--:--
点击播放，这里会跟随视频显示当前句的中英字幕。
Why the Markets Are Pricing AI Wrong | Gavin Baker
节目发布 2026-08-04 · Invest Like The Best
加文·贝克
帕帕特里克·奥肖内西EDITED TRANSCRIPT · 依据现场录音编译整理，可划线生成便签
编者按：本文是投资人加文·贝克（Gavin Baker）与播客主持人的一次对谈，录制于 2026 年夏末，地点在硅谷风险投资机构 Benchmark 的办公室。贝克是 Atreides Management 的创始人兼首席投资官，长期专注科技股，此前两个月里正在硅谷走访各家实验室、推理云与初创公司。谈话发生在一个 AI 相关股票普遍腰斩、市场信心动摇的月份，贝克逐一拆解了造成恐慌的几条叙事，并解释为什么他认为基本面反而在加速。本文依据现场录音编译整理，仅删去口语枝节与广告，论证、例证与语气均按原样保留。
压力测试：找不到负面指标
主持人： 加文，距离上次对谈才两个月。模型发布周期在缩短，我们节目之间的间隔也在缩短，你我现在基本是按模型发布节奏在录节目。
贝克： 之前有人指出，我们的播客每次都恰好赶上市场的阶段性高点，我对这种批评还挺敏感的。经过这一次，没人能再这么说了。
主持人： 这个月太疯狂了。你在想什么？
贝克： 我会把七月形容成「压缩进一个月的 2022 年」。当然有几个基本面上的负面因素，我们应该谈。但总体上看，基本面的天平我认为在明显改善。一大批 AI 股票从高点跌了五六成，姑且说一个月之内直线下跌四到六成。开录之前我问过你，你整个夏天都在这边，有没有听到过哪怕一个关于 AI 的负面量化指标？
主持人： 没有。一个减速的例子都没有。
贝克： 一个都没有。事实上，每一项指标都在加速。
主持人： 而且不是 AI 信徒的盲目乐观，是各人从各自的位置拿出的数据。
贝克： 完全如此。不管怎么切，看 GPU 的可得性也好，看 GPU 的零售价格也好，看这个月 DRAM 的现货价格也好，看 token 增速也好，全都在加速。我认为问题的很大一部分在于，市场看不见 Anthropic 和 OpenAI，也看不见那些在美国把推理变现的开源推理云，比如 Fireworks、Baseten、Modal、Together。一旦你看到这些，整幅图景就完全不同了。
私有公司盲区
贝克： 开源因为 GLM 5.2 和 Kimi K3 而大幅加速。英伟达的 Nemotron 也在稳步推进，美国这边还出了一个很小但很好的开源模型。OpenAI 加速了，Anthropic 继续高速增长，而且几乎可以肯定正在产出可观的自由现金流。大家都在看那张图：半导体公司的自由现金流一路向上，超大规模云厂商的自由现金流一路向下。可那张图漏掉了这些私有公司。
贝克： 那张图还漏掉了一件非常重要的事。在 2024 年和 2025 年，哪怕你是最乐观的人，你对 GPU 价格的预期也不过是缓慢下降；悲观的人则认为会断崖式下跌。我不认为 2024 年或 2025 年有任何人预料到，旧 GPU 的价格会在 2026 年垂直上涨。于是所有人都觉得自己很聪明，签了长期合约。某种程度上很多新型云厂商也不得不这么做，因为他们需要包销协议来给 GPU 融资。结果就是，已经装机的算力，其合约价格相对当前现货市场存在巨大折价。随着这些合约到期，算力会被重新定价到更高的水平。就算现货价格回落，算力也照样会往上重新定价。我认为你会看到大量的加速，它会回答那些关于投资回报的疑问。
三大云厂现金流加速
贝克： 这个季度你已经开始看到苗头了，前提是看经营性现金流而不是自由现金流。微软、Meta 和亚马逊已公布的经营性现金流增速从 28% 加速到 32%。这里面有一些相当大的一次性项目，这些超大规模云厂商似乎总有几十亿美元的「非经常性」法务开支，主要是交给欧盟的罚款，但这个季度一次性项目的规模异常地大。调整之后，增速是从 28% 到 35%。在这个体量上，这是实质性的加速。而这还是在他们点亮 Rubin 之前，Rubin 会以明显的溢价出租；也是在这些合约重新定价之前。
七月复盘：Meta出租算力
贝克： 这个月确实难熬，或许有必要把这个月怎么走到今天梳理一遍。首先是 Meta 要对外出租算力。市场把这当成极度利空：他们产能过剩了，要削减资本开支，这是一场灾难。事情完全不是这样。他们刚刚公布财报，并没有削减资本开支。真实情况是，他们看到 SpaceX 拥有巨大的算力装机量，把几个为训练优化的大集群卖进市场，价格比那些合约价高出一大截。至少分析师是这么看的：Meta 看到了机会。也有很多猜测说他们要融资。也许他们的想法是，先拿一小块产能证明我们能做出很强的内部收益率，然后去募股权资本，之后就放开手脚，大概率还会上调资本开支。
贝克： 看起来他们实际做的并不是这个。但无论如何，市场因为把它解读为利空而抛售。而我非常确定这不是利空。关于 Meta 的资本开支计划有大量的遥测数据，没有任何一项发生变化，如果说有变化，那也是继续变得更激进。紧接着他们发布了很长时间以来最好的模型 Muse 1.1，真的是一个非常好的模型。它被 Grok 4.5 的光芒盖住了，但确实是好模型，比他们过去两年的任何东西都强得多。所以他们绝无可能松油门。
Kimi与开源冲击：token就是token
贝克： 然后 Kimi 出来了，随之而来的是对开源的巨大恐慌。同一时间，Silicon Data 的 token 指数出现下探并走平。这两件事是相连的。那个 token 指数捕捉到的是结构变化。他们看不到全部的 token，但因为 GLM 5.2 和 Kimi 的相继落地，数据里出现了一个结构性的迁移：从昂贵的前沿模型 token，迁移到开源 token。前沿模型 token 的推理毛利，我们可以争论是 80%、90% 还是 95%，总之极高。
主持人： 对。
贝克： 不知为何，市场认为这是利空。但现实是，token 就是 token。在其他条件相同的情况下，生成一个 token 需要的算力完全一样，同样的浮点运算量，同样的内存，同样的瓦数。当然，token 之间并不完全等价。但大体上讲，开源夺取份额所做的，只是把利润从前沿模型那一层拿走，而由于需求弹性的存在，这反过来会推高 token 需求。你需要更多的算力。至于利润率，Anthropic 也好，开源也好，跑在同一批底层云厂商上，云厂商对算力收一样的价钱。所以你实际上只是把前沿模型的利润拿走，把更多的利润美元推进了 AI 基础设施这一层。
主持人： 这一串事情就是导火索。
贝克： 对。你想想，黄仁勋是全世界最大的开源支持者。他是个非常理想主义的人，是个爱国的美国人，我认为他总是做正确的事。可如果开源对他的生意有害，他会是全世界最大的开源支持者吗？如果是对世界正确的事，他也许仍会支持。但大概不会把它当成标志性议题。顺便说一句，我认为开源真的很重要。一个只有一两家占主导地位的前沿模型、收着 90% 毛利的世界，对人类不好，对社会可能也不好。我们之前讨论过，我们想要很多模型。
贝克： 市场消化了这个，缓过来了。然后中国有了 DUV 光刻机，这引发了半导体设备股的大规模抛售。
真实利率与信用利差，唯一的实质风险
贝克： 接下来就到了我认为在很多方面都算真正的担忧：真实利率上升了。这说得通，我们在投入大量资金为这场建设融资，信用资金确实在其中占越来越大的比重，尽管大部分仍然来自经营性现金流。真实利率上升，利差走阔。Meta 上周发了一笔债，定价并不在你想象中 Meta 债券该有的位置。这说明信用市场的态度。
主持人： Meta 的信用违约互换（CDS）也在飙。
贝克： 所有人的 CDS 都在飙。一些非常聪明的私募资本人士说，这恰恰是你该预期的：银行在对冲自己的承诺。但不管怎么说，这看起来不好，而且这些都是无法否认的事实。CDS 上升，利差走阔，真实利率上升。如果这场建设需要靠债务来融资，这会非常非常可怕。这就是为什么我认为，装机算力的现货价格与合约价格之间的差距如此重要。
主持人： 理解未来六个月左右的融资环境太重要了。
贝克： 关键在于这场建设需要多大程度上依赖信用。
主持人： 对。那就是经典的资本周期：我们开始靠债务把自己撑得过大，事情就从那里变得危险。
贝克： 百分之百。靠债务驱动的建设要求立即偿还。一旦供需稍微失衡，事情可以非常非常快地崩解。互联网泡沫就是这么破的。所以，如果一个人像我这样相信，不管对错，而且经过这个月我已经非常开放了，我一直在压力测试所有这些假设，尤其在信用上下了很深的功夫，因为这是真的，无法否认，如果我们需要信用来支撑这场建设，那就是重大利空。
按Ampere折价建模，少了七千亿信用需求
贝克： 但如果你把它建出模型来，看一致预期里超大规模云厂商准备上线的吉瓦数，这些是 Blackwell 和 Rubin 的吉瓦。Rubin 是英伟达的下一代芯片，Blackwell 是当前这一代。一致预期基本上是按 Ampere 的变现速率来给它们建模的，Ampere 落后两代，连 Hopper 都不是。这样算出来，超大规模云厂商的经营性现金流大约是 1.3 到 1.4 万亿美元。我认为他们按 Ampere 的速率变现的可能性极低，为什么可以展开讲，其中一部分理由就来自在这里亲眼看到的真实需求数据。但就算只假设他们按比当前 Blackwell 打折的价格变现，经营性现金流也更接近 2 万亿美元，这就把 7000 亿美元的信用需求拿掉了。
贝克： 然后很讽刺的是，随着装机算力重新定价，所有的信用比率都会改善。我们会继续加速。一致预期建的是减速，我认为不太可能。信用指标变好之后，突然之间用信用融资也变容易了。至于他们会不会选择那么做，走着瞧。
Blackwell空档期与GPU租金半年涨五成
贝克： 这多少回到我们上上次谈到的 Blackwell 空档期风险：你花几千亿美元买 Blackwell，一开始主要用于训练，训练不产生回报，这可能是个风险。一季度我们确实看到了这个风险。两个月前那期我之所以对这个风险释然了，是因为 Anthropic 那边的表现实在惊人，市场会看穿这个空档。四月、五月、六月，市场确实看穿了。到了七月，因为这一连串事情的叠加，市场不再看穿它了。而恰恰就在这时，经营性现金流开始真正加速。这就是事实，它在大规模地加速。微软六月上线了一大块产能，二季度财报里甚至都还没体现。
贝克： 所以归根结底一句话：你是否相信，在硅谷从私有公司那里看到的量化需求信号会持续下去，使得装机算力在合约到期后重新定价到更高水平。
主持人： 于是经营性现金流上升。
贝克： 经营性现金流上升，那么这场建设大部分可以从经营性现金流里出，也许全部都能。如果按当前价格重新定价，未来几年大概全部都能内部消化。
贝克： 这确实是市场里非常不寻常的一段。我们也应该谈谈我说的那些正在变好的基本面到底是什么。技术派会说，2022 年市场担心的是衰退、加息、通胀，你确切知道它在担心什么。DeepSeek 那次，你知道它在担心什么。「解放日」那次，你也知道。有一个非常清晰的东西，这在某种奇怪的意义上反而让人安心。而这次，我们谈了很多具体的事情，但除了信用之外，所有这些具体的事情都有点荒谬。可它还在跌。技术派会说，这有点吓人，按定义，打中你的永远是你没看见的那颗子弹。我以前说过，投资里最重要的三个词不是「安全边际」，而是「我不知道」。
贝克： 但你在这里待了两个月，我也在。我今天早上刚跟一家公司谈过，他们租了一个几千张 Blackwell 的集群，这是那种所有人都想跟它做生意的、最炙手可热的初创公司之一。当时的价格，姑且说每 GPU 小时两美元五左右。现在他们要续租同一个集群，同样的规模，B200，方方面面完全一样，对方说「没有任何差别」。七个月之后，他们指望能以略低于四美元的价格租到。这就是今天你会听到的事。这挺疯狂的，因为按常理，价格温和下降就已经算利好了。结果我们看到的是，取决于起点，六七个月里涨了 50% 到 60%。这样的例子太多了。有一家推理云，我记得是 Baseten，不太确定，他们上播客说，合约到期后我们打算为 Blackwell 多付 100% 的价钱。这意味着所有超大规模云厂商的盈利都被严重低估了。
Anthropic曲线争议与英伟达十年最低PE
贝克： 我这周来这里的主要任务就是压力测试。
主持人： 压力测试。
贝克： 对。找一些负面的东西告诉我。我问你的那个问题，「你听到过一个负面量化指标吗」，我这周逢人就问。
主持人： 大家说得最多的，是第三方数据显示 Anthropic 的增长曲线似乎稍微偏离了原来的轨迹。这是我听到的唯一一条。
贝克： 我认为这很可能是真的。但与此同时，OpenAI 和开源在大幅加速。你如果看那些数据，开源也许看起来差不多，我认为它可能已经加速了。宇宙学里讲暗物质，开源之于公开市场就有点像暗物质，公开市场很难度量它。但只要你跟踪这些推理云在说什么，这些是播客上的话、会议里的话，不是经审计的财报，但需求显然在加速。这说得通，因为 GLM 5.2 和 Kimi K3 带来了巨大的能力跃升，我认为这会继续。我认为你会看到英伟达把 Nemotron 稳步推向前沿。
贝克： 这是个非常难熬的月份。但同时我也惊叹于一点：我把每一个假设都压力测试了一遍，底层基本面在改善。而我们录这期节目的时候，英伟达的前瞻市盈率处在过去十年的最低点。
主持人： 疯狂。
贝克： 半导体股票唯一比现在更便宜的时候，是「解放日」和 DeepSeek，而那两次都是 V 形底。
主持人： 这在你看来意味着，市场认定它们的盈利被严重高估了？
贝克： 对，市场百分之百认为它们严重超额盈利。我们得谦逊。
主持人： 也许确实如此。
贝克： 也许确实如此。但我这周的任务就是拼尽全力找负面数据点。平时我来硅谷，听到的总是好坏参半，这里有点负面，那里有点正面，总体偏正面，科技长期创造价值。可这次我找不到一条量化的负面指标。除了 Anthropic 那份第三方数据，而那份数据在 Anthropic 的股东那里争议极大，他们急不可耐地想告诉你他们知道的情况，又非常害怕话传回公司说是他们讲的「其实一切都好」，会拿不到 IPO 配额。你能看出来 Anthropic 的股东都憋着一句「那不是真的」。
贝克： 我很难相信开源和 OpenAI 加速到这个程度，而 Anthropic 显然仍然处在领跑位置。另外，从第三方数据也能看到，Grok 和 Cursor 在七月经历了一个转折性的月份，Grok 4.5 和 Grok Build 出来了。所以这是个棘手的月份。我在富达（Fidelity）有个朋友说，过去三年在市场里活下来的办法，就是尽可能快地做那件最蠢、最肤浅的事，然后在这些事之间轮换。
主持人： 那现在这件事是什么？
贝克： 就是整个月都在削减风险，以回应这些除了信用之外事实上并不成立的叙事。而我们做的研究让我认为，随着算力重新定价，信用也不会成为问题。退一步说，就算你真的需要信用来建我们要吃的那些浮点运算，如果信用不在了，那只意味着已经存在的算力会更值钱。
所有人都用Claude解读新闻
贝克： 有人给我发了一篇很有意思的文章。我们以前谈过迈克尔·莫布森（Michael Mauboussin）的理论，多样性的崩溃是泡沫和崩盘的成因。而现在，我认识的所有做公开市场股票投资的人，不管散户还是机构，每一条新闻都立刻喂给 Claude，Claude Code 也好，Claude 的智能体也好。Claude 是概率性的，但它解读新闻的方式大概没有那么大的差异。
贝克： 这几乎像是回到了，股票市场其实从未有过这种状态，但人们谈论媒体的碎片化时会说，过去只有克朗凯特（Walter Cronkite）一个真理之声，如今没有了。而 Claude 有点像股票市场的克朗凯特，所有人都相信它。
主持人： 它说什么就是什么。
贝克： 它说什么就是什么。顺便说，它非常聪明，但并不总是对的，它的解读并不总是正确。而股票市场从根本上讲，是对未来的一种概率性的、贝叶斯式的解读。所以市场里的感觉就是：来了一条新闻，喂进 Claude，Claude 这样解读，一大群人按 Claude 的看法交易。你已经看到后果了。有个叫 TBU 的人，半导体圈里的一位，他贴了一张日本电容器股票的惊人图表，说我们在六周里走完了一整个电容器周期。确实如此，那些股票是翻倍、翻三倍还是翻四倍我不知道，总之垂直上去，然后「嗖」地下来。真实的基本面甚至还没到，而正常情况下需要三年的周期，已经在六周里走完了。
持续学习若突破，会砸掉训练需求吗
主持人： 你在这里的感受里，我尤其好奇一点：这里正在进行的、提升推理服务与模型训练每一个环节效率的创新，长期会怎样影响公开市场？你有没有了解到什么长周期的创新，让你格外兴奋或好奇？
贝克： 我非常好奇。似乎有很多人觉得自己非常接近解决持续学习和样本高效学习了，我们以前谈过这个。如果这两者被解决，有可能造成一种暂时的不连续。有人告诉我，人脑大约相当于在 200 亿个 token 上训练的，而这些模型是在 300 万亿个 token 上训练的。如果你能在 10 万亿个 token 上训练一个东西，然后把它放到世界上让它以样本高效的方式学习，这对训练需求听起来不太妙。但训练在半导体需求和算力需求中的占比，本来就会渐近到一个不趋近于零、但很小的数字。我要说这是最有意思的一点，至于是长周期还是短周期的事，谁也说不准。SSI 说他们八月会发布模型。有整整一代新实验室专注于这个方向。
主持人： 这对世界是好事。
贝克： 对世界是极大的好事，太棒了。
主持人： 我们都想要这个。
贝克： 我们都想要。它对世界是好事，而我很难相信它会真的对 AI 基础设施需求构成利空。但我还是要非常非常开放。这大概算是这趟最大的科学或技术层面的收获。
主持人： 我们仍然不知道。
贝克： 对，而且英伟达深度参与了所有这些初创公司。
什么会真正让你害怕
主持人： 如果非要你构造一组情形，能让你彻底转向、真正害怕，会是什么？是不是就是经营性现金流这件事没有兑现，于是整场建设都得靠债务融资？
贝克： 现金流不再加速，那是利空。这在一定程度上取决于 Anthropic、OpenAI、Grok、Cursor 和开源的表现。如果 GPU 价格出现相当剧烈而且持续的收缩，市场会立刻反应，那就令人担忧。如果 GPU 变得很容易拿到，你听过有谁说自己 GPU 太多吗？
主持人： 一个都没有。
贝克： 一个都没有。而且不是「没有」，是正相反，听起来像毒品市场，真的，太狂野了。显而易见的利空还有一长串。如果 Anthropic、OpenAI 这些实验室的总和进入平台期或者开始下滑，那就非常负面，除非原因只是开源 token 在夺取份额而没有做大蛋糕。我确实认为未来是多模型的。
路由器与开源微调：省钱不等于少算力
贝克： 尤其是 AI 原生公司，他们会想拿一个开源模型，这些推理云在有监督微调和强化学习上都已经非常擅长了。你可以拿自己的数据，定制一个开源模型，放到一个路由器后面。路由器通常先把请求发给你自己的模型，然后由 Claude、Grok 这样的前沿模型来核验，很多情况下你能以一半的成本得到略好的结果。
贝克： 但「一半的成本」这句话，很多人一听就觉得对 AI 需求是坏事。其实完全不是。用户付的成本只是 token 利润率的函数，你只不过是把 token 从毛利 90% 的昂贵 token，转移到毛利大概 30% 的 token。省下的钱来自这里，但生产这些 token 消耗的算力是一样的。而且所有这些事在不同的周期里同时发生。那些大型上市公司都在喊：天哪，我的 AI 开支涨了二十倍，三个月就烧光了预算。于是他们架一个路由器，这确实削减了他们的 AI 开支，但对算力没什么影响，甚至可能通过转向更便宜的开源 token 而增加了他们生成的 token 总量，那就是更多算力。一家公司在「什么任务用什么模型」上变聪明了，可能让它的开支企稳甚至下降，但这与它在路由器后面那些模型层实际消耗的 GPU 算力小时没有关系。随着转向更便宜的 token，你可以用得更多，GPU 算力小时大概率是上升的。
贝克： 这发生在最前沿的一批上市公司身上。然后你有一整波 AI 原生公司，他们全力押注，不雇人，钱主要花在 token 上，他们没有放慢。然后是美国东海岸那些几乎还没开始采用 AI 的公司，不在沿海、没那么前沿的公司，还有欧洲，正忙着在用 AI 之前先想清楚怎么监管它。
贝克： 所以是几波差异化的采用浪潮同时在发生。而有一个念头我一直挥之不去，上次可能也说过：有人估计，全世界大概只有 50 万人，也许 25 万人，在用智能体式 AI（agentic AI）。而我们已经处在算力短缺中。这个星球上有七八十亿人。从 50 万人变成
主持人： 变成 1%。
贝克： 变成一亿，五亿，会发生什么？
25万亿知识工作待替代
贝克： 还有一点很有意思。在 X 上发帖看反驳挺有帮助的。很多人说：好，我们接受你的论点，超大规模云厂商盈利被低估，算力会重新定价，他们的经营性现金流会加速，也许能自己出钱。但这些经营性现金流从根本上要从哪里来？客户在哪里？按定义，它只能来自两处：要么是生产率带来更快的经济增长，就像纳德拉说的，要么我们开始以 10% 的速度增长，要么不会；要么就是劳动力替代。在很多 AI 原生公司里，你确实看到劳动力替代，不是因为他们在裁人，而是他们雇的人少得多。每个全职员工的毛利润，a16z、ICONIQ 和一批公司做过这项研究，是垂直上升的，尤其是相对于过去几代初创公司。
贝克： 你有没有对你投的公司做过调查，他们的 token 开支和人力开支的比例？
主持人： 做过。通常按 token 开支占总薪酬开支的百分比来报。
贝克： 你看到的范围是多少？
主持人： 在那些真正全力投入的公司里，会很高，20%、25% 这个量级。
贝克： 我们的朋友迪伦·帕特尔（Dylan Patel）是个超级智能极大化主义者，他那里是 30%。
主持人： 那大概是我听过最高的了。
贝克： 我听过 50% 的。知识工作是 25 万亿美元。就取你的 20%，那是 5 万亿美元，它要么来自劳动力替代，要么来自更快的经济增长。我们非常非常希望它来自更快的经济增长。
主持人： 今天早上我从一位顶尖的科技公司 CEO 那里听到一件有意思的事，他创办过好几家公司。他说，如果你看创始人控制的公司，再剔除疫情期间过度招聘的因素，没有谁在真正裁员。这些人本该是最快用 AI 提效的，可他们并没有搞什么大规模裁员，这大概说明他们认为还有大量机会需要保留人手。
贝克： 百分之百。是增长，不是劳动力的减少。乐观情形是，你看过 Cognition、Ramp 和 Stripe 的图表，在 AI 上花钱最多的公司增长明显更快。
主持人： 我很喜欢 Cognition 那个指数。
贝克： Cognition 那个指数很惊人。怀疑者会指出，它没有控制行业变量，这是对的。但如果你往下挖，我记得其中一个例子是个水管工还是暖通空调承包商，蓝领工人因为 AI 都过得很好。
内存长约博弈：毁约就等于丢掉份额
贝克： 有件事我觉得应该谈，现在或者稍后都行：所有人都在引用那些长期协议（LTA）。现在所有东西都短缺。如果说有什么疲软，那只是因为吉瓦通电的速度不够快。吉瓦会通上电的，监管政策在朝好的方向走，燃气轮机厂商、柴油发电机组厂商都在扩产。人们把旧飞机上的涡轮拆下来翻新，改作他用。疯狂的事情正在发生，资本主义非常擅长这个。
贝克： 但我认为市场里最重要的问题之一，也是我看错的一个转变，是我们正在转向，尤其是内存行业，从短期内碾压业绩预期，变成用短期上行空间去换那些叫供应链协议、长期协议的东西。形式很多，本质上是客户预付款，有价格下限也有上限。
贝克： 这又回到劳动力那个话题。疫情期间很多公司裁人裁过了头，之后就非常不愿意再裁，几年前人们讲的「囤积劳动力」，你记得吗？我们来想一想撕毁一份长期协议的博弈论。真正上规模的公司有四家：有 Trainium 的亚马逊，有 TPU 的谷歌，AMD，还有比其他所有人加起来都大得多的英伟达。假设现在是 2027 年。必须认识到一点，内存是关键：在一个给定的算力单元里，每单位浮点运算配的内存越多，产出的 token 越多。这是提高单位算力 token 产出最重要的一件事。而这按定义会降低成本，所以需求根本没有负面反应，没有任何弹性，因为这是压倒其他一切的那根轴。
贝克： 从某种层面上，这就像这些公司之间的一场《权力的游戏》。好，2027 年或者 2028 年，你隐约想撕毁一份长期协议，压一压价格。但未来几年市场份额在很大程度上由供应链的配额决定，由你预先买下了什么决定。假设我们没有处在严重的供过于求里，其实就算严重供过于求，这套博弈逻辑也几乎同样成立。如果你撕毁了协议，而在接下来两三年里，不管什么原因，筹码又转回了内存厂商手里，你就出局了，结束了。比如说谷歌撕毁一份协议，假设 2028、2029 年供过于求，我是随口编的。他们撕毁协议，大概意味着供过于求，价格在下跌，产能自然收缩。那你觉得谷歌的配额会怎样？这是一个周期性行业，供过于求之后就是供不应求。你觉得内存厂商下一次会怎么对待他们的配额？
贝克： 既然这是一切围绕转动的轴心，撕毁一份长期协议可能会炸掉你的整个生意和整个品牌。这在以前从来不是问题。苹果，谁在乎，他们没有竞争对手，是压倒性的最大买家。回到三四五年前，他们知道自己可以为所欲为而不承担后果，因为量太大了，就算把海力士（SK Hynix）狠狠坑一把，美光当然照样接。现在不同了，你至少有四家玩家，还有所有的初创公司，你还投了 Etched。如果你撕毁一份协议，对方只需说：好啊，你毁了价格协议，我们就毁了数量协议，去你的，我们把量给你的竞争对手。你就丢了份额。
英伟达的信用包装与收入分成
贝克： 所以我认为，当前环境对英伟达的有利程度，让我很难理解它为什么以这么低的估值倍数交易。换句话说，如果你需要为芯片融资，而你确实需要，那没有什么比英伟达 GPU 更容易融资，没有。如果你需要土地和电力，他们在这盘棋和牵线搭桥上做得非常好。然后他们又推出了一个非常聪明的新商业模式，我会把它描述成一种信用包装：如果 GPU 价格高于一个下限，就分一份收入。这可能让他们通过特许权使用费的方式，非常快地做出一个巨大的云业务。这也是缓解现金流错配的另一条路：嘿，钱都是我们赚的。
贝克： 这不完全是供应商融资，因为不是他们借钱给买家，借钱给 GPU 买家的是别人。他们仍在做股权投资，但不是那种「你把钱给某人，其中一部分被用来买你的芯片」。英伟达说他们在所有股权投资里都写明这笔钱不能用来买英伟达芯片，但显然钱是可替代的。
主持人： 有趣的小细节。
贝克： 对。但我想在某种层面上，它大概让所有人都更安心。
主持人： 如果你是内存厂商，比如海力士的 CEO，你会怎么做？
贝克： 我会做英伟达现在正在做的一模一样的事。我会去找 GPU 的买家，AMD 也好，谁都好，说：我要参与英伟达那种信用包装。内存的生意本身没那么稳定、没那么可预测，所以也许他们先预付一些现金，这样他们不必承担责任，我是随口编的。但总之，趁你现在有钱、信用市场在造反，去做点类似的事。我敢肯定黑石和阿波罗的朋友们正在向内存公司建议某种变体：我们从今天的现金流里拿出一笔钱，付出去就没了，作为一种担保，让放贷的人安心，但我们也要在持续收入里分一杯羹。这百分之百是我会做的。它几乎是长期协议的逻辑延伸，长期协议是用上行空间换持久性，这里你实际上拿到了经常性收入的分成。
贝克： 这正是英伟达在做的事。我认为它被严重误解了，英伟达应该好好把它解释清楚。第一，他们非常看好 AI。他们每一次没有入股什么东西，事后都证明是错的。他们基本上入了所有东西的股，除了内存公司，还有很长时间没入股 Anthropic，后来也入了。既然你有现金流，又看好 AI，为什么不呢？黄仁勋看得到每一个实验室，知道所有的进展，那些持续学习的实验室，SSI 现在也在跟他们合作。他什么都看得到，而他看到的东西让他看多。所以，一是拿到股权的上行空间，二是拿到收入分成，同时你产出几千亿美元的自由现金流，帮着架桥跨过这段显然存在的缺口：在经营性现金流加速到能内部融资之前，所有人的自由现金流都转负了。
贝克： 这非常机会主义，而且是好的那种，它显著提高了他们每吉瓦的收入。它还巩固了竞争地位。你我都投了初创芯片公司，好，用那家初创公司的芯片，可他们付给台积电的价格是多少？比英伟达高。他们买 HBM 内存的价格？更高。这些芯片能像英伟达的那样容易融资、拿到同样的利率吗？不能。所以负担是真实存在的，尤其如果你用 HBM 内存，你就正好处在这场博弈的靶心。除非你做了非常不同的架构选择。正在发生的一切对做了那种选择的公司反而不错。
算力军备竞赛的博弈论
贝克： 再回到博弈论。Anthropic 如果当初在算力上像 OpenAI 那样激进，早就把比赛跑没了。而现在 OpenAI 回到了赛场上，我认为 Grok 也在赛场上。这几家是帕累托前沿上的公司。
主持人： 而且他们有算力。
贝克： 看完这一幕，你觉得还有谁会松油门？大概四个月前，达里奥（Dario Amodei）谈过这个，是一段非常深思熟虑的评论，他说这真的非常难：在这种规模上，算力买多了你可能破产，买少了你可能输掉。
主持人： 我们眼看着它发生了。
贝克： OpenAI 刚刚杀回来了，SpaceX 凭着 Grok 4.5 和 Cursor 大举入局。看完这些，从博弈论的角度，有谁会在近期退缩？尤其当这可以从经营性现金流里出钱的时候。
我竟是硅谷最看空的人
主持人： 你在这边走访时，有没有遇到比你看多得多的人？如果有，他们相信什么而你不信？
贝克： 老兄，这里基本上所有人都比我看多。我读了德瓦凯什（Dwarkesh Patel）写的一篇东西。
主持人： 算力价格涨三倍那篇？
贝克： 我忘了具体数字。
主持人： 不，好像是十五倍之类的。
贝克： 对，大意是租一张 H100 一年会花 25 万美元，相当于一份薪水。
主持人： 那是现货价的十五倍左右。
贝克： 正是。哇，这根本不在我的贝叶斯预期结果空间里。甚至不在我「考虑过但认为极不可能」的那一栏里。而这个人非常聪明，消息非常灵通。他还指出，算力的利润率在上升，算力的数量在上升，推理的利润率在上升，三者相乘，就是各实验室加上开源的总和出现这种疯狂加速的原因。当然开源的利润率并没有真的上升。
主持人： 所以所有人都更看多。
贝克： 对。我看着股市里发生的事，觉得自己像个傻乎乎的乐观派。然后我跟人聊，不管是实验室的人还是这个生态里的任何人，我相对于几乎所有人都是看空的。这真是一种奇怪的状态。
中国DUV：相变式进步，仍需二十年
主持人： 你怎么看中国 DUV 光刻机的消息？我看到的反应横跨整个光谱，一端说这相当于 ASML 在 2001 年的水平，另一端说不，这是关于全球尖端算力供应的一个新故事的开篇。
贝克： 我认为两者可能都对。打个比方，DUV 光刻机像螺旋桨飞机，EUV 像喷气发动机，或者随便什么类比。以前他们没有，现在据说有了。这是一次相变，像从液体变成固体。但那个固体，那台喷气发动机或者螺旋桨飞机，落后二十五年。它仍然重要，不该被轻视。但你在股市里看到的很好笑：市场大幅过度反应，而这件事如果真的会打到 ASML 的订单，可能是五年之后，到那时市场早已把它忘了，然后担心，忘了，担心，忘了，一路反复好几次。所以我认为这大概是过度反应，但我们也不该轻视它。对中国来说，这真的非常重要。有报道说一台 EUV 光刻机被走私进了中国。要是真的，那可是间谍史上的壮举，那东西是巨型机器，大得惊人。我不知道真假，有些传闻。中国人非常非常优秀，非常非常聪明，工作极其刻苦，他们把这看成国家的头等大事。但他们能从 2001 年一步跨到 2026 年，甚至 2030 年吗？
主持人： 这是干中学。
贝克： 是干中学。你没法加速「干」这一步，没法瞬移到未来，你必须真正走过那些学习周期。所以，它重要吗？重要。市场过度反应了吗？大概是。但作为一个美国人，要真正理解中国正在发生什么，并对此有完全的确信和清晰，是非常难的。不管好坏，我们正在脱钩。这个过程已经启动，到现在几乎在两边各自自我强化。这很不幸，但现实如此。他们不会停，我们也不会。
开源对软件业是福音
主持人： 对美国其他公司有什么看法？我感觉现在市场就是十家公司，其中几家还是私有的。
贝克： 上个月，除了 AI 之外的一切都是垂直上涨。而我认为，开源逼近前沿，加上 Fireworks 这样的公司让定制模型变得极其容易，让你在某些情况下能以明显更低的成本拿到比前沿模型更好的性能，这对软件行业是天赐之福。对所有那些 AI 原生公司也是。我们的朋友埃里克·维什里亚（Eric Vishria）大概两年前说过，他从没见过这么多公司在九个月之类的时间里从创立做到年收入五千万美元并且产生现金流。当时很难知道它们是否持久，很多人把它们贬为「ChatGPT 套壳」。现在有了开源，你就积累了你那个用例、你那个垂直领域独有的数据。
贝克： Fireworks 出了一个很酷的产品叫 Nexus。你用 Claude Code、OpenAI Codex 或 Grok Build，只需要三行代码，大概二十个词，Fireworks 就会用你的数据对一个模型做强化学习，再配一个路由器分发请求。效果惊人。这是每一家 AI 原生公司的解法，所以你看到 Harvey 在被收购之前、Cursor 都在重仓这个方向，Harvey、Legora，全都是。如果你能从只用一两三个前沿模型，变成
主持人： 什么最优用什么。
贝克： 前沿模型只承担 30% 到 60% 的 token 消耗，剩下的用你自己的强化学习模型，那你突然就不是套壳了，你的护城河深得多。
主持人： Cursor 那个东西让我很感兴趣。我记得是 Cursor，它像是用 AI 快速复刻了我们在人类身上学到的东西：用前沿模型做规划，然后把任务分包给更笨的模型。
贝克： 百分之百。
主持人： 效率高十五倍，大概是这个数。
贝克： 很讽刺的是，可能正是那些略微落后于前沿、利润率较低的开源 token，我们有朋友相信，一旦前沿模型达到递归自我改进，它在每一个智能层级上的服务成本都会通过蒸馏而大幅降低，那开源就没有立足之地了。我会说这是 Anthropic、OpenAI、Grok 极大化主义者的观点。我们不该轻视任何东西，我不知道，什么都有可能，要非常谦逊，尤其是经过我这个月之后。但这看起来不太可能。
主持人： 为什么？
贝克： 第一，有这么多 AI 原生公司已经积累了相当数量的领域专有数据。在开源迎来这个时刻、推理云和路由器真正成熟之前，你没得选，服务条款是什么你就接受什么。而现在如果你能跳下那台跑步机，你就有了某种程度的独立性，也许还有持久性和安全感。回到你的观点，也许这些便宜的 token 反而会大幅抬高最尖端的前沿 token 的价值。假设今天你有智商 120 的开源模型，我随口编的，而且跑起来很便宜，那能指挥它们的智商 160 的模型不是更值钱了吗？上次我们谈到，我一直很惊讶经济回报有这么大一部分归到了前沿模型。现在这在改变，你看看这些推理云，Together、Modal、Baseten，资本效率极高。这些商业模式令人震惊的地方在于，它们的增速几乎赶上早期的前沿实验室，却几乎不烧钱。相当了不起。用软件即服务（SaaS）那套「40 法则」的老指标来看，这些数字疯狂。
主持人： 你觉得一家组织内部的薪酬分布是不是有启发？CEO 挣的是中位数员工的多少倍，也许那就是前沿 token 与开源 token 的关系。
贝克： 绝对是。也许就像我们上次讨论的：蛋糕在飞快变大，前沿 token 可能继续拿走绝大部分经济价值，但不再像过去那样全拿；而开源 token 可能成为被处理的 token 的大多数。再说一遍，这对基础设施需求是好事，因为 token 就是 token，生产它需要同样的浮点运算、瓦数、空间和冷却。
最大风险是监管：讲好数据中心的故事
主持人： AI 可能发生的最坏的事是什么？监管？还是别的？
贝克： 我认为监管必然是最大的风险，这是最显而易见的风险。这也是我这周很高兴来这里的原因之一：我想被吓一吓。我不想觉得自己像个疯子，看着这些股票越来越便宜，还认为预期的远期回报在上升，而在基本面上，七月相对于六月甚至有实质性的改善。但我走下来仍然觉得，监管必然是最大的风险。你没法忽视纽约州搞数据中心禁令。我们活在一个奇怪的、后事实、后逻辑的政治世界里。我认为 AI 行业的公关做得糟糕透顶。
主持人： 至少他们现在意识到了。
贝克： 对，就算没修好，也意识到了。华盛顿的叙事，很多普通美国人接受的政治叙事是：数据中心会抬高你的电价，抢光你的水，拿走你的工作。而现实是，按照现在谈的那些协议，一座数据中心建起来，周边所有人的电价通常反而会下降，因为有表后（behind the meter）发电协议。这就是特朗普让大家签的那份数据中心承诺书。过去数据中心开发商只需要给警察局、消防局配几辆新车、几套新防弹衣就行。现在是：我们给你建一所医院、一所学校、一个新警察局和一个消防站，还降低你的电费，怎么样？而且这些工作岗位是持续的，因为你需要水管工、电工、暖通空调承包商。数据中心在很多方面是我这辈子见过的对蓝领工资最好的事。可名义上代表蓝领工人的民主党，却在把他们的工作拿走。
贝克： 还有一件事很疯狂。那句话怎么说的？谎言已经绕地球一圈
主持人： 真相还没起床。
贝克： 对，真相还没起床。有位作者在书里犯了个错，把数据中心的用水量高估了一万倍。不是一点点，不是一个数量级，不是两个，也不是三个。她多次承认了这个错误，「我完全错了」，这事已经被彻底澄清过了。
主持人： 大力水手效应，你听过吗？大力水手吃菠菜，起因是一本学术著作里小数点点错了位。菠菜的铁含量并不比别的东西高，就是那一个来源，然后一路传开，到今天人们还说菠菜含铁多。
贝克： 我真的以为菠菜含铁多。太疯狂了，我真的一直这么以为。每天都能学到新东西。
主持人： 同一回事。
贝克： 同一回事。所以，得有人把真相说出来。我觉得这个行业，如果没别人做，也许我来做。得有某种基金会，或者一个政治行动委员会，在「最终四强」赛、NFL 比赛、大学橄榄球赛、世界大赛期间投广告：数据中心是干什么的。一座签了这份承诺书的数据中心落户你的社区，你的电价会下降，它几乎肯定会实质性地回馈社区，你会看到大量高薪蓝领岗位涌入，而且会持续。很多人以为那是一次性的，其实不是。确实会有一个高峰，然后转到下一座数据中心，但这些数据中心持续需要返修和升级，技术一直在变。所以你会有更多工作，更便宜的电，更富裕的社区，对水没有影响，对环境没有影响，把数据中心建在城外十英里也很容易。
贝克： 这个故事需要被讲出来。同时还有，我记得上次谈过，AI 越来越多地在救人命、治愈罕见病。我记得是今年的美国临床肿瘤学会年会（ASCO），会上的气氛是：这是我们在单一会议上见过最多的科学突破，其中一部分肯定归功于 AI。我们需要讲这些故事。如果你有一个生病的孩子、生病的父母、生病的亲人，AI 实实在在地提高了他们康复的几率。所有人都需要去讲这个。而在硅谷的人看来这一切太显而易见了。
主持人： 他们以为别人早就知道了。
贝克： 他们无法理解，这个真实的观点与大多数美国人的看法相去甚远。所以这个行业真的需要把自己的故事讲好。纽约感觉只是许多州里的第一个，甚至一些深红州，他们极其支持增长，也在说：你们没把自己的故事讲好，我们就没法替你们讲。你们讲出来，我们可以复述，但你们才是专家。如果你不说出自己的真相，没有人会替你说。
SRAM 加速器与解耦推理
主持人： 我们还漏了什么？
贝克： 关于算力，我觉得所有讨论里都缺了一块：当那些基于 SRAM 的加速器进来之后会发生什么。它们不受 HBM 内存的制约，常常用较老的制程节点制造，不跟最新最强的 GPU 抢产能。把推理解耦（disaggregate）之后，人们谈预填充（prefill）和解码（decode），而解码又分两部分：注意力（attention）和前馈网络（feed forward network）。终极圣杯是：预填充在一块芯片上做，大概不带 HBM；注意力在一块带 HBM 的高性能芯片上做；前馈网络则在一块 SRAM 芯片上做。把 SRAM 加速器加到现有和新建的算力里，投资回报率极高。我们看到的是效果就是更好，尤其在前馈网络上，你就是打不过 SRAM。而且不管你多努力去调芯片上算力、HBM 与 SRAM 的比例，负载总在变，负载各不相同。能把它拆成这三部分，我认为对 AI 的投资回报会非常非常正面。
权力游戏里的黑马
主持人： 我突然想到一个有趣的问题，因为我喜欢你那个《权力的游戏》的说法。你能想象某个现在不在大家视野里的玩家，跃升到那种级别的重要性吗？比如美光突然变得像 Anthropic、OpenAI、微软、亚马逊、英伟达、SpaceX 那样重要？
贝克： 权力游戏里的黑马？智谱大概算一个。Fireworks 的乔琳（Lin Qiao），她是个绝对的狠角色。我们的朋友斯科特·吴（Scott Wu），Cognition。
主持人： 这个我举双手赞成。
贝克： 这几个是最明显的名字。
SpaceX：被市场低估的算力玩家
主持人： SpaceX 呢？眼看着它被公开市场消化，至少是初期，感觉如何？你觉得市场理解它是一家什么样的公司吗？这可是最重要的新上市公司。
贝克： 感觉并不理解。在我看来，它上市以来所有的基本面都在变好。Grok 4.5，收购 Cursor，Cursor 显然明显加速了。过去三年他们证明了自己能比任何人更快、以更低的价格把更多算力带上线。现在我们知道，就算调整现货与合约的价差，他们的巨大优势在于进场时正好踩在现货高点上。奇怪的是，对算力最利好的事之一，恰恰是他们一夜之间往市场里投放了海量算力，连个水花都没有，市场把它完全吸收了，一点都没有卖不出去。Substack 上有家 AI 研究号认为 SpaceX 打算上线 8 吉瓦的算力。我永远不会跟马斯克对赌，但那会是一个真正难以置信的壮举。
贝克： 自从他们签上一批合约以来，价格是涨的，不是跌的。他们每吉瓦的变现大约是 500 亿美元。一致预期明年的收入是 730 亿美元。那么，别管星链 V3，别管星链直连手机，Grok 4.5 加 Cursor 的年化经常性收入大概很快就能到 100 亿美元，这些也都别管，连星链的核心业务也别管。如果他们能上线接近那个规模的算力，8 吉瓦乘以每吉瓦 500 亿，而一致预期是 730 亿。当然这些不可能在 2027 年初全部点亮，我觉得非常不可信，那份报告我几乎不信。但到今天为止，一年内上线过 500 兆瓦以上电力的公司，只有超大规模云厂商、CoreWeave、Crusoe 和 SpaceX。而 SpaceX 上得最多、最快、成本最低。而且大家确实喜欢他们的集群。不过市场需要亲眼看到。
主持人： 这不会是市场今天对 SpaceX 的解读。
贝克： 不是。感觉纽约有家大对冲基金在做空它，逻辑是：算力现货价格会跌 90%，你把这么多算力带上线，收入远不及预期。也许吧。但我要说清楚，我这些年亲眼见过马斯克的公司做成非常惊人的事。那份 8 吉瓦、18 个月的报告，我引用它只是因为它是公开的，人人都能看到。
贝克： 马斯克有句话，我们的专长是让不可能的事迟到。
主持人： 我没听过，太棒了。
贝克： 这句话里有很多真实的成分。我只是认为，就他们可能带上线的算力规模而言，这只股票里几乎什么都没有计入。再说一遍，我不认为接近 8 吉瓦。这会非常难，给这些 GPU 通电非常难，但他们一直做得很好，而这既不在预期里，也不在人们的思考里。
主持人： 我想起那个梗：「SpaceX，数据中心公司？」
贝克： 对，绝对是。另外，我在星际基地（Starbase）待了不少时间，轨道算力一天比一天真实。
主持人： 前几天看到星舰着陆，非常酷。
贝克： 非常酷。有意思的是，我们在 Benchmark 的朋友投了 Star Cloud，一家轨道算力公司，SpaceX 在跟它合作，我想是让他们用星链的激光技术，这对轨道算力非常重要。这是一个不错的理智检验。据我所知 Benchmark 的人相当聪明，他们完全不来自马斯克的生态，却选择以不低的估值投了一家轨道算力公司，而这家公司还不享有 SpaceX 内部的发射资源。对我来说这就是一个「我疯了吗」的检验。也许我疯了，也许马斯克疯了，也许 Benchmark 也疯了，SpaceX 的工程师也都疯了。但老兄，这似乎不太可能。我们是不是该说一下我们在谁的办公室里？
主持人： 我们正坐在那张著名的桌子中间。
贝克： 对，这是他们那张著名的桌子，用来办那些著名的晚餐。谢谢 Benchmark 为这一期提供场地，谢谢埃里克，他帮我做的协调，谢谢所有合伙人。
贝克： 一年之后我们再看这些股票在哪里。好在时间会给出答案，人们会被证明是对的或错的，未来是概率性的，但这是一个激动人心的时刻。
主持人： 如果我们继续按模型发布周期录节目，那几周后见。
贝克： 太疯狂了。跟你聊总是很开心。
本期讲者
加文·贝克Atreides Management 创始人、管理合伙人兼首席投资官，专注科技与消费股。此前在富达（Fidelity）任职十余年，曾管理 Fidelity OTC 基金。
帕特里克·奥肖内西播客 Invest Like the Best 主播，Positive Sum 创始人。曾任 O'Shaughnessy Asset Management 首席执行官，该公司 2021 年被富兰克林邓普顿收购。
章节 · 点击跳转视频
0:00
开场：想被吓到却找不到负面数据
2:02
市场看不见的私营公司与算力重定价
5:07
七月复盘：从 Meta 到信贷利差
13:02
经营现金流能否覆盖建设
18:13
硅谷实地：GPU 租金不降反涨
24:12
Claude 共识与持续学习的冲击
30:05
路由器与开源 token 为何不减算力
36:49
内存长约博弈与英伟达新模式
47:33
算力军备赛、中国 DUV 与看多硅谷
55:14
AI 原生公司与推理云的崛起
1:01:40
最大风险是监管与叙事失败
1:08:25
解耦推理芯片与 SpaceX 算力
本期论点
本期回应
14:28
只要算力变现效率不倒退两代，近 2 万亿经营性现金流就足以抵消 7000 亿信贷需求
14:54
AI 算力需求更可能继续加速，而市场一致预期假设它会减速
19:33
GPU 租金在半年里涨了五到六成，说明超大规模厂商的算力投入普遍严重不足
54:27
先进光刻是边做边学的过程，学习周期无法加速，中国无法直接跳到当下最前沿水平
其他论点
12:49
靠债务推动的算力建设要求立刻还款，供需稍有失衡就会极快崩塌
21:41
英伟达远期市盈率处于十年最低，说明市场认定它当前的盈利严重超常且不可持续
24:21
散户和机构都把消息喂给同一个 Claude 解读，市场的解读多样性正在崩塌
27:39
持续学习和样本高效学习一旦解决，训练在算力需求中的占比会趋向一个很小但非零的数字
32:15
企业 AI 支出趋稳甚至下降，与其背后消耗的 GPU 算力小时数无关，后者仍在上升
34:12
支付 AI 账单的现金流只能来自生产率提升带来的经济增长，或来自劳动力替代
38:48
每 FLOP 配的内存越多产出的 token 越多，这是提升单位算力产出最重要的一件事
39:38
未来几年 AI 芯片的市场份额由供应链分配决定，取决于预先采购了多少产能
1:01:43
监管是 AI 行业面临的最大风险，超过技术或需求本身的问题
01开场：想被吓到却找不到负面数据
0:00
I want to be scared, you know, I don't want to feel like a lunatic watching these stocks get more cheaper thinking the expected forward returns are growing up. I may be kind of missing out here this week. >> pressure test? >> Yeah. >> Yeah. >> Find tell me something negative, but I haven't been able to find one that is like a quantitative metric. The underlying fundamentals are improving. Uh and stocks Nvidia's actually has we record this at its lowest forward PE of the last 10 years. The market 100% thinks they're significantly overvalued.
我想要感到害怕，你懂吧，我不想像个疯子一样看着这些股票越来越便宜，还觉得预期的未来回报在往上走。我这周可能有点错过机会了。>> 压力测试？>> 对。>> 对。>> 你给我说点负面的，但我一直找不到一个像量化指标那样的东西。基本面本身在改善。呃，而且我们录这期节目的时候，英伟达的远期市盈率其实处于过去10年的最低点。市场100%认为它们被严重高估了。
0:48
>> Gavin, [music] it's only been 2 months like the model release cycles, the gap between our podcast episodes are shortening. >> [laughter] >> We're basically you and I are basically on a model release cadence at this point. >> Well, I was I was sensitive to criticism that um that I think somebody pointed out that um our podcasts were coincident with like local market peaks. >> [laughter] >> And nobody can say that after this. >> What's on your mind? It's been a crazy crazy month. >> Yeah, I would describe um July as 2022 in a month.
>> Gavin，[音乐] 才过了两个月，就像模型发布周期一样，我们播客两期之间的间隔越来越短了。>> [笑] >> 我们现在基本上，你和我基本上是按模型发布的节奏在更新了。>> 嗯，我之前对一个批评还挺在意的，就是有人指出，我们的播客总是正好赶上局部的市场顶部。>> [笑] >> 这回之后没人能这么说了。>> 你最近在想些什么？这个月真是疯狂至极。>> 是啊，我会把七月形容成「浓缩成一个月的2022年」。
1:23
>> Yeah. >> There are some fundamental negatives which you which we should talk. But like on on the whole, the balance of fundamentals I think is improving significantly. Loads of AI names are down 50 60% from their highs. We'll call it 40 to 60% in a month in a straight line. And I asked you before we started, you've you've been out here for the summer. Have you heard a single negative quantitative metric about AI? >> Yeah. >> A single instance of deceleration. >> Nothing. >> Nothing. In fact, every metric is accelerating.
>> 是啊。>> 确实有一些基本面上的负面因素，我们应该聊聊。但总体来看，我觉得基本面的天平正在显著改善。一大堆AI相关的股票从高点跌了50%、60%。就算是40%到60%吧，一个月内一路直线下跌。我们开始之前我问过你，你整个夏天都在外面跑。你有没有听到过哪怕一个负面的关于AI的量化指标？>> 对。>> 哪怕一次减速的迹象。>> 完全没有。>> 完全没有。事实上，每一项指标都在加速。
02市场看不见的私营公司与算力重定价
2:02
>> And to your point, not just blind optimism from people excited about AI. Here's some data that they can show you and from their different vantage points. >> Absolutely. I mean, however you cut it, whether you cut GPU availability, whether you cut GPU retail pricing, I mean, whether you cut like the spot price of DRAM this month, token growth, everything is actually accelerated. And I do think a big part of the problem is one, the market does not have visibility into Anthropic, OpenAI, and then I would say these open-source inference clouds that monetize inference here in America, Fireworks, Baseten, Model together.
>> 而且正如你所说，这不只是那些对AI兴奋的人的盲目乐观。他们能拿出实打实的数据，而且是从各自不同的视角。>> 绝对是这样。我的意思是，不管你怎么看，无论你看 GPU 的供应情况，还是看 GPU 的零售价格，我是说，哪怕你看这个月 DRAM 的现货价格，token 的增长，一切其实都在加速。我确实认为问题很大一部分在于，第一，市场看不到 Anthropic、OpenAI 内部的情况，然后我要说的是那些在美国靠推理变现的开源推理云，Fireworks、Baseten、Together。
2:44
>> [snorts] >> And the picture looks very different when you see that. Because open-source is accelerated massively because of GLM 5.2 KiB K3. And then, you know, Neurotron continues to kind of chug along. We had a great, you know, very small American open-source model release. OpenAI has accelerated. Anthropic continues to grow really strongly and is almost certainly pumping out significant amounts of free cash flow. And I just think if, you know, there's this chart that everybody looks at of semiconductor cash flow going like this and hyperscale cash free cash flow going like that, and is you're missing these private companies. And then, I also think that that chart, um, misses something very important, which is just that you have everyone in '24 and '25 thought, even if you were really bullish, you thought that GPU prices, if you were really bullish, you thought they would to price around a GPU would, you know, decline slowly. You know, if you're bearish, you thought it would decline precipitously.
>>（哼笑）>> 当你看到这些，画面就完全不一样了。因为开源已经大幅加速了，得益于 GLM5.2、Kimi K3。然后，你知道，Nemotron 也还在稳步推进。我们有一个很棒的、你知道，非常小的美国开源模型发布。OpenAI 也加速了。Anthropic 继续强劲增长，而且几乎可以肯定正在产生大量自由现金流。我就是觉得，你知道，大家都在看的那张图，半导体的现金流是这样往上走，而超大规模厂商的自由现金流是那样走，而你漏掉了这些私营公司。另外我还觉得，那张图，嗯，漏掉了一件非常重要的事，那就是在 24 年和 25 年，所有人——哪怕你非常看多，你也只是认为 GPU 价格，如果你非常看多，你会认为一块 GPU 的价格会缓慢下滑。你知道，如果你看空，你会认为它会断崖式下跌。
3:51
I I think anyone in 24 or 25 thought that the prices of old GPUs would still be would be going vertical in 2026. Yeah, and so everybody thought hey, we're going to be smart. We're going to sign these long-term contracts. And to some degree like a lot of the deal clouds had to do that because they needed an off-take agreement to finance the GPUs. And so essentially you have the contracted base of installed compute trading at a massive discount to the current spot market. And has those contracts roll off and compute gets repriced higher.
我想 24 年或 25 年没有任何人认为旧 GPU 的价格到 2026 年还会垂直往上走。是啊，所以大家都想，嘿，我们要聪明一点，我们要签这些长期合同。而且某种程度上，很多 neocloud 不得不这么做，因为他们需要一个包销协议来为GPU 融资。所以本质上，你已装机算力的合约基础，相对于当前现货市场是在以巨大的折价交易。随着这些合同到期，算力就会被重新定价到更高的水平。
4:35
It's spot could decline and compute will still get repriced higher. You know, I think you're going to see a lot of acceleration that's going to answer these ROI questions. You've started to see that this quarter if we look at operating cash flow, not free cash flow. Operating cash flow from Microsoft, Meta, and Amazon has reported accelerated from 28 to 32. There were some actually pretty big unusual items now like these hyperscalers they always seem to have like billions of dollars of legal expenses that are unusual.
现货价可能下跌，但算力仍然会被重新定价得更高。你知道，我认为你会看到很多加速，这会回答这些 ROI 的问题。这个季度你已经开始看到了，如果我们看经营性现金流，而不是自由现金流。微软、Meta 和亚马逊已公布的经营性现金流增速从 28% 加速到了 32%。其实这次还有一些相当大的非经常性项目，像这些超大规模厂商，他们似乎总是有数十亿美元的法律费用，属于非经常性的。
03七月复盘：从 Meta 到信贷利差
5:07
Mostly fines to the EU. But there was an unusual amount of one-timers this this quarter. If you adjust for that, we went from 28 to 35 and that's that's a that's a material acceleration at this scale. And that's really before like they start to light up the Rubins which will come at a meaningful premium before these contracts reprice. It's been a it's been a it's been a challenging month that it's almost um you know, like is it helpful to kind of like walk through the month? How we got here? You know, so first there's Meta is going to rent out compute. And this is seen as like very bearish. They have excess capacity. They're going to cut CapEx.
大多是对欧盟交的罚款。但这个季度一次性项目的数量异常地多。如果你把这些调整掉，我们是从 28% 到 35%，而这在这个体量上是一个实质性的加速。而且这其实还是在他们开始点亮 Rubin 之前——Rubin 会带来可观的溢价——也在这些合同重新定价之前。这是一个、是一个很有挑战的月份，几乎可以说，嗯，你知道，要不要把这个月捋一遍会有帮助？我们是怎么走到这一步的？你知道，首先是 Meta 要把算力租出去。这被视为极度看空的信号。他们有过剩产能，他们要削减资本开支。
5:50
This is a disaster. This is not at all what it was. They just reported. They didn't cut CapEx. What it was is they saw SpaceX have a big installed base of compute and sell some big trading optimized clusters into the market at a truly massive premium to these contracted rates. And you know, at least the at least at least analysts like that, they saw an opportunity. There's a lot of speculation they're going to raise capital. So, like you know, maybe what they're thinking is like, "Hey, we will show on a small chunk of capacity that we can generate really strong IRRs. Then we'll go raise equity capital and and we'll and we'll be off to the and we'll be off to the races and probably raise CapEx."
这是场灾难。但完全不是这么回事。他们刚刚公布了财报，他们并没有削减资本开支。真实情况是，他们看到 SpaceX 拥有很大的装机算力基础，并把一些针对交易优化的大集群卖进市场，相对于这些合约价拿到了真正巨额的溢价。而且你知道，至少分析师喜欢这个，他们看到了一个机会。市场上有很多猜测说他们要去融资。所以，你知道，也许他们的想法是，“嘿，我们先用一小部分产能证明我们能产生非常强的 IRR，然后我们去募股权资本，然后我们就可以大干一场了”，而且很可能会提高资本开支。
6:37
It doesn't look like what that's what they're doing. But nevertheless, the market sold off because it interpreted this very negatively. And I was really sure it wasn't negative. You know, there's a lot of telemetry into Meta's CapEx plans. None of that telemetry had shifted at all. If anything, it was, you know, continuing to or continued to get more aggressive. And then shortly after that, they released their best model in a long time, use 1.1, which is actually really a very good model. I mean, it was overshadowed by Grok 4.5, but it was a good model.
看起来他们做的并不是这个。但不管怎样，市场还是抛售了，因为它把这件事解读得非常负面。而我当时非常确定这并不是负面的。你知道，关于 Meta 的资本开支计划，我们有很多信号。那些信号没有一个发生过变化。如果说有变化，那也是在继续变得、或者说持续变得更激进。紧接着不久，他们发布了很长时间以来最好的模型，Llama 1.1，其实真的是一个非常好的模型。我是说，它被Grok 4.5 抢了风头，但确实是个好模型。
7:12
Um way better than anything in two years. So, just no chance they're taking their foot off the gas. Then Kimi comes out. And then there's this huge freak out about open source. And at the same time, this silica data token index kind of dips and flattens. And the two are connected. What the silica data token index captures is mix. And they don't see all the tokens, but because of GLM 5.2 and and and and then Kimi, all of it took a while to layer in. There's kind of a a mix shift in this data from more expensive frontier tokens, which probably have an inference margin we can debate whether it's 80, 90, or 95.
嗯，比过去两年的任何东西都好得多。所以完全不可能是他们松油门了。然后 Kimi 出来了。接着就是对开源的巨大恐慌。与此同时，Silica Data 的 token 指数有点下探然后走平。这两件事是有关联的。Silica Data 的 token 指数捕捉到的是结构变化。他们看不到所有的 token，但因为 GLM 5.2、还有 Kimi，所有这些都花了一段时间才逐步体现出来。这份数据里出现了一种结构性转移，从更贵的前沿模型 token——它的推理毛利率我们可以争论是 80%、90% 还是 95%，
7:55
>> Yeah. >> But super high. Towards open source tokens. And for whatever reason, the market thought this was negative, but the reality is a token is a token, and you need the exact same amount of compute to make a token all else equal. Takes the same amount of flops, the same amount of memory, the same amount of watts. Now, tokens are not equal, but broadly speaking, all open source taking share does is kind of take margin dollars out of the uh frontier model layer and effectively by thereby, you know, there there is elasticity, thereby driving token demand.
>> 是啊。>> 但总之非常高。——转向开源 token。不知道为什么，市场认为这是负面的，但现实是，token 就是 token，在其他条件相同的情况下，生成一个 token 需要的算力是完全一样的。需要同样多的浮点运算、同样多的内存、同样多的瓦特。当然，token 并不都是等价的，但大体上说，开源抢占份额所做的，无非是把毛利美元从前沿模型这一层拿走，并且实际上因此，你知道，这里存在弹性，从而推动了 token 需求。
8:37
You need more demand for compute. And the margins, you know, Anthropic and open source, they all run on the same underlying cloud providers who charge the same amount of compute. You know, so you're literally just um taking margin from frontier models and essentially driving more margin dollars into the AI infrastructure layer. And like I think that's >> That was the catalyst. This this combination of things. >> Well, yeah, then it kept it it's it's like Jenson is the world's largest supporter of open source.
你需要更多的算力需求。而利润率，你知道，Anthropic 和开源，它们都跑在同样的底层云服务商上，这些云收取同样的算力费用。你知道，所以你只不过是把利润从前沿模型那里拿走，本质上把更多的毛利美元推进了 AI 基础设施这一层。我觉得那就是——>> 那就是催化剂。就是这一连串的事情。>> 嗯，是啊，然后它就一直持续，这就好比，Jensen 是全世界最大的开源支持者。
9:11
Do we really And he's like a super idealistic guy. He's a patriotic American. I think he always does what's right. But is it does it really stand to reason that Jenson would be the world's biggest supporter of open source if it was bad for his business? >> [laughter] >> He'd still support if it was the right thing for the world. >> Yeah. >> But maybe it wouldn't be a signature issue. >> Yeah. >> And by the way, I think open source is really important to world where there's just one or two dominant for tier bottles that charge like 90% margins.
我们真的要——而且他是个超级理想主义的人，他是个爱国的美国人。我认为他总是在做正确的事。但如果开源对他的生意不利，Jensen 会成为全世界最大的开源支持者，这在道理上说得通吗？为了他的生意？>>（笑声）>> 如果那对世界是正确的事，他还是会支持。>> 是啊。>> 但也许那就不会成为他的标志性议题了。>> 是啊。>> 顺便说一句，我认为开源真的很重要，在一个只有一两个占主导地位的前沿模型、收着 90% 毛利率的世界里。
9:44
It's not good for humans. It might not be good for society. And I think we want a lot of bottles as we've discussed before. So then it's like, okay, the market digests that comes through with it. Then China has a DUV machine and this causes, you know, everybody said these baskets has caused a huge sell off in semi-cap equipment. And then we get to what I think is in a lot of ways, um like the real concern, which is real yields have gone up, which makes sense, you know, we're investing a lot to fund this investment and for sure credit is an increasing part of it even if the majority is still funded over what the majority is still funded out of operating cash flows.
那对人类不好，对社会可能也不好。我认为我们想要很多种模型，就像我们之前讨论过的那样。所以接着就是，好吧，市场消化了这件事。然后中国有了 DUV 光刻机，这导致，你知道，所有人都说这些——这在半导体设备股里引发了一波巨大的抛售。然后我们就到了我认为在很多方面才是真正令人担忧的地方，那就是实际收益率上升了，这也说得通，你知道，我们正在投入大量资金来为这轮投资融资，而且可以肯定信贷占的比重在上升，即使大部分仍然是靠——大部分仍然是靠经营性现金流来支撑的。
10:26
And so real yields go up and spreads widen. Meta Meta priced, um a bond last week and, you know, it it it did not price where you would think a meta bond would price. And this just shows that the credit market >> And the media CDS was was blowing >> All of these CD CDS for everybody is is blowing out. And, you know, very smart private capital people just like that, hey, this is just exactly what you would expect. These are just banks, you know, kind of hedging hedging their commitments. But nonetheless, it doesn't look good and these are undeniable facts. CDS is up, spreads are widened, real yield real yields are up.
所以实际收益率上升，利差扩大。Meta 上周定价了一笔债券，你知道，它的定价并不在你以为一笔Meta 债券该有的位置上。这就说明信贷市场——>> 而且 Meta 的 CDS 在往外走——>> 所有人的 CDS 都在大幅走阔。而且，你知道，非常聪明的私募资本的人就会说，嘿，这完全是你应该预料到的。这只是银行，你知道，在对冲它们的承诺敞口。但尽管如此，这看起来不好看，而且这些是不可否认的事实。CDS 上升了，利差扩大了，实际收益率上升了。
11:07
And that is that would be really really scary if we needed debt to finance this build out. And that's where I think it's this differential between spot and contract pricing for the installed base of compute is so important. >> Ramp is the only platform built to make your finance team leaner, faster, and better, saving businesses 5% annually [music] on average, so you can stay focused on growth. Ramp customers grew revenue 3.2 times faster than the average American business. [music] Visa, Foursquare, Cursor, Stripe, Notion, 11:FS, Shopify, and 70,000 other businesses all run on Ramp. Mine does, [music] too, and so should yours. Learn more at ramp.com/invest.
如果我们需要债务来给这轮建设融资，那这真的会非常非常可怕。这也正是我认为，已装机算力的现货价和合约价之间的这个价差为什么如此重要。>> Ramp 是唯一一个为让你的财务团队更精简、更快速、更出色而打造的平台，平均每年为企业节省 5%（音乐）的开支，让你可以专注于增长。Ramp 客户的营收增速是美国企业平均水平的 3.2 倍。（音乐）Visa、Foursquare、Cursor、Stripe、Notion、11:FS、Shopify，以及另外 7 万家企业都在用 Ramp。我的公司也在用，（音乐）你的也应该用。访问 ramp.com/invest 了解更多。
11:51
>> The best AI and software companies from OpenAI to Cursor to Perplexity use WorkOS to become enterprise-ready overnight, not in months. Visit workos.com [music] to skip the unglamorous infrastructure work and focus on your product. >> Felix by Rogo is a personal finance agent that turns a single prompt into finished client-ready [music] work using your firm's own templates, contacts, and standards. Send Felix an email like, [music] "Take these comments and turn them for me." Or, "Update my tracker with the context of these emails." Or, "Run the ability to [music] pay math on this buyer." And Felix sends back finished PowerPoint decks, Excel models, and sourced [music] research. Felix works the way your team already does, delivering work quickly and accurately around the clock. Learn more at rogo.ai/felix.
>> 从 OpenAI 到 Cursor 再到 Perplexity，最好的 AI 和软件公司都用 WorkOS 来实现企业级就绪——一夜之间，而不是几个月。访问 workos.com（音乐），跳过那些不光鲜的基础设施工作，专注于你的产品。>> Rogo 出品的 Felix 是一个个人金融智能体，它能把一句提示变成可直接交付给客户的（音乐）成果，用的是你自己公司的模板、联系人和标准。给 Felix 发一封邮件，比如，（音乐）“把这些意见处理一下。”或者，“用这些邮件里的内容更新我的追踪表。”或者，“给这个买家算一下（音乐）支付能力的测算。”然后 Felix 会把做好的 PowerPoint 演示、Excel 模型和标注来源的（音乐）研究发回给你。Felix 按你团队本来的方式工作，全天候快速而准确地交付成果。访问 rogo.ai/felix 了解更多。
12:32
[music] >> It's so important to understand what the financing will be like for the next 6 months or something. >> The degree to which this buildout is going to require credit. >> Right. >> Which would be the classic like capital cycle, and then we start to overextend ourselves with debt, and that's where things get dicey. >> 100% to the, you know, debt-fueled buildouts, you know, they demand immediate repayment. Yeah. So, if supply and demand get a little bit out of whack, things can unwind very, very, very quickly. That's what happened to the internet.
（音乐）>> 理解接下来 6 个月左右的融资环境会是什么样，这非常重要。>> 也就是这轮建设到底会在多大程度上需要依赖信贷。>> 对。>> 那就会是经典的资本周期，然后我们开始用债务过度扩张，那时候事情就变得危险了。>> 百分之百，你知道，靠债务推动的建设，它们要求立刻还款。是啊。所以，如果供给和需求稍微失衡一点，事情就可能非常非常、非常快地崩掉。互联网泡沫就是这么回事。
04经营现金流能否覆盖建设
13:02
And so, if one believes as I do, rightly or wrongly, and I'm like, after this month, I'm super open, you know, I'm I'm looking like I've been pressure testing all of these, and like, I really went deep on credit because, hey, this is real, it's undeniable. And if we need credit to fund this buildout, this is like a significant negative. >> [snorts] >> And if you model it out, has if you look at the amount of gigawatts that are supposed to come out in consensus estimates for hyperscalers, they're effectively modeled and these are gigawatts of Blackwell and Rubin.
所以，如果一个人像我这样认为——不管对错，而且经过这个月之后，我是非常开放的，你知道，我一直在给所有这些做压力测试，就像，我在信贷这块真的钻得很深，因为，嘿，这是真实的，是不可否认的。而如果我们需要信贷来为这轮建设融资，这就是一个重大的负面。>>（哼笑）>> 而如果你把它建个模，如果你看预计要上线的吉瓦数量在超大规模云厂商的一致预期里，他们实际上已经把这些东西模型化了，而且是以吉瓦为单位的 Blackwell 和 Rubin。
13:47
Rubin being Nvidia's next chip, Blackwell being the current chip. They are essentially modeled to monetize roughly at the rate of Ampere, which is two generations behind. Not at Hopper, but Ampere. So, there's 1.3 trillion and 1.3 to 1.4 trillion in hyperscale operating cash flow. If you just assume that they they're not I I think it's very unlikely they monetize at the rate of Ampere and we can we can go into why. Some of it comes from just, you know, seeing what is happening on the ground with demand here for real quantitative metrics.
Rubin 是英伟达的下一代芯片，Blackwell 是现在这一代。而模型里对它们变现效率的假设，基本上等同于 Ampere 的水平，也就是落后两代的东西。不是 Hopper，是 Ampere。所以，有 1.3 万亿，以及 1.3 到 1.4 万亿的超大规模厂商经营性现金流。如果你只是假设他们不会——我觉得他们以 Ampere 那种效率变现的可能性非常低，我们可以聊聊为什么。其中一部分就来自于，你懂的，实地看到这边的需求情况，那些真实的量化指标。
14:24
But, like, let's just say they monetize at a discount to current Blackwells. Then it's more like 2 trillion of operating cash flow and that kind of takes 700 billion of credit demand out. Um and, you know, and then obviously these, you know, ironically, has, you know, that improves all the credit ratios, has these installed bases of compute reprice. We're going to continue accelerating. Consensus is modeling in a deceleration, which I think is unlikely. Um then the credit metrics look better and then all of a sudden it gets easier to finance with credit.
但是，比方说，就算他们的变现比现在的 Blackwell 打个折扣。那也差不多是 2 万亿的经营性现金流，这一下就把 7000 亿的信贷需求给抵掉了。嗯，而且，你知道的，然后很明显这些，讽刺的是，这会改善所有的信用比率，让这些已装机的算力重新定价。我们还会继续加速。市场一致预期里假设的是减速，我觉得这不太可能。嗯，那样的话信用指标看起来会更好，然后突然之间用信贷来融资就变容易了。
15:03
Now, whether they whether they they choose to do that or not, we'll see, but this this is all a little bit, um you know, I think we spoke >> 2 months ago. >> No, but the time before that about kind of the risks of a Blackwell air pocket where you're spending hundreds of billions of dollars on Blackwells. they're mostly being used for trading initially. Trading does not generate, you know, a return. And that this could be a risk. And we actually really saw that kind of it, you know, in the first quarter.
至于他们会不会选择那么做，我们走着瞧，不过这些多少有点，嗯，我记得我们聊过——>> 两个月前。>> 不，是再之前那一次，聊的是 Blackwell「空档期」的风险，就是你花了几千亿美元买Blackwell，一开始主要用来做训练。而训练是不产生回报的。这可能是个风险。而我们确实在第一季度多少看到了这种情况。
15:38
And I think one reason, you know, like to the podcast two months ago, I I got comfortable with that risk was just that you were seeing such incredible things out of Anthropic. And then it's like, "Okay, well, the market's kind of going to look past this." And it did look past it in April, in May, in June. And then in July, because of this kind of confluence of things, stopped looking past it. Just as the operating cash flow started to really accelerate. And it's this is just a fact. It is accelerating at big scale.
我觉得有一个原因，就像两个月前那期播客说的，我之所以对这个风险比较放心，是因为你看到 Anthropic 那边做出了那么惊人的东西。然后就会觉得，「好吧，那市场大概会忽略掉这一点。」事实上四月、五月、六月市场确实忽略了。然后到了七月，因为各种因素凑到了一起，市场不再忽略了。而恰恰就在这个时候，经营性现金流开始真正加速。这就是个事实，它在很大的体量上还在加速。
16:14
Um and you know, like Microsoft, they brought out a huge slug of capacity in the month of June. That didn't even show up in the second quarter. So essentially, what this all comes down to is do you believe that the kind of quantitative demand signals seeing on the ground here in Silicon Valley from from private companies are going to continue such that the installed base of compute reprices higher as contracts roll off. >> Operating cash flows go up. >> Yeah, operating cash flows go up and you can fund this out of most of this out of operating cash flows. Maybe all of it.
嗯，还有你看微软，他们在六月份放出了一大批产能。那甚至都还没体现在第二季度里。所以说到底，这一切归结为：你相不相信在硅谷这边从私营公司那里看到的那些量化需求信号会持续下去，以至于随着合同到期，已装机算力的定价会往上走。>> 经营性现金流上升。>> 对，经营性现金流上升，这样你就能用经营性现金流覆盖掉大部分支出。也许是全部。
16:53
Like if it reprices at current rates, you could probably fund all of it for the next several years. It's so it's it's been a it has been a very un- usual episode in the market. And you know, in some ways, the fact that and we should talk about what the fundamentals are that are getting better that I'm talking about. You know, technicians would say it's actually in '22, okay, the market is worried about a recession, rates going up, you know, inflation. That's what the market was worried about in '22. You knew exactly what it was. Okay, deep seek, you know what it's worried about.
如果按现在的价格重新定价，未来几年你大概能全部靠这个来覆盖。所以这真的是一段非常不寻常的市场行情。而且某种程度上说，我们应该聊聊我说的那些正在变好的基本面到底是什么。你知道，技术派会说，其实在 22 年，好，市场担心衰退、利率上行、通胀。那就是 22 年市场担心的东西。你很清楚是什么。好，DeepSeek那次，你也知道它在担心什么。
17:33
Liberation day, you know what it's worried about. Uh there's something very clear and it in a weird way that's that is comforting reassuring. And here, you know, we talked about a lot of specific things, but it just feels all those specific things with the exception of credit like are are just kind of ridiculous. Um it's to the fact that it is still going down. You know, a technician would say, "Hey, that's that's a little scary. You know, it's definitionally the bullet you don't see that gets you." You know, I think we've talked before about how like I think the three most important words in investing aren't margin of safety, but I don't know.
「解放日」关税，你也知道担心什么。那种东西非常明确，而且很奇怪的是，这反而让人觉得踏实、安心。而现在，我们聊了很多具体的点，但感觉那些具体的点，除了信贷之外，都有点站不住脚。嗯，问题就在于它还在往下跌。技术派会说：「嘿，这有点吓人。按定义讲，打中你的是你没看见的那颗子弹。」我想我们以前聊过，投资里最重要的三个词不是「安全边际」，而是「我不知道」。
05硅谷实地：GPU 租金不降反涨
18:13
But just, you know, I've you've you've been out here for 2 months. I've been out here, you know, I literally spoke to a company this morning who rented a cluster of several and this is one of, you know, kind of sexiest startups that people want to be in business with. And they had rented a cluster of several thousand block wells. And we'll just call it, you know, somewhere in the mid $2 per GPU hour. They're renting the exact same cluster, exact same size cluster, essentially identical in every way, B200s. Said, "No no differences."
但就是说，你在这边待了两个月了，我也来了一段时间，我今天早上还跟一家公司聊过，他们租了一个集群，好几千卡，这还是那种大家都抢着跟它做生意的最热门创业公司之一。他们租了一个好几千张 Blackwell 的集群。价格大概就算它每 GPU 小时 2 块多美元吧。他们现在要租的是一模一样的集群，规模完全相同，各方面都基本一样，B200。人家说「没有任何差别」。
18:47
And they're hoping 7 months later to pay just under $4. Like you know, just you hear this today. And that's like that's pretty crazy because again, you would just you would expect a really like a gentle decline in prices would be bullish. Instead, you know, we're up, you know, depending on the starting point 50 to 60% in 6 or 7 months. And it just there've been so many anecdotes like that. Like I think one of the inference clouds I think it was based in I'm not sure. They went on a podcast and they essentially said we are planning to pay 100% more for Blackwell's when our contract expires. And that just means that essentially all the hyperscalers are under running.
而七个月之后，他们希望能谈到略低于 4 美元的价格。就是说，你今天就能听到这种事。这挺疯狂的，因为按常理你会预期价格温和下降才算是利好。结果呢，取决于你从哪个时点算起，六七个月里价格涨了 50% 到 60%。类似这样的案例太多了。我记得有一家推理云厂商，具体在哪儿我不太确定。他们上了一期播客，基本上说：我们合同到期后，打算为 Blackwell 多付 100% 的价钱。这就意味着，所有超大规模厂商本质上都投得不够。
19:38
And I haven't like my main kind of mission out here this week >> It's like pressure test? >> Yeah. >> Yeah. >> Find tell me something negative, you know? Like you know, the question I asked you, have you is there one negative quantitative metric you've you've heard? Has been what I've been asking everyone. >> The main thing people are saying is the Anthropic like the third-party data suggests that the Anthropic like curve started to go off of its trajectory a little bit. That's like the only thing that I >> I think I think that's I think that may very well be true, but then you have OpenAI and open source massively accelerating.
而我这周来这边的主要任务就是——>> 相当于压力测试？>> 对。>> 嗯。>> 就是「跟我说点负面的」，懂吧？就像我问你的那个问题：你有没有听到过哪怕一个负面的量化指标？这就是我一直在问每个人的问题。>> 大家主要提的是 Anthropic，第三方数据显示 Anthropic 那条曲线开始稍微偏离原来的轨迹了。基本上就只有这一条我——>> 我觉得，我觉得这很可能是真的，但另一边你有OpenAI 和开源在大幅加速。
20:17
And if you look at the sub it is not accelerating. Maybe I don't know that it looks the same. I think it may have accelerated. Like I think open source is a little bit of a you know, they talk about dark matter in the universe. Like open source is kind of dark matter to the public markets. You know, it's hard for public markets to measure it. But like if you just track what these inference clouds are saying and you know, these are people saying things on podcasts or people saying things in meetings they're not you know, audited financials but like demand is clearly accelerating which makes sense cuz you had this huge capability leap which you learn 5.2 and KBK3 which I think we're going to see continue. I think you're going to see Nvidia bring Neobtron steadily closer to the frontier. It has been a very like it's been a hubbly challenging month and but just it's also like wow, I've kind of pressure tested every assumption.
而如果你看订阅数据，它并没有在加速。也许我不确定它是不是看起来一样。我觉得它可能加速了。我觉得开源有点像——人们说宇宙里有暗物质，开源对公开市场来说就有点像暗物质。公开市场很难去度量它。但如果你就跟踪这些推理云厂商说的话——当然，这些都是人们在播客上说的，或者在会上说的，不是经过审计的财报——但需求明显在加速，这也说得通，因为你经历了一次巨大的能力跃迁，比如 5.2 和 K3，我觉得这还会继续。我觉得你会看到英伟达把 Nemotron 稳步推向前沿。这一个月确实挺难熬的，但同时也是，哇，我基本上把每一个假设都压力测试了一遍。
21:16
The underlying fundamentals are improving. Uh and stocks Nvidia's actually as we record this at its lowest forward PE of the last 10 years. >> Crazy. >> The only time the Sibbys have been cheaper were liberation day deep seek and that was those were kind of uh V bottoms. Um >> And that means to you just that the market thinks they're significantly over earning? >> Yeah, the market 100% thinks they're significantly over earning. And you we need to be humble. >> Maybe they are. >> Maybe they are. Um but like my kind of mission out here this week was to look for negative data points as hard as I could. I normally come to Silicon Valley and you know, there's a mixture of like okay, here's here's something negative, here's something positive, da da da. On balance, it's positive, you know, tech it creates value over time.
底层基本面在改善。而股价方面，就在我们录这期节目的时候，英伟达的远期市盈率是过去十年的最低点。>> 太离谱了。>> 唯一比这更便宜的时候是「解放日」和 DeepSeek 那两次，而那两次基本上都是 V 型底。>> 那对你来说意味着市场认为它们的盈利严重超常？>> 对，市场百分之百认为它们的盈利严重超常。而我们也得保持谦逊。>> 也许确实是。>> 也许确实是。嗯，但我这周来这边的任务，就是尽我所能去找负面的数据点。我平时来硅谷，通常会听到各种混杂的东西，好，这里有点负面的，这里有点正面的，等等。总体上是正面的，你懂的，科技长期在创造价值。
22:13
But I haven't been able to find one that is like a quantitative metric. Other like that that Anthropic third-party data, I would say that seems to be a hotly contested by the um by the Anthropic shareholders who are like >> [laughter] >> who are bound or kind of like chopping at the bit to tell you what they know. They're also very scared they're not going to get an IPO allocation >> [laughter] >> if it gets back to the company that they're the ones who said, "Actually things are great." You know, you can just see Anthropic shareholders like they want to be like, "It's not true." You know.
但这次我一个都没找到，一个像样的量化指标都没有。除了 Anthropic 那个第三方数据，而我要说，那个数据似乎被 Anthropic 的股东激烈地质疑，他们恨不得——>> [笑]>> 他们简直是摩拳擦掌，急着要把自己知道的告诉你。但他们也很怕拿不到 IPO 的份额，>> [笑] >> 万一传回公司说是他们说的「其实情况好得很」。你能想象 Anthropic 的股东那种心情，他们特别想说「这不是真的」。
22:45
>> [laughter] >> I mean, it's hard for me to believe that um open source and open AI have accelerated to the extent they did and but yeah, Anthropic is clearly, you know, kind of in the in the pole position. And oh, by the way, you know, Grok and Cursor have also, you can see from third-party data, like July was a pretty transformational month with um Grock 4.5 Grock builds coming out. So, it has been a tricky month and um and I have I have a friend um I have a friend at Fidelity who just says the way to have navigated like the last 3 years is just do the dumbest, most superficial thing as quickly as possible and just cycle between them.
>> [笑] >> 我是说，我很难相信开源和 OpenAI 加速到那种程度，但同时——是啊，Anthropic 显然还是处在领跑位置。而且顺带一提，Grok 和 Cursor也是，从第三方数据能看出来，七月是相当有转折意义的一个月，Grok 4.5、Grok 的构建功能都出来了。所以这是很棘手的一个月。嗯，我有个朋友，我有个在富达的朋友，他就说，过去三年最好的应对方式就是：用最快的速度做最蠢、最表面的那件事，然后在这些事之间来回切换。
23:36
>> What is that? What is that now? >> Well, that's just that has been to cut risk all month in response to these kind of narratives that just like factually except for credit are not true and the work we've done makes me think that credit just isn't going to matter has this reprice. Let's just say you do need credit to like build the flops we eat. Well, if credit's not there, it just means the flops that are there are going to be even more valuable cuz there is an interesting like essay that got sent sent to me.
>> 那现在这件事是什么？>> 嗯，那就是整个月都在降风险，去响应这些叙事——而这些叙事在事实层面上，除了信贷那条，都是不成立的。而我们做的研究让我觉得，一旦重新定价，信贷根本就不会成为问题。就算你确实需要信贷才能堆出我们要的算力，那好，如果信贷不到位，那就意味着已经存在的算力会更值钱。有一篇挺有意思的文章有人发给我。
06Claude 共识与持续学习的冲击
24:12
You know, I think we've talked before about Mike Mauboussin's theory that like a breakdown of diversity is kind of what leads you know, to bubbles and crashes. And essentially everyone I know in the public equity investment business, whether retail or institutional everything immediately, every piece of news gets fed into Claude. And Claude Claude code, sometimes, you know, a Claude agent and you know, Claude it's probabilistic. There's probably not that much variation in the way it's interpreting this news.
我想我们以前聊过 Mike Mauboussin 的理论，就是「多样性的崩塌」多少会导致泡沫和崩盘。而现在，我认识的公开股票投资行业里的每一个人，不管是散户还是机构，任何一条消息，立刻就被喂给 Claude。喂给 Claude、Claude Code，有时候是一个 Claude agent，而 Claude 是概率性的。它解读这条新闻的方式，大概不会有太大差异。
24:46
It's uh it's almost like we're back to um you know, in stock market terms like the like there's never really been this way in the stock market before, but people talk about the fragmentation of media and how it used to be like Walter Cronkite was the only voice of truth and now we don't have that anymore. It's like Claude is kind of Walter Cronkite for the stock market and everybody just believes >> Whatever it says. >> Whatever it says. [laughter] And this is leading to like >> really and and by the way, it's really smart, but it's not always right. It's not um it's interpretation isn't always correct. And with the stock market, you are fundamentally dealing about, you know, a probabilistic Bayesian interpretation of the future.
这就好像我们回到了——用股市的话说，股市上其实从来没有出现过这种情况，但人们会谈媒体的碎片化，说以前只有沃尔特·克朗凯特一个真相之声，现在再也没有那种东西了。而现在 Claude 有点像是股市的沃尔特·克朗凯特，所有人就都相信——>> 它说的一切。>> 它说的一切。[笑] 这就导致——>> 是真的，而且顺便说，它确实非常聪明，但它并不总是对的。它的解读不总是正确的。而在股市里，你本质上打交道的是对未来的一种概率性的、贝叶斯式的解读。
25:34
And so it just it feels like in the market, there is this Here's this piece of news. It gets fed through Claude. Claude interpreted this way. 90 a huge chunk of people trade on Claude's view. Um And so you've seen stuff. There's this guy uh TBU, TBU. He's like part of like the autonomous semiconductor mafia, but he posted this amazing chart of Japanese capacitor stocks. And he said we've had a capacitor an entire capacitor cycle in 6 weeks. And it's true, you know, the stocks like whether they double, triple, or quadruple, I don't know, but like vertical.
所以感觉市场上就变成这样：这里有一条新闻，喂进 Claude，Claude 这样解读——的方式。90% 的人会照着 Claude 的观点去交易。嗯，所以你也看到了一些东西。有个人叫 TBU，TBU。他算是那个「自主半导体黑帮」的一员，他发了一张特别精彩的图，是日本电容股的走势。他说我们在 6 周里走完了一整个电容周期。这是真的，你知道，那些股票到底是翻倍、三倍还是四倍我不知道，但基本上就是垂直拉升。
26:14
And then whoosh, you know what I mean? Like the actual fundamentals haven't even hit. And yet you've already had what probably would have normally been a 3-year cycle in like 6 weeks. >> What's your sense of being out here especially it makes me especially curious about this, the innovation that is going on here to improve the efficiency and every aspect of serving inference, of training models, et cetera, and how that will affect like public markets over time. Like have you learned anything interesting about like the long lead time innovation type stuff that has you especially excited or or curious?
然后嗖的一下，你懂我意思吧？实际的基本面都还没兑现呢。可你已经走完了一个正常情况下大概要三年才能走完的周期，只用了 6 周。>> 你在这边（硅谷）的感受是什么？这一点让我特别好奇，就是这里正在发生的那些创新，用来提升效率、提升推理服务和模型训练等等各个环节的效率，以及这些会怎样影响公开市场的长期走势。你有没有学到什么有意思的东西，关于那种长周期、长交付期的创新，让你特别兴奋或者特别好奇的？
26:51
>> Yeah, I am very curious. It was like all there seemed to be like a lot of people seem to feel like they're very close to solving continual learning and sample-efficient learning, which we've talked about before. And it is possible that if those are solved that, you know, could that be like a temporary like kind of like discontinuity? You know, it demand if it's that of, you know, having to like I think somebody told me that the uh like I was traded on effectively 20 billion tokens, and then it's like these models are traded on 300 trillion tokens. And if, you know, you could trade something on 10 trillion tokens and then let it out into the world and learn sample efficiently, you know, that that doesn't sound good for trading demand, but like trading as a percentage of semiconductor demand and compute is going to asymptote to something not approaching zero, but very small. But I would say that is the most kind of interesting, and you know, who knows if it's long horizon or short horizon.
>> 有，我非常好奇。好像所有人……好像很多人都觉得他们已经非常接近解决持续学习和样本高效学习的问题了，这个我们之前也聊过。而且有可能如果这些真的被解决了，你知道，那会不会形成一种暂时的、某种断层式的跳变？你知道，需求方面，如果是那种……我记得有人跟我说过，我这个人实际上只用了 200亿个 token 就「训练」出来了，而这些模型是用 300 万亿个 token 训练的。如果你能用 10 万亿个 token 训练出点什么，然后把它放到真实世界里去、让它以样本高效的方式学习，你知道，那听起来对训练需求不是好事，但训练占半导体需求和算力的比重会渐近趋向于某个不是零、但很小的数字。不过我会说那是最有意思的一点，而且谁知道呢，它到底是长周期还是短周期。
27:58
You know, SSI says that they're going to come out, you know, with their their model in in August. You know, there's this whole generation of new labs that are focused on this. >> And this would be good for the world. >> This would be amazing for the world, yeah. This would be awesome for the world. >> Yeah, we all want we want this, right? >> Yeah, we want this. It would be amazing for the world, and it's just it's hard for me to believe that that would actually be negative for AI infrastructure demand. But again, trying to be really, really open-minded. I I would say that was probably like the biggest like what what do we call it? Scientific or technical takeaway.
你知道，SSI 说他们会在 8 月份推出他们的模型。你知道，现在有一整代新的实验室都在专注做这件事。>> 而这对世界来说会是好事。>> 这对世界来说会非常棒，是的。这对世界来说会特别棒。>> 是啊，我们都想要，我们都希望这样，对吧？>> 对，我们想要这个。这对世界会非常棒，而且我很难相信这件事真的会对 AI 基础设施需求构成利空。但话说回来，我努力让自己保持非常非常开放的心态。我会说那大概是最大的……我们该怎么称呼它？科学上或技术上的收获。
28:34
But it's just, you know, it's also >> We still don't know. >> Well, yeah, and also like Nvidia is heavily involved with all of these startups. >> So, what would like if I was forced to if If just forced to come up with the set of circumstances that would really switch you around and get you really scared. Is it would it just be that this operating cash flow thing doesn't play out and therefore we just need to debt finance this whole thing? >> cash flow does not continue to accelerate. That that would be negative.
但这个嘛，你知道，也就是 >> 我们还是不知道。>> 嗯，对，而且 Nvidia 跟所有这些创业公司都有很深的绑定。>> 那么，如果非要我……如果非要我说出一组什么样的情况会真的让你态度反转、让你真的感到害怕。是不是就是经营性现金流这件事没有兑现，因此我们只能靠举债来给整件事融资？>> 现金流不再继续加速。那会是利空。
29:03
Um it that to some degree is going to be a function of how Anthropic, Open AI, Grok Cursor, we should call it Grok, and open source you know, if like if there was a pretty dramatic like contraction in GPU prices that was kind of sustained, I mean the market would react to that instantly. That would be worrisome. If it started to get to be really easy to get GPUs, I mean have you heard anyone say they have too many GPUs? >> [laughter] >> Like not not a single person. And it's not like it's the opposite. It sounds like a drug market or something. It really does. It's just wild. But yeah, I mean I think there's a long list of pretty obvious things. You know, if like Anthropic, Open AI, if the sum of these labs plateaus or you know, starts to decline, that's really negative unless it's just because open source tokens are not growing the pie and taking share. Uh and I do really think the future is like multi multi model.
嗯，这在某种程度上取决于 Anthropic、OpenAI、Grok、Cursor（我们就叫它 Grok 吧）还有开源的表现，你知道，如果 GPU 价格出现相当剧烈的下跌而且持续了一段时间，我是说，市场会立刻做出反应。那会让人担心。如果开始变得特别容易拿到 GPU——我是说，你听谁说过他们的 GPU 太多了吗？>> [笑] >> 一个人都没有。而且完全是反过来。这听起来简直像毒品市场之类的。真的就是这样。太疯狂了。不过是的，我觉得能列出一长串挺显而易见的信号。你知道，比如Anthropic、OpenAI，如果这些实验室加总起来开始见顶或者开始下滑，那是真的很利空——除非那只是因为开源 token 在做大蛋糕、在抢份额。呃我确实认为未来是多模型并存的。
07路由器与开源 token 为何不减算力
30:05
Yeah, I think particularly for the AI natives, they're going to want to take an open source model. It's got, you know, all these inference clouds have got really good at um you know, at supervised fine-tuning and reinforcement learning. So you can take your data, customize an open source model, and then get something that you can put behind a router, and the router routes it to often first your model, and then Claude, a frontier model, whatever, Claude, Grok, um checks it, and you can in a lot of cases get slightly better outcomes at half the cost.
是的，我觉得尤其是那些 AI 原生公司，他们会想拿一个开源模型，你知道，现在所有这些推理云在监督微调和强化学习上都做得非常好了。所以你可以拿自己的数据，定制一个开源模型，然后得到一个可以放在路由器后面的东西，路由器通常会先把请求路由到你自己的模型，然后再让 Claude、某个前沿模型，随便什么，Claude、Grok，嗯，来检查一遍，很多情况下你能以一半的成本拿到略好一点的结果。
30:45
But again, that half the cost, I think a lot of people hear that, they're like, "That's bad for AI demand." It's actually not at all because the cost the user pays has you know, it's just a function of the margin on the tokens, and you're literally just shifting tokens from really expensive tokens with like 90% gross margins to tokens with maybe, let's call it a 30% gross margin. And that's where the savings are coming from, but the tokens cost the same amount of compute to produce. And then also all these things are kind of happening at kind of um at different cycle times.
但还是那句话，「一半的成本」，我觉得很多人一听到这个，就会说，「那对 AI 需求是利空啊。」其实完全不是，因为用户支付的成本，你知道，它只是 token 上毛利率的函数，你本质上只是把 token 从毛利率 90% 的那种非常昂贵的 token，转移到毛利率也许……就叫它 30% 的 token 上。省下来的钱是从这儿来的，但生产这些 token 消耗的算力是一样的。而且这些事情都是在不同的周期节奏上发生的。
31:24
You know, all these, you know, big public companies are like, "Oh my god, my AI spend is 20x. I've burned my budget in 3 months." So, they set up a router, and that actually cuts their AI spend, but it doesn't really impact. It may actually increase the amount of tokens that they are generating just by shifting them to these cheaper open-source tokens, and that's just more compute. So, you know, a company getting smarter about which model to use for which task, that you know, that may lead to a a a stabilization of their spend or even a decline, but it actually has nothing to do with the amount of you know, GPU compute hours they are effectively consuming behind you know, these the these model layers of this router. The GPU compute hours probably are going up as you, you know, shift to these cheaper tokens you can use more of.
你知道，所有这些大型上市公司都在说，「我的天，我的 AI 支出涨了 20 倍。我三个月就把预算烧完了。」于是他们搭了一个路由器，那确实削减了他们的 AI 支出，但并没有真的产生影响。实际上它可能还增加了他们生成的 token 数量，只是把这些 token 转移到了更便宜的开源 token 上，而那就是更多的算力。所以，你知道，一家公司变得更聪明地去选择哪个任务该用哪个模型，你知道，那可能会让他们的支出趋于稳定甚至下降，但这其实跟他们在这些模型层、这些路由器背后实际消耗的 GPU 算力小时数完全没有关系。GPU 算力小时数很可能还在上升，因为你知道，你转向了这些更便宜的 token，所以你能用得更多。
32:24
So, and then, you know, that's happening to like a cutting-edge of public companies, and then you have this whole wave of AI natives, and like they're leading into this so hard, and they're not hiring humans. They're just really putting it mostly into tokens. And so, they're not slowing down. And then you have companies on the East Coast of America who have like barely adopted AI, companies, you know, broadly speaking, on other, you know, not on the coast who maybe aren't as cutting and then Europe who's like just trying to figure out how to regulate AI, >> [laughter] >> before using it.
所以，然后，你知道，这是发生在最前沿的那批上市公司身上的，然后你还有 AI 原生公司这一整波，他们在这件事上押得特别狠，而且他们不招人。他们基本上就是把钱投到 token 上。所以他们不会放慢。然后你还有美国东海岸的那些公司，他们几乎还没怎么用上 AI，还有，你知道，广义上说，不在两岸的那些公司，可能没那么前沿，然后还有欧洲，他们还在琢磨怎么监管 AI，>> [笑] >> 然后才用它。
33:02
>> Yeah, so just like there's kind of these differential differential kind of waves of adoption all happening at the same time. But the thought I can't get out of my mind is like I think I said it maybe last time, but just Yoc's estimate like I don't know, 500,000 people in the world, 250,000 maybe are using a genetic AI. And we're in a cute compute shortage. That's Do you know there's seven or eight billion people on the planet? What happens when we go from 500,000 >> to 1% >> 100 billion, you know, to 500 billion?
>> 是啊，就是说，这些不同层次的采用浪潮同时在发生。时间。但我脑子里挥之不去的一个念头是，我可能上次就说过，就是 Yoc 的那个估算，我不知道，全世界大概 50 万人，也许 25 万人在用 agentic AI。而我们已经处在严重的算力短缺里了。这……你知道地球上有七八十亿人吧？当我们从 50 万人 >> 涨到 1% >> 涨到 1 亿、5 亿人的时候会怎样？
33:39
And then I do think it's it it it is interesting, you know, a lot of people are just like, okay, well, you know, I I do think it's like helpful to post on X to see the pushback. And a lot of people are saying, well, you know, where fundamentally is the Okay, we accept your argument that hyperscalers are under earning in this compute re-prices. Are their operating cash flows going to accelerate and maybe we can fund this, but like who Where is that operating going to come from? Where is the customer?
然后我确实觉得这挺有意思的，你知道，很多人就会说，好吧，你知道，我确实觉得在 X 上发帖挺有用的，能看到别人的反驳。很多人会说，那从根本上讲，在哪儿……好，我们接受你的论点，就是超大规模云厂商在这轮算力重定价中是在少赚。那他们的经营性现金流会加速吗？也许我们能靠这个来给这一切买单，但那笔经营现金流到底从哪儿来？客户在哪儿？
34:09
And kind of definitionally it has to either come from, you know, faster economic growth through productivity kind of Satya's comments like either we're going to start growing 10% or we're not. Or labor substitution. And for sure, I think in a lot of these AI natives, you're seeing labor substitution, but not because they're firing people, they're just not hiring nearly as many humans. You know, the gross profit dollars per FTE and you know, A16Z iconic, a bunch of companies that have done this work, you know, they're you know, they're they're vertical uh particularly relative to past generations of startups.
从定义上讲，它要么来自通过生产率提升带来的更快的经济增长——就像 Satya 说的那样，要么我们开始以 10% 的速度增长，要么就没有。要么来自劳动力替代。我确实认为在很多这些 AI 原生公司身上，你看到了劳动力替代，但不是因为他们在裁员，而是他们招的人远远少了。你知道，每个全职员工创造的毛利润金额，还有 a16z 那张经典的图，一堆公司都做过这方面的研究，你知道，那条曲线是垂直向上的，尤其是相比过去几代创业公司。
34:48
And then it is interesting, you know, like are you kind of doing any surveys of your companies that their tokens bid relative to labor spend? >> Oh, yeah. I mean, it's always reported as a percent of percent tokens as a percent of like total comp spend or something like that. >> what are the ranges you've seen? >> I mean, like in the really pill companies, like it gets really high. 20%, 25%, something like that. >> Well, our our your our friend Dylan Patel at Jasper >> [laughter] >> So, he's an ASI maxi, but he's at 30%.
然后挺有意思的是，你知道，你有没有对你投的公司做过调研，看他们的 token 支出相对于人力支出是多少？>> 哦，有。我是说，一般都是以百分比的形式汇报的，token 占总薪酬支出的百分比之类的。差不多就是这样。>> 你看到的区间是多少？>> 我是说，在那些真正「吃了药」的激进公司里，这个比例会变得非常高。20%、25%，差不多这个量级。>> 嗯，我们的……你的、我们的朋友 Dylan Patel 在 Jasper >> [笑] >> 他是个 ASI 死忠，不过他那边是 30%。
35:19
>> Yeah. >> Uh >> He probably that's probably the highest one I've heard. >> Uh I've actually heard a 50. And there's 25 trillion dollars in knowledge work. And so, let's you know, let's say that that's you know, let's take your 20% number. That's 5 trillion and that either comes out of labor substitution or faster economic growth. And we really, really, really want to have, you know, humans it to come from faster economic growth. >> One interesting thing I heard this morning from one of the great like leading technology CEOs has founded several companies that if you look at the founder letting controlled companies and adjust for some of the like COVID era, you know, over hiring, like nobody's really laying people off. Like these are the people that would probably be most quick to adopt AI to you know, become more efficient or whatever. Like they're not really doing jack aside like huge scale layoffs, which probably tells you something about where they think there will be lots of opportunity to
>> 是啊。>> 呃 >> 那大概是我听到过的最高的了。>> 呃我其实听到过 50% 的。而知识工作的总规模有 25 万亿美元。所以，你知道，我们就……就用你说的 20% 这个数字吧。那就是 5 万亿美元，而这要么来自劳动力替代，要么来自更快的经济增长。我们真的真的真��希望，你知道，人类能让它来自更快的经济增长。>> 今天早上我从一位非常顶尖的、创办过好几家公司的科技公司 CEO 那里听到一件有意思的事：如果你去看那些由创始人掌控的公司，再把疫情时期那种过度招聘的因素调整掉，你会发现其实没人真在裁员。这些人本该是最会迅速采用 AI 来，你知道，提高效率之类的那批人。结果他们基本上没搞什么大规模裁员，这大概说明他们认为未来还会有大量机会需要人，再加上 >> 百分之百，多头的逻辑就是
36:14
still have people plus >> 100% well, the bull case >> So, growth not labor not labor growth. >> the bull case and you know, you've seen charts from Cognition, Ramp and Stripe that the companies that are spending the most on AI are growing growing meaningfully faster. >> Yeah, I love that cognition index. >> Yeah, the cognition index is wild. Now, all the skeptics will point out rightfully, it's not really controlling for industry, but then if like you dig down into it, you know, I think one of them gave an example of I forget if it was a plumber or an HVAC contractor, but like, you know, and everybody who's a blue-collar workers doing great cuz of AI.
>> 所以是增长，不是靠裁员换增长。>> 多头的逻辑，而且你知道，你也看过 Cognition、Ramp 和 Stripe 出的那些图，就是在 AI 上花钱最多的公司，增长明显更快。>> 是啊，我特别喜欢 Cognition 那个指数。>> 对，Cognition 那个指数太夸张了。当然，所有怀疑论者都会指出——而且说得对——它并没有真正控制行业变量，但如果你深入去看，你知道，我记得其中有人举了个例子，我忘了是水管工还是暖通空调承包商，但你知道，所有蓝领工人因为 AI 都过得很好。
08内存长约博弈与英伟达新模式
36:49
By the way, something that I think we should touch on and we we could do it now or later is just everybody is citing these LTAs. So, that we're everything is at a shortage. Everything is at a shortage right now. You know, if if there's weakness, it's just cuz we can't energize the gigawatts fast enough. The gigawatts are going to get energized like it, you know, regulatory policy is moving in a in a good way. You the turbine manufacturers, the diesel gen set manufacturers, you know, they're ramping up.
顺便说一句，有件事我觉得我们应该聊聊，现在聊或者一会儿聊都行，就是所有人都在引用那些长期协议（LTA）。所以说，现在什么都短缺。现在一切都处于短缺状态。你知道，如果说有什么疲软，那也只是因为我们没法足够快地把那些吉瓦级电力接进来。这些吉瓦最终都会通上电，你知道，监管政策是往好的方向发展。涡轮机制造商、柴油发电机组制造商，你知道，他们都在扩产。
37:15
>> You're you're you're ripping turbines off old airplanes and, you know, reconditioning them and then repurposing them. There's crazy things happening. Capitalism is very, very good at this. But I do think one of the most important questions in the market and like a transition of the market that like I got wrong is we are shifting, particularly for particularly for memory more than anything else, from, you know, crushing numbers in in the short term to they are trading short-term upside for these, you know, what do they call them? Supply chain agreements, long-term agreements, LTAs, where they essentially, you know, agree there's there's many flavors, but the customer prepays and it's, you know, there's a floor and a ceiling.
>> 你们甚至在从旧飞机上拆涡轮机下来，翻新之后再重新利用。现在真是各种疯狂的事都在发生。资本主义在这方面非常非常擅长。但我确实觉得，市场上最重要的问题之一——也是我判断错了的一个市场转变——就是我们正在转变，尤其是存储这块比其他任何环节都明显，从短期内赚得盆满钵满，转向他们愿意用短期的上行空间去换取那些，怎么说来着？供应链协议、长期协议，也就是 LTA。本质上就是双方约定，有很多不同的形式，但基本上是客户预付款，然后设一个价格下限和上限。
38:03
And this comes back to the point about labor because, you know, a lot of people after um you know, after kind of like firing, you know, too many people, were, you know, you during during COVID, were really reluctant to lay people off. And that, you know, they talked about labor hoarding if you remember a few years ago. You remember this? I'm just Let's just think about the game theory of breaking an LTA. So, there's four companies that like matter at scale. There's Amazon with their trade ups. There's Google with their TPUs.
这又回到了劳动力那个话题，因为很多人在经历了那种大规模裁员之后，你知道，在疫情期间裁掉了太多人，后来就变得非常不愿意再裁员了。所以几年前大家都在讲"劳动力囤积"，你还记得吧？我们不妨从博弈论的角度想想毁约 LTA 意味着什么。真正有规模的公司就四家。亚马逊有他们的自研芯片。谷歌有 TPU。
38:39
There's [snorts] AMD. And then there's Nvidia, who's like much bigger than everybody else combined. You know, let's just say it's 2027. It's very important to realize memory is The more memory you put with flop for a given unit of compute, the more tokens you get out. It's the single most important thing you could do to increase kind of token output per unit of compute. And then that obviously definitionally actually lowers costs, which is why the demand hasn't responded at all negatively. There's been no elasticity just because it's like kind of the only It's the axis that is dominating all others.
还有 [吸鼻子声] AMD。然后是英伟达，体量比其他所有家加起来还大得多。我们就假设现在是 2027 年。很重要的一点是要意识到，存储……在给定算力单位下，你配的内存越多、每 FLOP 配的内存越多，你产出的 token 就越多。这是提升单位算力 token 产出最重要的一件事。而这在定义上其实就是在降低成本，这也是为什么需求完全没有出现负面反应。根本没有弹性，因为这基本上是压倒其他所有维度的那个轴。
39:19
Um And this is like at some level like a giant Game of Thrones or Imposters between these companies. And okay, it's 2027. You're like or 2028. You're vaguely tempted to break one of these LTAs. Try and get a lower price. But to a large degree, market shares are I think for the next several years are going to be determined by supply by supply chain allocations and kind of what you have kind of pre-purchased. So, if you break the LTA and you And this is This is assuming we're not in a severe oversupply situation.
嗯，这在某种程度上就像这些公司之间的一场大型《权力的游戏》或者《狼人杀》。好，现在是 2027 年。或者 2028 年。你有点想毁掉其中一份 LTA，想压低价格。但很大程度上，我认为未来几年的市场份额将由供应链分配决定，取决于你预先采购了多少。所以如果你毁约……这是假设我们不处于严重供过于求的状态。
40:00
And but all the logic almost the game theory even holds in a severe oversupply situation. If you break your LTA and then in the next two or three years for any reason leverage shifts back to the memory guys, you're out of business. It's over. You know, like let's let's just say Google breaks an LTA. You know, there's there's an over there's an over supply of making this up in 28, 29. They break their LTAs. Well, if they're breaking their LTAs, it probably means, you know, you're over supply, prices are coming down. And then, you know, capacity that naturally contracts.
不过就算在严重供过于求的情况下，这套博弈论逻辑基本也成立。如果你毁了 LTA，然后在接下来两三年里因为任何原因，议价权又回到了存储厂商手里，你就完蛋了。彻底玩完。比如我们就说谷歌毁了一份 LTA。假设 28、29 年出现了供过于求——我随口编的。在28、29年编出来的。他们毁了 LTA。那如果他们毁约，很可能意味着当时供过于求、价格在跌。然后产能自然就收缩了。
40:36
Well, like what do you think's going to happen to Google's allocations? And then, you know, this is a cyclical industry and over supply is followed by under supply. What do you think they think is going to happen to their allocations next time? So, I just think given that this is like the access around which kind of everything is revolving, man, like you might blow up your entire business and your franchise by breaking an LTA. And that was never the case before, you know, Apple, who cares, you know, they're buying they don't have a competitor.
那你觉得谷歌的分配额度会怎样？而且这是个周期性行业，供过于求之后就是供不应求。你觉得下一轮他们的分配额度会是什么下场？所以我觉得，既然这是所有事情围绕转动的那个核心，天哪，你可能会因为毁掉一份 LTA 而把整个业务和你的护城河都炸掉。而这在以前是从来不会发生的。比如苹果，谁在乎呢，他们采购的时候没有竞争对手。
41:10
They're the over they're overwhelmingly the largest purchaser. They know they can do whatever this is, you know, going back three, four, five years. They know they can do whatever they want with no consequences cuz their volume is so big, you know, that even if they like super screw Hynix, Micron will of course take them. This is this is just different, you know, you have at least four players. Then you have all the startups. You're an investor in Etched. And if you break an LTA, and that they just say, "Okay, fine. You know what? Great. You broke the price agreement.
他们是压倒性最大的采购方。他们知道自己可以为所欲为——这是说三四五年前的事。他们知道自己想干什么就干什么，不会有任何后果，因为他们的量太大了，就算他们把海力士坑得很惨，美光当然也还是会接他们的单。现在完全不一样了，至少有四个玩家。另外还有那一堆创业公司。你就是 Etched 的投资人。如果你毁了 LTA，对方大可以说："行啊，没问题。你违反了价格协议，
41:46
We're going to break the volume agreement. And, you know, screw you. We're going to give the volume to your competitor." You just you just lost share, you know? That's so I think the the you know, it like I think, you know, Nvidia's dominance I think is uh like I think the current environment they state to which it favors Nvidia, like it is a hard for me to understand why it's trading at such a low multiple. You know, in other words, like if you need to be able to finance the chips and you do, nothing's more financeable than an Nvidia GPU. Nothing.
那我们就违反供货量协议。去你的，我们把货给你竞争对手。"你的份额就这么没了，懂吗？所以我觉得，英伟达的主导地位……我觉得当下这个环境对英伟达有利的程度，让我很难理解它的估值倍数为什么这么低。换句话说，如果你需要给这些芯片做融资——你确实需要——那没有什么比英伟达 GPU更容易融资的了。没有。
42:24
If you need to get, you know, land and power, well, they're doing a very good job of playing that chess game and and matchmaking. And then they've kind of rolled out this really clever, you know, new business model, which I would describe as kind of like a credit wrapper um with a revenue share if GPU prices are above a floor. Yeah. Um and this could lead to them like having a really giant cloud business effectively through royalties really quickly. And it is another way of kind of alleviating this um you know, cash flow mismatch. Like, hey, we're making all the cash.
如果你需要拿到土地和电力，他们在这盘棋上下得非常好，还在做撮合。匹配机制。然后他们还推出了一种非常聪明的新商业模式，我会把它描述成一种信用包装，再加上在 GPU 价格高于某个下限时的收入分成。是的。这可能让他们很快通过版税的方式实际上拥有一块巨大的云业务。而且这也是缓解现金流错配的另一种方式。就是说，嘿，钱都是我们在赚。
43:03
Yeah, and and like this isn't this isn't really vendor financing cuz they're not loading them the money. Somebody else is loading the GPU buyer the money. So, it's not quite vendor it's not vendor financing. It's um it's, you know, they're still making equity investments, but it's not it's not like you're just putting money into someone in return for them, you know, you know, and then some of that money, you know, is used to buy your chips, even though, you know, Nvidia said that they write into all their you know, equity investments that um you know, the money can't be used to buy Nvidia chips, but obviously money is fungible.
对，而且这其实算不上供应商融资，因为不是他们把钱借出去。是别人把钱借给 GPU 买家。所以这不太算是供应商融资。他们仍然在做股权投资，但这不像是你单纯把钱投给某个人，换取他们……然后那笔钱里有一部分被拿来买你的芯片——尽管英伟达说他们在所有股权投资里都写明了这笔钱不能用来买英伟达的芯片，但钱显然是可以互相替代的。
43:37
And um >> Funny thing. >> What's that? >> like a funny little thing. >> Yes. [laughter] Um Yeah. But, you know, I think at some level it probably makes everybody feel better. Um >> What would you do if you were the memory like if you were the CEO of Hynix? >> I'd do the exact same thing Nvidia's doing right now. >> Which is? >> I I would be going >> I say, all right, I'm going to say >> be going to the buyers of GPUs, Radeon, and whoever and say, I'll participate in the Nvidia credit wrapper. Now, their business is just inherently less stable and predictable, but in some way, and maybe they just put up some cash up front, so it's like they're not on the hook. You know what I mean? I'm just making this up.
嗯，>> 有意思。>> 什么？>> 就是个挺好笑的小细节。>> 是的。[笑] 嗯，是啊。但我觉得这在某种程度上大概能让所有人心里都好受一点。嗯 >> 如果你是存储厂商，比如你是海力士的 CEO，你会怎么做？>> 我会做和英伟达现在一模一样的事。>> 也就是？>> 我会去…… >> 我就说，好，我会说 >> 去找 GPU 的买家们，Radeon，以及其他任何人，然后说：我要参与英伟达那个信用包装。当然他们的业务本身没那么稳定、没那么可预测，但换个方式，也许他们就先拿出一笔现金，这样他们就不用一直担责。懂我意思吧？我这都是随口编的。
44:19
But like, do something like you can because you have money now, and credit markets are revolting. There many, you know, like, you know, the the people, I'm sure the, you know, our friends at, you know, Blackstone and Apollo are suggesting some variant of this to the memory companies. But hey, we will like put up some amount of money from our cash flow today, and then it's gone. It's, you know, surety uh that they, you know, makes the the the person who's extending the debt feel better, but we want a some sort of a cut of the ongoing revenues as well. Right.
但就是做点类似的事，因为你现在有钱，而且信贷市场正在闹脾气。我敢肯定，我们在黑石和阿波罗的那些朋友正在向存储厂商推销这类方案的各种变体。就是说，嘿，我们今天从现金流里拿出一部分钱放进去，这笔钱就出去了，算是一种担保，让出借方心里踏实一些，但我们也想从持续的收入里分一杯羹。对。
45:01
Like that is like 100% what I would do. And it's almost like a logical extension of, you know, the LTAs where they're kind of trading upside for durability. Here, you know, you can you know, you can effectively get a royalty on recurring revenues. And that is that is what Nvidia is doing. And I do think that is very misunderstood. And I think it would serve Nvidia well to really explain this. One, they're really bullish on AI. Um Essentially, every time they haven't taken an equity stake in something, it's been a mistake.
我百分之百会这么做。而且这几乎就是 LTA 的逻辑延伸——LTA 是用上行空间换确定性。而在这里，你实际上可以从经常性收入里拿到一份版税。这正是英伟达在做的事。我确实觉得这一点被严重误解了。我觉得英伟达好好解释一下这件事对他们有好处。第一，他们非常看好 AI。基本上，每一次他们没有入股某家公司，事后都证明是个错误。
45:39
You know, I mean, they've taken equity stakes in everything essentially except the memory companies that for a long while Anthropic that they took an equity stake in Anthropic. But like, why not if you have cash flow and you're bullish on AI? And Jensen because he sees every lab. He knows all the advances, you know, like all these continual learning labs, you know, safe superintelligence is now working with them. You know, he he sees everything and like what he sees makes him bullish. Um So, what have some equity upside and then two, have a revenue share and you're generating hundreds of billions of dollars of um of free cash flow um and a helping to kind of bridge, you know, what what is clearly kind of a gap at least, you know, given everybody's gone free cash flow negative until the operating cash flow accelerates enough that you can internally fund this.
我是说，他们基本上什么都投了，除了存储公司——还有很长一段时间没投Anthropic，后来他们也入股了 Anthropic。但你既然有现金流又看好 AI，为什么不投呢？而黄仁勋因为他能看到每一家实验室，他知道所有的进展，比如所有这些做持续学习的实验室，Safe Superintelligence 现在也在和他们合作。他什么都看得到，而他看到的东西让他非常看好。所以，一方面拿到一些股权上行，另一方面拿到收入分成，同时你每年产生几千亿美元的自由现金流，帮着弥合一个明摆着的缺口——毕竟现在所有人的自由现金流都转负了，直到经营性现金流增长到足以让你自己内部供血为止。
46:33
It's almost like I mean it's um very opportunistic and it like significant and in a good way and it significantly increases their revenue per gigawatt. And then it also strengthens their competitive position. You know, that's you know, you and I, we both have startups, but okay, that's that's that's great. Use that startups chip. Um well, what prices are they paying big atomic to me? Higher than Nvidia and all these guys. What prices are they paying for um HBM D-Ram? Higher. Um Can you finance those chips easily at the same rate as Nvidia? No. And so it's always like, you know, there's there's a real burden, particularly if you use HBM D-Ram, like you're just you're in the crosshairs of this. Um unless like actually you you know, maybe actually like they made really different architectural choices. Everything that's happening is actually pretty good for hip. Just are going back to game theory.
这几乎就是……我是说，这非常机会主义，而且影响很大，是好的那种大，还大幅提高了他们每吉瓦的收入。同时也强化了他们的竞争地位。你看，你和我都有创业公司，好，那很棒，用那家创业公司的芯片吧。那他们给大宗采购的定价是多少？比英伟达和这些公司都高。他们为 HBM 内存付的价格是多少？更高。那这些芯片能像英伟达的一样容易、以同样的利率融到资吗？不能。所以总是会有一个实实在在的负担，尤其是如果你用 HBM 内存，你就正好在这波冲击的枪口上。除非说，其实他们做了很不一样的架构选择。眼下发生的一切其实对他们挺有利的。回到博弈论。
09算力军备赛、中国 DUV 与看多硅谷
47:33
Entropic if they had been as aggressive on compute as OpenAI had been, they would have run away with it. And so now OpenAI is back in the game. I think Grok is in the game. Those are the companies on the Pareto frontier. >> And they have the compute. >> And do you think after watching that anyone is going to let off the gas? Cuz you just you know, it was I think 4 months ago that Dario was talking about how you know, it was a real it was a really thoughtful commentary, but he's like, it's really, really hard because, you know, if you buy too much compute, you could go bankrupt at the scale of these things.
Anthropic 如果在算力上像 OpenAI 那样激进，他们早就一骑绝尘了。所以现在 OpenAI 又回到了牌桌上。我觉得 Grok 也在牌桌上。这些就是处在帕累托前沿的公司。>> 而且他们有算力。>> 那你觉得看完这一幕之后，还会有人松油门吗？因为我记得大概四个月前，Dario 还在讲，那其实是一段非常深思熟虑的评论，他说这真的非常非常难，因为如果你买太多算力，以现在这个体量，你可能会破产。
48:12
But if you don't buy enough, you could lose. Well, >> We saw it happen. >> OpenAI just got back into the game, and now SpaceX is in the game in a big way of Grok 4.5 and Cursor. And like, after watching that, from a game theory perspective, is anybody going to back off anytime soon, especially if it can be funded out of operating cash flow? >> Vanta automates security and compliance for over 16,000 fast-moving companies like Ramp, Cursor, and Harvey, keeping them audit ready around the clock. It's the number one agentic trust platform, [music] and it now helps companies like yours watch for the risks that show up between audits across your [music] vendors, your AI tools, and your whole environment.
但如果你买得不够，你可能会输掉。是啊，>> 我们已经亲眼看到了。>> OpenAI 刚刚重返赛场，现在 SpaceX 也以 Grok 4.5 和 Cursor 的方式大举入局。看完那些之后，从博弈论的角度看，短期内会有谁选择退出吗，尤其是在这些投入可以靠经营性现金流来支撑的情况下？>> Vanta 为超过 16,000 家快速发展的公司实现安全与合规自动化，比如 Ramp、Cursor 和 Harvey，让它们全天候保持随时可审计的状态。它是排名第一的 agentic 信任平台，[音乐] 现在它还能帮助像你这样的公司，监控两次审计之间出现的风险，覆盖你的[音乐]供应商、你的AI 工具，以及你的整个环境。
48:51
Every new tool your team signs up for, every vendor that turns on AI features is an opportunity for something to go wrong, [music] and most security programs weren't built for AI's pace of growth. The Vanta agent works like a 24/7 [music] GRC engineer in the background finding issues, drafting fixes for you, and cutting vendor assessment time by up to 50%. Whether you're [music] a fast-growing startup or a global enterprise, Vanta helps you earn and prove trust. Invest Like the Best listeners get a special offer for $1,000 off at vanta.com/invest.
你的团队注册的每一个新工具，每一家开启 AI 功能的供应商，都可能带来出问题的风险，[音乐] 而大多数安全体系并不是为 AI 这种增长速度而设计的。Vanta agent 就像一个 7×24 小时[音乐]待命的 GRC 工程师在后台运行，发现问题、替你起草修复方案，并把供应商评估的时间缩短最多50%。无论你是[音乐]高速成长的初创公司，还是全球化的大企业，Vanta 都能帮你赢得并证明信任。Invest Like the Best 的听众可以在 vanta.com/invest 获得 1,000 美元优惠的专属福利。
49:23
Ridgeline is the first end-to-end system of record with embedded AI for investment management [music] firms, running portfolio accounting, reconciliation, reporting, trading, and compliance >> [music] >> on one unified platform. Firms are moving off legacy technology and onto Ridgeline because of how far ahead Ridgeline's AI features are compared [music] to anything else in investment management software. I've been hearing from a lot of investment managers about AI, and they fall roughly into two camps, with some unsure of where to even start, and others convinced they can build their own order management system over a weekend. [music] The reality is that running an investment firm will always require governance controls and a single source of truth for your data, and no amount of AI enthusiasm changes [music] that requirement. Ridgeline is built on exactly that foundation, which is why I believe that firms that come out ahead in the AI era will be the ones running on Ridgeline's unified [music] platform.
Ridgeline 是第一个面向投资管理[音乐]公司的端到端记录系统，并内嵌了 AI，在一个统一平台上运行投资组合会计、对账、报告、交易和合规 >> [音乐] >>。各家公司正在从老旧技术迁移到 Ridgeline，因为 Ridgeline 的 AI 功能相比[音乐]投资管理软件里的其他任何产品都领先太多。我听过很多投资管理人谈 AI，他们大致分成两派：有些人不知道该从哪儿入手，另一些人则确信自己能在一个周末里搭出自己的订单管理系统。[音乐] 现实是，运营一家投资公司永远都需要治理控制和数据的单一可信来源，再多的AI 热情也改变不了[音乐]这个需求。Ridgeline 正是建立在这样的基础之上，所以我相信，在 AI 时代脱颖而出的公司会是那些跑在 Ridgeline 统一[音乐]平台上的公司。
50:11
If you're serious about your firm's AI strategy, Ridgeline should be part of that conversation. You can request a demo at ridgeline.ai. [music] >> Have you met anyone in your travels out here that you would say is like way more bullish than you, and if so, what do they believe that you don't? >> I mean, essentially everyone out here is more bullish than me, man. >> [laughter] >> I like You know, I read this thing that Dworkesh wrote, and I was like >> The 3x compute price thing or whatever? >> Yeah, well, he was I forget what it was.
如果你真的重视公司的 AI 战略，Ridgeline 应该被纳入这场讨论。你可以在 ridgeline.ai 预约演示。[音乐] >> 你在外面跑的这些日子里，有没有遇到过比你更看多的人？如果有，他们相信什么是你不相信的？>> 我是说，基本上外面每个人都比我更看多，兄弟。>> [笑] >> 我挺喜欢，你知道，我读了 Dworkesh 写的那篇东西，我当时就想 >> 那个算力价格 3 倍什么的？>> 是啊，他是——我忘了具体是什么了。
50:37
>> No, no, it was like 15x or something. >> Yeah, but no, but just basically that um you know, renting an H100 for a year would cost $250,000. You know, um the salary >> that's 15x the current spot or something. Yeah. >> Exactly. Like, wow, you know, that was just like that >> in my book. >> That wasn't in my you know, forget my like Bayesian probability space of expected outcomes. That wasn't even in my >> [laughter] >> considered but dismissed as totally unlikely outcomes. You know, and then that guy is, you know, he's very Dworkesh, he's a very smart guy, he's very plugged in.
>> 不不，好像是 15 倍还是多少。>> 对，不过重点是，基本上就是说，租一块 H100 一年要花 25 万美元。你知道，那个价格 >> 大概是现在现货价的 15 倍之类的。对。>> 没错。就像，哇，你知道，那种事在我这儿 >> 完全>> 那根本不在我的，你知道，不在我那套贝叶斯概率的预期结果空间里。它甚至都没进到我那个 >> [笑] >> 考虑过但被当作极不可能而排除的结果清单里。你知道，而且那个人，Dworkesh 特别有他的风格，他非常聪明，消息也非常灵通。
51:15
And um and you know, and then he pointed out that like, hey, the you know, something like I think he just said margins on compute are going up, the amount of compute is going up, and inference margin's going up, and if you multiply those three, that's how you're getting this crazy acceleration in the sum of the labs plus open source, although obviously open source the margins on open source are not really going up. But I mean >> So, everyone's more [laughter] bullish >> Yeah, yeah, I like you know, I just I look at what's happening in the stock market and I feel like a foolish optimist.
然后，你知道，他还指出，嘿，大概是——我记得他说的是，算力的利润率在上升，算力的总量在上升，推理的利润率也在上升，如果你把这三个乘起来，那就是为什么各大实验室加上开源的总和会出现这种疯狂的加速，当然，开源那边显然利润率并没有真的在上升。不过我是说 >> 所以，大家都更[笑]看多>> 是啊，是啊，我，你知道，我就是看着股市里正在发生的事，我觉得自己像个愚蠢的乐观主义者。
51:51
And then when I talk to people whether it's people at the labs, whether anyone in this ecosystem like I'm like bearish relative to essentially everyone. >> [laughter] >> Which is just a strange state of affairs. What do you make of the DUV news out of China where I've seen reactions really along a spectrum of like this is the equivalent of like what ASML had in 2001 or something. Or like no, this is actually the first bit of news in a a new story for how we should think about the global supply of cutting-edge compute.
可等我跟人聊天，不管是实验室里的人，还是这个生态里的任何人，我都觉得自己相对基本上所有人都偏空。>> [笑] >> 这状态挺奇怪的。你怎么看中国那边关于 DUV 的消息？我看到的反应跨度非常大，有的说这相当于 ASML 在 2001 年左右的水平。也有的说，不，这其实是一个新叙事的第一条消息，关于我们该如何看待全球尖端算力的供给。
52:25
>> I think both could be true. You know, it's just like um like let's just make an analogy. Like let's just say a DUV machine was a jet turbine and now like an EUV machine is like warp drive. Um you know, or what whatever it's going to be, you know, a DUV machine is like a propeller plane EUV is like a jet turbine. Um but like they didn't have it before and now they allegedly do. And that is like a phase transition, you know, you it's like you've gone from like liquid to solid. Now that solid that you know, jet engine, prop plane, whatever is 25 years behind but still it's important and I don't think should be dismissed, but I also you know, it's kind of >> funny you just see this in the stock market, you know, it's like the stock market massively overreacts and then like if this ever hits ASML's orders, maybe it hits it in 5 years and like the market has forgotten about it, got worried about it, forgotten about it, got worried about it, forgotten about it multiple times um along the way. Um so I
>> 我觉得两者可能都对。你知道，就像，打个比方吧。比如说 DUV 机器是喷气涡轮发动机，而现在EUV 机器就像曲速引擎。嗯，或者随便怎么说吧，DUV 机器像螺旋桨飞机，EUV 像喷气涡轮。但关键是，他们以前没有，而现在据称有了。这就像一次相变，你知道，就像从液态变成了固态。现在那个固态的东西，你知道，喷气发动机也好、螺旋桨飞机也好，落后 25 年，但它仍然很重要，我不觉得应该被轻视，不过我也，你知道，这挺 >> 有意思的，你在股市上就能看到，就像股市会大幅过度反应，然后如果这件事真的影响到 ASML 的订单，可能是 5 年后才影响到，而那时市场早就把它忘了，先是担心，然后忘掉，再担心，再忘掉，一路上反复好几次。所以我确实觉得那大概是一次
53:36
do think that was probably an overreaction, but we shouldn't dismiss that either. And if you're China, like this is like really important to you. Um and they're you know, there are some reports that like an EV machine had been smuggled into China. Um and I mean, what a feat of espionage cuz those things are like giant machines. They're they're [laughter] huge. Um I don't know if that's true. You know, there's some noise about it, but um you know, China, they're really really good. They're really really smart. They work brutally hard.
过度反应，但我们也不该轻视它。而如果你是中国，这件事对你来说真的非常重要。嗯，而且有一些报道说，有 EUV 机器被走私进了中国。我是说，那得是多厉害的谍报手段，因为那些东西是巨型机器。它们非常非常[笑]庞大。我不知道是不是真的。你知道，有一些传言，但，你知道，中国人真的非常非常厉害。他们真的非常非常聪明。他们拼命地干活。
54:12
And you know, they see this is super important for them as a country. Um but are they going to go from the year 2001 to 2026 or even 2000 and you know, 30? Are they going to it it cuz it really is it is >> It's a learning by doing. >> It's it's a learning by doing. And you kind of have to Yeah, if like like you can't you can't accelerate the doing. You can't you can't teleport into the future. You actually have to go through those learning cycles. So, is it significant? Yes. Did the market overreact? Probably. But like I think a lot of like I think it's it's very hard as an American to really understand what is happening in China and like have like total conviction and clarity, you know, like for for better or worse, like we are decoupling. And um just that is a process that is been set in motion.
而且你知道，他们把这件事看成对国家极其重要的事。但他们会从 2001 年一下跳到 2026 年吗？哪怕是跳到 20、你知道，30 年？他们能做到吗？因为这真的是 >> 这是一个边做边学的过程。>> 对，这是一个边做边学的过程。而且你必须——对，你没法加速那个“做”的过程。你没法直接瞬移到未来。你真的得一轮一轮地走完那些学习周期。所以，它重要吗？重要。市场反应过度了吗？大概是的。但我觉得，作为一个美国人，其实很难真正理解中国正在发生什么，也很难有那种完全的笃定和清晰，你知道，不管是好是坏，我们正在脱钩。而这已经是一个被启动的过程。
10AI 原生公司与推理云的崛起
55:14
And at this point it almost feels like it's kind of self-reinforcing on each side. And you know, that's that's unfortunate. Um but we are where we are. And they're not they're not going to stop, neither are we. >> Any commentary on like every other company in America? Like I feel like right now it is 10 companies, couple private. >> Well, that that last month, I mean, everything but AI was vertical. And I do think open you know, open source getting closer to the frontier and companies like Fireworks making it really easy to customize a model such that you can get in some cases better than frontier for performance for meaningfully lower cost.
到了现在这个阶段，几乎感觉它在双方各自内部都在自我强化。你知道，这挺不幸的。但我们就处在这个位置。他们不会停下来，我们也不会。>> 对美国其他所有公司有什么看法吗？我感觉现在就是 10 家公司的天下，其中几家还是非上市的。>> 呃，上个月，我是说，除了 AI 之外的所有东西都是直线往上的。而且我确实认为，开源越来越接近前沿，加上像 Fireworks 这样的公司让你非常容易地定制模型，在某些情况下你能拿到比前沿模型更好的性能，成本却低得多。
56:01
That is a godsend for the software industry. And it's also a godsend for all these like, you know, there's there's a lot of AI natives and like all these AI natives, you know, it's like our friend Vashria, I think he said 2 years ago, I've never seen more companies go from like being founded to like $50 million a year in revenue and generating cash flow in like whatever it is, 9 months. And it's hard to know if any of them are durable because like back then, like it's like, hey, you know, these are a lot of people would dismiss them as chat GPT wrappers.
这对软件行业来说简直是天赐之物。而且对所有那些，你知道，有很多 AI 原生公司，这些 AI原生公司，就像我们的朋友 Vashria，我记得他两年前说过，我从没见过这么多公司，从成立到做到一年 5,000 万美元收入并且产生正现金流，只用了大概 9 个月。而且很难判断它们中有没有哪个是可持续的，因为在那会儿，就像，嘿，你知道，很多人会把它们贬为 ChatGPT 套壳。
56:34
Well, now with open source, you've actually you you've generated some data that's unique to your your use case, whatever your vertical you're going after has a wrapper is. Fireworks, they did come out with a really cool product called Nexus. And if you're using cloud code, open AI codex, grok build, it is literally three lines of code, like 20 words. And um Fireworks adjust your data kind of, you know, they can RL a model and then there's a router that sends the query and they've had amazing results.
但现在有了开源，你其实已经积累了一些对你的使用场景独有的数据，不管你切入的垂直领域里“套壳”指的是什么。Fireworks 确实推出了一个非常酷的产品，叫 Nexus。如果你在用 Claude Code、OpenAI Codex、Grok Build，那就真的只需要三行代码，大概 20 个词。然后 Fireworks 会处理你的数据，你知道，他们可以对一个模型做 RL，然后有一个路由器把查询分发出去，他们的效果非常惊人。
57:10
Um and this is kind of the solution for every AI native and that's why you saw, you know, Harvey uh before it was acquired, um Cursor leads so heavily into this. Harvey, Legora, all of them. Because if you can go from just using one, two, two three frontier models to use a >> Whatever's optimal. >> those frontier bottles for whatever it is, 30% 60% of your token consumption and then use your own RL bottle, all of a sudden you're not a rapper. You're way more defensible. Um >> I was so interested by that cursor thing that came out. I think it was cursor where it's sort of like a AI speed running like what we've learned amongst humans, which is you could use the frontier model to plan and then farm out tasks to the dumber models.
这基本上就是每一家 AI 原生公司的解法，所以你才会看到，你知道，Harvey 在被收购之前，还有 Cursor 都非常重地押在这上面。Harvey、Legora，全都是。因为如果你能从只用一个、两个、三个前沿模型，变成用 >> 用最优的那个。>> 让那些前沿模型承担，比如说，30% 到 60% 的 token 消耗，然后其余的用你自己 RL 出来的模型，一下子你就不是套壳了。你的护城河要深得多。嗯 >> 我对 Cursor 出的那个东西特别感兴趣。我记得是 Cursor，它有点像是 AI 在加速复现我们人类之间学到的东西，就是你可以用前沿模型来做规划，然后把任务分包给更笨的模型。
58:00
>> 100% >> And and it's 15 times more efficient or whatever the metric was. >> It it it may be that like if this is like super ironic, um but it may be that like lower margin open source tokens that are just a little bit behind the frontier and you know, we have a we have friends who believe that you know, frontier [clears throat] once a frontier model hits RSI it will actually have a dramatically lower cost >> to serve the local >> at at every at every level of intelligence by kind of distilling this and then there's no place for open source. I would say that's like a you know, a um Anthropic OpenAI Grok maximalist view. And you know, we should we shouldn't dismiss anything. I don't know or really important. Anything is possible. Like you know, we we we would have like be be very humble. I particularly want to be humble after the month I've had.
>> 百分之百 >> 而且效率高 15 倍，或者指标是多少来着。>> 有可能——这会非常讽刺——但有可能，那些低毛利的开源 token，只比前沿落后一点点，而且你知道，我们有些朋友相信，你知道，前沿模型[清嗓子]一旦达到 RSI（递归自我改进），它的服务成本反而会大幅下降 >> 在本地 >> 在每一个智能水平上，通过这种蒸馏方式，那样开源就没有立足之地了。我会说那算是一种，你知道，Anthropic / OpenAI / Grok 极大化派的观点。而且你知道，我们不应该轻视任何可能性。我不知道，这真的很重要。一切皆有可能。就像，你知道，我们应该非常谦逊。尤其是在我经历了这样一个月之后，我特别想保持谦逊。
59:00
But that doesn't seem that likely to be and >> Why? >> Well, um one because there are so many of these AI natives that have actually generated a decent amount of domain specific proprietary data. >> Yeah. >> And kind of before like open source had this moment to these inference clouds and these routers really developed like you kind of didn't have a choice. Like whatever the terms of service were, you accepted them. But if you can now kind of get off that treadmill, um that gives you a degree of independence, maybe durability, safety. But kind of going back to your point, it may be that these cheaper tokens just massively inflate the value of the most cutting-edge frontier tokens.
但那种情况看起来不太可能 >> 为什么？>> 嗯，一是因为有太多这类 AI 原生公司，它们确实积累了相当数量的领域专有数据。>> 是的。>> 而且在开源迎来这个时刻、这些推理云和路由器真正发展起来之前，你其实没有选择。就是不管服务条款写的是什么，你只能接受。但如果你现在能从那台跑步机上下来，那就给了你一定程度的独立性，也许还有持久性和安全性。但回到你刚才那个点，也有可能这些更便宜的 token 反而会极大地抬高最尖端前沿 token 的价值。
59:55
Because if like today if you have I you know I'm going to make this up, you know, 120 IQ open-source models, um and they're really cheap to run, well, doesn't that make a 160 IQ model that can orchestrate them more valuable? And so just we talked last time about how I've been really surprised that you know so much of the economic returns have accrued to the frontier. Now that that is changing with what we're seeing with these kind of inference clouds. Um together, Modal, um uh Base 10 in a very cash-efficient way.
因为如果像今天，如果你有——你知道，我随口编一个数——有 120 智商的开源模型，嗯，而且它们运行成本真的很低，那这不就让一个能调度它们的160智商的模型更值钱了吗？所以我们上次就聊过，我一直很意外，你知道，经济回报有那么大一部分都归到了前沿实验室手里。而现在这一点正在改变，我们从这些所谓的推理云身上就能看到。比如 Together、Modal，还有 Base 10，而且它们的现金效率非常高。
1:00:30
What's shocking about those business models is they're growing almost as fast as the frontier labs in the early days, but burning very little cash. Like it's it's pretty extraordinary, you know, from like you know look to go back to silly SaaS metrics like the you know the rule of 40 perspective. Like these are crazy numbers. >> Do you think there's a lot of instruction in just like the distribution of pay inside of an organization? Like the CEO makes X times more than the median person at a company and maybe that's frontier tokens versus, you know, open-source tokens.
这些商业模式让人震惊的地方在于，它们的增长速度几乎跟前沿实验室早期一样快，但烧的钱非常少。这真的挺不可思议的，你知道，拿那些老掉牙的 SaaS 指标来说，比如所谓的"40法则"来看，这些数字简直离谱。>> 你觉得这是不是有点像一个组织内部的薪酬分配结构？比如 CEO 拿的是公司中位数员工的 X 倍，也许这就对应着前沿模型的 token 和开源模型的 token 之间的关系。
1:01:04
>> Absolutely. Yeah. >> Something simple like we can >> Yeah, it may be that what we discussed last time where you know frontier tokens I think they may lose a like the pie is growing really really fast. They may continue to capture the overwhelming majority of economic value, but kind of not all of it the way they have been, and open source tokens might be the majority of token source tokens processed. It just again, going back, that's great for infrastructure demand because a token is a token and it takes the same amount of flops, watts, space, cooling to make.
>> 完全正确，是的。>> 类似这么简单的类比 >> 是的，可能就像我们上次讨论的那样，前沿 token我觉得它们可能会失去一部分——但整个蛋糕在飞快变大。它们可能还是会拿走绝大部分的经济价值，但不像过去那样几乎全拿，而开源 token 可能会占到所处理 token 的大多数。处理完毕。再说一遍，回到那个点，这对基础设施需求是好事，因为 token 就是 token，生成它需要同样多的算力、电力、场地和散热。
11最大风险是监管与叙事失败
1:01:40
>> What's the worst thing that could happen in AI? Is it regulatory? Is it some sort of like >> I think regulatory has to be the biggest risk. Um I mean, it's the most obvious risk. And so that was kind of one reason I was excited to be here this week was to just like I I want to be scared, you know? I like I don't I don't want to feel like a lunatic, you know, watching these stocks relative to um you know, get more cheaper thinking the expected forward returns are going up. You know, while you know, it feels like the on the ground fundamentals have like pretty materially improved.
>> AI 领域可能发生的最糟糕的事是什么？是监管吗？还是某种 >> 我觉得监管肯定是最大的风险。嗯，我是说，这是最明显的风险。所以这也是我这周很想来这儿的原因之一，我就是想被吓一吓，你懂吗？我不想觉得自己像个疯子，你知道，看着这些股票估值变得更便宜，还觉得预期的远期回报在上升。而与此同时，你知道，地面上的基本面感觉已经有了相当实质性的改善。
1:02:19
Um in July relative to even June, but I still came come away thinking like you know, regulation, it just has to be the biggest risk. Like you just can't ignore New York making a data center moratorium. It just like we are we're living in this weird post-factual, post-logical political world. It you know, and I mean, I think the AI industry it has done a terrible job of PR, and I do think they're >> it at least realizes that now. >> Yeah. >> Maybe if not fixed it, it realizes it. >> Yeah, but like kind of the narrative in Washington, you know, that that that the political narrative, you know, I think amongst a lot of ordinary Americans is like data centers, they're going to raise your electricity prices, they're going to take all your water, and they're going to take your job.
嗯，七月相比六月都是如此，但我最后还是会觉得，你知道，监管就是最大的风险。你没法忽视纽约州搞数据中心暂停令这件事。就感觉我们生活在一个后事实、后逻辑的怪异政治世界里。你知道，我是说，我觉得 AI 行业在公关上做得糟透了，而且我确实觉得他们 >> 至少现在意识到了这一点。>> 是的。>> 就算还没解决，至少意识到了。>> 对，但华盛顿那边的叙事，你知道，那种政治叙事，我觉得在很多普通美国人眼里就是：数据中心要抬高你的电费，要抢走你所有的水，还要抢走你的工作。
1:03:11
>> [laughter] >> And the reality is like given the deals that are being cut now, when a data center goes in, electricity prices actually generally go down for everyone around there because of behind the meter deals. This is that like data center pledge that kind of Trump asked people to to sign. Generally, the data center developer, you know, it used to be they just had to build a like, you know, whatever. They had to get the police department or the fire departments like, you know, new trucks and new cars and, you know, new body armor or whatever.
>> [笑] >> 但现实是，就以现在谈成的那些协议来说，数据中心一落地，电价对周边所有人来说其实通常是往下走的，因为有"表后"供电协议。这就是特朗普让大家签的那个所谓数据中心承诺。一般来说，数据中心开发商，你知道，以前他们只需要建个什么，你知道，随便什么。他们得给警察局或者消防局搞点，你知道，新卡车、新车，还有新防弹衣之类的。
1:03:40
Now, it's like, well, we're going to build you a hospital, a school, a new police station, and a fire station. And we're going to lower your power bills. How does that sound? And by the way, the jobs are ongoing because it turns out that you kind of need these plumbers, electricians, you know, HVAC contractors. And this is like data centers are like are in a lot of ways the best thing to happen for blue-collar wages in my lifetime. And yet, you have the Democrats who ostensibly represent the, you know, the blue, you know, these blue-collar workers taking their jobs away.
现在变成了：我们要给你们建一所医院、一所学校、一个新警察局，还有一个消防站。而且我们还要把你们的电费降下来。你觉得怎么样？顺便说一句，这些工作岗位是持续性的，因为事实证明你一直需要这些管道工、电工，还有暖通空调承包商。所以数据中心在很多方面是我这辈子见过的对蓝领工资最有利的一件事。然而，你却看到民主党人——名义上代表这些蓝领工人的那群人——在抢走他们的饭碗。
1:04:20
Um And so, um it it also like it's it's just kind of wild how like, what is the phrase? Like a lie can go around the world >> Faster than the truth gets out of bed, yeah. >> Yeah, faster than truth gets out of bed. But an author made a mistake in a book and overestimated the amount of water usage in data centers by 10,000 X. Not a little bit. Like not one order of magnitude. Not two orders of magnitude. Not three, you know? Um and um she's admitted that mistake many times. I was completely wrong.
嗯，所以，这事儿也挺离谱的，就像，那句话怎么说来着？谎言能绕地球一圈 >> 而真相还没起床，对。>> 对，谎言绕地球一圈，真相还没起床。但有位作者在书里犯了个错，把数据中心的用水量高估了一万倍。不是差一点点。不是差一个数量级，不是两个数量级，也不是三个，你懂吧？嗯，而且她已经多次承认了那个错误。说自己完全错了。
1:04:57
It's like been super debugged. >> uh Popeye effect. Did you Did you hear that example? >> No. >> The the, you know, Popeye eats spinach. The reason was same deal in an academic in an academic book. They placed the decimal two things wrong. So, spinach does not have more iron than everything else. It was just this one source and then that propagated through people still say it has more iron. >> I literally had I thought it had more iron. [laughter] I mean that's wild. That's wild. I literally thought spinach had more iron.
这事儿已经被扒得很清楚了。>> 呃，大力水手效应。你听过这个例子吗？>> 没有。>> 就是，你知道，大力水手吃菠菜那个。起因是一模一样的情况，出在一本学术著作里。他们把小数点点错了两位。所以菠菜的含铁量并不比其他东西都高。就是这么一个来源，然后传开了，人们到现在还说它含铁量最高。>> 我真的就一直以为它含铁量更高。[笑] 我是说，这太离谱了。太离谱了。我真的以为菠菜含铁量更高。
1:05:24
>> [laughter] >> That's amazing. >> Yeah, you learn something new every day. >> Same thing though. >> Uh yeah, it's the same thing and it's just So somebody just needs to tell the truth. Like like I I feel like the industry and I thought like jeez, maybe if nobody else is going to do it like I'll do it. Like there needs to be some sort of foundation. Maybe it's a pack that runs ads during the final four, during the NFL games, during college football games. >> Here's the virtues. >> World Series. Here's what a data center does. Your power a data center that signed this pledge in your community.
>> [笑] >> 太神奇了。>> 是啊，每天都能学到新东西。>> 不过是一回事。>> 呃，对，是一回事，就是所以得有人把真相讲出来。我感觉这个行业，我当时想，天哪，要是没别人来做，那我来做吧。就是说得有个什么基金会。也许是搞个政治行动委员会，在"最终四强"赛期间投广告，在 NFL 比赛期间、在大学橄榄球赛期间投。>> 讲讲好处。>> 世界大赛期间也投。讲讲数据中心到底是干什么的。你所在社区里签了这份承诺的数据中心。
1:05:56
>> Yeah. >> Your power prices are going to go down. They're almost certainly going to um you know like contribute to the community in a material way. You're going to see a massive influx of super high paying blue collar jobs that are going to persist and I think a lot of people thought that they were one time and they're just not. Like there's for sure a spike and then that moves to the next data center but there is an ongoing kind of you know, need for kind of RMA and then upgrades at these data centers and and technology is changing. So you're going to have more jobs, you're going to have cheaper power.
>> 对。>> 你的电费会降下来。它们几乎肯定会以实实在在的方式回馈社区。你会看到大量高薪蓝领岗位涌入，而且会持续下去。我觉得很多人以为这些岗位是一次性的，其实根本不是。当然会有一个高峰，然后转移到下一个数据中心，但这些数据中心还有持续的需求，你知道，要做 RMA 返修、要升级，而且技术还在变。所以你会有更多工作机会，电也会更便宜。
1:06:31
You're going to have a wealthier community. Um there's going to be no impact on on water, no impact on the environment. You know, and it's easy to build the data center 10 miles out of town, you know. And so like that story needs to be told along with, you know, like there are you know, we we we we we heard a we we heard a story I think we talked about it last time about how AI is increasingly really saving lives, curing rare diseases. Like we um you know, I think I can't remember if it was I I think it was at ASCO this year.
你的社区会更富裕。嗯，对水没有影响，对环境也没有影响。而且你知道，把数据中心建在离城镇十英里外是很容易的事。所以这样的故事得讲出来，还有，你知道，我们之前听到过一个——我想我们上次聊过——关于 AI 正越来越多地拯救生命、治愈罕见病的故事。比如我们，嗯，我记不清了，我想应该是今年的 ASCO 大会上。
1:07:01
You know, the kind of vibe, you know, the the vibe was like hey, we've this is the most scientific breakthroughs we've ever seen at a single conference. And for sure some of that is due to AI. And so we need to like tell those stories. Like, you know, if you have a sick child, you know, a sick parent, uh a sick loved one, like AI meaningfully increases the odds of them recovering. Like if we just we we need it's everybody needs to tell this. And I think people out here it's all of this is so blindingly obvious to them that they they >> seems everyone else already knows.
你知道，那种氛围，当时的感觉就是：嘿，这是我们在单场会议上见过的最多的科学突破。而且其中肯定有一部分要归功于 AI。所以我们得把这些故事讲出来。你知道，如果你家有个生病的孩子，或者生病的父母，或者生病的亲人，AI 能实实在在提高他们康复的几率。所以我们真的需要——每个人都得把这个讲出来。我觉得这边的人，这一切对他们来说太过显而易见了，以至于他们 >> 好像觉得别人早就知道了。
1:07:45
>> They they can't Yeah, they can't process that this is a true but wildly divergent view from most Americans. And so like I think the industry really needs to tell its story better. Cuz this is like New York, it just feels like it's the first of many and even in some of these deep red states, they're super pro-growth. They're just like, "Hey, you guys are not doing a good job telling your story. Then we can't we can't tell your story. If you tell your story though, we can retell it, but like you're the experts."
>> 他们没法，对，他们没法理解这是一个虽然正确、但跟大多数美国人的看法天差地别的观点。所以我觉得这个行业真的得把自己的故事讲得更好。因为像纽约这样的，感觉只是众多案例中的第一个，甚至在一些深红州，他们本来是非常支持发展的。他们就是会说："嘿，你们自己讲故事讲得不怎么样。那我们也没法替你们讲。但如果你们把故事讲出来，我们可以帮你们转述，可你们才是专家啊。"
12解耦推理芯片与 SpaceX 算力
1:08:25
Um you know, if you like like something I've if you do not speak your own truth, no one else will. >> [clears throat] >> Yeah. What have we missed with >> I think something that is missing from all of this conversation about compute is what is going to happen when you put these SRAM-based accelerators that are not constrained by HBM DRAM and are often made on older nodes that are not competing with like the latest and greatest GPUs. You can whether you there's when you disaggregate Fritz, there's people talk about prefill and decode, but decode has two parts, attention and feed forward network, and like the ultimate holy grail is if you could do pre-fill on one chip.
嗯，你知道，就像我一直觉得的，如果你不为自己发声，就没人会替你发声。>> [清嗓子] >> 是的。我们还漏了什么 >> 我觉得在所有关于算力的讨论里，缺失的一块是：当你把这些基于 SRAM 的加速器用上去，会发生什么？它们不受 HBM DRAM 的限制，而且往往是在较老的制程上制造的，不用去跟那些最新最强的 GPU 抢产能。你可以，不管你——当你做解耦的时候，人们会谈预填充和解码，但解码本身有两部分：注意力和前馈网络，而终极的圣杯是：你能不能在一块芯片上做预填充，
1:09:08
Um it probably doesn't have HBM DRAM. Do the attention on a super high-powered chip with HBM DRAM, and then do the feed forward network on one of these SRAM chips. But like the ROI on adding these SRAM accelerators, uh to the existing install base of compute and and new compute. But like what we're seeing is like you do better. You just can't beat SRAM in particular for that feed forward network. And And you just almost you can't, no matter how much you try to get the ratio of compute to HBM DRAM to SRAM on the chip correct. Like the workloads are always changing, and there's different workloads.
嗯，那块芯片可能不需要 HBM DRAM；在一块带 HBM DRAM 的超高性能芯片上做注意力；然后在这些 SRAM 芯片上做前馈网络。而把这些 SRAM 加速器加到现有算力装机量以及新增算力上，那个投资回报率——我们看到的情况是，效果确实更好。尤其是在前馈网络这块，你根本比不过 SRAM。而且你几乎做不到，不管你多努力去把芯片上算力、HBM DRAM 和 SRAM 三者的比例调到正好。因为负载一直在变，而且有各种不同的负载。
1:09:55
And like being able to disaggregate it to these three parts, uh like I think this is this is going to be really really positive for the ROI of AI. >> For some reason I just thought of a funny question, which I love their framing of Game of Thrones versus all these people. Can you imagine a player that is not currently on everyone's mind becoming relevant at like the major Game of Thrones scale? Like that could be like Micron all of a sudden, you know, it'd be like a sample answer to the question of someone that becomes as important as Anthropic, OpenAI, Microsoft, Amazon, you know, Nvidia, SpaceX.
所以能把它解耦成这三部分，呃，我觉得这对 AI 的投资回报率会是非常非常正面的一件事。>> 不知怎么我突然想到一个有意思的问题，我很喜欢他们那个把这些玩家比作《权力的游戏》的说法。你能想象有哪个现在还没进入所有人视野的玩家，将来会重要到《权力的游戏》那种量级吗？比如说可能突然冒出来的美光，你知道，这算是一个示例答案——那种会变得像 Anthropic、OpenAI、微软、亚马逊，你知道，英伟达、SpaceX 一样重要的角色。
1:10:30
>> Yeah. So like a dark horse Game of Thrones player? >> that come to mind. Um like Li Bu is probably a dark horse. Um I do think um Lin at Fireworks, she is like a just an absolute killer. Um I think uh you know, our friend Scott Wool. >> Mhm. >> You know, cognition is kind of like uh >> Here, here to that one. >> Yes. Uh uh I think those are uh the most obvious names. >> What about SpaceX? What's it been like watching that be digested by public markets at least initially? >> Uh uh >> Do you think the market understands it as a company? The most important new company to be public?
>> 是的。那算是《权力的游戏》里的黑马选手吗？>> 能想到的那种。嗯，比如 Li Bu 大概算是一匹黑马。嗯，我确实觉得，Fireworks 的 Lin，她绝对是个狠角色。嗯，我觉得，你知道的，还有我们的朋友 Scott Wu。>> 嗯哼。>> 你知道，Cognition 有点像是……>> 这个我举双手赞成。>> 是的。呃，我觉得这些是最显而易见的名字了。>> 那 SpaceX 呢？看着它被公开市场消化，至少是初期阶段，感觉如何？>> 呃，呃……>> 你觉得市场真的理解这家公司吗？理解它是最重要的新上市公司？
1:11:23
>> It doesn't really feel like it it does because it's kind of like such a it's such a like everything to me is the fundamentals have gotten better since it IPO'd. Like Rock 4.5, the Cursor acquisition. You know, Cursor uh has clearly accelerated meaningfully. And then they have showed that they could, you know, they've they've showed over the last 3 years they could bring out more compute faster than anyone at lower prices. And now we know that they could even adjusting for the spot first contract gap, like their big advantage was they came into the market you know, and just hit those spot highs.
>> 感觉并没有，因为这就像是……对我来说所有东西都在于，自从它 IPO 以来，基本面其实变得更好了。比如 Grok 4.5、收购 Cursor。你知道，Cursor 显然已经明显加速了。然后他们证明了自己可以，你知道，他们在过去三年里证明了，他们能比任何人更快、以更低的价格推出更多算力。现在我们知道，即使把现货价和首批合约价之间的差距考虑进去——他们最大的优势就是进入市场时，你知道，正好赶上了那些现货高点。
1:12:05
Uh uh And in a strange way, like one of the more bullish things for compute is like, you know, they put a vast amount of compute into the market overnight. And it wasn't even really a blip. It was like the market just utterly absorbed it, you know? Like just the free trade didn't sold out at all. Uh uh But you know, a you know, a Substack writer Wolfund Fund AI, they think that SpaceX is going to try and bring out 8 gigawatts of compute. I will never bet against Elon. But I mean that would be a truly incredible feat.
呃，而且在某种奇怪的意义上，对算力更看多的一点是，你知道，他们一夜之间往市场里投放了海量的算力。一夜之间。结果市场连一点波动都没有。就像市场把它彻底吸收了，你知道吗？就是说，自由交易市场根本没被撑爆。呃，不过你知道，有个 Substack 作者 Wolfund Fund AI，他们认为 SpaceX 要尝试推出 8 吉瓦的算力。我永远不会赌 Elon 会输。但我是说，那真的会是个难以置信的壮举。
1:12:46
And they are rates have gone up since they signed those last contracts, not down. And they're monetizing at something like 50 billion a gig. And ConsenSys estimates for next year are 73 billion. So, forget Starlink V3. Forget Starlink direct to cell. Grok 4.5 and Cursor, the sum of that probably hits a $10 billion ARR pretty quickly. For Forget all of that. Um you know, forget like the core base Starlink business. If they bring out anywhere near that, the ConsenSys estimate is 73 billion. And that's 8 gigs at 50 billion a gig.
而且自从他们签下最后那批合同以来，价格是涨了，不是跌了。他们的变现能力大概是每吉瓦 500 亿美元。而 ConsenSys 对明年的预期是 730 亿。所以，先别提 Starlink V3。也先别提 Starlink 直连手机。Grok 4.5 加上 Cursor，这两块加起来大概很快就能做到 100 亿美元的 ARR。把这些全都忘掉。嗯，你知道，连 Starlink 的核心基础业务也先忘掉。如果他们能推出接近那个数量的算力，ConsenSys 的预估是 730 亿。那就是 8 吉瓦乘以每吉瓦 500 亿。
1:13:28
And obviously, that would not all be lit up at the beginning of '27. And it seems very implausible to me. Like, I almost don't believe the Funder report. Um but you it To this day, the only companies that have brought out more than 500 megawatts of power in a year are the hyperscalers, Coreweave, Crusoe, and SpaceX. And SpaceX has kind of brought out the most the fastest at the lowest cost. And then people do actually really like their clusters. Um but again, it's kind of like the market is going to need to see that.
而且显然，那些并不会在 2027 年初就全部点亮。在我看来这非常不现实。我几乎不相信那份 Funder 的报告。嗯，但你……直到今天，一年内推出超过 500 兆瓦电力的公司只有那几家超大规模云厂商、CoreWeave、Crusoe，还有 SpaceX。而 SpaceX 大概是推出最多、最快、成本最低的。而且大家确实非常喜欢他们的集群。嗯，但话说回来，市场还是得亲眼看到这一点才行。
1:14:13
>> That would not be the market's interpretation of SpaceX today. >> No, no. Um and it does feel like, you know, there's this There's There's a big New York hedge fund short case on it. And I think they think, you know, oh, the spot price for compute's going to go down 90% and, you know, you're going to bring out all this You're going to bring all this on all this compute. It's not going to generate, you know, nearly as much revenue as you think. Maybe, but also want to be really clear like Like, I have seen those I have seen Yolt's companies, you know, do really impressive things over the year. Pretty the Funder AI report of 8 gigawatts at 18 months.
>> 那绝不是今天市场对 SpaceX 的解读。>> 不是，完全不是。嗯，而且确实感觉，你知道，有这么个……纽约有家大对冲基金在做空它，有一套做空逻辑。我觉得他们的想法是，你知道，哦，算力的现货价会跌 90%，而且，你知道，你要推出这么多……你要把这么多算力全部上线，它产生的收入根本不会有你想的那么多。也许吧，但我也想说得非常清楚，我这些年见过 Elon 的公司做出过真正令人印象深刻的事。就说那份 Funder AI 的报告，18 个月内 8 吉瓦。
1:14:48
I'm I'm just quoting that cuz it's public. It's available to everyone. Ooh, like that. >> yeah. >> That Yes. Um you know, I think one of Elon's phrases is we specialize in making the impossible late. >> [laughter] >> I've never heard that. That's great. >> Yeah. Um it you know, there's like kind of a lot of truth to that. Yeah, yeah. Um but I just think very little is built in from my perspective to that stock for the amount of compute that they might be able to bring on. And again, I don't think it's anywhere near eight.
我只是在引用它，因为它是公开的，所有人都能看到。哦，这个不错。>> 是啊。>> 那个……对。嗯，你知道，Elon 有句话是：我们的专长是把不可能的事做迟到。>> [笑声] >> 我从没听过这句。太妙了。>> 是啊。嗯，你知道，这话还挺有几分道理的。对，对。嗯，但我就是觉得，从我的角度看，那只股票几乎没有把他们可能上线的算力规模计入定价。而且再说一次，我不认为那个数字能接近 8 吉瓦。
1:15:30
Um and it's going to be really hard and energizing these GPUs is really hard. But they've been good at it and it doesn't feel like that's in estimates or really in people's thinking. >> I'm thinking about that funny meme that says SpaceX, the data center company? >> [laughter] >> You said yes, absolutely. Um and then I would also just say like from Yeah, I did spend a lot of time at Starbase and um orbital compute feels more real every day. >> Pretty cool to see that Starship landing the other day.
嗯，而且这会非常难，给这些 GPU 通上电真的非常难。但他们一直做得很好，而这一点感觉既不在预期里，也不真的在人们的思考范围内。>> 我想到那个搞笑的梗图，说 SpaceX，那家数据中心公司？>> [笑声] >> 你说过，是的，完全没错。嗯，然后我还想说，从……对，我确实在 Starbase 待了很长时间，轨道算力这件事感觉一天比一天真实。真实。>> 前几天看到 Starship 着陆，挺酷的。
1:16:02
>> Pretty cool to see the Starship landing and then it's you know, it is funny. There's our friends at Benchmark. They funded Star Cloud and I don't know last time Star Cloud is an orbital compute company that like SpaceX is kind of partnering with. Um they're going to I think let them use the Starlink laser technology, which is really important for orbital compute. And like but I do think that's like kind of a good sanity check. Last time I checked, you know, the Benchmark guys were pretty smart.
>> 看到 Starship 着陆确实挺酷的，然后你知道，有意思的是，我们在 Benchmark 的朋友们，他们投了 Star Cloud，我不知道上次……Star Cloud 是一家轨道算力公司，SpaceX算是在跟他们合作。嗯，我想他们会让 Star Cloud 使用 Starlink 的激光技术，这对轨道算力来说非常关键。而且我确实觉得这是一个挺好的现实检验。上次我了解的时候，你知道，Benchmark 那帮人是挺聪明的。
1:16:27
And they're not coming from the Elon ecosystem at all. And they chose to fund an orbital compute company at like you know, a decent valuation without the internal launch that SpaceX gets. And that's just to me that's a good like, "Hey, am am I >> crazy?" >> Am I crazy? And it's like, well, maybe I'm crazy and maybe Elon's crazy and maybe Benchmark is also crazy and maybe the SpaceX engineers are also crazy. But man, that just doesn't seem that probable to me. And I mean, we we should say should we say whose offices we're in?
而且他们完全不是从 Elon 的生态圈里来的。他们选择以一个相当不错的估值去投一家轨道算力公司，而且这家公司并没有 SpaceX 那种内部发射能力。没有。对我来说这就是个很好的检验：“嘿，我是不是疯了？”>> 我是不是疯了？然后就是，好吧，也许我疯了，也许 Elon 也疯了，也许 Benchmark 也疯了，也许SpaceX 的工程师们也都疯了。但老兄，这在我看来概率实在太低了。我是说，我们是不是该说一下我们现在在谁的办公室里？
1:17:08
>> Yeah, we're sitting in the middle of the famous table. >> Yes, this is their famous table. [laughter] Yes, this is their famous table for their famous dinners. Um so thank you Benchmark. Thank you Benchmark for this episode. Yes, thanks Eric. Um and and she we should thank them all. >> Um Eric Eric coordinated for me so he gets a special shout out. >> Thank you, Eric. >> Thank you all of the partners. >> Thank you, Eric. Well, you know, just you know, we will see where all of these stocks are in a year.
>> 对，我们就坐在那张著名的桌子中间。>> 没错，这就是他们那张著名的桌子。[笑声] 对，这就是他们办那些著名晚宴用的著名桌子。嗯，所以谢谢 Benchmark。谢谢 Benchmark赞助这一期。对，谢谢 Eric。嗯，我们应该谢谢他们所有人。>> 嗯，Eric，Eric 帮我协调的，所以他值得特别点名感谢。>> 谢谢你，Eric。>> 谢谢所有的合伙人。>> 谢谢你，Eric。好吧，你知道，我们就看一年后这些股票都在什么位置吧。
1:17:32
And the great thing is, you know, time will tell. You know, people are going to be right or wrong. You know, the future's probabilistic, but we are at like it's an exciting moment. >> Well, if we keep doing this on the the model release cycle, I'll see you in a couple weeks. >> [laughter] >> Yeah, it's crazy. >> As always a blast to do with you. >> You know how small advantages compound over time? That's true [music] in investing and just as true in how you run your company. Your spending system is your capital allocation strategy.
最棒的一点是，你知道，时间会给出答案。你知道，人们要么对，要么错。你知道，未来是概率性的，但我们正处在一个……这是个激动人心的时刻。>> 好吧，如果我们继续按模型发布的节奏来录，那我几周后就又见到你了。>> [笑声] >> 是啊，太疯狂了。>> 和你一起录永远都很尽兴。>> 你知道微小的优势是如何随时间复利累积的吗？这一点在投资中是真的 [音乐]，在你如何经营公司这件事上同样成立。你的支出体系就是你的资本配置策略。
1:18:02
Ramp makes it smarter by default. Better data, better decisions, better economics over time. See how [music] at ramp.com/invest. As your business grows, Vanta scales with you, automating compliance and giving you a single source of truth for security and risk. Learn more [music] at vanta.com/invest. Ridgeline is redefining asset management technology as a true partner, [music] not just a software vendor. They've helped firms 5x in scale, enabling faster growth, smarter operations, and a competitive edge. Visit ridgelineapps.com to see what they can unlock for your firm.
Ramp 让它默认变得更聪明。更好的数据、更好的决策、随时间积累的更好的经济效益。了解详情 [音乐] 请访问ramp.com/invest。随着你的业务成长，Vanta 与你一同扩展，自动化合规流程，为你的安全与风险提供唯一可信的数据来源。了解更多 [音乐] 请访问 vanta.com/invest。Ridgeline 正在重新定义资产管理技术，成为真正的合作伙伴 [音乐]，而不只是一个软件供应商。他们已经帮助多家公司将规模扩大 5 倍，实现更快的增长、更聪明的运营和竞争优势。访问 ridgelineapps.com看看他们能为你的公司释放什么潜力。
1:18:32
>> [music] >> The best AI and software companies from OpenAI to Cursor to Perplexity use Work OS to become enterprise ready overnight, not in months. >> [music] >> Visit workos.com to skip the unglamorous infrastructure work and focus on your product.
>> [音乐] >> 从 OpenAI 到 Cursor 再到 Perplexity，最优秀的 AI 和软件公司都在使用 WorkOS，让自己一夜之间具备企业级能力，而不是花上好几个月。>> [音乐] >> 访问 workos.com，跳过那些不起眼的基础设施工作，专注于你的产品。
视频总结 · 一句话概括与核心要点
一句话概括
盖文·贝克认为 2026 年 7 月 AI 股票 40–60% 的暴跌是市场误读所致：除信用利差走阔外，所有量化基本面指标（GPU 现货价格、token 用量、超大规模云厂商经营现金流）都在加速改善，装机算力随合同到期重新定价将让资本开支主要靠经营现金流覆盖，而英伟达却处于十年来最低的远期市盈率。
核心要点
- 7 月是「压缩成一个月的 2022 年」，但找不到一个负面量化指标。 大量 AI 股票从高点直线跌去 40–60%；贝克在硅谷两个月专门「找恐惧」，逐一问遍业内人，没人能拿出一个减速数据。GPU 可得性、GPU 零售价、DRAM 现货价、token 增长全部在加速。唯一的反例是第三方数据显示 Anthropic 增长曲线略偏离轨迹，但 OpenAI 与开源推理云同期大幅加速，且该数据被 Anthropic 股东激烈质疑。
- 合同价与现货价的巨大价差是被忽视的核心变量。 2024–25 年无论多头空头都预期旧 GPU 价格会下跌，于是纷纷签长期合约（新云也需承购协议才能融资）。结果 2026 年老 GPU 价格反而垂直上涨：一家明星初创公司 7 个月前以每 GPU 小时约 2.5 美元租 B200 集群，续约同样集群报价近 4 美元；一家推理云公开表示合约到期后计划为 Blackwell 多付 100%。这意味着整个超大规模云厂商的装机算力都在「低赚」，合约滚动到期时将向上重定价。
- 经营现金流已在加速，信用融资的需求可能被高估。 微软、Meta、亚马逊本季度合并经营现金流从 280 亿升至 320 亿；剔除异常多的一次性项目（主要是欧盟罚款）后为 350 亿，而这还在 Rubin 上线、合约重定价之前。共识模型把超大规模厂商的 Blackwell/Rubin 吉瓦按落后两代的 Ampere 货币化率估算，得出约 1.3–1.4 万亿经营现金流；若仅按低于当前 Blackwell 的价格估算，则约 2 万亿，可减少约 7000 亿的信用需求，并同步改善信用比率。
- 信用市场恶化是唯一「真实」的负面信号。 实际收益率上行、利差走阔、Meta 新发债定价明显偏弱、各家 CDS 全面走阔。若这轮建设必须靠债务融资，这就是重大风险：债务驱动的扩张要求即时回报，供需稍失衡便会迅速崩解，这正是互联网泡沫的死法。贝克的判断是：如果装机算力按现价重定价，未来几年资本开支几乎可全部由经营现金流覆盖；即便信贷缺位，现有算力只会更值钱。
- 开源崛起不是利空，而是把利润从模型层转移到基础设施层。 GLM 5.2 和 Kimi K3 引发市场对开源的恐慌，恰逢 token 指数走平——两者相关，因为指数捕捉的是从毛利 80–95% 的前沿 token 向开源 token 的「结构性转移」。但 token 就是 token：同样的 FLOPs、内存、瓦特。开源夺取份额只是压缩前沿实验室的利润率，通过需求弹性反而拉高算力消耗。企业用路由器把前沿 token 换成毛利约 30% 的开源 token，账面 AI 支出稳定或下降，实际消耗的 GPU 小时却在上升。旁证：黄仁勋是全球最大的开源支持者，若开源有损其生意，不会将其作为标志性议题。
- Meta「出租算力」被误读为削减资本开支。 实情是 Meta 看到 SpaceX 把训练优化集群以远高于合约价的溢价卖入市场，想借机在小块产能上证明高 IRR，可能为之后融资、进一步提高资本开支铺路。Meta 财报未削减资本开支，随后发布的模型（Muse 1.1）是两年来最强，没有松油门的迹象。
- Claude 成了股市的「克朗凯特」，多样性崩塌导致周期被压缩。 引用莫布森理论：观点多样性的崩塌导致泡沫与崩盘。如今公募机构与散户把每条新闻都喂给 Claude，其概率性解读差异不大，大量资金按同一解读交易。案例：日本电容器股票在 6 周内走完了通常需要 3 年的完整周期——先垂直拉升再暴跌，而基本面甚至还没兑现。
- 长期协议（LTA）的博弈论使违约几乎不可能。 内存是提升单位算力 token 产出的最重要杠杆，行业已从短期「碾压业绩」转向客户预付、设定上下限的长期协议。四家规模级买家（亚马逊 Trainium、谷歌 TPU、AMD、体量远超其余总和的英伟达）中若有谁在 2027–28 年因短期过剩撕毁协议，供应商会反手撕毁量的承诺、把产能给其竞争对手；周期性行业过剩之后必是短缺，届时分配份额将决定市场地位。这与过去苹果作为唯一超大买家可以随意压价的时代完全不同。
- 英伟达的「信用包装＋收入分成」是被误解的新商业模式。 由第三方向 GPU 买家放贷，英伟达提供担保并在 GPU 价格高于底价时分享收入——这不是厂商融资，却能借「特许权使用费」快速建起等效的云业务，显著提高每吉瓦收入并强化竞争壁垒（初创芯片公司的晶圆和 HBM 采购价更高、融资更难）。贝克若是海力士 CEO 会做完全相同的事。英伟达几乎每次没拿股权都是错误，而其远期市盈率处于十年最低，只在解放日和 DeepSeek 两次 V 型底更便宜过，说明市场 100% 认为它在严重「超赚」。
- 需求的终极来源是劳动替代或生产率增长，而非「找不到客户」。 知识工作市场规模 25 万亿美元；AI 原生公司 token 支出占人力总成本已达 20–25%，Dylan Patel 报 30%，有人报 50%。按 20% 算即 5 万亿，必须来自劳动替代或更快经济增长。目前替代表现为「不再多招人」而非裁员，创始人控制的公司剔除疫情超额招聘后几乎无人大规模裁员。Cognition、Ramp、Stripe 数据显示 AI 支出最高的公司增速显著更快。全球仅约 50 万人在用智能体 AI 就已造成算力短缺，扩张到 1% 人口时的需求难以想象。
结论与值得注意的细节
- 贝克自评「相对硅谷所有人都算空头」：所有实验室与生态内人士都比他更乐观。Dwarkesh Patel 提出 H100 年租金可达 25 万美元（约当前现货 15 倍），此前根本不在他的概率空间内。他的 Fidelity 朋友总结近三年的生存法则：「尽快做最愚蠢最肤浅的事，然后在这些事之间轮换」。
- 让他真正转空的条件：经营现金流停止加速；GPU 价格出现持续性大幅收缩；有人说自己 GPU 太多（目前无一例，「像毒品市场」）；各实验室总和增长停滞或下降且并非因开源扩大蛋糕。最大宏观风险是监管——纽约州数据中心暂停令只是开端。
- AI 行业公关彻底失败：华盛顿叙事是「数据中心抬高电价、抢水、抢工作」，现实是 behind-the-meter 协议使周边电价下降，开发商承诺建医院学校，数据中心是他一生所见对蓝领工资最有利的事。某作者书中把数据中心用水量高估了一万倍，已多次公开认错却仍在传播（类比菠菜含铁量小数点错位的「大力水手效应」）。他考虑亲自资助一个在 NFL、四强赛期间投放广告的组织来讲这个故事。
- 持续学习与样本高效学习接近突破：多人认为很快能解决。若模型从 300 万亿 token 训练降到 10 万亿再放入世界自学，训练需求占算力比重将渐近极小值，但他不信这对基础设施总需求是负面。SSI 称 8 月发布模型。
- SRAM 加速器是被遗漏的话题：将推理拆解为 prefill、attention、feed-forward 三段，分别用不同芯片处理，feed-forward 用不受 HBM 约束、可在旧制程制造的 SRAM 芯片，ROI 极高，对 AI 投资回报是显著正面。
- SpaceX 是「数据中心公司」：过去三年比任何人更快、更便宜地上线算力，一夜之间向市场投放海量算力却被完全吸收。Wolfund AI 报告称其拟在 18 个月内上线 8 吉瓦，按每吉瓦约 500 亿美元货币化，远超明年 730 亿共识——贝克不信能到 8 吉瓦，但认为股价几乎没有计入此项。纽约某大对冲基金做空逻辑是算力现货价将跌 90%。轨道算力「一天比一天真实」，Benchmark 在非马斯克生态下投资 Star Cloud 是一个理性校验。马斯克名言：「我们专长于让不可能的事迟到」。
- 中国 DUV 突破：相当于从无到有的相变，但落后约 25 年，学习曲线无法瞬移。市场大概率过度反应，但不应忽视；中美脱钩已自我强化，双方都不会停。
- 黑马玩家：Fireworks 的 Lin Qiao（其 Nexus 产品三行代码即可让 AI 原生公司用私有数据 RL 出自有模型并路由，是 Cursor、Harvey、Legora 摆脱「套壳」标签的关键）、Cognition 的 Scott Wu、Libu。推理云（Fireworks、Together、Modal、Baseten）增速接近早期前沿实验室，却几乎不烧现金。
- 廉价开源 token 可能反而抬高前沿 token 价值：若 120 智商的开源模型极便宜，能够编排它们的 160 智商模型只会更值钱——类似公司内 CEO 薪酬与中位数员工的分布。
- 本期在 Benchmark 办公室著名的晚餐长桌上录制；两人调侃播客节奏已与模型发布周期同步，且此前几期恰逢局部市场高点，「这次没人能这么说了」。
核心句型 · 9
1. however you cut it, whether you cut X, whether you cut Y
“However you cut it, whether you cut GPU availability, whether you cut GPU retail pricing”
cut 在此指「切分、分析数据」。先用 however you cut it 概括「怎么看都一样」，再用 whether you cut 逐一列举维度。适合汇报多角度验证同一结论。
2. does it really stand to reason that … if …?
“Does it really stand to reason that Jenson would be the world's biggest supporter of open source if it was bad for his business?”
stand to reason 意为「合乎情理」。用反问句质疑对方结论与前提的一致性，是英语辩论中常见的归谬法开头。
3. That's not at all what it was. What it was is …
“This is not at all what it was. What it was is they saw SpaceX have a big installed base of compute”
先全盘否定流行解读，再用 What it was is 引出真实情况。第二句的 is 后面直接接完整句子，口语中常见。适合纠正误读。
4. If anything, it was …
“None of that telemetry had shifted at all. If anything, it was continuing to get more aggressive.”
if anything 意为「如果说有变化的话，反而是…」，用来在否定之后加一个反方向的补充。仿写：Sales didn't fall. If anything, they rose.
5. It's definitionally X that Y
“It's definitionally the bullet you don't see that gets you”
强调句 It's … that … 加上副词 definitionally（按定义），把一句格言说得像定理。适合概括一个必然成立的规律。
6. Nothing's more X than Y. Nothing.
“Nothing's more financeable than an Nvidia GPU. Nothing.”
最高级否定句后单独重复 Nothing 作为一句话，是口语中加重语气的手法。书面写作可少用，演讲和论辩中效果很强。
7. Forget X. Forget Y. … Forget all of that.
“So, forget Starlink V3. Forget Starlink direct to cell. … Forget all of that.”
连续祈使句 Forget 把次要因素一一排除，最后用 Forget all of that 收束，突出「只看这一项就够了」。适合做保守估算时排除其他上行因素。
8. I will never bet against X. But …
“I will never bet against Elon. But I mean that would be a truly incredible feat.”
bet against 意为「赌某人失败」。先表明基本立场，再用 But 引出保留意见，既尊重对方又不放弃判断。
9. let's just say (that) …
“Let's just say they monetize at a discount to current Blackwells. Then it's more like 2 trillion”
let's just say 引出假设性情景，后接 then 给出推论。用于敏感性分析或思想实验，比 suppose 更口语，比 if 更有「退一步讲」的意味。
词汇精讲 · 153 · 按出现顺序
lunatic
/ˈluːnətɪk/
n.
0:00
疯子；此处指「感觉自己像个疯子」的自嘲
forward PE
phr.
0:00
远期市盈率，按未来 12 个月预期盈利计算
pressure test
phr.
0:00
压力测试；对论点反复检验找漏洞
cadence
/ˈkeɪdns/
n.
0:48
节奏；周期（release cadence 发布节奏）
coincident with
phr.
0:48
与…同时发生；正好赶上
deceleration
/ˌdiːseləˈreɪʃn/
n.
1:23
减速；增速放缓
vantage points
/ˈvæntɪdʒ pɔɪnts/
n.
2:02
视角；观察的有利位置
however you cut it
phr.
2:02
不管怎么看、怎么切分数据
spot price
phr.
2:02
现货价（对应 contract 合约价）
chug along
phr.
2:44
缓慢而稳定地推进
pumping out
phr.
2:44
大量产出（现金流、产品）
hyperscale
/ˈhaɪpərskeɪl/
adj.
2:44
超大规模的；hyperscaler 指微软、亚马逊等云巨头
precipitously
/prɪˈsɪpɪtəsli/
adv.
2:44
陡然地；断崖式地
off-take agreement
phr.
3:51
包销协议，买方承诺购买未来产出
installed compute
phr.
3:51
已装机算力（installed base 装机存量）
roll off
phr.
3:51
（合同）到期终止
repriced
/riːˈpraɪst/
v.
4:35
重新定价
unusual items
phr.
4:35
非经常性项目（财务术语）
one-timers
/ˈwʌnˌtaɪmərz/
n.
5:07
一次性项目（口语化财务表达）
material
/məˈtɪriəl/
adj.
5:07
实质性的、重大的（财务用语）
light up
phr.
5:07
点亮、启用（新算力）
premium
/ˈpriːmiəm/
n.
5:50
溢价（at a premium to 高于…）
IRRs
n.
5:50
内部收益率（internal rate of return）
off to the races
phr.
5:50
顺利启动、大干一场
sold off
phr.
6:37
（市场）抛售
telemetry
/təˈlemətri/
n.
6:37
遥测数据；引申为可追踪的信号
overshadowed
/ˌoʊvərˈʃædoʊd/
v.
6:37
使黯然失色；被抢风头
taking their foot off the gas
phr.
7:12
松油门；放缓投入
freak out
phr.
7:12
恐慌；惊慌失措
mix shift
phr.
7:12
结构性转移，组合构成的变化
layer in
phr.
7:12
逐步体现、叠加进（数据）
all else equal
phr.
7:55
其他条件相同的情况下
elasticity
/ˌiːlæˈstɪsəti/
n.
7:55
（价格）弹性，降价带来的需求增加
stand to reason
phr.
9:11
合乎情理、顺理成章
signature issue
phr.
9:11
标志性议题
semi-cap equipment
phr.
9:44
半导体设备（semiconductor capital equipment）
real yields
phr.
9:44
实际收益率，名义利率减通胀
spreads widen
phr.
10:26
信用利差扩大
blowing out
phr.
10:26
（利差、CDS）大幅走阔
CDS
n.
10:26
信用违约互换，为债券违约买的保险
differential
/ˌdɪfəˈrenʃl/
n.
11:07
差额；价差
buildout
/ˈbɪldaʊt/
n.
12:32
建设；基础设施扩建
overextend
/ˌoʊvərɪkˈstend/
v.
12:32
过度扩张（尤指负债）
dicey
/ˈdaɪsi/
adj.
12:32
危险的、不确定的
out of whack
phr.
12:32
失衡、失常
unwind
/ʌnˈwaɪnd/
v.
12:32
崩解；（头寸）平仓
consensus estimates
phr.
13:02
市场一致预期
monetize
/ˈmɑːnətaɪz/
v.
13:47
变现
credit ratios
phr.
14:24
信用比率（如负债/现金流）
air pocket
phr.
15:03
气穴；引申为收入空档期
confluence
/ˈkɑːnfluəns/
n.
15:38
汇合；多种因素叠加
look past
phr.
15:38
忽略、看穿（短期问题）
slug of capacity
phr.
16:14
一大批产能（slug 一大块）
technicians
/tekˈnɪʃnz/
n.
16:53
技术分析派（看图表的交易员）
margin of safety
phr.
17:33
安全边际（价值投资术语）
anecdotes
/ˈænɪkdoʊts/
n.
18:47
轶事；个案证据
under running
phr.
18:47
投入不足、少赚（under-earning 之意）
dark matter
phr.
20:17
暗物质；比喻看不见但存在的量
audited financials
phr.
20:17
经审计的财务报表
V bottoms
phr.
21:16
V 型底，急跌后急速反弹
over earning
phr.
21:16
盈利超常、盈利不可持续
hotly contested
phr.
22:13
激烈争议的
chopping at the bit
phr.
22:13
摩拳擦掌、迫不及待（标准写法 champing at the bit）
allocation
/ˌæləˈkeɪʃn/
n.
22:13
配售份额；分配额度
pole position
phr.
22:45
领跑位置（赛车术语）
cut risk
phr.
23:36
降低风险敞口、减仓
breakdown of diversity
phr.
24:12
多样性崩塌（市场参与者观点趋同）
probabilistic
/ˌprɑːbəbɪˈlɪstɪk/
adj.
24:12
概率性的
fragmentation
/ˌfræɡmenˈteɪʃn/
n.
24:46
碎片化
Bayesian
/ˈbeɪziən/
adj.
24:46
贝叶斯的，按新证据更新概率
continual learning
phr.
26:51
持续学习，模型部署后继续学习
sample-efficient
adj.
26:51
样本高效的，用少量数据学会
discontinuity
/ˌdɪskɑːntɪˈnuːəti/
n.
26:51
断层；不连续的跳变
asymptote
/ˈæsɪmptoʊt/
v.
26:51
渐近趋向（此处作动词用）
plateaus
/plæˈtoʊz/
v.
29:03
见顶；停滞
growing the pie
phr.
29:03
做大蛋糕
supervised fine-tuning
phr.
30:05
监督微调
reinforcement learning
phr.
30:05
强化学习
gross margins
phr.
30:45
毛利率
leading into
phr.
32:24
全力投入（leaning into 之口误）
agentic
/eɪˈdʒentɪk/
adj.
33:02
智能体式的，能自主执行任务的
labor substitution
phr.
34:09
劳动力替代
FTE
n.
34:09
全职当量员工（full-time equivalent）
comp spend
phr.
34:48
薪酬支出（compensation）
maxi
/ˈmæksi/
n.
34:48
极大化派、死忠（maximalist 缩写）
HVAC
n.
36:14
暖通空调（heating, ventilation, air conditioning）
energize
/ˈenərdʒaɪz/
v.
36:49
通电；给设施接入电力
reconditioning
/ˌriːkənˈdɪʃnɪŋ/
v.
37:15
翻新
repurposing
/ˌriːˈpɜːrpəsɪŋ/
v.
37:15
改作他用
a floor and a ceiling
phr.
37:15
价格下限与上限
labor hoarding
phr.
38:03
劳动力囤积，宁可多养人也不裁
game theory
phr.
38:03
博弈论
leverage
/ˈlevərɪdʒ/
n.
40:00
议价权（此处非「杠杆」）
cyclical
/ˈsɪklɪkl/
adj.
40:36
周期性的
franchise
/ˈfræntʃaɪz/
n.
40:36
核心业务、护城河业务
multiple
/ˈmʌltɪpl/
n.
41:46
估值倍数
financeable
/faɪˈnænsəbl/
adj.
41:46
可融资的、可抵押的
matchmaking
/ˈmætʃmeɪkɪŋ/
n.
42:24
撮合
credit wrapper
phr.
42:24
信用包装，为他人贷款提供信用支持
royalties
/ˈrɔɪəltiz/
n.
42:24
版税；权利金
vendor financing
phr.
43:03
供应商融资，卖方借钱给买方买自己的产品
fungible
/ˈfʌndʒəbl/
adj.
43:03
可互换的、不可区分的
on the hook
phr.
43:37
承担责任、被套牢
revolting
/rɪˈvoʊltɪŋ/
v.
44:19
反叛、闹脾气
surety
/ˈʃʊrəti/
n.
44:19
担保
a cut of
phr.
44:19
分一杯羹、分成
durability
/ˌdʊrəˈbɪləti/
n.
45:01
持久性、确定性
equity stake
phr.
45:01
股权
opportunistic
/ˌɑːpərtuːˈnɪstɪk/
adj.
46:33
机会主义的、见机行事的
in the crosshairs
phr.
46:33
处于枪口下、成为目标
run away with it
phr.
47:33
一骑绝尘
Pareto frontier
phr.
47:33
帕累托前沿，无法被全面超越的最优集合
plugged in
phr.
50:37
消息灵通、人脉深
warp drive
phr.
52:25
曲速引擎（科幻中超光速推进）
phase transition
phr.
52:25
相变，性质的跳跃式转变
feat
/fiːt/
n.
53:36
壮举
espionage
/ˈespiənɑːʒ/
n.
53:36
谍报活动
learning by doing
phr.
54:12
边做边学、干中学
conviction
/kənˈvɪkʃn/
n.
54:12
笃定；坚定的判断
decoupling
/ˌdiːˈkʌplɪŋ/
v.
54:12
脱钩
godsend
/ˈɡɑːdsend/
n.
56:01
天赐之物、及时雨
wrappers
/ˈræpərz/
n.
56:01
套壳（在 API 外包一层界面的产品）
defensible
/dɪˈfensəbl/
adj.
57:10
有护城河的、可防御的
farm out
phr.
57:10
外包、分派给下级
distilling
/dɪˈstɪlɪŋ/
v.
58:00
蒸馏，把大模型能力压缩到小模型
maximalist
/ˈmæksɪməlɪst/
n.
58:00
极大化主义者
proprietary
/prəˈpraɪəteri/
adj.
59:00
专有的、私有的
treadmill
/ˈtredmɪl/
n.
59:00
跑步机；比喻无法脱身的循环
orchestrate
/ˈɔːrkɪstreɪt/
v.
59:55
调度、统筹
accrued to
phr.
59:55
归于、累积到
cash-efficient
adj.
59:55
现金效率高的、烧钱少的
rule of 40
phr.
1:00:30
40 法则，增速加利润率不低于 40%
moratorium
/ˌmɔːrəˈtɔːriəm/
n.
1:02:19
暂停令
post-factual
adj.
1:02:19
后事实的，不以事实为依据的
behind the meter
phr.
1:03:11
表后（供电），不经公共电网
pledge
/pledʒ/
n.
1:03:11
承诺书
ostensibly
/ɑːˈstensəbli/
adv.
1:03:40
表面上、名义上
order of magnitude
phr.
1:04:20
数量级
propagated
/ˈprɑːpəɡeɪtɪd/
v.
1:04:57
传播、扩散
influx
/ˈɪnflʌks/
n.
1:05:56
涌入
blindingly obvious
phr.
1:07:01
显而易见到刺眼
divergent
/daɪˈvɜːrdʒənt/
adj.
1:07:45
分歧的、相差甚远的
disaggregate
/dɪsˈæɡrɪɡeɪt/
v.
1:08:25
解耦、拆分
holy grail
phr.
1:08:25
圣杯；终极目标
feed forward network
phr.
1:09:08
前馈网络（Transformer 的一部分）
dark horse
phr.
1:10:30
黑马
absolute killer
phr.
1:10:30
狠角色、顶尖高手
blip
/blɪp/
n.
1:12:05
小波动、短暂异常
ARR
n.
1:12:46
年度经常性收入（annual recurring revenue）
implausible
/ɪmˈplɔːzəbl/
adj.
1:13:28
难以置信的、不太可能的
short case
phr.
1:14:13
做空逻辑
sanity check
phr.
1:16:02
现实检验、合理性检查
a blast
phr.
1:17:32
极其尽兴的经历
下载便签
手机：长按图片也可保存
← 上一期 · NO.210Eric Weinstein: The State of American Science, Breakthrough Coverups, and the Danger of Physics
下一期 · NO.212 →What Happens When the AI Boom Runs Out of Money
苏菲周报 · THE WEEKLY
每周一封，
追问一个大问题。 苏菲拉底的每周来信，写这一周在追问的问题和看到的回应。
追问一个大问题。 苏菲拉底的每周来信，写这一周在追问的问题和看到的回应。
苏菲拉底
ASK THE BIG QUESTIONS · THINK DEEPLY · SEE THE WORLD DIFFERENTLY
微信公众号
© 2026 苏菲拉底 · 内容仅供学习
[email protected]