any startups, you know, they'll tell you that they're building on top of these three and the world's a huge place and the economy is a huge place that that they'll be able to differentiate within those. I'm so curious then what you think all of this means for software. Um when I look through your portfolio, I don't see a ton of uh big software companies, enterprise software companies. I I don't know if you once had them and sold them or or how you thought about it, but it's hard to have the experience of building really useful, cool little tools, even if they're still toys, and not have the thought of, wow, you know, like if I spend enough time on this, even if I'm not technical, maybe I could build a, you know, an ERP equivalent replacement or something for my company. There [snorts] doesn't seem to be a fundamental reason why that's not possible and and then those companies could be in lots of trouble. Seems like everyone has a a strong view on this one way or the other. I'm curious how you've
任何创业公司,你知道,他们会告诉你,他们都是在这三家之上做开发的,而且世界这么大,经济体量这么大,他们能够在这几家的基础上做出差异化。那我就很好奇,你觉得这一切对软件行业意味着什么。嗯,我翻看你们的持仓时,没看到多少大型软件公司、企业软件公司。我不知道你们是不是曾经持有过又卖掉了,或者你们是怎么看的,但很难在亲手搭出那些真的很有用、很酷的小工具之后——哪怕它们还只是玩具——不产生这样的念头:哇,你知道,比如如果我在这上面花足够多的时间,就算我不是技术出身,也许我也能给公司做一个相当于 ERP 的替代品之类的东西。似乎[嗤笑]没有什么根本性的理由说明这做不到,而一旦如此,那些公司可能就有大麻烦了。看起来每个人在这件事上都有很鲜明的观点,不是这边就是那边。我很好奇你们是怎么看待这类公司的,毕竟你们好像并没有持有多少
便签笔记
40:46
approached those sorts of companies given that you don't seem to own a ton of them. >> We were at certain points maybe 5 years ago, we might have had 40 or 50% of our portfolio in software. And early on in in our April 2023 seminar, we said definitely invest in chips first and and we said but at the application layer initially we thought [snorts] these companies are huge. They have huge sales forces. They can take these AI APIs and and build products and they have the data. This is going to be amazing for software.
这类股票。>> 我们在某些时候,大概五年前吧,可能有 40% 到 50% 的仓位在软件上。而在早期,在我们 2023 年 4 月的研讨会上,我们说一定要先投芯片,同时我们说,但在应用层,我们最初以为[嗤笑]这些公司体量巨大,有庞大的销售团队,他们可以拿这些 AI API 去做产品,而且他们手里有数据。这对软件来说将会是件了不起的事。
便签笔记
41:21
And pretty quickly we realized their AI products were not very good. They weren't moving the needle. Nobody could charge for them. We basically sold almost all of our software, almost all of our application software. We still have one or two small ones, but entering this year, we were actually net net short. And uh it really helped us in the first quarter. There's so many layers. The old way of software is like using a pen and paper or it's like a horse and buggy. The new way of software is like a jet engine or frankly like the transporter from Star Trek. It's so revolutionary changing that it feels like it has to be disruptive now.
但我们很快意识到,他们的 AI 产品并不怎么样,没能带来实质性的推动,谁也没法为它收费。我们基本上把软件股几乎全卖了,应用软件几乎全清了。我们还留着一两只小的,但进入今年时,我们其实是净空头。这在第一季度确实帮了我们大忙。这里面有太多层次了。旧的软件方式就像用纸笔,或者像马车。新的软件方式则像喷气发动机,坦白说,更像《星际迷航》里的传送装置。它的变革性太强了,让人感觉现在就必然会带来颠覆,就是现在。
便签笔记
42:14
even if it's not disruptive now or right away. Uh so the software companies have another problem which is their list on the to-do list or priority list of any CIO has fallen a lot. So even if AI is not going to be disruptive, they're spending it on anthropic tokens because there's faster ROI there. Um, second, um, if they're spending all that money over there, it pushes on the budget, so that hurts them. Third, a lot of software companies were able to raise price every year. Um, [clears throat] and now they're probably nervous about doing that. Then fourth, we'll see what happens with jobs cuz I don't, you know, there's smart people on both sides of that, but we are seeing some companies really gut their jobs or whatever, >> freeze hiring and so that hurts on seats. Maybe in terms of them building their own apps, maybe it just um >> you know, if you want to be optimistic, it's it's taken them a while to do that.
即便它现在或者说马上还没形成颠覆。呃,所以软件公司还有另一个问题:它们在任何一位 CIO 的待办清单或优先级清单上的位置都大幅下滑了。所以就算 AI 不会带来颠覆,他们也在把钱花在 Anthropic 的 token 上,因为那边的回报更快。嗯,第二,如果他们把那么多钱花在那边,预算就被挤压了,这会伤到软件公司。第三,很多软件公司过去每年都能提价。嗯,[清嗓]而现在他们大概会对这么做感到心虚。第四,就看就业会怎么样了,因为我不知道,你懂的,这件事的两边都有聪明人,但我们确实看到一些公司在大幅砍岗位之类的,>> 冻结招聘,这会打击席位数。至于他们自己去做应用,也许只是……>> 你知道,如果你想乐观一点,那就是他们花了不少时间才做到这一步。
便签笔记
43:24
We talked about how early the primitives of AR are. So maybe they have just taken a while to get to something they can commercialize, but you know, they might not have the right people. They might not know it's a different selling motion from selling a fixed system versus, you know, if you're installing something that does human work, you got to be right at the side to make sure it's really getting done. So you need the FDE for deployed engineers and they might not have the right people internally to do that. Then of course there's the the risk of you can build it yourself. The bulls will say, well, they're never going to build their own ERP system. And that's probably right. And it is true that technology, old tech is very sticky. Like mobile video games didn't hurt console games and uh the tablet didn't hurt the PC and the smartphone didn't hurt the PC and uh there's a lot of integrations and work that goes into these software. So that's all true and companies do like to buy from they don't
我们聊过 AI 的那些基础组件还处在多早期的阶段。所以也许他们只是需要时间才能做出可以商业化的东西,但你知道,他们可能没有合适的人。他们可能没意识到,这和卖一套固定的系统是完全不同的销售方式——如果你部署的是一个替人干活的东西,你必须待在客户身边,确保事情真的被完成了。所以你需要 FDE,也就是前线部署工程师,而他们内部可能没有合适的人来做这件事。当然还有一个风险,就是你可以自己造。多头会说,他们绝不会自己去搭一套 ERP 系统。这大概是对的。而且技术确实很有黏性,老技术非常顽固。比如手游没有伤到主机游戏,平板没有伤到 PC,智能手机也没有伤到 PC,而且这些软件背后有大量的集成和工作量。所以这些都成立,而且企业确实更愿意采购,不太愿意自己造。所以这些都对,但你不能想象不出
便签笔记
44:30
like to build themselves that much. So that's all true but you can't imagine a world where in 1 2 3 4 5 years um you could have a brand new AI native company going after each one of these very strong incumbents and it might their data advantage could get obiated. it might be easy to take it out and put the new one in with AI and such. So, what's good if you like so is the valuations are very high and everybody knows they're under pressure. Some people are tempted to buy these, but the AI um coding tools are just getting better and better. So, we'll we'll have to wait and see. And we're we're watching these software companies very closely to see if they're getting any revenue that can change that trajectory. But it's hard because if you're a company like Salesforce, you've got 40 billion in sales and now you you might have 500 of ARR 700 of AR of AI. So you've got this huge base. Now maybe this starts to work but it takes a while. And in software there's the rule of 40 which is your growth rate plus your operating margin.
这样一种可能:在一两三四五年内,会出现一家全新的 AI 原生公司,去正面进攻这些非常强大的在位者,而它们的数据优势可能被抹平。用 AI 之类的手段,把旧的拆掉、换上新的,可能会变得很容易。所以,要说有什么好处的话,就是这些公司的估值都很高,而且所有人都知道它们承受着压力。有些人会忍不住去抄底,但 AI 的编程工具正在变得越来越好。所以我们只能拭目以待。我们在非常密切地关注这些软件公司,看看它们能不能拿到足以改变这个轨迹的收入。但这很难,因为如果你是 Salesforce 这样的公司,你有 400 亿美元的营收,而现在你的 AI 可能只有 5 亿 ARR、7 亿 ARR。你有这么庞大的基数。也许这开始奏效了,但需要很长时间。软件行业有个「40 法则」,就是增长率加上经营利润率。
便签笔记
45:44
And if you've got a 20% growth rate and 20% that's good. For AI, we have a new kind of rule of 40. We call it well, it's really for chip investing. But if what percent of your sales are AI, say 30%, and what's your market share in that category? Say 30%. You'd be 60. That's a great place to look because you've got exposure and you've got a strong market position. Problem with software is their AI is 1 or 2% at this stage and it's a long way to go. Um, one thing we are picking up though now lately and this is halfbaked, but AI could make some of these software platforms more important because what's the first thing you do with claude? You plug it into Slack. If that can become a key repository, that will make Slack a permanent fixture within the organization. And so maybe these agents, maybe the next wave of AI will be these agents that use tools and they might operate inside of the existing incumbent software tools to use them like a human being would.
如果你有 20% 的增长率加 20% 的利润率,那就不错了。对 AI,我们有一个新版的 40 法则。我们把它叫做,嗯,其实主要是用在芯片投资上的。就是看你营收中有多少百分比来自 AI,比如 30%,再看你在那个品类里的市场份额,比如 30%,加起来就是 60。那是个很值得关注的位置,因为你既有敞口,又有很强的市场地位。软件的问题在于,它们的 AI 现阶段只占 1% 或 2%,路还长着呢。嗯,不过我们最近也在琢磨一件事,虽然还不成熟,那就是 AI可能会让其中一些软件平台变得更重要,因为你拿到 Claude 之后第一件事是什么?你会把它接进 Slack。如果 Slack 能成为一个关键的信息库,那它就会成为组织内部的永久性设施。所以也许这些智能体,也许 AI 的下一波浪潮就是这些会使用工具的智能体,而它们可能就运行在现有在位者的软件工具内部,像人一样去使用这些工具。
便签笔记
46:49
>> Just to pull in that thread, uh it seems like the commonality of the tools they might use that are the most sticky would be network-based tools. Uh so Slack is a great obviously a great example of the software in Slack itself is I don't know leaves something to be desired. It's not the software is not the special part. It's that everyone is there, >> right? But I'm curious yeah what kinds of things you would want. Is it just network you know the presence of a network effect? Is that the only thing that really matters?
>> 顺着这个思路,看起来它们可能会用到的那些最有黏性的工具,共同点似乎都是基于网络的工具。呃,Slack 显然是个很好的例子——Slack 本身的软件,我说不好,是有点让人不太满意的。特别之处并不在软件本身,而在于所有人都在上面,>> 对吧?但我很好奇,你会想要什么样的东西。是不是只看网络,你知道,只看有没有网络效应?这是唯一真正重要的东西吗?
便签笔记
47:14
>> It's still early in our thinking here but I don't know even even even maybe you know workday or the HR systems or um the big systems of record you know the agents may be running on top of on top of them. CRM is going headless or they're making a headless version and that's sort of the bare case too that you get relegated to just being a database but you know there's a human interface to it then they need to make the AI interface which is no interface it's just them going right into the data and so you know you lose that customer interaction but if if the agents are going right to right to CRM and doing the work inside of CRM M that that will solidify CRM so you won't have to think it's going away.
>> 我们在这方面的思考还很早期,但我不知道,甚至可能,你知道,Workday 或者那些 HR 系统,或者那些大的记录系统,你知道,智能体可能就跑在它们之上。CRM 正在走向无头化,或者说他们正在做一个无头的版本,这也算是那种最悲观的情形——你被降格成一个纯粹的数据库,但你知道,还是有一个人类界面在那儿的;然后他们就得去做 AI 界面,而所谓 AI 界面就是没有界面,就是 AI 直接去读数据。所以你就失去了那种客户互动。但如果这些 agent 是直接进到 CRM 里、在 CRM 内部干活的,那这反而会巩固 CRM 的地位,所以你就不用担心它会消失了。
便签笔记
10数据中心硬件的「去商品化」
48:07
>> Can we talk about chips? You've referenced them a few times. >> Inf infrastructure chips, you know, everything around the data center maybe. I don't know how you conceive of it. >> Why is this so interesting to you? I love the the modified rule of 40 for percentage that's AI and percentage market share in the category. That's an interesting stat. >> What companies shine on that today? What are what are lagards, you know, that are surprising? For the past 40 years, nothing has changed in the data center.
>> 我们能聊聊芯片吗?你前面提了好几次。>> 基础设施芯片,你知道的,围绕数据中心的一切吧。我也不知道你是怎么划分这一块的。>> 为什么这个领域让你这么感兴趣?我很喜欢你那个改良版的「40 法则」——AI 占比加上这个品类里的市场份额占比。这个指标挺有意思的。>> 按这个标准,今天哪些公司表现亮眼?哪些是落后者,有没有让人意外的?过去 40 年里,数据中心其实什么都没变。
便签笔记
48:33
Even with cloud, we're basically Intel x86. It became the data center chip sometime in the '9s. And um and compute grew in the cloud era and it grew compute workloads grow you know 25 to 40% every year but Moore's law is improving at that rate. So it didn't require tremendous innovation and there really was almost no growth in hardware for years and years and years and the whole industry basically commoditized every part every chip every part of the server the printed circuit board to the memory to the enclosures to the networking you know there was no innovation you would go from one gig to 10 gig That would take 7 years. And when you do switch in the first year, it does take some innovation to get to 10 gig and would create a little cycle, but then it would commoditize.
即便有了云,我们基本上还是英特尔 x86。它大概在九十年代成了数据中心芯片。然后在云时代算力增长了,算力工作负载每年增长大概 25% 到 40%,但摩尔定律也在以同样的速度改进。所以并不需要什么了不起的创新,硬件方面其实很多很多年几乎没有增长,整个行业基本上把每个零件都商品化了——每颗芯片、服务器的每个部件,从印刷电路板到内存、到机箱、到网络设备,你知道,完全没有创新。你从 1G 升到 10G,那得花七年。而且第一年切换的时候,做到 10G 确实需要一点创新,会形成一个小周期,但之后就又商品化了。
便签笔记
49:38
And now you go to AI and the workloads are growing 10x every year and they're pushing every single aspect of this hardware to the physical limits of what it can do. And so, not only are you creating tremendous unit growth, but the industry, we call it the decommoditization of the hardware industry. And I I met with Shawn Maguire like 3 years ago, and he said, "I wish I could come back and be a a hardware hedge fund because all the companies are public and they all have powerful IP." And Sequoia made some of their best investments back in the hardware day with Apple and Cisco and others. And we're in this renaissance of chips. So not only do you have tremendous unit growth, but you it's requiring tremendous innovation and what that means, you know, at every aspect of the server. And so you know memory which used to be a pure commodity, this high bandwidth memory is stacked 10 chips on top. you know the input outputs are 10x what they were before like took Samsung for years to do it and it's a critical critical
现在到了 AI,工作负载每年增长 10 倍,它们把这套硬件的每一个环节都推到了物理极限,逼到能力的尽头。所以你不但创造了巨大的出货量增长,而且整个行业——我们管这叫硬件行业的「去商品化」。我大概三年前见过 Shawn Maguire,他说:「我真希望能回去做一只硬件对冲基金,因为这些公司全都上市了,而且个个手里都有很强的知识产权。」红杉当年在硬件时代做过一些他们最好的投资,比如苹果、思科等等。而我们现在正处在芯片的文艺复兴里。所以你不但有巨大的出货量增长,而且它还要求大量的创新,这意味着服务器的每一个环节都得变。所以你看内存,以前是纯粹的大宗商品,现在这种高带宽内存是十颗芯片叠在一起的,输入输出是以前的 10 倍,三星花了好几年才做出来,而且它是极其关键的一环,还在不断迭代升级,所以他们得跟英伟达
便签笔记
50:58
piece and then that is constantly upgrading so they're on the same you know they've got to be working with Nvidia for three or four generations in advance we we had this with Celestica Celestica was a contract manufacturer and this has been a disaster industry since 1999. It went all offshore to China. It was commodity, but they hung on and they they kind of kept Celestica's heritage was IBM supercomputing and they kept all that talent and skill. [snorts] And then we noticed they were the sole supplier of the Google TPU server. We're [snorts] like, "Oh my god, this was like three years ago. The stock was trading at eight times earnings." And they had this whole and then they also had this whole business of selling Ethernet white box which is code word for commodity white box Ethernet switches into the clouds.
提前三四代一起做研发。我们在 Celestica 上就碰到过这种情况——Celestica 是做代工制造的,这个行业从 1999 年以来一直是个灾难。全都搬到中国去了,全是大宗商品化的,但他们撑住了,而且他们某种程度上保住了 Celestica 的传承——它源自 IBM 的超级计算业务,他们把那些人才和技术都留住了。(吸鼻子)然后我们注意到,他们是谷歌 TPU 服务器的独家供应商。我们(吸鼻子)当时就想:「我的天。」这大概是三年前的事,那时股票的市盈率才 8 倍。他们还有一整块业务是把以太网白盒——白盒就是大宗商品的代号——白盒以太网交换机卖给各家云厂商。
便签笔记
51:58
It it turns out that these are excellent businesses. Not only do they have tremendous growth, but to do an AI uh server computer, it's it's liquid cooled. It's running so much hotter and you know it's two or $300,000 piece of machinery whereas an old server was $5,000. If it breaks you just throw it away. If this thing breaks the whole thing goes down. So you become like critical infrastructure like selling a critical part on a plane. You'll never get swapped out. And then they they it turned out they were quite good at liquid cooling and you know a lot of other people tried to do it and failed and so they've retained that position.
结果发现这些其实都是极好的生意。它们不但增长惊人,而且要做一台 AI服务器,那是液冷的,运行温度高得多,而且这是一台二三十万美元的设备,而老式服务器才 5000 美元。坏了直接扔掉就行。但这玩意儿一坏,整套系统就停摆。所以你就变成了关键基础设施,就像卖飞机上的关键零件一样,你永远不会被换掉。然后事实证明他们在液冷上相当在行,你知道很多别的公司试过都失败了,所以他们守住了这个位置。
便签笔记
52:42
Then it also turned out that the Ethernet market was because you were in the old days you would go from 100 gig to 400 to 800. It would be a 7-year cycle to upgrade. Now they're upgrading every year and that's really hard to do. Then there's a whole software layer, the open source sonic layer. The the guys at at Celestica invented were some of the people that wrote that open- source software. They work very closely with Broadcom. So what we thought was just a great growth driver turned out to be great competitive advantages and they have like 50 60% share of the cloud Ethernet switch market which is a crucial market for um AI because AI is incredibly network intensive. And then even something like the printed circuit board. I mean a regular server you need 10 layers. These AI servers you need a 40 layer and there's very few PCB suppliers that can make this. And um there's all kinds of complexities in there. And we also own Elite Materials which makes the leading ingredient which is copper clad laminate which goes into
再然后又发现,以太网市场也是——因为过去你从 100G 升到400G 再到 800G,升级周期要七年。现在他们每年都要升级一次,这非常难做到。再往上还有一整层软件,就是开源的 SONiC 那一层。Celestica 的那些人里,就有当年参与写这套开源软件的人。他们和博通合作非常紧密。所以我们原本以为只是一个很棒的增长驱动力,结果发现那是极强的竞争优势,而且他们在云以太网交换机市场占了大概 50% 到 60% 的份额,这对AI 来说是个至关重要的市场,因为 AI 对网络的消耗极其巨大。然后哪怕是印刷电路板这种东西——普通服务器你需要 10 层,这些 AI 服务器你得要 40 层,而能做出来的 PCB供应商非常少。而且里面还有各种各样的复杂工艺。我们也持有 Elite Materials(台光电),他们做的是最核心的原料——覆铜板,就是用在这些板子上的。所以 PCB 的出货量在涨,层数
便签笔记
53:52
these boards. And so the PCB uh units are growing, the layer counts are rising. So you've got like a 50 to 60% keer just in the units and then the ASPs are rising and then the gross profits are rising and your visibility which used to be hey we'll call you next week if we need you to like hey we need you for the next four years to be like designing this road map with us. So you've you've gone from a 5% grow or low margin to you know a 35% 40 50 topline kager for the next four years with rising margins and then on top of that there's shortages of everything. So even if it is a commodity it's going to be a great cycle. So we see that up and down the supply chain. You find these companies like Corning like they make the fiber.
也在涨。所以光是出货量你就有 50% 到 60% 的复合增速,然后平均售价在涨,毛利也在涨,而且你的能见度——以前是「嘿,需要你的时候我们下周给你打电话」,现在变成「嘿,我们需要你在接下来四年里跟我们一起设计这条路线图」。所以你从一个 5% 增长、低毛利的生意,变成了未来四年 35%、40%、50% 的营收复合增速,同时毛利还在上升,而且在这之上,什么东西都短缺。所以哪怕它是个大宗商品,这也会是个很棒的周期。所以我们在整条供应链上下都看到了这种情况。你会找到这类公司,比如康宁,他们做光纤。
便签笔记
54:49
Um they've got some ridiculously high share of the fiber. I was reading this uh Microsoft data center they just built. There's enough fiber to circle the world four and a half times in that one thing. And their fiber is thinner and more bendable and can be specially manufactured to the exact specs. and it's higher margin and it's the fastest growing part of their business. And then they're doing, you know, in networking there's scale out which is kind of connecting all the server racks together. Then there's scale across which is connecting the data centers together. And when you want to build one of these huge clusters and you can't get all the power in one place for training, you want to wire them together. But the wires you need like 10x the wire has to be so much thicker. So that's creating huge growth. And where the real kicker comes in is when you do scale up. That's connecting every GPU in the rack to the other ones. That's done over copper.
嗯,他们在光纤上的份额高得离谱。我在读微软刚建的一个数据中心的资料,光那一个项目里的光纤就足够绕地球四圈半。而且他们的光纤更细、更能弯折,还能按精确规格定制生产,毛利更高,是他们业务里增长最快的部分。然后他们在网络上还做,你知道,有 scale out,就是把所有服务器机架连在一起;还有 scale across,就是把数据中心之间连起来。当你要建一个这种超大集群,而训练所需的电力没法全放在一个地方时,你就得把它们连起来。但那些线缆,你需要 10 倍的量,而且线得粗得多。所以这带来了巨大的增长。而真正的爆发点在于 scale up,也就是把机架里每一颗 GPU 和其他 GPU 连起来。这现在是用铜做的。
便签笔记
55:54
Eventually that'll be done over fiber. when that happens that two to three X's Corning's opportunity. So you just have at every layer of of the rack, >> everyone's overwhelmed. >> Everyone's overwhelmed. But the story like in the power supplies, every Nvidia chip or rack uses n 50 to 125% more power. And like literally that drives the ASPs of Delta and Advanced Energy. I just I think it's it's I can't believe these stories when I hear I'm like wait so your ASPs are going to like go up 40% for the next four years in a row and it's higher margin. The broader picture is like we're going to be the AI demand if we're right with this L curve. We're already short, you know, the DRAM market, the NAN market, the PCB. We're already like 30, we're 30% short all these things as we are now.
最终这会改成用光纤做。那一天到来的时候,康宁的市场机会会翻两到三倍。所以在机架的每一层你都有——>> 每个人都应接不暇。>> 每个人都应接不暇。但就像电源这块的故事,英伟达每一代芯片或机架的耗电量都要多出 50% 到 125%。而且这实实在在地推高了台达和 Advanced Energy 的平均售价。我就是……我听到这些故事的时候简直不敢相信,我心想:等等,所以你的平均售价接下来连续四年每年都要涨 40%,而且毛利还更高?更大的图景是,如果我们对这条 S 曲线的判断没错,AI 需求会……我们现在已经短缺了,你知道,DRAM 市场、NAND 市场、PCB,我们现在就已经短了 30%,就按当下的情况算。
便签笔记
56:59
>> As your business scales up, everything gets more complex, especially your compliance and security needs. With so many tools offering band-aids and patches, it's unfortunately far too easy for something to slip through the cracks. Fortunately, Vanta is a powerful tool [music] designed to simplify and automate your security work and deliver a single source of truth for compliance and risk. There's a reason that Ramp, Cursor, and Snowflake all [music] use Vanta. It frees them to focus on building amazing differentiated products, knowing that compliance and security are under control. Invest like the best listeners get a special offer of $1,000 off Vanta when you go to vanta.com/invest.
>> 随着你的业务规模扩大,一切都会变得更复杂,尤其是合规和安全方面的需求。市面上那么多工具都只是打补丁、贴创可贴,不幸的是,出纰漏实在太容易了。好在,Vanta 是一个强大的工具(音乐),它专为简化和自动化你的安全工作而设计,为合规和风险管理提供单一可信数据源。Ramp、Cursor 和 Snowflake 都在用 Vanta(音乐),是有道理的。它让他们能专心打造出色的、有差异化的产品,同时确信合规和安全都在掌控之中。《Invest Like the Best》的听众可以享受专属优惠:访问 vanta.com/invest,Vanta 立减 1000 美元。
便签笔记
57:33
I know firsthand how complex the tech stack is for asset management firms. And seemingly every new tool and data source makes the problem even worse, adding more complexity, more headcount, [music] and more risk. Ridgeline offers a better way forward. One unified platform that automates away that complexity across portfolio accounting, reconciliation, reporting, [music] trading, compliance, and more, all at scale. Ridgeline is revolutionizing investment management, helping ambitious [music] firms scale faster, operate smarter, and stay ahead of the curve. See what Ridgeline can unlock for your firm. Schedule a demo at ridgeline.ai. [music] >> The measure of percent AI, percent market share. Do you care more about the absolute or the rate of change of those metrics?
我亲身了解资产管理公司的技术栈有多复杂。而似乎每加一个新工具、新数据源,问题就更糟一分,带来更多复杂度、更多人力(音乐)、更多风险。Ridgeline 提供了一条更好的路:一个统一的平台,把投资组合会计、对账、报表(音乐)、交易、合规等等环节的复杂度自动化掉,而且能规模化运行。Ridgeline 正在变革投资管理,帮助有野心的(音乐)公司更快扩张、更聪明地运营、始终领先一步。看看 Ridgeline 能为你的公司释放什么价值。到 ridgeline.ai 预约演示。(音乐)>> 关于 AI 占比、市场份额占比这两个指标,你更在意的是绝对值还是变化率?
便签笔记
11框架为何难复制,以及风险清单
58:11
>> It's good because I took we did this presentation in two 2024 where we actually listed everybody's market share and everybody's and then I I asked Claude to plot plot it to a thing and it actually didn't get it right because what it didn't get is the rate of change. So the rate of change is important and that's incredible too because you go from 10% to 30% and your growth rate accelerates and your margins accelerate. So rate of change is very important. >> Why don't more people get this right in public markets? Like if your whole framework is S-curve, competitive advantage, underappreciated earnings power. It feels like the movie's been played out a lot over the last 25, 30 years.
>> 这问题问得好,因为我们在 2024 年做过一个演示,里面实际列出了所有人的市场份额,然后我让 Claude 把它画成图,结果它其实没画对,因为它没抓住变化率这一点。变化率是很重要的,而且这也很惊人,因为你从 10% 涨到 30%,你的增长率会加速,毛利率也会加速。所以变化率非常重要。>> 为什么公开市场上没有更多人搞明白这一点?如果你的整个框架就是 S 曲线、竞争优势、被低估的盈利能力,感觉这部电影在过去二十五、三十年里已经放过很多遍了。
便签笔记
58:52
>> My mom said, "Why do you tell everyone your secret? [laughter] It's like it's why does the casino teach people how to play blackjack? It it's harder. It's really hard to do. It's it's you have to have a a deep you have to be comfortable investing. You know, we've been doing I've been doing tech for 20 years at Whale Rock. We've got a team that's been doing this, covered many cycles. We know the different. So, very few people, no one's paid attention to hardware and chips at all. So, you've got all these newbies coming into it.
>> 我妈说:「你干嘛把自己的秘诀告诉所有人?」(笑)这就像……赌场为什么要教人怎么玩二十一点?因为更难。这真的很难做。你必须要有很深的……你得能坦然地做这种投资。你知道,我们一直在做……我在 Whale Rock 做科技已经二十年了。我们有一个团队一直在做这个,经历过很多轮周期。我们知道其中的差别。所以,真的没什么人——根本没人关注过硬件和芯片。所以进来的全是一堆新手。
便签笔记
59:23
>> You and Gavin, that's it. and Gavin's done a great job. people weren't comfortable with it and it's it's harder to do than it seems and the chart you know a lot of these companies their charts are up so it's scary can I buy and then you also have to have the holistic view because if you don't have conviction so every you know every time with Nvidia over the last four years it's like oh they had a great year oh my god it's got to be a bubble and then they had another great year and it's like 6 months of marking time it's got to be a bubble this is like getting out of hand this is pretty scary like and the the bare cases are not like totally without merit but if you can see the whole picture and understand how these things are unfolding and gain conviction in that frankly if you're just a semi- analyst so many semi analysts missed it because they didn't see what was really happening at the foundational model layer so it helps to have have the big picture it helps to have you know decades and scores of scurves that
>> 就你和 Gavin,就这两个人。Gavin 也做得非常好。大家对这块不自在,而且它比看上去要难做,还有那些图表——你知道很多这类公司的股价图都是一路往上的,所以挺吓人的,我还能买吗?然后你还得有一个整体的视角,因为如果你没有信念……你知道,过去四年里英伟达每一次都是这样:哦,他们有了很好的一年,我的天,这肯定是泡沫;然后他们又有了很好的一年,然后横盘六个月,这肯定是泡沫,这有点失控了,挺吓人的;而且那些看空的理由也不是完全没道理,但如果你能看到全局,理解这些事情是怎么一步步展开的,并因此建立起信念——坦白说,如果你只是个半导体分析师,很多半导体分析师就错过了,因为他们没看到基础模型层真正在发生什么。所以有大局观是有帮助的,有几十年的经验、看过成百上千条 S 曲线、知道它们在
便签笔记
60:25
you're looking at and and where it plays in different things. >> What what in this whole picture, you know, I would describe your your stance so far in the first hour discussion as like very bullish on on the impact that AI is going to have and the returns available as a result. What makes you the most concerned or uncertain or is it just the rate at which all this stuff changes and like what keeps you worried amidst what seems like pretty extreme bullishness? I mean, one thing that bothers me is there's a lot of negativity in the general population about AI and there's a lot of negativity in some aspects of the government. You know, I think Maine just banned data centers and 80 only 20% of the people are optimistic about AI and potential for negative regulation. But I do think kind of the genie is out of the bottle. Another risk is that if AI sort of slows down in its improvements, I think there's a whole lot of AI adoption to happen even if the models didn't improve. But Jensen said this, you know, years ago when he was
不同领域怎么演绎,这都有帮助。>> 在这整幅图景里,你知道,我会把你在这第一个小时讨论里的立场描述成对 AI 将带来的影响、以及由此可获得的回报都非常看多。那什么让你最担心或最不确定?还是说就是这一切变化的速度?在这看起来相当极端的乐观之下,有什么是让你放心不下的?我是说,有一件事让我不安:普通大众对 AI 有很多负面情绪,政府某些方面也有很多负面情绪。你知道,我觉得缅因州刚刚禁掉了数据中心,而且只有 20% 的人对 AI 持乐观态度,还有出台负面监管的可能。但我确实觉得这瓶子里的精灵已经放出来了。另一个风险是,如果 AI 的进步速度慢下来——我觉得即便模型不再进步,也还有一大堆 AI 应用的普及要发生。但黄仁勋很多年前说过,当时他
便签笔记
61:31
talking about his GP crap, just the graphics chips. If good enough is good enough, I won't have a business. Now every year he made the graphics a little bit better and people always wanted the best in AI. If anthropics sort of hits a wall and stops improving or open AI then the open source models will catch up and um and then it might be a race to the bottom and it might be you know it won't be good for the stocks probably. It could be good for the chip companies. chip companies don't care >> who's winning tokens, right?
讲的是他的图形芯片这摊「破事」:如果「够用就行」真的够用了,我就没生意了。而他每年都把图形做得更好一点,人们也总是想要最好的。在 AI 里也一样,如果 Anthropic 撞上了天花板、不再进步了,或者 OpenAI 也是,那开源模型就会追上来,然后可能就是一场逐底竞争,那样的话,对股票来说大概不是好事。但对芯片公司可能是好事。芯片公司不在乎 >> 谁在 token 上赢,对吧?
便签笔记
62:08
>> Who wins. >> So, that's another positive and they'll benefit if if open source, you know, Jensen really wants open source to like take off. It's all he kept on mentioning at at his last GTC. So, that could be a risk. Another thing is if one or two of the players falters and loses its position and can't compete, that could be like a lot of compute that they don't need in the future. Now, if AI is so big, somebody else will suck that up. And we saw that with, you know, Oracle cancelled a big deal and then Meta went right in. But let's just say Meta decided not to be involved with AI. Hey, we can't keep up. It's just going to be a waste of our resources. So, we we watch that very carefully and um in general, we see more, you know, more more companies truly going after this and even Microsoft going trying to build their own. So I think those are those are some of the key risks.
>> 谁赢。>> 所以这是另一个利好,如果开源真的起来了,他们也会受益,你知道的,黄仁勋非常希望开源能够真正发展起来。这也是他在上一次 GTC 上反复提到的。所以这可能是一个风险。另一件事是,如果一两家玩家失利、丢掉自己的位置、竞争不下去了,那可能意味着未来会空出大量他们用不上的算力。当然,如果 AI 真有那么大,别人也会把这些算力接过去。我们也确实看到了,你知道,甲骨文取消了一笔大单,然后 Meta 马上就顶了上去。但假设 Meta决定不再涉足 AI,说'嘿,我们跟不上了,这就是在浪费我们的资源'。所以我们会非常仔细地关注这一点,而且总体上我们看到越来越多的公司真正在往这个方向发力,连微软都在试图自己造芯片。所以我认为这些就是一些关键风险。
便签笔记
63:05
>> Seems like you really have done very little in the application layer of AI. Historically the apps ended up being most of the market cap you know not not the infrastructure and there wasn't really a model layer in the past. I guess you could say it was the clouds or something. >> Yeah. >> Why focus so much on the bottom layers of Jensen's five layer cake versus things in the application layer that are actually getting used by consumers? Well, we do, you know, part of OpenAI is they have chatbt which which is an application, but we think the application layer well a it always comes later. So, you know, the first three or four years of the iPhone and then the applications really took time. So, maybe it's just starting. Um but to date um we found that area to be pretty risky because where does the where does the foundational model end and where does the application begin and can can the applications build enough of a moat um where they can fend off and um and build and build businesses in that. and um and
>> 看起来你们在 AI 的应用层几乎没怎么布局。从历史上看,最后大部分市值都落在应用上,而不是基础设施,而且过去也并不真的存在一个模型层。我想你可以说那时候相当于是云厂商之类的吧。>> 是的。>> 为什么这么专注于黄仁勋'五层蛋糕'的底层,而不是应用层那些真正被消费者用起来的东西?其实我们也有涉及,你知道,OpenAI 有一部分就是 ChatGPT,那就是一个应用,但我们认为应用层——首先,它总是来得更晚。你知道,iPhone 头三四年,应用真正起来是花了时间的。所以也许现在才刚刚开始。嗯,但到目前为止,我们觉得那个领域风险相当大,因为基础模型的边界在哪里结束、应用又从哪里开始?应用能不能建立起足够深的护城河,让它们能抵御竞争、并在此之上做成生意?嗯,我们本以为会在一些
便签笔记
64:10
we we thought we would see it in some of the incumbents like a a CRM and they're starting and maybe just a matter of time but we really haven't seen it in the enterprise world and there there are some you know very good startup application companies out there but the ecosystem is not clear you know like when we started the ecosystem and chips was clear when we started the foundational model ecosystem wasn't clear now it's clearer to us and at the application layer it's still kind of unclear and a little bit dangerous because um but there will be great application companies built you know we really were watching Brett Taylor at Sierra Brett was CEO of CRM he wrote Google Maps he was CIO of Facebook and uh he he's building this fantastic company called Sierra we're not involved but that's where the rubber hits the road will he be able to turn this into a huge company and he's doing quite well.
现有巨头身上看到,比如 CRM 类公司,它们已经开始了,也许只是时间问题,但我们在企业级市场还真没看到。当然,外面是有一些非常优秀的创业型应用公司的,但生态还不清晰,你知道,就像我们当初进入芯片领域时,生态是清晰的;我们进入基础模型时,生态还不清晰,现在对我们来说更清晰了;而在应用层,现在仍然有点模糊,也有点危险,因为——不过肯定会有伟大的应用公司诞生的,你知道,我们一直在关注 Sierra 的 Brett Taylor,Brett 当过 CRM(Salesforce)的 CEO,他写过 Google Maps,还当过 Facebook 的 CTO,呃,他正在做一家非常棒的公司叫 Sierra,我们没有投,但那就是真正见真章的地方——他能不能把它做成一家巨型公司,而他目前做得相当不错。
便签笔记
12研究机器、产品线与父亲
65:09
We'll see. It's a matter of timing when these things really start to to to come in into their own and prove they're sustainable. It usually doesn't start in the first 3 or 4 years. It comes a little bit later. >> At your office, you have this this giant uh award wall for the research. I can't remember what it's exactly. It's for the best research job or project of the year given to an analyst. And I think you won it. you gave it self awarded in their own when you're by yourself, but you've got this now long 20 year history of a year one or one or more people, you know, put their name on this wall for having done the best job on a research project that year. I'm so curious about the nature of that research and how it's changing as a result of all of this.
我们拭目以待。这些东西真正开始成气候、并证明自己可持续,是个时机问题。通常不会在头三四年就开始,而是会晚一些才来。>> 在你们办公室里,有一面巨大的研究获奖墙。我记不清具体叫什么了。是颁给分析师的年度最佳研究项目之类的奖。我记得你自己拿过一次,是你自己给自己颁的,在他们的自己一个人的时候……但你们现在已经有了长达20年的历史,每年会有一位或几位人把名字刻在这面墙上,表彰他们当年在某个研究项目上做得最出色。我特别好奇那种研究的本质,以及在这一切(AI)的影响下它正在如何变化。
便签笔记
65:51
Say, you know, the person that's going to win the award this year and the sort of work that that requires a human to do when so much of the work that probably would have won you the award in, I don't know, 2009 or something could probably be fully automated or done in an hour with cloud code or something today. How is the nature of research and what gets you on that whale rock award wall changing in real time? I would like to say that we're so advanced in our AI systems that it's a huge change so far.
比如说,今年会拿这个奖的那个人,以及那种工作到底需要人做些什么,因为在2009年前后大概能让你拿奖的很多工作,今天可能已经能被完全自动化,或者用Claude Code之类的东西一小时就搞定。研究的本质、以及能让你登上Whale Rock荣誉墙的东西,正在实时发生怎样的变化?我很想说我们的AI系统已经先进到带来了巨大的改变,但目前还没到那一步。
便签笔记
66:20
I mean, it's it's helping us get up to speed and we have a handful of of great apps, but it's not yet it's not supplanting the job of the analysts. And so much of what we're doing is we're meeting with as many companies as humanly possible. We're developing relationships with with the management teams that we cover. We're talking to the competitors. The system we use is right out of common stocks and uncommon profits which was written by Philip Fish Fischer in the 1950s. And it's the scuttlebutt approach. It's growth investing. It's it's get out there and talk to suppliers, uh, customers, competitors, looking for the key characteristics of these leading companies and really developing conviction in them. Now, if it's a new complicated area like ABF substrates or PCBs, we're able to get up to speed on those things quickly, but it can't pick stocks for you in any kind of a way. I will say that, you know, if you're an analyst who's good at the blocking and tackling and there's a role for that,
我是说,它确实帮我们更快上手,我们也有几个很棒的应用,但它还没有取代分析师的工作。我们做的事情里,很大一部分是尽可能多地去见公司。我们在和我们覆盖的管理层建立关系。我们和竞争对手交流。我们用的这套方法直接来自《怎样选择成长股》,那是费雪(Philip Fisher)在1950年代写的。就是所谓的“闲聊法”(scuttlebutt)。这是成长股投资。就是走出去,去和供应商、客户、竞争对手聊,寻找这些领先公司的关键特质,真正建立起对它们的信念。现在,如果碰上像ABF载板或PCB这种复杂的新领域,我们能很快把这些东西搞明白,但它没法以任何方式替你选股。我想说的是,如果你是一个擅长做基础功、扎实执行的分析师,这当然有它的价值,
便签笔记
67:24
but that role is you need to have obviously the insight on top. So, we're now like using AI to write notes uh or you know review the quarter or and those notes are much better but there better be a really good paragraph on top which is the wisdom. What does this mean? How does this deal with our thesis? Um what changed? You know, don't just be a reporter. Um so the AI can be a great reporter. It can't it can't quite pick into the future. And like the job that the guys did on app 111 two years ago. I mean I think we got two of the best adte guys around and they you know they convinced me to buy I knew adte I started actually nearby here in New York at at that internet advertising startup and after I did banking.
但这个角色的前提是,你显然还得在上面加上洞察。所以我们现在会用AI来写纪要,或者做季度回顾之类的,那些纪要质量好多了,但最上面那一段最好得非常出色,那才是智慧所在。这意味着什么?这和我们的投资逻辑有什么关系?有什么变了?别只当一个记录员。AI可以是个很棒的记录员,但它还没法真正预见未来。就像两年前我们几个人在AppLovin上做的那个工作。我觉得我们有业内最好的两位广告科技分析师,他们说服我买入。我本来就懂广告科技——我做完投行之后,其实就在离这儿不远的纽约一家互联网广告创业公司起步的。就是我做完投行之后。
便签笔记
68:16
So I knew internet advertising and ad tech which is historically a terrible industry. Um, but Michael and Sam really figured out the Apploven story like before anybody and they followed it when it was private. They know all the competitors. They know all the intricacies of, you know, there's all this terminology and um they, you know, Sam went to the Las Vegas app advertising conference and we went to con and, you know, we talked to scores and scores of of people. So um and they did the work on the model and developed a great relationship with Adam Ferogi.
所以我了解互联网广告和广告科技,而这历来是个很糟糕的行业。但Michael和Sam真的比所有人都更早看懂了AppLovin的故事,它还在私有阶段时他们就一直跟踪。他们了解所有竞争对手,了解所有细枝末节,你知道,这行有一堆专门术语,而他们……Sam去了拉斯维加斯的应用广告大会,我们也去了那个会,跟几十上百号人聊过。所以他们在模型上下了功夫,还和Adam Foroughi建立了很好的关系。
便签笔记
68:54
He's one of the best managers out there. And um I don't see AI doing that. >> What role does talking to other investors outside of your firm play in your life? Like >> I one of the great things is just the friendships I've built with so many smart investors and and frankly Philip Fischer said part of his process was like get to know a good 10 or 15 like-minded people around the country and share ideas and um and a you know they're great great friends to make a lot of them have been on your podcasts and uh and and you develop good friendships and and you you share ideas, talk ideas. It's important that it's a two-way street. Um, I call it the tripod. When I like something and then my analyst likes it and then somebody who I really respect also likes it. That's three legs of the stool can really help the conviction.
他是业内最优秀的管理者之一。这种事,我看不出AI能做到。>> 和公司之外的其他投资人交流,在你的生活里扮演什么角色?>> 一件很棒的事,就是我和这么多聪明的投资人建立起的友谊。坦白说,费雪也讲过,他流程的一部分就是结识全国各地十几个志同道合的人,互相分享想法。而且他们也是很好的朋友,其中不少人上过你的播客。你会建立起很好的友谊,互相分享想法、讨论想法。重要的是这必须是双向的。我把它叫做“三脚架”。当我看好某个东西,我的分析师也看好,然后一个我非常尊重的人也看好,这三条腿撑起来,真的能极大增强信念。
便签笔记
69:54
>> What have you learned about shaping the products that you offer your investors across the history of the firm? It's not just one monolithic structure anymore. >> There's there's several things that if I'm an investor and I want to give you money, I can there's a couple ways I can do that. >> How did you arrive at those things? And and how do you how could you turn that experience into um advice for other investors that are trying to provide their LPs with the right set of options? >> For the first 15 years, it was a long short fund and we you know, you want to be focused and if you defocus that can be hard. So we we grew that and we got that to the scale that we wanted to.
>> 在公司这些年里,关于如何设计你提供给投资人的产品,你学到了什么?现在已经不是单一的一种结构了。>> 有好几种方式,如果我是投资人,想把钱交给你,我可以有几种不同的做法。>> 你是怎么走到这些产品形态的?你又能把这些经验变成什么样的建议,给其他想为自己的LP提供合适选项的投资人?>> 头15年我们只有一只多空基金。你知道,你需要专注,如果分散注意力,会很难做好。所以我们把它做大,做到了我们想要的规模。
便签笔记
70:32
We're 20 years old, maybe 10 years in, people started to ask for a long only product. And so in 2020, we we launched uh the long only fund. So we're 6 years on that. And um that's now larger than the long short. The bulk of the assets are in these two. In maybe 2015, we we formalized that we might be doing privates. And so we gave investors the option to opt in or opt out and you could do 15% or 25%. So, but we didn't break the seal on the privates until 2020. In 2021, we offered um a hybrid fund that could be 80% uh into privates. sort of similar approach but if you wanted more exposure to privates. Um and then very recently we launched the whale rock meggaap tech fund and we just think there's a huge structural underweight of the largest tech companies in the world because a we also realize that a a lot of our performance over the years was from some of the largest companies whether it be Apple or Amazon or Tesla and and people just it's hard to overweight these to the to the amount. And so a lot of our
我们成立20年了,大概第10年的时候,开始有人要求做纯多头产品。于是在2020年,我们推出了纯多头基金。到现在6年了。而且它现在的规模已经比多空基金还大。绝大部分资产都在这两只里。大概2015年,我们把可能会做一级市场投资这件事正式化了。我们给投资人选择加入或退出的权利,你可以选15%或25%的比例。但我们直到2020年才真正开始做一级投资。2021年,我们推出了一只混合基金,最高可以有80%投向一级市场,思路类似,但适合那些想要更多一级市场敞口的人。然后就在最近,我们推出了Whale Rock大型科技股基金,因为我们认为全世界最大的那些科技公司存在巨大的结构性低配。一方面,我们也意识到我们这些年很大一部分业绩,其实来自一些最大的公司,无论是苹果、亚马逊还是特斯拉,而人们很难把它们超配到应有的比例。所以很多最大的资金池,比如捐赠基金之类的,
便签笔记
71:54
largest pools of capital endowments or what have you, they realize they they have been massively underweight, the largest tech companies in the world for the last because they only have, you know, they have a lot of privates. They don't have a ton of public and then maybe half the public is international. And then of their public bucket, they don't want to they there's a belief that there's no alpha in large cap. So they underweight large cap and they have a lot of small and mid managers that are stock pickers because it's intuitive that large cap can't have alpha. Um and then in their hedge fund portfolio, even if it's long bias, they're not going to have 15% and Nvidia and all these other things. And we realize that there's a huge that this people are worried that there's these big companies. This is just a product of the digital economy in that, you know, in tech, the leader usually grows bigger and wins and develops very high market share quickly and and there's great competitive advantages and and they're also selling
他们意识到自己这些年来对全球最大的科技公司严重低配,因为他们的配置里一级市场占了很多,二级市场并不多,而且二级市场里可能有一半是国际资产。然后在他们的二级市场那块里,他们又相信大盘股没有阿尔法。所以他们低配大盘股,转而配了一大堆做选股的中小盘管理人,因为直觉上会觉得大盘股不可能有阿尔法。而在他们的对冲基金组合里,即便是偏多头的,也不会拿15%的英伟达之类的仓位。我们意识到这里存在一个巨大的……人们担心这些大公司太大了。但这其实就是数字经济的产物——在科技行业,领先者通常会越长越大、赢家通吃,很快形成极高的市场份额,而且有很强的竞争优势,同时还在
便签笔记
72:57
around the globe. So, this is going to lead to massive profit pools and massive market caps and it's just going to happen into the future. And so, most endowments are betting against this. They're they're because they're completely underweight this. And finally somebody came to us and said you know what should we do which index we got to and I'm on the board of Hamilton College and they were trying on their investment committee they were trying to figure out this and so we kept on hearing it and finally uh one of our clients was like we said we'll we'll do this for you because there's a lot of alpha to be had and the mag 7 or the fang or whatever it's going to be different and you know in 2022 they all rallied but like last year they were very divergent and this year they're down and so we created the the Whale Rock Mega Cap Tech Fund which is the top 30 the universe is the top 30 market caps globally and then we pick you know the 12 or 13 that are the best and I think there's tremendous alpha in
全球范围内销售。所以这必然带来庞大的利润池和庞大的市值,而且未来还会继续如此。因此,大多数捐赠基金实际上是在做空这个趋势,因为他们对此完全低配。终于有人来找我们说,我们该怎么办、该用哪个指数。我也在汉密尔顿学院的董事会,他们的投资委员会也在琢磨这个问题。我们不断听到这样的声音,最后有一位客户提出来,我们就说,我们可以为你做这件事,因为这里有很多阿尔法可赚。而且“七巨头”或者“FANG”之类的组合,未来构成会不一样。你知道,2022年它们全都上涨,但去年它们的走势分化很大,今年又在下跌。所以我们创立了Whale Rock大型科技股基金,它的股票池是全球市值前30的公司,然后我们从中挑出最好的12到13家。我认为最大市值的公司里有巨大的阿尔法,因为你想想,
便签笔记
74:00
the largest cap because if you think about it a small cap it just takes one person to to figure out it's good and move it up but it takes a hundred people, 100 diversified PMs to realize Google's not a loser, it's a winner. And can we figure that out before 95% of those generalist PMs do it? And you know, we've been able to do it. >> We like your odds in that. >> Yeah, we like your odds in that. And so there is alpha to be had there. And then as an asset category, it's great because these companies by definition have wonderful modes and maybe they're not the super S-curve, but sometimes they are. I mean, Nvidia sure is, and TSM is really levered to it, and Heinix is extremely levered to it, and ASML is levered to it. So, it's a great um so that's a new we're four months into that one. And so the right way maybe to think about it, it it it sort of sounds like really what you've built is a research machine to understand the world through the lens of companies and that the thing you're constantly trying to improve is
小盘股只需要一个人发现它好、把它买上去;但要让一百个人、一百位分散化的基金经理意识到谷歌不是输家、而是赢家,这就难多了。那我们能不能在那95%的通才基金经理之前想明白这件事?事实证明我们做到了。>> 这方面我们看好你们的胜算。>> 是的,我们看好你们的胜算。所以那里确实有阿尔法可赚。而且作为一个资产类别,它非常好,因为这些公司从定义上就有极好的护城河,也许它们不处在那种陡峭的S曲线上,但有时候确实是。我是说,英伟达肯定是,台积电和它高度绑定,海力士也极度受益,ASML也一样。所以这是个很棒的……这是新产品,我们才做了四个月。所以,也许正确的理解方式是——听起来你真正打造的其实是一台研究机器,用公司的视角去理解世界,而你一直在努力改进的就是
便签笔记
75:10
that research machine and then the way that you would then express that through products is multiplied. But if I was to try to understand Whale Rock, it would be to investigate the research machine first and foremost. We call it the whale rock learning machine and it's a group of 10 highly experienced individuals that you know Warren Buffett reads books and we read books and we read blogs and we but we're also in tech you got to go out and talk to people. So we do 2500 3,000 facetoface meetings with management teams and you know Mer and Buffett talk about compounding knowledge. We've been compounding that knowledge for 20 years. you know, there's changes to the team, but broadly there's a lot of um consistency to it.
这台研究机器,然后你通过产品把它的成果放大表达出来。但如果我要理解Whale Rock,首先要研究的就是这台研究机器。我们把它叫做Whale Rock学习机器,它由10位经验非常丰富的人组成。你知道,巴菲特读书,我们也读书、读博客,但我们身处科技行业,你必须走出去和人交流。所以我们每年做2500到3000场与管理层的面对面会议。芒格和巴菲特讲知识的复利,我们已经把这种知识复利积累了20年。团队会有人员变动,但整体上有很强的连续性。
便签笔记
75:53
Um Andrew and Michael have been with me for 19 and 18 years and the average experience level on the team is 10 or so years and that includes some of the new newer people. And uh yeah, that research engine can support all these all these products and it's the same people that do publiclix and the private. So, we're not going to scour the world and turn over every A B. But when we see something that fits into our system, we're able to act on it. >> It's so much fun to do this with you. When I do this, I ask the same traditional closing question of everybody. What is the kindest thing that anyone's ever done for you?
Andrew和Michael已经跟了我19年和18年,团队的平均从业经验大概是10年左右,这还包括一些较新的成员。是的,这台研究引擎可以支撑所有这些产品,而且做二级和一级投资的是同一批人。所以我们不会去满世界搜罗、翻遍每一块石头。但当我们看到符合我们体系的东西时,我们能够立刻行动。>> 能和你聊这些真的太有意思了。每次做这个节目,我都会问每位嘉宾同一个收尾的老问题:别人为你做过的最善良的事是什么?
便签笔记
76:30
>> Well, I got to say it's definitely my father who, you know, I was super lucky. My father um graduated Cornell a double e electrical engineering, pivoted to Wall Street and um uh had a great career at Goldman Sachs and he was um he ran corporate finance in the 80s and then ran private equity as chairman in the 90s and uh he was just whips smart but he he had such humility and was such a great gentleman and uh when I started Whale Rock, you know, friends and family, he was the first call, but he said, you know, I've been at Goldman for for 41 years. How about I come and join you? I'll be the gray hair. I'll be the oversight. I'll be the chairman. You do what you do. You build the firm in uh Boston. Build the team, run the money, I'll help raise some money. And we got to work together for 6 years until he passed away in 2011. But I just feel so lucky to have worked with him. You know, it's not easy running a fund. We never raised our voice. And he was just an amazing mentor to so many people. And
>> 嗯,我得说,那肯定是我父亲,你知道,我真的特别幸运。我父亲从康奈尔大学电气工程专业毕业,后来转行去了华尔街,在高盛有过一段非常出色的职业生涯。八十年代他负责公司金融业务,九十年代又以董事长的身份掌管私募股权业务。他非常非常聪明,但同时又特别谦逊,是一位真正的绅士。当我创办 Whale Rock 的时候,你知道,找亲朋好友募资嘛,他是我第一个打电话的人。但他说,我在高盛已经待了 41 年了,不如我来加入你吧?我来当那个'白头发',我来做监督,我来当董事长。你就做你擅长的事,在波士顿把公司建起来,把团队搭起来,管好资金,我来帮忙募一些钱。我们一起共事了 6 年,直到他 2011 年去世。但我真的觉得能和他一起工作太幸运了。你知道,管理一只基金并不容易。我们从来没有对彼此提高过嗓门。他还是那么多人的一位了不起的导师。他去世之后,我收到了非常多的来信,人们说:'你父亲对我影响太大了。
便签笔记
77:43
when he passed away, um I got so many letters from people who said, "Your father was just such an influence on me. He was such a gentleman. He was such a great mentor to me." And so I just feel so lucky uh to have worked with him. And if I could be half the person that he is, I'd be completely winning. And >> how did he do that? How did he What was his method? Why did so many people say that? >> Um I don't know. He He just He was He was modest. He was whipsmart. He was wise. He was also known as a um a great investor, which isn't the most common thing at a lot of investment banks. He also was on their commitments committee and kept him out of a lot of tougher situations and yeah he was very warm and he people would could go into his office with with with problems and he handled it handled it with grace and um whether it's a personal problem or a work issue or what have you and he just had this soft way and he also had a great sense of humor.
他是一位真正的绅士,是我非常好的导师。'所以我真的觉得能和他共事非常幸运。如果我能做到他一半那样的人,我就完全算赢了。>> 他是怎么做到的?他有什么方法?为什么那么多人都这么说?>> 嗯,我也说不好。他就是……他很谦逊,非常聪明,也很有智慧。他还以出色的投资人著称,这在很多投资银行里并不算常见。他还是公司承诺委员会(commitments committee)的成员,帮公司避开了不少棘手的局面。是的,他非常温暖,人们可以带着问题走进他的办公室,而他总能从容优雅地处理好,不管是个人的烦恼还是工作上的事,或者其他什么。他就是有那种温和的方式,而且他还很有幽默感。
便签笔记
78:48
>> Lucky. >> Yeah. I'm so lucky. So [music] Alex, thanks so much for your time. >> Thanks so much. Your finance team isn't losing money on big mistakes. It's [music] leaking through a thousand tiny decisions nobody's watching. Ramp puts guardrails on spending before it happens. Real-time limits, automatic rules, zero firefighting. Try it at ramp.com/invest. [music] As your business grows, Vant scales with you, automating compliance and giving you a single source of truth for security and risk. Learn more at vanta.com/invest. [music] Ridgeline is redefining asset management technology as a true partner, not just a software vendor. They've helped firms 5x [music] and scale, enabling faster growth, smarter operations, and a competitive edge. Visit ridgelineapps.com [music] to see what they can unlock for your firm.
>> 太幸运了。>> 是啊,我太幸运了。那么 [音乐] Alex,非常感谢你抽出时间。>> 太感谢了。你的财务团队并不是在大失误上亏钱,而是[音乐]在成千上万个没人盯着的小决策中一点点流失。Ramp 在支出发生之前就设好了护栏:实时额度限制、自动规则、零救火。访问 ramp.com/invest 试用。[音乐] 随着业务增长,Vanta 也随你一起扩展,自动化合规流程,为你的安全与风险管理提供单一可信数据源。访问 vanta.com/invest 了解更多。[音乐] Ridgeline 正在重新定义资产管理技术——它是真正的合作伙伴,而不只是一个软件供应商。他们帮助多家公司实现了 5 倍[音乐]增长与规模扩张,带来更快的成长、更聪明的运营和竞争优势。访问ridgelineapps.com [音乐] 看看他们能为你的公司释放什么潜力。
便签笔记
79:38
Every investment firm is unique, and generic AI doesn't understand your process. Rogo does. It's an AI platform built specifically for Wall Street, connected to your data, understanding your process, and producing real outputs. Check them out at rogo.ai/invest. The best AI and software companies from OpenAI to cursor to Perplexity use work OS to become enterprise ready overnight, not in months. Visit works.com to skip the unglamorous [music] infrastructure work and focus on your product.
每家投资机构都是独一无二的,通用的 AI 并不理解你的流程。Rogo 懂。它是一个专为华尔街打造的 AI 平台,连接你的数据,理解你的流程,产出真正可用的成果。了解详情请访问rogo.ai/invest。从 OpenAI 到 Cursor 再到 Perplexity,最优秀的 AI 与软件公司都在用 WorkOS,一夜之间就具备企业级能力,而不用花上好几个月。访问 workos.com,跳过那些不起眼的[音乐]基础设施工作,专注于你的产品。
便签笔记
视频总结 · 一句话概括与核心要点

一句话概括

Whale Rock 创始人 Alex 用"S 曲线 + 竞争优势 + 被低估的盈利能力"三段式框架解释:AI 采用曲线陡峭到只能叫"L 曲线",企业级 AI 渗透率不到 1%、算力全球短缺,因此基础设施与硬件层、以及已在基础模型层胜出的三家寡头(Anthropic、OpenAI、Google)才是这轮最高确定性的机会,而应用软件股反而是受损方。

核心要点

  • 投资顺序是"先芯片、后模型、暂缓应用",理由是需求最先到达且赢家最易辨认。 2022 年 11 月 ChatGPT 发布后,团队按 Jensen 的五层栈(电力 → 芯片 → 云 → 基础模型 → 应用)拆解,2023 年初判断"不管上层谁赢,算力都是刚需",因此先押基础设施;直到 2025 年才对模型层收敛出结论。
  • 基础模型层从 60 家收敛为三家寡头,结构类比云计算的三家撑起整个 SaaS 世界。 创业公司几乎全部出局,Amazon"没真正出现",Meta 强势入场后受挫需要彻底重启;Anthropic 作为黑马专攻企业侧,OpenAI 拿下消费侧,Gemini 不可低估。开源(尤其中国模型)的风险被评估为可控:从 80 分benchmark 到 85 分是巨大解锁,而开源方算力不足,能逼近前沿但无法越过,随后掉队。
  • 模型不是商品,差异化是真实护城河。 Anthropic 在私募股权与金融类任务上突出,Google 擅长吞 PDF;路由器在多模型间切换的表象掩盖了训练方法与技能差异。更关键的是围绕 API 建的生态——SDK、Claude for Cowork、编排层、"harness"(榨干模型的外围软件)——这与 AWS 2013 年被误判为"仓库里的商品服务器"、实则靠自建产品慢慢锁死客户是同一剧本。
  • 代码是 AI 的真正解锁点,仅编码一项就是半万亿美元市场。 第一代工具(如 $20/月的 Copilot)只能改语法、找 bug、写一段代码;Anthropic 的产品能 agentic 运行。Anthropic 内部有人一天烧 $100 token,年化 2–3 万美元 × 全球约 2000 万程序员 = 5000 亿美元 TAM,而这还是基于 7–9 个月前的技术。佐证:Karpathy 与 Torvalds 一年内立场翻转——过去 20% 由工具写、80% 手写,现在 Karpathy 除英文外不再写一行代码。
  • 投资 Anthropic 的具体路径:错过 600 亿轮,180 亿估值那轮进场。 60 亿美元那轮时因不了解公司、毛利为负、尚未看到编码爆发而放弃。后来见了 Dario,判断管理团队优秀、几乎零流失、代码质量高;用 Claude Code 扒全网关于编码市场的反馈做了 90 页 PPT 递给公司,拿到超出体量的配额。收入路径:1 亿 → 10 亿 → 冲 90 亿,"从 1 亿到 10 亿是一回事,做到 90 亿是另一回事"。
  • 渗透率数据是全篇看多的支点:企业 AI 不足 1%,基础设施约 10%,真正深度用 AI 的知识工作者只有 10 个基点。 Anthropic 约 1400–1500 万 DAU,但其中只有小部分在真正"用 AI";多数人用的还是 AI 1.0(加强版搜索引擎)。预计四年内从 10bp → 1–2% → 3–5% → 15%。而算力已经全球短缺——Anthropic 目前只拿到所需的一半,这还是在大规模上量之前;Andreessen 认定未来四年唯一确定的事就是算力不够。
  • AI 采用速度像收音机而非洗碗机。 历史 S 曲线里收音机 7 年接近满渗透,洗碗机因为要接管道所以极慢;B2B/SaaS/云属于"洗碗机"型(要接进既有系统,30–50% 增速),而 AI 打开浏览器就能用,所以是垂直向上的"反 L 型"。风险对照组:B2B 互联网的牛市论调因底层基础设施不到位而推迟了 20 年才由 SaaS 兑现;AI 同样面临安全顾虑、IT 部门阻力等文化障碍(云也是靠 CIA 和 Capital One 背书才起飞)。
  • 应用软件股被从 40–50% 仓位清到净做空。 逻辑链:软件公司的 AI 产品不够好、卖不上价、不动指标;CIO 优先级下滑,预算转向 ROI 更快的 Anthropic token;提价能力受限;裁员/冻结招聘冲击席位数。Salesforce 有 400 亿营收但 AI ARR 只有 5–7 亿,基数太大。半熟的反向观点:Agent 需要工具,Claude 第一件事就是接 Slack,若成为关键仓库反而会把 Slack 变成组织内的永久设施;CRM 走向 headless,虽失去用户界面,但 agent 直接在 CRM 里干活反而巩固其地位。
  • 硬件的"去商品化"是这轮最被忽视的利润池。 过去 40 年数据中心几乎无变化(x86 主导,摩尔定律自动消化 25–40% 的负载增长,整链条商品化,1G→10G 要走 7 年)。现在负载每年 10 倍,每个环节被逼到物理极限:HBM 十层堆叠、IO 提升 10 倍且要提前三四代与 NVIDIA 协同;PCB 从 10 层升到 40 层,单位数量与层数、ASP、毛利同时上行;能见度从"下周需要再打给你"变成"未来四年一起做路线图";每代 NVIDIA 机架多耗电 50–125%,直接推高 Delta、Advanced Energy 的 ASP。案例:Celestica 当年 8 倍 PE,独供 Google TPU 服务器,兼具液冷能力与 50–60% 云以太网交换机份额;AI 服务器 20–30 万美元一台且液冷、坏了整机停摆,等于飞机上的关键件,不会被换掉。DRAM、NAND、PCB 目前均短缺约 30%。
  • 筛选指标是"AI 版 rule of 40":AI 收入占比 + 该品类市场份额,且变化率比绝对值更重要。 30% + 30% = 60 是好位置;软件公司的 AI 占比只有 1–2%,差得远。份额从 10% 走到 30% 时增速与利润率同时加速——这正是他让 Claude 作图时模型没抓住的那一层。
  • 买点靠直觉与实地证据,卖点看渗透率到 30–40%。 引 Andy Grove:"战略拐点时不能信数据",要靠轶事与视觉线索——在中国看到 12 岁小孩用大屏手机玩重度游戏(手游 S 曲线)、Gartner IT Symposium 上 Splunk/VMware/AWS 的会场连续几个时段爆满(企业需求可见化)。渗透率到 30–40% 后指数增长结束、卖方模型追上、不再有大幅超预期。承认的错误:2012 年美国智能机渗透 50% 时卖出 Apple,此后靠 App Store 30% 抽成仍能复合 20%,但"0 到 50% 那段才是每年 50–70% 的大年"。允许迟到——Peter Lynch 的建议是"把图表涂掉,一切关乎未来",只要曲线顶部够高,错过前 100% 也无妨。

结论与值得注意的细节

估值锚点很反直觉。 他强调好机会常以极低 PE 出现:2023 年买 NVIDIA 约 4 倍市盈率、2019 年为汽车 S 曲线买 Tesla 约 5 倍、Apple 约 4 倍、买 Amazon 时 AWS 基本白送。世界不做指数思考,只盯下一季度,这正是三年期预测能创造 alpha 的原因。

他自己列出的风险,值得单独记。 一是公众与政治阻力(缅因州禁数据中心、仅 20% 民众对 AI 乐观、监管风险),但"精灵已出瓶";二是模型进步撞墙——Jensen 的名言"如果够好就够好了,我就没生意了",若 Anthropic/OpenAI 停止改进,开源追上就是逐底竞争,对股价不利但对芯片公司无所谓(芯片商不在乎谁赢 token);三是若某家玩家出局,其算力需求可能落空(Oracle 取消大单后 Meta 立刻接盘,但若 Meta 也退出就不同了)。他同时指出,即使模型不再进步,仅现有能力的采用空间就还很大。

关于 AI 对研究工作的影响,他的回答相当克制。 AI 目前不能替代分析师:写纪要、复盘季度可以,"但最上面那段体现智慧的话必须是人写的——这对我们的论点意味着什么、什么变了,别只当记者"。核心方法仍是 Fisher 1950 年代《怎样选择成长股》的 scuttlebutt——每年 2500–3000 次管理层面对面会议(其中 10–15% 是私募公司),跑展会、访遍供应商客户竞争对手。他称之为"whale rock learning machine":10 人团队,两位核心成员跟了 19 年和 18 年,平均从业 10 年,同一批人同时做公开与私募。

私募配置的能力是逐步长出来的,不是天生的。 2020 年才真正破例,第一笔是 Stripe——因为深研 Adyen 必须把 Stripe 摸透("这是可口可乐和百事"),科普期间朋友来电,凭 TPV 超 5000 亿的披露、Adyen 25–30bp 对比 Stripe 40–50bp 的费率、员工人数反推盈利能力,在 350 亿估值上做了 1 亿美元 block;事后证明费率更高、TPV 接近 1 万亿。卖方喜欢他们的一点是:VC 会卖,而他们会一路持有到公开市场(Nubank 也是如此)。

产品线扩张的底层洞察值得注意:捐赠基金系统性低配全球最大科技公司。 前 15 年只有多空基金,2020 年推纯多头(现已超过多空规模),2021 年推最高 80% 私募的混合基金,最近推出 Whale Rock Mega Cap Tech Fund(宇宙为全球市值前 30,选出 12–13 只)。理由是"大盘股无 alpha"是直觉性偏见——小盘股一个人发现就能推涨,而要让 100 位分散型 PM 意识到 Google 是赢家而非输家,需要时间差,这个时间差就是 alpha。

最后一段与投资无关但真诚。 被问及别人对他最大的善意,他讲父亲:康奈尔电气工程出身转投华尔街,在高盛 41 年,80 年代管公司金融、90 年代任私募股权董事长;Whale Rock 创立时本来只是"友情募资第一通电话"的对象,却主动提出加入当"白头发"和董事长,共事六年直到 2011 年去世,"我们从没提高过嗓门"。他说若能做到父亲一半,就算完胜。

核心句型 · 8
1. It's one thing to …, but it's another to …
“It's one thing to grow from, you know, 100 to a billion, but it's another to do nine.”
用于对比两件表面同类、难度却不同的事,后半句是真正的重点。写作中常用来抬高门槛或戳破类比,如 It's one thing to build a demo, but it's another to run it in production.
2. It wasn't until … that …
“AI has been out hidden inside of these companies, but it wasn't until Chachi PT took it public and ignited what it was.”
强调「直到某事发生,局面才改变」。前半句铺垫长期潜伏,后半句给出触发点。本篇的核心叙事结构就靠它——技术早已存在,拐点另有其时。
3. Not only … but (also) …
“Not only are you creating tremendous unit growth, but the industry, we call it the decommoditization of the hardware industry.”
注意 not only 置于句首时后接部分倒装(not only are you…)。适合叠加两层论据:第一层是常识,第二层才是新意,用于让论点升级。
4. It's okay to …. It's okay to miss the first …
“It's okay to be late. It's okay to miss the first one, two, three years in a lot of cases”
连用 It's okay to 反复宽慰,语气平实却有力。适合在给建议、破除焦虑时使用,配合后面的理由从句效果最好。
5. No matter who / what …, we know …
“No matter who wins above, which we weren't sure at the time, we know we're going to need tremendous amounts of compute.”
经典的「不确定中找确定」句式:先承认某个变量无解,再指出无论结果如何都成立的结论。在做分析、写投资或战略判断时极为实用。
6. It's not just X. It's how big / how tall …
“It's not just, oh, it's taken off now. It's how tall, how big is this S-curve”
先否定一个粗浅的关注点,再把注意力引向更关键的量级问题。仿写时后半句常接 how much / how long / how far 引导的名词性从句。
7. The problem with X is that …
“Problem with software is their AI is 1 or 2% at this stage and it's a long way to go.”
口语里常省略 The 和 that。用于在承认某事物优点后,一句话点出致命短板,是分析类表达里最省力的转折方式。
8. That's where the rubber hits the road.
“That's where the rubber hits the road will he be able to turn this into a huge company”
习语,意为「到了真正见真章的时刻」。用于把讨论从理论拉回可验证的现实检验点,比 that's the key 更生动、更口语。
词汇精讲 · 177 · 按出现顺序
exponentially /ˌekspoʊˈnenʃəli/ adv. 0:00
指数级地(与 linearly 线性地相对,本篇反复对照使用)
penetrated /ˈpenətreɪtɪd/ v. (pp.) 0:00
(市场)被渗透的;less than 1% penetrated 指渗透率不足 1%
linearly /ˈlɪniərli/ adv. 0:00
线性地、成正比地增长
conviction /kənˈvɪkʃn/ n. 0:53
(投资上的)坚定信念;highest conviction position 指信念最强的持仓
anecdote /ˈænɪkdoʊt/ n. 0:53
轶事、个案;投资语境中常指非量化的第一手见闻
deep dive n. / phr. 0:53
深度研究、深挖;do a deep dive into… 为固定搭配
paradigm /ˈpærədaɪm/ n. 0:53
范式;compute paradigm 计算范式
oligopoly /ˌɑːlɪˈɡɑːpəli/ n. 1:55
寡头垄断(少数几家主导市场)
commodity /kəˈmɑːdəti/ n. / adj. 1:55
大宗商品;引申为「无差异化、只能拼价格的产品」
fell away phr. v. 1:55
(逐渐)掉队、消失、脱落
dark horse n. 3:05
黑马;出人意料的竞争者
faltered /ˈfɔːltərd/ v. 3:05
(努力、势头)受挫、动摇、停滞
underpin /ˌʌndərˈpɪn/ v. 3:05
支撑、构成…的基础
leading edge n. 3:05
最前沿(技术);at the leading edge 处于最前沿
counted out phr. v. 3:05
被排除在外、被认定出局;can never be counted out 永远不能小看
leapfrog /ˈliːpfrɔːɡ/ v. 4:26
跳跃式超越、蛙跳式反超
runway /ˈrʌnweɪ/ n. 4:26
(增长/资金的)跑道、可持续空间
thesis /ˈθiːsɪs/ n. 4:26
投资逻辑、核心论点(复数 theses)
kicker /ˈkɪkər/ n. 4:26
额外的推动力、意外加成因素
skeptical /ˈskeptɪkl/ adj. 4:26
持怀疑态度的
agentic /eɪˈdʒentɪk/ adj. 4:26
具备自主行动能力的(AI 语境新词,指能自己规划并执行任务)
unfettered /ʌnˈfetərd/ adj. 5:42
不受约束的、无限制的
mind you phr. 5:42
请注意、话说回来(插入语,用于补充一个易被忽略的限定)
taken off phr. v. 6:50
(业务、需求)迅速起飞、腾飞
stickiness /ˈstɪkinəs/ n. 8:15
黏性(用户难以离开的程度)
differentiation /ˌdɪfəˌrenʃiˈeɪʃn/ n. 8:15
差异化(与竞品拉开区别的能力)
ingesting /ɪnˈdʒestɪŋ/ v. 8:45
摄取;技术语境指「导入并解析(数据/文件)」
franchise /ˈfrænˌtʃaɪz/ n. 8:45
(商业上的)特许经营权;引申为「稳固的业务基本盘」
harness /ˈhɑːrnəs/ n. 8:45
原意为马具/挽具;AI 语境指围绕模型 API 的软件层、脚手架
orchestration /ˌɔːrkɪˈstreɪʃn/ n. 8:45
编排、调度(协调多个组件协同工作)
lock in n. / phr. 8:45
锁定效应(客户切换成本高而难以离开)
on steroids phr. 9:58
加强版的、被大幅强化的(口语,略带夸张)
primitives /ˈprɪmətɪvz/ n. 9:58
(技术)基础原语、最小构件
tinkerers /ˈtɪŋkərərz/ n. 11:05
爱折腾的人、修补者(技术采用曲线中最早的一批用户)
early adopters n. 11:05
早期采用者(技术扩散理论中的第二梯队)
basis points n. 11:05
基点,1 基点 = 0.01%;口语常缩为 bips
juice /dʒuːs/ v. 12:17
(口语)催高、拉升(数据、指标)
busy work n. 12:17
琐碎无价值的杂活
deliverables /dɪˈlɪvərəblz/ n. 13:31
交付物、可交付成果
auditable /ˈɔːdɪtəbl/ adj. 13:31
可审计的、可追溯核查的
allocation /ˌæləˈkeɪʃn/ n. 14:09
配额、分配额度(一级市场中指获准认购的份额)
legacy /ˈleɡəsi/ n. 14:09
传统、遗留背景;此处指「原本的出身与打法」
gross margins n. 14:57
毛利率
turnover /ˈtɜːrnoʊvər/ n. 14:57
人员流失率(另有「营业额」义,需看语境)
play out phr. v. 14:57
(计划、局面)逐步展开、兑现
scour /ˈskaʊər/ v. 15:54
彻底搜遍、翻查(scour the internet 扒遍全网)
home run n. 15:54
棒球术语「本垒打」,引申为「大获全胜的一笔」
unicorn /ˈjuːnɪkɔːrn/ n. 15:54
独角兽(估值超 10 亿美元的未上市公司)
due diligence n. 16:57
尽职调查
like the back of your hand phr. 16:57
了如指掌(know sth like the back of one's hand)
take rate n. 16:57
抽佣率、平台佣金比例
financials /faɪˈnænʃlz/ n. 18:06
财务数据、财务报表
underwrote /ˌʌndərˈroʊt/ v. (pt.) 18:52
underwrite 的过去式;此处指「按假设做投资测算/承销定价」
upsize /ˈʌpsaɪz/ v. 18:52
加码、扩大(交易规模)
modest /ˈmɑːdɪst/ adj. 18:52
(数字上)保守的、说得偏低的;也可指谦逊
predicated on /ˈpredɪkeɪtɪd/ phr. 19:26
以…为前提、建立在…之上
nitty-gritty /ˌnɪti ˈɡrɪti/ n. / adj. 19:26
(口语)最实质的细节、干货核心
nuance /ˈnuːɑːns/ n. 19:26
细微差别、微妙之处
underappreciated /ˌʌndərəˈpriːʃieɪtɪd/ adj. 20:13
被低估的、未被充分认识到的
crucial /ˈkruːʃl/ adj. 21:14
决定性的、至关重要的
ignited /ɪɡˈnaɪtɪd/ v. 21:14
点燃、引爆(需求、热潮)
clunky /ˈklʌŋki/ adj. 21:14
笨重难用的(形容早期产品)
range anxiety n. 22:24
里程焦虑(电动车用户担心续航不足)
churn out phr. v. 22:24
大量生产、批量制造
inflection /ɪnˈflekʃn/ n. 22:24
拐点(inflection point 曲线转向的那一点)
deflationary /dɪˈfleɪʃəneri/ adj. 23:32
通缩的;此处指「使单位价格下降的」
line item n. 23:32
报表中的单列科目
stay on top of it phr. 24:28
持续跟进、掌控住局面
sell side n. 24:28
卖方(券商研究部门,与买方基金相对)
ancillary /ˈænsəleri/ adj. 25:35
辅助的、附属的(ancillary business 周边业务)
multiple /ˈmʌltɪpl/ n. 25:35
(估值)倍数,如市盈率倍数
flatline /ˈflætlaɪn/ n. / v. 26:14
横盘、毫无起色的平直阶段
pitfalls /ˈpɪtfɔːlz/ n. 26:14
陷阱、易犯的错误
anecdotal evidence n. 26:14
轶事性证据(个案见闻,非统计数据)
connecting the dots phr. 27:03
把零散线索串联起来、看出整体图景
standing room only phr. 27:35
座无虚席、只剩站位(形容爆满)
virtualized /ˈvɜːrtʃuəlaɪzd/ v. 27:35
(把硬件)虚拟化
pattern recognition n. 27:35
模式识别(此处指从历史案例中辨认重复结构)
commissioned /kəˈmɪʃnd/ v. 28:49
委托(某人做研究、创作)
slope /sloʊp/ n. 28:49
斜率(此处指采用曲线的陡峭程度)
evangelists /ɪˈvændʒəlɪsts/ n. 30:28
布道者、内部积极推动者
exhaustive /ɪɡˈzɔːstɪv/ adj. 32:24
穷尽式的、无遗漏的
exposure /ɪkˈspoʊʒər/ n. 32:24
(投资)敞口、对某主题的暴露程度
disruption /dɪsˈrʌpʃn/ n. 32:24
颠覆(式变革)
network effect n. 33:13
网络效应(用户越多产品越有价值)
lithography /lɪˈθɑːɡrəfi/ n. 34:13
光刻(半导体制造核心工艺)
rolled into one phr. 34:46
集…于一身、合而为一
bulls /bʊlz/ n. 34:46
多头、看涨者(对应 bears 空头)
susceptible /səˈseptəbl/ adj. 35:44
易受…影响的(susceptible to)
erosion /ɪˈroʊʒn/ n. 35:44
侵蚀、逐步削弱
in the fullness of time phr. 35:44
假以时日、时机成熟之时(略正式)
pure plays n. 35:44
纯粹标的(业务单一、只押注某一主题的公司)
cash cows n. 37:10
现金牛(持续产生大量现金的成熟业务)
escape velocity n. 37:10
逃逸速度;商业上指增长快到摆脱竞争压制的临界点
capital intensive adj. 37:10
资本密集型的
recursive /rɪˈkɜːrsɪv/ adj. 37:52
递归的(recursive improvement 自我递归改进)
liftoff /ˈlɪftɔːf/ n. 37:52
起飞(火箭离地的瞬间,引申为业务腾飞)
gazillion /ɡəˈzɪljən/ n. 38:54
(口语夸张)无数、海量
eyeballs /ˈaɪbɔːlz/ n. 38:54
(互联网黑话)用户注意力、流量眼球数
moving the needle phr. 41:21
带来实质性的改变(needle 指仪表指针)
horse and buggy n. / adj. 41:21
马车(时代);喻指过时落后的方式
gut /ɡʌt/ v. 42:14
大幅削减、掏空(gut their jobs 大砍岗位)
seats /siːts/ n. 42:14
(SaaS)席位数,按人头计费的订阅单位
selling motion n. 43:24
销售动作/销售模式(B2B 行业术语)
incumbents /ɪnˈkʌmbənts/ n. 43:24
在位者、现有市场主导者
trajectory /trəˈdʒektəri/ n. 44:30
轨迹、发展路径
tempted /ˈtemptɪd/ adj. 44:30
被诱惑的、忍不住想…的
halfbaked /ˌhæfˈbeɪkt/ adj. 45:44
不成熟的、想得还不周全的(half-baked)
repository /rɪˈpɑːzətɔːri/ n. 45:44
存储库、信息汇集地
permanent fixture n. 45:44
永久固定设施;喻指「拔不掉的存在」
leaves something to be desired phr. 46:49
尚有不足、不尽人意(委婉批评的地道说法)
commonality /ˌkɑːməˈnæləti/ n. 46:49
共同点、共性
systems of record n. 47:14
记录系统(企业中存放权威数据的核心系统)
headless /ˈhedləs/ adj. 47:14
无头的(软件术语:剥离前端界面,只提供数据与 API)
relegated /ˈreləɡeɪtɪd/ v. 47:14
被降格、被贬到(次要位置)
solidify /səˈlɪdɪfaɪ/ v. 47:14
巩固、使更牢固
commoditized /kəˈmɑːdətaɪzd/ v. 48:33
(被)商品化,即失去差异化只剩价格竞争
enclosures /ɪnˈkloʊʒərz/ n. 48:33
机箱、外壳(服务器硬件部件)
decommoditization /ˌdiːkəˌmɑːdətaɪˈzeɪʃn/ n. 49:38
去商品化(本篇自造词:从无差异化重新变回有定价权)
renaissance /ˈrenəsɑːns/ n. 49:38
复兴、再度繁荣
contract manufacturer n. 50:58
代工制造商
offshore /ˌɔːfˈʃɔːr/ adv. / v. 50:58
迁往海外(生产);go offshore 产能外移
heritage /ˈherɪtɪdʒ/ n. 50:58
传承、历史积淀(此处指技术与人才的沿袭)
sole supplier n. 50:58
独家供应商
code word n. 50:58
暗语、委婉说法(code word for… 「…的另一种说法」)
swapped out phr. v. 51:58
被替换掉、被换下
ASPs n. 53:52
Average Selling Prices 平均售价
visibility /ˌvɪzəˈbɪləti/ n. 53:52
(业务)能见度,即对未来订单的可预见程度
topline /ˈtɑːplaɪn/ n. / adj. 53:52
营收(利润表最上面一行),对应 bottom line 净利
ridiculously /rɪˈdɪkjələsli/ adv. 54:49
高得离谱地、荒唐地(口语强调词)
specs /speks/ n. 54:49
规格、技术参数(specifications 的缩略)
clusters /ˈklʌstərz/ n. 54:49
(算力)集群
overwhelmed /ˌoʊvərˈwelmd/ adj. 55:54
应接不暇的、被压垮的
slip through the cracks phr. 56:59
从缝隙中溜走、被疏漏掉
single source of truth n. 56:59
单一可信数据源(企业数据治理术语)
reconciliation /ˌrekənsɪliˈeɪʃn/ n. 57:33
对账(财务术语);日常义为「和解」
stay ahead of the curve phr. 57:33
领先一步、抢在趋势前面
newbies /ˈnuːbiz/ n. 58:52
新手、菜鸟(口语)
holistic /hoʊˈlɪstɪk/ adj. 59:23
整体的、全局性的
marking time phr. 59:23
原地踏步、横盘等待(军队原地踏步走)
without merit phr. 59:23
毫无道理、站不住脚(not without merit 并非全无道理)
getting out of hand phr. 59:23
失控、脱轨
bullish /ˈbʊlɪʃ/ adj. 60:25
看多的、乐观的(反义 bearish)
the genie is out of the bottle phr. 60:25
精灵已出瓶——事已发生、无法收回
race to the bottom n. 61:31
逐底竞争(不断降价降标准的恶性竞争)
hits a wall phr. 61:31
撞墙、遇到瓶颈无法再进
suck that up phr. v. 62:08
(把闲置产能)吸收掉、接下来
fend off phr. v. 63:05
抵御、挡开(竞争、攻击)
moat /moʊt/ n. 63:05
护城河(巴菲特术语,指可持续的竞争优势)
the rubber hits the road phr. 64:10
到了真正见真章的时刻(轮胎接触路面)
supplanting /səˈplæntɪŋ/ v. 65:51
取代、顶替(正式用词)
up to speed phr. 65:51
跟上进度、快速上手(get up to speed on sth)
scuttlebutt /ˈskʌtlbʌt/ n. 66:20
闲谈、小道消息;费雪投资法中特指遍访产业链取证的调研法
blocking and tackling phr. 66:20
(源自橄榄球)扎实的基本功、基础执行力
substrates /ˈsʌbstreɪts/ n. 66:20
基板、载板(半导体封装材料)
humanly possible phr. 66:20
人力所能及的(as many as humanly possible 尽人力之所能)
intricacies /ˈɪntrɪkəsiz/ n. 68:16
复杂细节、错综之处
scores and scores phr. 68:16
数十上百(score = 二十,叠用表极多)
two-way street n. 68:54
双向的关系(强调互惠、不能只索取)
like-minded /ˌlaɪk ˈmaɪndɪd/ adj. 68:54
志同道合的
monolithic /ˌmɑːnəˈlɪθɪk/ adj. 69:54
单一整块的、铁板一块的
break the seal phr. 70:32
开封、破例迈出第一步
underweight /ˌʌndərˈweɪt/ adj. / v. 70:32
低配(持仓比例低于基准),反义 overweight 超配
endowments /ɪnˈdaʊmənts/ n. 71:54
(大学等机构的)捐赠基金
alpha /ˈælfə/ n. 71:54
阿尔法:超越市场基准的超额收益
stock pickers n. 71:54
选股型基金经理
divergent /daɪˈvɜːrdʒənt/ adj. 72:57
分化的、走势各异的
rallied /ˈrælid/ v. 72:57
(股价)大幅反弹、上涨
levered to /ˈlevərd/ phr. 74:00
高度受益于、对某趋势有杠杆式敞口
generalist /ˈdʒenrəlɪst/ n. / adj. 74:00
通才(非专精某行业的基金经理)
odds /ɑːdz/ n. 74:00
胜算、几率(like your odds 看好你的胜算)
act on it phr. 75:53
据此采取行动、立刻执行
pivoted /ˈpɪvətɪd/ v. 76:30
转向、转型(职业或业务方向的转换)
humility /hjuːˈmɪləti/ n. 76:30
谦逊
oversight /ˈoʊvərsaɪt/ n. 76:30
监督、把关(另有「疏忽」义,需看语境)
whipsmart /ˈwɪpsmɑːrt/ adj. 77:43
极其聪明、机敏过人(whip-smart)
with grace phr. 77:43
从容优雅地(处理棘手局面)
理解自测 · 11 题 · 是真懂了,还是以为自己懂
1. 讲者把 AI 技术栈分成哪几层?他们最先投的是哪一层,理由是什么?

自下而上是:电力、芯片、云、基础模型、应用(黄仁勋常说的「五层蛋糕」)。2023 年初他们决定先投芯片和基础设施。理由有两条:一是底层最先拿到需求,二是当时上层谁会赢并不确定,但无论谁赢都需要海量算力——这是「不确定中的确定项」。这一选择也解释了全篇后半段为何花大量篇幅讲硬件与数据中心供应链,而不是应用公司。

2. 他用什么数据推算出「光编程就是半万亿美元市场」?

两个数字相乘:一是 Anthropic 内部有人每天在 token 上花约 100 美元,折算下来一年约 2–3 万美元;二是全球程序员约 2000 万人。二者相乘得到约 5000 亿美元的市场规模。他特意补充这还是基于「七八九个月前的技术」算出的。需要注意这是自下而上的 TAM 推算,人均支出与程序员总数两个假设都很敏感,属于框架性估算而非精确预测。

3. 收音机和洗碗机的 S 曲线差别说明了什么?

收音机约七年就接近全渗透,是史上最快的采用曲线之一;洗碗机则极慢,因为它必须接入房屋的后端管线。差别在于部署摩擦:需不需要改造既有基础设施。讲者由此推出 B2B 慢于 B2C 的一般规律,并用它解释自己在 2000 年前后看多 B2B 互联网为何失败——底层基础设施未就位,同一判断早二十年就是错的。这套研究由他们委托 Horace Dediu 完成。

4. 他们为什么几乎清空了应用软件持仓,甚至一度转为净空头?

2023 年 4 月他们原本判断应用层软件会大受益,因为这些公司有庞大销售队伍、数据和渠道。但很快发现它们的 AI 产品做不好、卖不上价、不推动业绩。随后他列出四重压力:CIO 优先级下降、预算被 Anthropic 的 token 挤占、不敢再年年提价、裁员冻招压缩按席位收费的基数。以 Salesforce 为例,400 亿美元营收基数下 5–7 亿的 AI ARR 无法改变轨迹。结果是几乎清仓,年初净空头,第一季度因此获益。

5. 他为什么把当前的企业 AI 叫「L 曲线」而不是 S 曲线?

S 曲线的典型形态是长期潜伏、然后陡升、最后走平,潜伏期的原因是采用障碍。他的论点是这一次几乎没有部署摩擦——「至少对消费者甚至企业来说,你只要打开浏览器它就在那儿」。对照 1998 年「知道该建网站却建不出来」和云计算因安全顾虑增长仅 30–50%,AI 的接入成本低得多。加上企业端今年像开关被打开,所以曲线不是缓起而是一路直上,他称之为反过来的 L。

6. 「硬件去商品化」为什么是反直觉的判断?它的论据链是什么?

直觉上硬件是典型的大宗商品生意——低毛利、拼价格、无护城河。他的反驳分两步:过去四十年之所以如此,是因为算力负载年增 25–40%,而摩尔定律以同样速度改进,硬件不需要创新;现在负载年增 10 倍,摩尔定律追不上,每个环节都被逼到物理极限。于是 HBM 要十层堆叠、PCB 从 10 层变 40 层、服务器改液冷、以太网从七年一代变成一年一代,能做出来的供应商极少,知识产权与协同研发周期重新成为壁垒。

7. 按他的框架,什么时候该卖出?苹果的案例说明了什么?

规则是渗透率到 30–40% 之后:指数增长结束,卖方分析师的预期跟上来,不再出现大幅超预期,超额收益的来源消失。苹果是他们 2012 年在美国智能手机渗透率约 50% 时卖出的。事后看苹果并未衰落——靠 App Store 30% 抽成等周边业务继续以约 20% 复利增长,所以他称之为「一个错误」;但他同时指出真正的大年(年涨 50–70%)确实集中在 0–50% 那一段。这说明该规则牺牲的是后段的稳健复利。

8. 既然 S 曲线是公开的常识,为什么公开市场上很少有人做对?

他的回答不是保密,而是执行门槛。第一,需要长期积累:他在 Whale Rock 做科技已二十年,团队跨越多轮周期,能分辨不同曲线的差异。第二,是人才真空——几乎没人认真研究硬件和芯片,进来的多是新手。第三,也是最关键的,是持仓时的心理成本:过去四年英伟达每一次大涨后都被喊泡沫,且看空理由并非全无道理,没有全局视角就扛不住。他还补充,很多半导体分析师错过,是因为看不到基础模型层在发生什么。

9. 他为什么认为 Anthropic、OpenAI 不会重蹈诺基亚、RIM 的覆辙?

他先承认反例的分量:RIM、Palm、诺基亚、HTC、LG、摩托罗拉都处在史上最好的 S 曲线上却归零,所以 S 曲线只回答市场多大、不回答谁赚到。他给出的差异是四点护城河:关键知识产权(各家训练方法与擅长领域不同)、企业品牌(问任何 CIO 第一个说出的是 Claude)、逃逸速度(10 倍营收增长加持续融资能力,足以对抗大厂现金牛)、以及递归式自我改进——用领先的编程能力回头改进自身模型。他也承认例外:范式转移时领先者仍可能像 AOL 那样掉队。

10. 如果有人反驳说「半万亿编程市场」是循环论证——token 消耗被融资补贴和算力短缺扭曲,讲者会如何回应?

按他全篇的论证方式,他大概率不会争辩单价,而是转向三条旁证。其一是行为反转证据:Karpathy 与 Torvalds 这样最懂行的人从「AI 只能写 20%」改口到「几乎不再手写代码」,需求真实性不依赖定价。其二是供给端的物理证据——在只有 0.1% 的人深度使用 AI 的情况下,DRAM、NAND、PCB 已短缺约 30%,短缺不会由补贴凭空造出。其三是他自己在 p27 引用的安迪·格鲁夫原则:拐点期本来就不能信数据,只能靠轶事与现场线索。但这恰恰是这套框架最难证伪之处:所有反证都可以被归为「你还没看到拐点」。

11. 这套 S 曲线框架放到供给与监管受限的场景(缅因州禁数据中心、电力瓶颈)还成立吗?

部分成立,但需要修正。框架本身已经内置了修正机制——他用电动车的例子承认 S 曲线未必走完:渗透率在 10–15% 撞墙,而非预期的 40–50%,因此「必须调整、必须盯紧」。他也把公众情绪(仅 20% 民众乐观)、缅因州禁建数据中心和监管列入了风险清单。但他的判断是「精灵已出瓶」,监管不足以逆转。真正的软肋在于:他的 L 曲线论证同时依赖「摩擦极低」和「供给极紧」,而电力与土地约束恰恰是一种会随规模变大而变强的摩擦——这两条论据在长期是互相拉扯的,他在这一小时里并未正面处理。

精读便签
下载便签 手机:长按图片也可保存
← 上一期 · NO.070Harvard Professor Teaches Writing (Michael Pollan Interview) 下一期 · NO.072 →Jim Keller: Moore's Law, Microprocessors, and First Principles | Lex Fridman Podcast #70
订阅苏菲周报 每周一封:本周入库的精读、一个值得带走的问题、一条苏菲按。免费,随时退订。
免费 · 每周一封 · 一键退订
苏菲拉底 THE SOPHIE LAB · ASK THE BEST MINDS THE BIG QUESTIONS 内容仅供学习 · thesophielab.com