《当 AI 泡沫的钱烧完了》中英对照逐字稿（80 分钟，带时间轴）

对谈人：Ben Thompson（Stratechery）· 主持：Patrick O’Shaughnessy（Invest Like The Best #487）· 录制：2026 年 8 月 · 全长 01:20（视频版）来源：What Happens When the AI Boom Runs Out of Money（Invest Like The Best 频道上传版）中文精读讲义：《当 AI 泡沫的钱烧完了》：Ben Thompson 谈大科技、中国与 AI 资本周期

说明：这份逐字稿是把这一期的官方播客音频交给本地 Whisper 转写的，不是官方字幕，也不是对视频音轨的转写。本机无法下载这段视频（YouTube 要求登录验证），所以改用同一期播客的音轨——同一场对话，但两版剪辑不同：播客版开头多约两分钟的赞助口播（播客 01:17 / 视频 01:20），因此本文的时间轴对应播客音频，放进视频里会有几分钟偏移。转写为机器输出，只对明显的人名、公司名与产品名做了校正（例如 TSMC、Trainium、 C.C. Wei、Larry Ellison、WorldCom、Jay Cooke 等），口语重复与语气词照录，个别词仍以音频为准；中文为本文翻译，英文为原话照录。

章节速览

01:54 开场：如果美国赢下 AI 竞赛
04:13 与中国竞争，和「我们有多依赖中国」
10:23 铁路、久期错配，与「钱可能不够」
13:31 伯克希尔、谷歌与绝对利润
16:22 可验证与不可验证的领域
23:52 从买到租：推理的真实成本
25:34 消费者不愿付费，所以广告是唯一解
29:04 广告飞轮、八千亿资本开支与时间错配
33:34 航运、内存与大宗商品市场的逻辑
39:13 「风险不会消失，只会转移」：台积电、英特尔与稀缺
45:37 亚马逊的 Graviton、Trainium 与苹果的聚合者位置
48:15 苹果会不会掉进微软的陷阱；五家前沿模型公司
55:38 微软的中间件剧本：90 年代的 IBM
57:56 界面即腹地：Meta、TikTok 与注意力
64:11 广告的社会价值、ATT，与「智能会不会变成大宗商品」
70:53 英伟达、电力约束，与「我们希望泡沫留下什么」

1 开场：如果美国赢下 AI 竞赛

To learn more, visit PSUM.VC. So, Ben, if you can believe it, how long it’s been since we last did this. The world was very different. No AI at the time. We talked about aggregation theory mostly. I thought a fun place to begin, since the world has changed so much, is to hear what you think it would mean for the U.S. to win the AI race.

要了解更多，请访问 PSUM.VC。Ben，你大概不敢相信，距离我们上一次做这档节目已经过去多久了。那时候世界完全是另一个样子，还没有 AI，我们主要聊的是聚合理论。既然世界变了这么多，我想开头先聊个有意思的问题：你认为「美国赢下 AI 竞赛」意味着什么？

I think it would be very problematic for the U.S. to win. Let’s say we take the most sort of fantastical scenario where if you control AI, you basically, your military is better than anyone else. Somehow it fixes their manufacturing, all these things that I don’t think AI is necessarily going to do because they sort of deal with the real world.

我认为美国赢下这场竞赛会非常有问题。我们设想一个最天马行空的场景：如果你控制了 AI，你基本上就拥有比任何人都强的军事力量。有人说它还能顺带把制造业修好——但我不认为 AI 一定会做这些事，因为制造业面对的是真实世界。

But in this world, what is the game theory optimal response of China to blow up TSMC? Game theory can get very sort of convoluted and complex. To me, this one actually isn’t that complicated. There is a fundamental disconnect that I have with a lot of the rhetoric coming out of Silicon Valley, coming out, I think, of one of the labs in particular, where if we get to a place where we have a meaningful superiority in terms of a military national security perspective, I think that’s very dangerous for the world. But in that state, how much does it extend beyond TSMC being blown up? Because in that state, I would assume we figured out how to build fabs here in the U.S., you know, to some degree and are less reliant on that one choke point.

但在这个场景里，中国「炸掉台积电」在博弈论上的最优反应是什么？博弈论可以变得非常绕、非常复杂，可在我看来这一局其实没那么复杂。我跟硅谷的很多论调——尤其某一家实验室的论调——有一个根本分歧：如果我们走到某个在军事和国家安全意义上拥有显著优势的位置，我认为那对全世界都非常危险。但在那种状态下，这件事能超出「台积电被炸掉」多少？因为在那种状态下，我会假设我们已经在美国本土学会了造晶圆厂，至少到某种程度，不再那么依赖那一个卡脖子点。

I think there’s a little bit of magical thinking, which I just invoked in terms of manufacturing and whether it be fabs, whether that be actuators, all these precursors. I think the degree to which we are dependent on China is underappreciated and is not something that is going to be fixed outside of a conflict. Just because fixing so many of these things is going to be dramatically dumb. If your competitor is sourcing from China and you’re going to start sourcing or getting things from the U.S., you’re going to be at such a disadvantage, relatively speaking, that you’re just not going to do it. So you do it when you have literally no choice.

我认为这里有一点我刚才说的那种「魔幻思维」，无论说的是晶圆厂、执行器，还是各种前置零部件。我们对中国的依赖程度被严重低估了，而且它不会在没有冲突的情况下被解决。因为要补上这么多环节，在经济上会蠢得离谱：如果你的竞争对手从中国采购，而你开始从美国采购，相对而言你会处于极大的劣势，你根本不会去做。所以只有在真的别无选择时你才会做。

And that works for very big headline items. Like you can browbeat Apple to move some of their iPhone manufacturing to India, for example. But even that is a good example because Apple is not truly moving out of China.

这招对非常大的、能上头条的项目有效。比如你可以施压苹果，让他们把一部分 iPhone 产能搬到印度。但这本身就是个好例子——因为苹果并没有真正离开中国。

4 与中国竞争，和「我们有多依赖中国」

They’re diversifying to an extent, but it would just cost so much. It’s like paying an insurance policy that if you don’t have to pay it and it’s astronomically expensive, you’re just not going to pay it. It’s one of those sort of hypotheses that I just have a hard time even grokking because the only world I see where we truly pull out and have no dependency on China such that if they want to blow up Taiwan, who cares? There’s no impact on us. It seems pretty fantastical to me. And I think there’s a bit of facing reality in this regard that is not present in these conversations.

他们确实在某种程度上分散风险，但代价实在太高了。这就像买一份保险：如果可以不买，而保费又贵得离谱，你就是不会买。这类假设我连理解都觉得困难，因为我想象不出一个我们真正脱钩、对中国毫无依赖的世界——那种世界里，中国想炸掉台湾就炸掉，谁在乎？对我们毫无影响。这在我看来相当魔幻。我认为这些讨论里缺少一点「面对现实」的成分。

Put yourself in their shoes. What do you think the motivations are? Everyone can use a good bogeyman. I think from the AI trade perspective, nothing works better than we have to beat China. And I do think we need to beat China. We need to be competitive. I despair at the extent to which over the last few years in particular, so many of our responses, particularly from a political perspective, has been to try to be like China. I think we should be going the other direction, more openness, more innovation, less top-down control, less restrictions on speech and things along those lines. America succeeds by being on the leading edge and by leading into that.

站在他们的角度想一想：你觉得动机是什么？每个人都需要一个好用的假想敌。从 AI 投资这条叙事看，没有什么比「我们必须击败中国」更好用了。我确实认为我们需要击败中国，我们需要有竞争力。但让我非常沮丧的是，尤其在过去几年里，我们太多的反应——特别是政治层面的反应——是在试图「变得像中国」。我认为我们应该往反方向走：更开放、更多创新、更少的自上而下控制、更少的言论限制，诸如此类。美国的成功之道是站在最前沿，并带头冲进去。

You said probably the U.S. being purely dominant in AI is not the right end state for the world. What is your ideal equilibrium for how this goes worldwide? There’s a bit where AI right now is kind of like the Taiwan situation in that the current status quo actually doesn’t seem so bad. The question is how sustainable is it? But maybe it’s sustainable for longer than we think. The way I think about it right now is I think OpenAI and Anthropic are clearly on the frontier. Who knows what’s happening with Google and then Grok and Meta are chasing them. Meanwhile, the Chinese are very capable, very smart, and also definitely distilling these models to sort of stay about six to nine months behind.

你说过，美国在 AI 上一家独大大概不是世界应有的终局。那在你心目中，全球格局的理想均衡是什么样的？有一点是，AI 现在的局面有点像台湾问题：当前的现状其实看起来并不坏，问题在于它能持续多久。不过它可能比我们以为的更能持续。我现在的看法是：OpenAI 和 Anthropic 显然在前沿，谷歌的情况谁也说不清，再往后是 Grok 和 Meta 在追。与此同时，中国人非常有能力、非常聪明，而且肯定在蒸馏这些模型，让自己大致保持在落后六到九个月的位置。

And it feels like a pretty good equilibrium that I think is generally favorable to the U.S. Now, the question is how long can it stay this way? And there’s lots of questions on there like can the Chinese actually pull ahead? I’m still a little skeptical for various reasons, whether it be from chips. Getting to the leading edge in that last six to nine months is very difficult. It’s going to be instructive how Meta and Grok do in terms of actually catching up, especially as we get into the world of AI improving itself, using AI to make the AI better, which I think is definitely a real thing.

这感觉是个相当不错的均衡，我认为总体对美国有利。现在的问题是，这种状态能维持多久？这里有很多疑问，比如中国人真的能跑到前面去吗？出于各种原因——包括芯片——我还是有点怀疑。要补上最后这六到九个月的差距、真正到达最前沿，是非常困难的。接下来 Meta 和 Grok 在「追赶」这件事上做得怎么样，会很有启发，特别是当我们进入「AI 自我改进」的世界——用 AI 让 AI 变得更好——我认为这绝对是真实存在的。

I think you see a real acceleration from both OpenAI and Anthropic recently, which was sort of theorized and it seems to be coming true. And to the extent that’s true, can you actually catch up? And I think the other question about this, by the way, is to what extent does that apply to cost to serve, to marginal costs? If you can apply AI to optimizing your stack, to figuring things out, to analyzing all the data, is your cost to serve structurally lower than anyone else? This is the thing about the open source models. The talk about them being free is bizarre to me because it’s marginal costs.

我认为最近你在 OpenAI 和 Anthropic 身上都能看到真实的加速，这原本只是个理论推演，现在似乎正在成真。而如果这是真的，你还追得上吗？顺便说，这件事的另一个问题是：它能多大程度作用于「服务成本」、也就是边际成本？如果你能用 AI 去优化自己的技术栈、去搞明白问题、去分析所有数据，你的服务成本是不是就结构性地低于别人？这就是开源模型的问题所在。说它们「免费」在我看来很荒谬，因为真正起作用的是边际成本。

You still have to run inference like GLM or Kimi. Kimi is very expensive to serve. The cost per answer is significantly higher. Everyone referring to these as free, it feels like in the narrative, it’s in people’s head that free is free. Now I can use AI for free. No, you can’t use AI for free. You’re not paying necessarily the R&D to create the AI, but you’re definitely paying the inference to sort of run it. So right now, I kind of like where we are. And the pushback would be, that’s right now. It’s not going to stay that way, which I think is fair pushback. But I don’t know, this way longer than we think.

你还是得跑推理，比如 GLM 或者 Kimi。Kimi 的服务成本非常高，每个答案的成本明显更高。大家把它们称为「免费」，是因为在叙事里、在人们脑子里，「免费」就是免费——现在我可以免费用 AI 了。不，你不可能免费用 AI。你或许不用为创造这个 AI 的研发付费，但你肯定要为运行它的推理付费。所以现在，我挺喜欢我们所在的位置。反驳会说：这只是「现在」，不会一直这样——我认为这个反驳是合理的。但我说不好，这个状态可能比我们以为的持续时间更长。

If you could know anything about the future of how this will go, to be more confident in where the equilibrium will end up, what is it? Is it like the length of the S-curve, like how far up the S-curve we are? At some point, these things presumably will level out, maybe not. What would be the thing you’d want to know that would give you a better sense of what the future might look like? I am concerned that with the scare around people freaking out about Mythos and this Hugging Face incident, that the actual implication of that is not that we reduce these dangers, but we just stop releasing stuff.

如果关于这件事的走向，你可以知道未来的任何一件事，从而对最终会落在哪个均衡更有信心，你会想知道什么？是 S 曲线的长度，还是我们已经在 S 曲线的哪个位置？到某个时点，这些东西大概会走平，也许不会。你最想知道的、能让你更看清未来的那一件事是什么？我担心的是：在人们对 Mythos 和这次 Hugging Face 事件的恐慌之下，真正的后果不是我们减少了这些危险，而是我们干脆不再发布东西。

We on the outside start to lose any sense of like where exactly, what is actually the frontier and where it is. There becomes sort of a false sense of security because right now everyone’s basing their understanding of Mythos on Fable, but how good is Fable actually relative to Mythos? That sort of gap is only going to, I think, increase over time. So I think that’s a real question that I’m not sure about. This question of the recursiveness and AI sort of making itself better, does that lead to some sort of takeoff? And at the end of the day, there’s timing questions in lots of different ways.

我们这些外部人于是开始失去判断力：真正的前沿到底在哪、是什么水平。这会带来一种虚假的安全感——因为现在所有人对 Mythos 的理解都建立在 Fable 之上，可 Fable 相对于 Mythos 到底有多强？我认为这种差距只会越来越大。所以这是个真实的问题，而我没有答案。「递归」这件事——AI 让自己变得更好——会不会导致某种起飞？归根结底，这里有很多个维度上的「时间点」问题。

I’m worried about the timing mismatch in terms of the actual return on investment producing enough revenue to fuel investment. We’re working our way down the capital curve. We started with free cash flow. The speed with which the tech companies blew through the debt markets is kind of incredible. It took like a year and now Google is issuing equity. NVIDIA is putting together this $500 billion thing to tap into like pension funds and insurance floats and things like that. What’s after that? Where does the money come after that? Well, ideally, we actually flip back to free cash flow funding this. But if there’s a gap there, if we don’t get there soon enough, then we could have a big blow up.

我担心的是时间上的错配：投资的实际回报能产生足够的收入来支撑投资吗？我们正在沿着资本曲线一层层往下走。起初用的是自由现金流；科技公司花掉债务市场的速度简直惊人，也就一年时间，现在谷歌开始发股票了。英伟达在搞这个 5000 亿美元的东西，去撬动养老金、保险浮存金之类的资金。那之后呢？再后面的钱从哪来？理想情况下，我们会转回到用自由现金流来支撑这一切。但如果中间出现缺口——如果我们没能及时走到那一步——那就可能出一次大爆炸。

But at the same time, even if we have this blow up, the AI is not going away. It’s not going to stop improving. It’s going to keep sort of progressing in a way that we look back on the dot-com era or we look back on the railroad era or we look back on whatever bubbles through history, ultimately immaterial in terms of the broad scope of humanity, even if they were very devastating. What can the railroads teach us to think? It’s not the last bigger build out, right? In terms of percentage GDP or getting there? I think we might be bigger at this point or it’s like it was the biggest.

但与此同时，就算真的爆了一次，AI 也不会消失，它不会停止进步。它会继续往前走，就像我们回头看互联网泡沫时代、回头看铁路时代、回头看历史上任何一次泡沫一样——即便那些泡沫的破坏力非常大，放在人类整体的尺度上，最终都无足轻重。铁路能教给我们什么？它不是最后一次更大的建设潮，对吧？从占 GDP 的比例看……我认为我们这一次可能更大，或者说，当年那次是最大的。

10 铁路、久期错配，与「钱可能不够」

In the ballpark. The railroads had a real duration mismatch. To build a railroad and make money off it was a decade or multiple decades long endeavor, whereas you had to issue money to pay for it in the short term. And the world ran out of money, right? And I think that is probably the aspect. I think that’s why people reach for the railroads, because everyone talks about, are we going to have enough compute? Are we going to have enough electricity? Maybe the nearest term question is, are we going to have enough money? Which is kind of a bizarre thing to think about. That’s what happened in the 1870s.

差不多是一个量级。铁路有一个真实的期限错配：修一条铁路、靠它赚钱，是十年乃至几十年的工程，而付钱却要在短期融资完成。然后世界没钱了，对吧？我想这才是关键所在。我想这就是为什么大家会拿铁路来类比：所有人都在谈，我们的算力够不够？电力够不够？也许最近在眼前的问题是——钱够不够？这个想法本身有点怪诞。而 1870 年代发生的就是这件事。

The world just ran out of money. The funny thing is, the railroads kept operating and they expanded the West. Their contributions to GDP were astronomical. They’re still contributing to GDP. Railroad money is what’s going into Google right now for Berkshire Hathaway. It’s very funny.

世界就是没钱了。有意思的是，铁路继续运营，并且开拓了西部。它们对 GDP 的贡献是天文数字，至今仍在贡献。而现在伯克希尔投进谷歌的钱，就是铁路的钱。这很有趣。

It’s quite literal. Berkshire Hathaway has this problem. To me, this NVIDIA deal is very much paired with the Google equity issuance, which I thought was shocking when it happened. Why was it shocking? Because it’s Google, they can’t raise money? Like, why are they issuing equity? Why are they reducing their upside if they believe so strongly in this? But the Berkshire comparison is interesting, because to a rough approximation, they have See’s Candies, famously, right? Tremendously high margin business. The problem with a lot of high margin businesses is the percentage profit you can make is very high, but the absolute profit you can make is capped.

这相当字面。伯克希尔就有这个问题。在我看来，英伟达这单交易和谷歌增发股票这件事是很配的一对，而谷歌增发发生时我觉得很震惊。为什么震惊？因为那是谷歌啊，他们还融不到钱吗？他们为什么要增发？如果他们如此坚信这件事，为什么要稀释自己的上行空间？但和伯克希尔的类比很有意思，因为粗略地说，伯克希尔有著名的喜诗糖果，对吧？利润率极高的生意。很多高利润率生意的问题是：你能赚到的利润率很高，但你能赚到的利润绝对值是有天花板的。

There’s no reinvestment runway. That’s right. Like, you’re just accumulating cash. The brilliance of the BNSF railway thing was basically they took the See’s Candy profits and said, here’s another industry whose margins are way worse, but the absolute dollar amounts are so large that those way worse margins result in absolute profits that are much larger. BNSF in 2025 or something, the amount of free cash they threw off in one year was more than See’s Candies had thrown off in its entire lifetime, even though you’re talking about a low margin business compared to a very high margin business. I think there’s an aspect from Berkshire Hathaway where once your capital gets so large, you start operating in a world of absolute numbers as opposed to percentage numbers.

没有可再投资的跑道。对，你只是在堆积现金。BNSF 那件事的高明之处在于：他们拿着喜诗糖果的利润说，这里有另一个行业，利润率差得多，但绝对金额太大了，以至于那些差得多的利润率带来的绝对利润要大得多。BNSF 在 2025 年左右的某一年里产生的自由现金流，比喜诗糖果整个生命周期产生的还多——尽管这是一个低利润率生意对比一个极高利润率生意。我认为伯克希尔身上有这样一个侧面：一旦你的资本规模大到一定程度，你就开始在一个「绝对数字」而非「百分比数字」的世界里运作。

And the reason why I thought that was so interesting, that story, is it seems to capture where Google itself might be going. And so there is very symbolic for them to invest in Google. Google has this unbelievable high margin business of search, one of the most perfect, beautiful business models of all time, and the purest aggregator of them all, like scales in every direction, doesn’t have to invest any money to do it. Everything’s zero marginal cost. It’s amazing. Meanwhile, there’s this AI opportunity, which requires just astronomical — it’s just incinerating cash. But you can imagine if AI is intelligence and its TAM is basically all white-collar work and eventually with robotics, everything potentially, the absolute profits available here, even if the margins are lower, is so much larger that will we look back on Google Search as See’s Candies?

我觉得这个故事如此有趣的原因是，它似乎正好刻画了谷歌自己可能正走向的地方。所以由伯克希尔来投资谷歌极具象征意味。谷歌拥有搜索这个难以置信的高利润率生意，是史上最完美、最优美的商业模式之一，也是所有聚合器里最纯粹的一个：向各个方向都能规模扩张，而且不需要投钱去做，一切边际成本为零，太惊人了。与此同时，还有一个 AI 机会，它需要天文数字的投入——简直是在烧钱。但你可以想象，如果 AI 就是智能，它的 TAM 基本上是所有白领工作，最终加上机器人可能是所有一切，那么这里可获得的绝对利润，即便利润率更低，也要大得多——将来我们回头看，会不会把谷歌搜索看作当年的喜诗糖果？

13 伯克希尔、谷歌与绝对利润

It feels like that’s what’s happening. In that world, you use all your free cash flow. They’ve done that. You tap the debt markets to the tune of hundreds of billions of dollars. They’ve done that. You issue equity. What does an equity issue do? It dilutes your interest and your interest to your shareholders. So you have a smaller percentage of the pie. Well, you have a smaller percentage of an astronomically larger pie. At the end of the day, no one’s going to be complaining. It is very symbolic. Berkshire being the symbol of that equity issuance in that, are they actually not just an investor in Google, but a model for Google and where they’re going?

感觉正在发生的就是这件事。在那个世界里，你会用掉全部自由现金流——他们已经这么做了；你去债务市场借上几千亿美元——他们也做了；你增发股票。增发做了什么？它稀释了你的权益，也稀释了你在股东中的权益，你占的饼更小了。不过，你占的是一个天文数字般更大的饼里更小的一块。归根结底，没人会抱怨。这非常有象征意义：伯克希尔成了这次增发的象征——它究竟只是谷歌的投资者，还是谷歌正在走向的那种模式的样板？

I’m curious. Setting aside the commercial and competitive components of this, like you’re describing, how AI pilled on the pure technology would you say you are relative to other people thinking about this space? I have a view that is both super bullish and less bullish in some respects. So I am not fully convinced about the generalizable argument. AI is clearly incredible at coding. It kind of blows my mind that people were doing this a year ago, like actually writing out code. It’s very good at math, obviously. But the obvious riposte is that these are sort of verifiable domains. What is the evidence, or where is the compelling evidence, that being very good at verifiable domains clearly translates to being very good at sort of unverifiable domains, or domains that have very long sort of verification loops?

我很好奇。先把商业和竞争层面的东西放到一边——就纯技术而言，你觉得你相对于其他思考这个领域的人，有多「AI 上头」？我有一个既极其乐观、又在某些方面没那么乐观的看法。我没有完全被「可泛化」这个论点说服。AI 在写代码上显然惊人——一年前人们还在真的一行行敲代码，这让我觉得不可思议。数学上它显然也很好。但最直接的反驳是：这些都算是「可验证」的领域。有什么证据、或者说有说服力的证据，能证明在可验证领域表现极好，就一定能迁移为在「不可验证」的领域、或者验证链条极长的领域也表现极好？

I think that’s still a little bit to be determined. And it’s interesting because I raised this question and there are some people at the labs that were on a panel and I was kind of annoyed at the answer because the answer took me for an AI bear. Oh, well, people thought we couldn’t solve chess or we couldn’t solve Go. We solved those easy enough. And I’m like, I thought we could solve chess. I thought we could solve Go because they’re knowable domains. Scale was the answer to both of those. But also both of those were bounded. What is the go-to example that’s not chess, that’s not Go, that is genuinely in a new space that’s sort of an unknowable space where it’s doing things that were not possible?

我觉得这一点还有待确定。有意思的是，我提出这个问题时，某场论坛上有几位实验室的人，我对他们的回答有点恼火，因为那个回答把我当成了看空 AI 的人。「哦，当年人们也觉得我们解决不了国际象棋、解决不了围棋，我们不也轻松解决了吗。」我心说：我从来没觉得解决不了国际象棋和围棋，因为它们是「可知」的领域，规模化就是这两件事的答案。但这两者也都是有边界的。那么，那个「不是国际象棋、不是围棋」的默认例子是什么？它必须真正处在一个新的、某种不可知的空间里，在做过去不可能做到的事。

That is sort of the I’m not fully convinced sense. However, AI trained at a rough approximation, trained on all the data of the Internet. All the data of the Internet, that’s distillation. It distilled all of the end state of human thought. The actual typing on Reddit. It doesn’t have the traces. It doesn’t actually have the thought, the emotion or whatever that went into typing that comment or typing and writing that essay. Say Neuralink, whatever. What if the actual payoff from Neuralink is actually capturing the traces of human thought that actually dramatically expands the capabilities of these models?

这就是我「没有完全被说服」的那一层。不过，粗略地说，AI 是在整个互联网的数据上训练的。整个互联网的数据，那本身就是蒸馏：它蒸馏了人类思考的最终结果。是 Reddit 上真正敲出来的那些字。它没有「痕迹」，它并没有那次评论、那篇文章背后真正的思考、情绪等等。比如说 Neuralink，或者别的什么——如果 Neuralink 真正的回报，其实是捕捉人类思考的「痕迹」，从而大幅扩展这些模型的能力呢？

16 可验证与不可验证的领域

In this world, my concerns about verifiability is like, well, we solve verifiability by getting more data. My sense is that a huge number of jobs, a huge amount of economic activity does not exist in these domains that I’m not convinced that AI is good at. Actually, there’s a lot of people in the world who are kind of like sentient AIs to a certain extent. They operate very well in verifiable domains. They’re given jobs. They do them. And it’s almost like a somewhat pessimistic view of humanity to a certain extent. But I think that market is so huge and so large that if the models did not improve at all from where they are right now, the economic opportunity is actually massive.

在这个世界里，我对「可验证性」的担心是：我们说那就靠更多数据来解决可验证性。我的感觉是，有大量的工作、大量的经济活动，并不属于「我不认为 AI 擅长」的那些领域。事实上，世界上有很多人某种程度上就像「有感知的 AI」：他们在可验证的领域里做得非常好，被分配任务，然后把任务做完。这在某种程度上算是对人类比较悲观的看法。但我认为那个市场太大了，以至于即便模型从今天起完全不再进步，经济机会本身依然极其巨大。

I wrote an article a while ago. There’s the whole like accelerationist movement. What I call myself was a reluctant accelerationist. I think we need to push forward because we can’t go back. And the worst thing we can do is get stuck where we are. So I’m very AI-pilled in terms of its impact on the economy. It’s sort of upside in terms of monetization. I’m not sure about the timing. What would be like the gradient towards it? Imagine law or medicine where I don’t know whether or not you would consider those verifiable. Like law is like a code of some sort. Medicine, we have a certain state understanding of things.

我之前写过一篇文章。现在有所谓「加速主义」运动，我称自己是一个「不情愿的加速主义者」。我认为我们必须往前推，因为回不去了，而最糟糕的情况就是卡在原地。所以从 AI 对经济的影响来说，我是非常「AI 上头」的。从变现的角度看，上行空间很大，只是我不确定时间点。通往那里的坡度会是什么样的？比如法律或医学——我不知道你会不会把这两个算作「可验证」。法律有点像某种代码；医学上，我们对很多东西有某种确定的理解。

I mean, I think medicine is by far one of the biggest opportunities. It’s both one of the biggest opportunities and also one of the most challenging ones because of all the regulations and all the access. Like if you could turn an AI, turn machine learning onto all the medical records, I think the number of discoveries and improved treatments we could come up with in a very rapid amount of time would be unbelievable. So that is a very optimistic view. On the flip side, like when is that going to happen, right? I think the optimistic frame I put on humans is our capacity to create needs is sort of unlimited.

我认为医学是最大的机会之一，而且是遥遥领先的机会之一。它既是最的机会之一，也是最难的机会之一，因为涉及大量监管和准入。比如，如果你能把 AI、把机器学习用到所有病历上，我相信在很短时间内能得到的发现和改进的治疗方案数量会难以想象。所以这是个非常乐观的看法。反过来说——这事什么时候才会发生？对吧。我对人类乐观的那一面是：我们创造需求的能力几乎是无限的。

So I think we’ll do a very good job of creating new opportunities and job serving the fullness of time. The sort of more pessimistic way to put it is our ability to create red tape and muck is also fairly unlimited. How much of our economy is actually we’ve managed to create more and more jobs that is just make busy and make slow to a certain extent. If I go back to the early 2010s, maybe the aggregation theory was stewing in your brain and then you published it in 2015, I think it’s fair to say like that theory, that idea, maybe you can just quickly remind people what it is, defined the winners and losers of that era of technology.

所以我想，在足够长的时间里，我们在「创造新机会、创造新工作」这件事上会做得很好。更悲观的说法是：我们制造繁文缛节和烂摊子的能力也几乎是无限的。我们的经济里，有多大一部分其实是不断创造出越来越多「让人瞎忙、让事情变慢」的岗位？回到 2010 年代初，也许聚合理论当时正在你脑子里发酵，然后你在 2015 年把它发表出来。可以说，那个理论、那个想法定义了那一代科技的赢家与输家。

I’m really curious how you’re thinking about what theory or principles will define this era of winners from like a financial perspective and market cap perspective. I go back and forth even just on the question of aggregation theory itself. How much does that apply in the current era? Just the same thing, yeah, yeah. Yeah, because like a pushback that people have is one of the key components of aggregation theory is zero marginal costs. And zero marginal costs shows in lots of ways. The one that I focused on at the beginning was distribution. And people say, oh, I don’t have distribution. I have to pay Google for ads.

我特别好奇你在想什么：从财务和市场市值的角度，什么样的理论或原则会定义这个时代的赢家？即便只就聚合理论本身而言，我也在来回摇摆：它在当下这个时代还有多大适用性？道理其实是一样的。人们的一个反驳是：聚合理论的关键要素之一是零边际成本，而零边际成本体现在很多方面。我一开始关注的是分发。人们会说，哦，我没有分发能力，我得给谷歌付广告费。

It’s like, oh, no, you have a website. Your problem isn’t that you have distribution. Your problem is you don’t have demand and you’re paying for demand when you’re paying for ads and things on those because the aggregators control demand. And they control demand because in a world of abundance, the hard problem is not distribution. It’s discovery. How do you actually find what you’re interested in? So the companies that solve discovery in their domain come to dominate that market. They get a virtuous feedback loop. That’s sort of aggregation theory in a nutshell. And the other thing is transaction costs. There’s no transaction costs. Google can scale to the whole world.

其实不是：你有网站，你的问题不是「没有分发」，你的问题是「没有需求」，而你付广告费的时候，你是在为需求付费——因为聚合器控制着需求。它们之所以控制需求，是因为在一个丰裕的世界里，难题不是分发，而是发现：你到底怎么找到自己感兴趣的东西？所以在自己的领域里解决「发现」问题的公司会主宰那个市场，并获得正向反馈循环。这就是聚合理论的大意。另一点是交易成本：当交易成本为零，谷歌就能把规模铺到全世界。

And they can scale to the whole world, not just on the user side, but also on the monetization side. The vast, vast, vast majority of advertisers on Google or Meta never talk to someone at Google or Meta. They just go up and they buy ads. It’s all done by computers. The perfect business. And those computers, from a business perspective, cost zero dollars. AI, obviously, that changes significantly. Inference costs are real. But then again, how real are they? They’re real right now. I don’t know. Are they? Depends on the company. But they’re way more real than those prior examples. Well, like if you look at gross margins.

而且它能铺到全世界——不只是用户侧，也包括变现侧。谷歌或 Meta 上绝大多数广告主从来没有跟谷歌或 Meta 的人说过话，他们只是上去买广告，全部由计算机完成。完美的生意。而那些计算机，从生意角度看，成本是零。AI 显然大幅改变了这一点：推理成本是真实存在的。但话说回来，它有多真实？它现在是真的，我不知道，是吧？取决于哪家公司。但它比之前那些例子要真实得多。你看毛利率就知道了。

For sure. But you have this incredible spread. So you have people, I think the vast majority of people who are using ad today are using it as basically a Google substitute or like a recipe maker or whatever it might be. And my suspicion is that the cost to serve those people is extremely low. And low indeed, basically similar to serving them a web page. I would imagine it’s marginally higher, but not that much higher. Then you have on the other extreme, people who are actually leveraging test time scaling. It used to be we just scaled by making the models bigger and bigger. Now you can scale as far as time.

当然。但这里存在一个巨大的价差带。今天绝大多数使用 AI 的人，其实是在把它当成谷歌的替代品，或者一个「菜谱生成器」之类的东西。我怀疑服务这些人的成本极低——低到基本等同于给他们提供一个网页，我想会略高一点，但高不了多少。而另一个极端是真正在用「测试时扩展」的人。过去我们靠把模型越做越大来扩展，现在你可以按「时间」来扩展。

How long do you think about the answer? Well, you could think about the answer for days or weeks or months. That is directly marginal cost. Every second longer you’re thinking is costing more money, which speaks to like we think about AI and inference as this one question. That’s I was pushing back on you. But actually, the marginal cost question for the different user, the user using free chat GPT and the user trying to solve a math theorem, they’re not even remotely in the same universe. I think you see this challenge actually in the enterprise in a very interesting way. Microsoft recently, they are shifting their enterprise plan.

你会花多久去想一个答案？你可以为一个答案想几天、几周甚至几个月——这是直接的边际成本：你多想一秒，就多花一分钱。这说明我们不该把 AI 和推理成本当成一个问题来看，这也是我刚才想反驳你的地方。事实上，不同用户的边际成本问题完全不在同一个宇宙里：用免费版 ChatGPT 的用户，和试图证明一个数学定理的用户，天差地别。我认为在企业市场上也能很有意思地看到这个挑战。微软最近在调整他们的企业套餐。

So they come out with like an E7 plan, $100 per user per month. That includes some amount of usage. But then they also are charging for usage on top of that. I think this is a kind of a fraught position for Microsoft to an extent, because the positive way to think about Microsoft is they do everything you need as a business. Every individual component might not be the best, but you get it all for one price and they all mostly work together. And if you’re particularly a small or medium sized business or even a large enterprise, there’s real value in that. That’s right. It makes life easy.

比如他们推出了 E7 套餐，每个用户每月 100 美元，包含一定用量的额度，但超出部分还要按用量再收费。我认为这对微软来说在某种程度上是个很棘手的处境，因为对微软正面的一种理解是：企业需要的一切他们都做。单独看每个组件可能都不是最好的，但你付一份钱就能全都拿到，而且大体上彼此协同。尤其对中小企业，甚至大企业来说，这里是有真实价值的。确实，它让日子变简单。

The moment you start having to think about how much you’re paying, it’s not just that that’s a new decision, number one, that is untethered from headcount.

可一旦你开始需要琢磨「我到底花了多少钱」，问题就不只是多了一个新决策——第一，这个决策跟人头数脱钩了。

23 从买到租：推理的真实成本

Microsoft got the benefit is when you were hiring a new employee, you would think about the cost of that employee and baked in the cost of that employee is $100 a month or $50 a month for their license. It was kind of a thoughtless revenue stream for Microsoft. Now, if you think about usage, you have to think every single month, how much do I want to spend? That introduces two problems. Number one, most companies aren’t set up to do this. They make budgets like once a year. This idea we’re going to be thinking about through our budgetary allotment on like a monthly basis doesn’t compute.

微软过去的好处在于：你雇一个新员工时，会去考虑这个人的成本，而这个成本里天然就包含了他每月 100 或 50 美元的软件许可费。对微软来说，这是一条不用动脑子的收入流。而现在，如果按用量计费，你每个月都得想：我到底想花多少？这带来两个问题。第一，大多数公司没有为此做好准备，他们的预算一年做一次。「我们要按月来思考预算分配」这件事，在他们的体系里根本算不过来。

There’s an aspect where they’re used to thinking about CapEx decisions or one-time cost. And there’s a bit where when I’m talking about this employee, like the loaded cost of employee, it’s not CapEx, but it’s kind of like CapEx. It’s like you make the decision up front and you don’t think about it anymore. The decision is sort of already made. But if you’re thinking about usage, you’re doing it again. And the final thing is if you’re every month looking at your Microsoft bill and how much did I use, you start thinking about what am I paying for? How good is each of these products? Should I actually just start thinking about and spraying this out?

另一面是，他们习惯了 CapEx 式的决策，或者说一次性成本：这就像——我说的是一个员工的「全负荷成本」，它不是资本开支，但有点像资本开支：你在前面做一次决定，之后就不再想了，决定等于已经做完。但如果要按用量算，你就得再做一次这个决定。最后一点是：如果你每个月都盯着微软账单看「我用了多少」，你就会开始想——我到底在为哪些东西付钱？每个产品到底有多好？我是不是该重新考虑、把它们拆开来逐个用？

And I think they had to do it because that extreme of user who uses a ton of tokens and is actually leveraging AI costs way more to Microsoft than $100 a month. They can’t support them, but they want to hold on to this set cost for the vast majority of employees who can fit in that because they need to ask their customers to think a little bit for those extreme employees. But they don’t want them to think too much because that breaks the model in very surprising ways. Are you surprised at all that the recipe builder user that is very low cost to serve, that there hasn’t been a great business model that’s emerged around them just yet?

我认为他们不得不这么做，因为那种极端用户——用掉海量 token、真正把 AI 用出杠杆的人——对微软来说成本远超每月 100 美元，他们养不起。但他们又想对绝大多数「装得进这个价格」的员工保住这个固定价格，所以他们需要让客户为那些极端用户多少想一想，但又不想让客户想太多，因为那会以非常意外的方式把这个模型弄崩。你会不会觉得意外：对那些成本极低的「菜谱生成器」用户，至今还没有出现一个很好的商业模式？

25 消费者不愿付费，所以广告是唯一解

Google and Facebook are sort of business perfected in this prior era. They haven’t seemed to figure this out at all. I am frustrated, but not surprised. This is obviously a market that should be supported by advertising. That is why advertising is always the consumer business model. Consumers don’t want to pay. There’s two things to understand about consumers that Silicon Valley has to relearn about every 10 years. Number one, consumers do not want to pay for software. And number two, consumers do not care about being productive. We went through this in early SaaS. The canonical company for this, in my mind, is Dropbox. So Dropbox, unbelievable product, like especially when it first came out.

谷歌和 Facebook 是上一个时代被做到极致的生意，但他们似乎完全没搞明白这件事。我很沮丧，但不意外。这显然是一个应该由广告来支撑的市场。这也是为什么广告永远是面向消费者的商业模式：消费者不想付钱。关于消费者，有两件事硅谷每隔十年就得重新学一遍。第一，消费者不愿为软件付费。第二，消费者不关心「效率」这件事。我们在早期 SaaS 阶段经历过这个。在我心里，这件事的典型公司是 Dropbox——那个产品好得难以置信，尤其刚推出的时候。

In business school, I was one of the first people who used Dropbox, and that went off like crazy. I have so much storage still, like my free Dropbox, because I gave out my code to like so many people. So Drew Houston makes this amazing product, so easy to use, just absolutely seamless. He was very clear about this. He wanted to build a consumer company. And there’s that famous story of him meeting with Steve Jobs. Apple was interested in acquiring Dropbox. And they’re like, oh, we want to build a company. And Steve’s, you know, you’re a feature, not a company, which that plain Jane, just file sync.

在商学院时，我是最早用 Dropbox 的人之一，它在同学里传得飞快。我的免费 Dropbox 到现在还有巨量空间，因为我把邀请码发给了太多人。Drew Houston 做出了这个了不起的产品，极易上手、完全无缝。他在这件事上非常明确：他想做一家面向消费者的公司。还有个著名的故事，他去见乔布斯。苹果当时有意收购 Dropbox，他们说「我们想做成一家公司」，乔布斯说：你们是一个功能，不是一家公司——说白了，就是文件同步而已。

Apple did make a feature as far as like sort of iCloud Drive. And with Dropbox, they grew very fast. And then they had like a two-year lull. And in that two-year lull, what they had to do was basically completely rebuild the app from the bottoms up. Because not enough consumers are going to pay for it. Enterprises could see the value they would pay. But if you want an enterprise, you need permissions, you need control, you need someone else to be able to set all these sorts of things. And their app wasn’t even created to do that at all. So they had to rebuild the whole thing and realize the only way we’re going to make money is by selling to companies.

苹果确实做了这样一个功能，比如 iCloud Drive。Dropbox 增长非常快，然后经历了大约两年的停滞。在那两年里，他们不得不做的是把整个应用推翻重建。因为愿意付费的消费者不够多，而企业能看到价值、愿意付钱。但要做企业客户，你需要权限、需要管控、需要能让别人来配置这些设置，而他们最初的应用架构根本没有考虑这些。所以他们只能重写整个东西，并意识到唯一能赚钱的路是卖给公司。

Why do companies pay? Because companies are paying employees. To the extent they can make their employees more productive, they’re getting a greater return on their investment. It’s the complete inverse of a consumer. A consumer is like, I spent all day working. Why do I want to come home and be more productive? I want to sit on the couch and watch reels. But you see that with AI. And you also have this overarching skepticism of advertising. I’ve gotten so much traction on Stratechery by being an advertising appreciator. And I go back and read my early articles about advertising that were kind of directionally correct, but also like were not very good at all.

公司为什么付钱？因为公司已经在给员工发薪水了；只要能让员工更高效，他们就能获得更高的投资回报。这和消费者完全相反。消费者想的是：我工作了一整天，为什么回到家还要变得更高效？我只想瘫在沙发上看短视频。而在 AI 这里也能看到同样的情况。此外还有一层对整个广告业的普遍怀疑。我在 Stratechery 上因为「为广告说好话」获得了极大的关注度。回头看我早期写广告的那些文章，方向是对的，但也写得实在不怎么样。

But I got so much traction doing it because I was the only person writing about advertising. In a world of everyone who wanted to have a blog, in Twitter, no one wanted to talk about advertising. But even now, there’s in Silicon Valley this sort of embarrassment about the fact that the Valley is in many respects monetized by advertising. And particularly during the last sort of eight years, there was a Facebook’s icky. Best engineers don’t want to go work on this problem. And so you literally had OpenAI replaying the Dropbox story, but at like 100x a size, being like, no, we’re going to sell subscriptions to consumers.

但它之所以获得这么大关注，是因为我是唯一一个写广告的人。在一个所有人都想写博客、都在用 Twitter 的世界里，没人愿意谈广告。即便到了现在，硅谷对「这个山谷在很多方面是靠广告变现的」这件事多少还有点羞耻感。特别是过去八九年里，有一种「Facebook 很脏」的情绪，最好的工程师不愿意去做这件事。于是 OpenAI 简直就是把 Dropbox 的故事重演了一遍，只不过体量大了 100 倍：不，我们要向消费者卖订阅。

They did. They sold a lot, but they didn’t sell enough. If you’re going to be in the consumer market, you have to be doing advertising. They’re doing advertising now. It’s a little weird. They finally pivoted to doing advertising at the same time. They’re like, oh, crap, we need to go for the enterprise because Anthropic is kicking our rear end. So I’m not quite sure what they’re doing there. They have been rolling out ad features very rapidly. Things like copy and the connections with retailers so you know if a purchase went through so you can do all the tracking and things like that. I’m very interested to see how that goes.

他们确实卖了不少，但卖得不够。如果你想进入消费者市场，就必须做广告。他们现在在做广告了，这有点奇怪：他们终于转向做广告，同时又发现「糟糕，我们得去做企业市场了，因为 Anthropic 正在把我们打得满地找牙」。所以我不太确定他们到底在干嘛。他们一直在非常快地铺广告功能：比如文案、与零售商的连接（这样你就知道购买有没有真的完成，从而能做全套追踪之类的事）。我很想看看这会怎么发展。

29 广告飞轮、八千亿资本开支与时间错配

There’s a bit where had they leaned into advertising immediately as soon as ChatGPT was a hit, I think they would have a killer ad product right now. I think that Google would be in much bigger trouble. I think Meta would be in much bigger trouble because if you have this flywheel, the thing about advertising with consumers is your ability to monetize the consumer is infinite because the advertiser is bearing the price increase. So there’s zero elasticity issues. If you’re charging consumers a price, if you want to raise the price, like Netflix, this is their problem with the subscription plan. How much can they raise prices before consumers rebel and drop a tear or give up the service entirely?

如果他们当时在 ChatGPT 一炮而红时就立刻扑向广告，我认为他们现在手上会有一个杀手级的广告产品，谷歌会麻烦得多，Meta 也会麻烦得多。因为如果你有这个飞轮，面向消费者的广告有一点很关键：你的变现能力是无限的，因为涨价是由广告主承担的，所以完全不存在弹性问题。而如果你是直接向消费者收费、想涨价——比如 Netflix，这就是他们订阅制的麻烦所在：涨到多少消费者会反弹、骂一句然后干脆退订？

Charging people money is hard. Giving people things for free is easy. And it’s very frustrating that OpenAI did not pursue this sooner. I know you’ve been spending time with some of the big money firms and sources of capital. What is your sense of their appetite right now and how they’re thinking about the future? Because I think this year it’s going to be $800 billion or something that we’re going to spend in CapEx. Next year is supposed to be $1.3 trillion, I think, is the current estimate. It’s going to keep going up from there. We’re burning through all the compute that gets installed basically immediately. It’s such a strange circumstance that we can use the capacity right away as soon as it’s online.

向人收钱很难，免费送东西很容易。OpenAI 没有更早去做这件事，实在让人沮丧。我知道你最近在和一些大型资金机构、资本来源打交道。你现在感觉他们的胃口如何、他们怎么想未来？因为今年我们在资本开支上的花费大概是 8000 亿美元，明年按目前的估算是 1.3 万亿，然后只会继续往上。而且新装好的算力几乎立刻就被消耗掉——这种局面很奇特：产能一上线，我们马上就能用满。

Well, that’s the thing, though. So there’s a few timing mismatches that are happening right now. We can’t use it right away. All the bulls on Twitter is always like, we don’t have enough compute. We don’t have enough compute. Well, we don’t have enough compute because there was insufficient investment made in 2023 and 2024, which, yes, absolutely. And by the way, if you think there’s not enough compute, TSMC decreased their rate of growth in 2023 and 2024 and 2025. Our shortage of compute is going to get worse in the next few years because a fab, the lead time is even greater than a data center. Today, when we say there’s not enough compute, it’s not like all the money that the companies are putting in today manifests in compute tomorrow.

但问题恰恰在这里。现在存在几个时间上的错配。我们并不能「马上」用上它。Twitter 上所有的多头都在说：我们算力不够、算力不够。可是我们算力不够，是因为 2023、2024 年的投资不足——这我完全同意。顺便说，如果你觉得算力不够，那要知道台积电在 2023、2024、2025 年反而下调了增长速度。未来几年我们的算力短缺会更严重，因为晶圆厂的交付周期比数据中心还要长。今天我们说算力不够，并不意味着公司今天投下去的钱明天就变成算力。

No, it all manifests in compute in 2028 and 2029. On the calls, you have both Andy Jassy and Cyanadella are out there saying, look, we’re just building data centers. Like, these are the shells. We might not use them now. Maybe we’ll use them in the future. And we only buy GPUs when we know there’s demand for them. That is a great story to tell. I’m not sure that I think is a lot of BS because the reality is if you’ve built the shell, that money is sitting there. You’re not going to let it just sit there. If you invest in a fixed cost, and this is the whole logic of commodity markets.

不，它要到 2028、2029 年才会变成算力。在财报电话会上，Andy Jassy 和 Sundar Pichai 都会说：你看，我们只是在建数据中心，这些是先搭好的壳子，现在可能用不上，也许将来会用；而 GPU 我们只在确认有需求时才买。这是个很好听的故事。但我不确定这是不是一堆废话，因为现实是：壳子建好了，钱就已经砸在那儿了，你不会让它就那么闲置着。只要你投了固定成本——而这正是大宗商品市场的全部逻辑。

I think tech in general doesn’t understand commodity markets. Tech is, by and large, focused on if I produce a highly differentiated product, and that differentiation could be like software, it could be a network in terms of developers, it could be a social network sort of thing where peer-to-peer, where I’m highly differentiated, then my ability to charge higher prices provides sort of my profit margin. So, the classic example is like Apple. They have their ecosystem, and they have their software, and they have third-party, and all those sorts of things. And so, they can charge 50% margins on their iPhone. Everyone looks at Apple as like the ideal business model.

我认为科技行业整体上并不理解大宗商品市场。科技公司大体上关注的是：如果我做出一个高度差异化的产品——差异可以来自软件，可以来自开发者网络，也可以是社交网络那种点对点效应——那么我就能靠更高的定价获得利润空间。经典例子就是苹果：他们有生态、有软件、有第三方，所以他们能在 iPhone 上拿到 50% 的利润率。所有人都把苹果看作理想商业模式的样板。

That’s how you run a business. But in a commodity market, the price is set by the marginal supplier. Cost of service is all that matters. That’s right. I had a good friend in Taiwan who is in shipping. Fascinating industry. It’s kind of like the airlines, too, another industry that I love to look at. You buy a ship, and the cost of that ship is depreciation. Your marginal cost is actually quite low. It’s the fuel to run the ship and the cost of the crew and like your port fees. Not that much. What that means is you are going to run that ship. As full as you really cost.

这才是经营之道。但在大宗商品市场里，价格是由边际供给者决定的，「服务成本」就是一切。没错。我在台湾有个好朋友是做航运的，那是个非常迷人的行业，也有点像航空业——另一个我很爱研究的行业。你买一艘船，船的成本体现为折旧，而你的边际成本其实相当低：就是烧的油、船员的工资，加上港口费之类，没多少。这意味着什么？意味着你一定会让这艘船跑起来。

No, you’re going to run it no matter what. And you’re going to bring down the price of a container as low as it needs to be to cover your marginal costs. Now, your paper losses in this situation might be very large because your accounting loss includes depreciation. But the depreciation is an accounting figment. You already paid the money. You’re going to run that ship at whatever the market will bear. And the container, the beauty of the container, it is a pure commodity. The cost of the market is going to be the marginal cost. Now, if it gets low enough, at some point, people will exit because their marginal costs, they’re actually losing money on a shipment.

不管怎样你都会让它跑。而且你会把集装箱的运价压到只要覆盖边际成本就行。在这种局面下，你的账面亏损可能非常大，因为会计亏损里包含折旧。但折旧只是会计上的虚构：钱你早就付了。你会以市场能承受的任何价格让这艘船跑起来。而集装箱这个东西的美妙之处在于，它是纯粹的大宗商品：市场价格就等于边际成本。如果价格低到一定程度，某个时点就会有人退出，因为他们的边际成本已经让他们每运一单都在亏钱。

Not just paper money, but like actual real money. They will exit, but then the supply is diminished. So then the price will go back up and you get this interplay of sort of coming in and off. But then let’s say the market is very high like it was during COVID. It’s like, wow, we’re making so much money right now because there’s not enough supply.

亏的不是账面数字，而是真金白银。他们会退出，供应随之减少，价格又涨回去，于是就有了这种进进出出的动态。但假设市场像新冠期间那样极度火爆：哇，我们现在赚翻了，因为供给不够。

33 航运、内存与大宗商品市场的逻辑

There wasn’t enough supply of ships. So containers went from usually being like $3,000, $4,000 to $17,000, $18,000. The amount of money that these shipping companies made in a very short amount of time was insane. What happens though? Well, imagine if we had more ships, right? The problem is it takes two years to build a ship. If everyone makes this decision simultaneously, you suddenly have a lot of ships, price plummets, et cetera. Where we see this is in components, in memory in particular. Memory, very famous for boom and bust cycles, people entering the market late. But to what extent are data centers going to be memory makers where right now everyone can see we don’t have enough compute?

那时候船的供给不足，集装箱运价从通常的三四千美元涨到一万七千、一万八千美元，这些航运公司在极短时间内赚到的钱惊人。但接下来会怎样？你可以想象，如果我们有更多船就好了——问题是造船要两年。如果所有人同时做这个决定，你会突然拥有大量船只，运价随之暴跌，依此类推。我们在零部件、尤其是内存上就看到过这一幕：内存以繁荣-萧条周期闻名，总有人在周期后段才进场。但数据中心会在多大程度上变成「内存厂商」？现在所有人都看得见算力不够。

So everyone’s like, we absolutely have to be investing because there’s so much money to be made. And look at our payback period. The problem is you’re measuring your payback period in a time of scarcity. Is that payback period going to hold in a time of abundance? And the sort of the bulls would say there’s never going to be a time of abundance. AI. We’re going to be short forever. We’re going to be short forever. Which maybe we will be. My concern is even if that’s right, we could still have an air gap in that there’s so much money going into it right now. And not enough has come online to actually make sufficient revenues to handle the situation where we run out of capital.

于是所有人都说：我们必须投，因为这里有太多钱可赚，看看我们的回本周期。问题是，你是在稀缺时期衡量回本周期的——这个回本周期在丰裕时期还成立吗？多头会说：永远不会出现丰裕时期，AI 我们会永远缺货、永远缺货。也许真是这样。我担心的是，即便这是对的，我们仍可能撞上一个「断档」：现在投进去的钱太多，而跑起来、产生足够收入的部分又太少，一旦资本耗尽就接不上了。

I believe in AI. I think it’s a real thing. I think the economic impact is going to be astronomical. I think all the concerns about societal impact are very real and are going to come to bear in a major way. You can believe all that and still be worried about are we going to make the bridge to this actually generating the level of returns necessary to continue to feel this sort of going forward. When you zoom in on TSMC and the component makers where fabs are involved and so far, at least my understanding is that they’ve been quite conservative in their willingness to expand capacity, build new fabs, meet the market’s demand with similar growth, which they have not done.

我相信 AI，我认为它是真实的东西，经济影响会是天文数字。我也认为所有关于社会影响的担忧都非常真实，并且会以重大方式显现。你可以同时相信这一切，同时仍然担心：我们能不能搭好这座桥，让它真正产生足以支撑后续投入的回报水平？当你把镜头对准台积电和那些涉及晶圆厂的零部件厂商时——至少据我理解，他们在扩产、建新厂、以同等速度匹配市场需求这件事上一直相当保守，而他们并没有做到。

If that just re-limits this whole thing and prevents us from getting one of these giant overbuilds? We can talk about a few different ones. We’ll start with memory. Memory used to have tons and tons of memory makers. Every time there’d be a boom, memory makers would sort of re-enter the market. New countries would come in. Like, Taiwan used to have, like, a memory market. But you would get these exact dynamics. If there’s a shortage of memory, there’s so much money to be made, you can’t bring capacity on immediately. It’s the same as shipping. It’s the same as what we’re seeing right now. That would spur people to come in the market.

如果这反而给整件事设了上限，让我们等不到那种巨大的过度建设呢？我们可以谈几个不同的行业。先从内存说起。内存行业过去有非常非常多的厂商；每次景气上行，内存厂商就会重新进场，新的国家也会进来——比如台湾曾经也有内存市场。但你会看到完全相同的动态：内存短缺时利润丰厚，可产能没法立刻上来，这和航运一样，也和我们眼下看到的情况一样。高利润会引诱人们进场。

You’d get too much capacity. Prices would plunge. And people would just get blown out. Because the issue is the upfront cost for these is so large. Just like buying a ship, like, building a fab is even more so. And memory now, like, the leading edges of memory are using things like EUV machines. So the costs are getting into the billions of dollars for these lines. What happens is every time with these boom and bust cycles, some people would enter, more people get washed out. You go through these famous historical moments for these memory cycles. Companies just get blown out. One of the most interesting, actually, memory stories is how Samsung sort of took over memory was they saw it as an opportunity and they had studied history.

于是产能过剩，价格暴跌，一批人被彻底打垮。问题在于这些生意的前期投入太大：买船如此，建晶圆厂更是如此。而今天内存的最前沿要用 EUV 光刻机这样的设备，这些产线的成本已经进入数十亿美元量级。结果是，每一轮繁荣-萧条周期里，总有人进场，更多人被洗出去。内存周期历史上那些著名的时刻，就是一批公司被彻底打爆。实际上最有趣的记忆体故事之一，是三星如何拿下内存：他们把它看作机会，而且研究过历史。

And they realized that actually the way to take over the market is to invest into downturns so that you’re ready when the next cycle comes around, which requires a ton of guts and a ton of discipline and a ton of money. But they did that. It basically wiped out the Japanese. That’s when the South Koreans generally took over the market in a major way. But it got down to three. And the problem is, three, it’s not a monopoly, but it’s kind of an oligopoly. And they all got a lot more discipline about let’s not make the mistakes of the past. And we’re not colluding, but we all are on the same page about let’s not do that.

他们意识到，拿下这个市场的办法其实是在下行期投资，这样当下一轮周期到来时你已经准备好了——这需要极大的胆量、极强的纪律和极多的钱。他们做到了，基本上把日本人打没了。那是韩国人整体上大举接管这个市场的时刻。但最后只剩下三家。问题在于：三家不是垄断，但算是一种寡头。他们都在「不要再犯过去的错误」这件事上变得非常有纪律。我们不是在串谋，但我们在「别那么干」这件事上是有共识的。

And I think that dynamic sort of ran head on to the current moment where it just took a while for them to realize, no, there is a secular shift in memory demand that didn’t exist for a very long time. I think the memory solution will be solved eventually. The other risk they run is Apple’s lobbying to get Chinese memory. What is the number one focus of like algorithmic changes? How can we use less memory? I think the memory makers probably screw themselves in the long run by creating such a massive target on their back. I’ve analogized memory makers to Iran. The issue with the Strait of Hormuz is it’s very effective.

我认为这种动态正好撞上了当下这一刻：他们花了一段时间才意识到——不，内存需求出现了很长时间里都不曾存在的结构性转变。我认为内存的问题最终会被解决。他们面临的另一个风险是苹果游说引入中国的内存。算法改进的第一号重点是什么？就是怎么用更少的内存。我认为内存厂商从长期看可能是在给自己挖坑，因为他们把自己变成了一个巨大的靶子。我曾经把内存厂商类比成伊朗。霍尔木兹海峡的问题在于，它确实非常有效。

It’s more effective if you don’t use it because then it’s always hanging out there as something you could do. Now they did it. Turns out it worked. But the UAE, Saudi Arabia, they’re going to build pipelines. They’re going to build new ports. They’re not going to let this happen again. It’s very painful right now, but say Iran wants to close the Strait of Hormuz in 2035, it’s not going to have any effect because it will have been built around. My concern for the memory makers is they might have done the same thing. No one’s going to let themselves get in this situation again as far as memory goes.

而如果你不用它，它反而更有效，因为它始终悬在那里、是一种你可以动用的威胁。现在他们真的用了，结果证明有效。但阿联酋、沙特会去修管道、建新港口，他们不会让这种事再发生一次。现在是非常痛苦，但假设伊朗想在 2035 年关闭霍尔木兹海峡，那将毫无影响，因为到时候大家已经绕开它建好了通路。我对内存厂商的担心是：他们可能干了同样的事。就内存而言，没人会让自己再陷入这种处境。

TSMC is arguably worse because there’s only one. There is one company on the leading edge. Obviously, Intel and Samsung are trying to get there. It’s the same thing. All markets carry risk. And a lot of the question is, who ends up holding the risk? What I think the way the tech companies didn’t fully appreciate is the extent to which TSMC has offloaded risk onto the big tech companies. And the way they’ve done that is the risk that TSMC is worried about is overcapacity. If we build too much, it’s not just that we built too much and we have all these fixed costs that are not being fully utilized.

台积电可以说是更糟，因为它只有一家。在最前沿上只有一家公司，显然英特尔和三星都在努力挤进去。道理是一样的：所有市场都带风险，而核心问题往往是——最后是谁在承担这个风险？我认为科技公司没有充分意识到的是：台积电把风险转嫁给了大型科技公司，转嫁的程度有多大。他们转嫁的方式是：台积电担心的风险是产能过剩。

But if we build a fab, we expect that fab to run for 30 years. We’ve like baked in too much capacity into the system for years and years and years. So they are very biased towards being much more conservative. There’s a little bit of a culture component to this, too. One of the most interesting TSMC stories, it’s kind of analogous, that Samsung story, was Morris Chang retired in like the late 2000s.

如果我们建太多，不只是建多了、有一堆固定成本没被充分利用那么简单；而是我们建一座晶圆厂，是expecting它要跑 30 年——等于我们把这个系统的产能提前锁死了很多很多年。所以他们天然极度保守。这里面也有文化成分。台积电最有趣的故事之一——和刚才三星那个故事有点像——是张忠谋在 2000 年代末退休。

39 「风险不会消失，只会转移」：台积电、英特尔与稀缺

New leadership took over. There’s the Great Recession. And so they pulled back their planned spending. He comes in, fires everyone. And he’s like, the iPhone just launched. This is the biggest opportunity we’ve ever seen. We need to be investing, not cutting. And they invested through the Great Recession and through that downturn. That’s what laid the foundation for them taking over sort of leading edge semiconductors in that time. Morris Chang is a one of one. On the Mount Rushmore, in my mind, of the greatest and most impactful tech executives of all time, the entire fabless model is so critical to what tech is and what it does.

新的管理层接手，赶上大衰退，于是他们砍掉了原计划的支出。张忠谋回来，把一众人换掉，说：iPhone 刚刚发布，这是我们见过的最大的机会，现在要投，不是砍。他们穿过大衰退、穿过那轮下行期继续投资——这就是他们在那段时间拿下最前沿半导体地位的地基。张忠谋是独一无二的。在我心里，史上最伟大、最有影响力的科技高管「总统山」上应该有他；整个 fabless（无晶圆厂）模式对科技产业之所以是今天的样子，实在太关键了。

And also just the guts to do that at that time, particularly in someone who lived there, a culture that doesn’t necessarily tend to make those sorts of bets. TSMC, they were pretty conservative, to be totally honest. So what happens, though? Where’d the risk go? TSMC’s like, we don’t want to take the risk. Risk doesn’t disappear. It just moves. The risk is right now where you have every single big tech company realizes if we had more compute, we could be making more money. So there’s lots of foregone revenue and foregone profits that is the manifestation of the risk that TSMC handed off to them. Risk doesn’t disappear.

而且还要有在当时那么做的胆量，尤其对一个在当地生活、身处一种未必习惯于下这种注的文化里的人而言。老实说，台积电其实相当保守。那么接下来发生了什么？风险去哪了？台积电说：我们不想承担这个风险。可风险不会消失，它只会移动。今天这个风险落在每一家大型科技公司身上：他们都清楚，如果有更多算力，就能赚更多钱。所以大量被放弃的收入、被放弃的利润，正是台积电转嫁出去的风险的具体表现。风险不会消失。

It just gets handed off. And sometimes that risk doesn’t manifest in losing money. It manifests in not making money. And there is money not being made right now because what happened was they were very excited about 5G. They did a big wave of, like, investment, expanding their fabs in around 2020, 2021, 22. And they’re like, oh, yeah, we’re good. Like I said, 2024, chance to beat us out in 2022. Big thing in tech in 2023. In 2024, their growth rate went down. In 2025, their growth rate went down. In 2026, it’s up now. It was very funny because I was writing about this a while ago.

它只是被转手了。而且有时候这种风险不表现为亏钱，而表现为「没赚到钱」。现在确实有该赚而没赚到的钱，因为当初他们对 5G 非常兴奋，在 2020、2021、2022 年前后掀起一波扩产晶圆厂的投资潮，然后觉得「行了，我们够了」。结果 2024 年、2025 年他们的增速一路下滑，到 2026 年才重新起来。这很有趣，因为我之前就在写这件事。

And then I think it was, like, one or two earnings calls ago. C.C. Wei, the CEO and chairman, is talking about, like, use cases for AI, the whole earnings call in a way he never had before. This is why the bearmakers are scared. Usually there’s, like, a bullwhip. And they’re worried about being at the end of the bullwhip where the demand happens and it works its way down the chain. And they’re at the end. And then they double down. It’s already too late. They’re wasting all their money. And I think the thing with AI is if it’s a bullwhip, it’s, like, the longest bullwhip of all time.

然后大概是最近一两次财报电话会上，董事长兼 CEO 魏哲家整场都在谈 AI 的应用场景，这是以前从未有过的。这就是为什么「造熊者」会害怕。通常存在所谓「牛鞭效应」：需求发生在一端，再沿着链条往上传导，而处在链条末端的人最怕——等他们加码下注时已经太晚了，钱全浪费了。而我认为 AI 这件事，如果真是牛鞭效应，那也是有史以来最长的一根牛鞭。

There’s still so much to be built. And it just took a while for Asia to get the message where these sort of companies are. I think they’ve, by and large, gotten it. But them getting the message, it then takes several years for that to actually materialize. Do you have a sense for how long you think it will take given the extreme shortage of compute? The interesting thing is what this means for Intel and Samsung’s sort of logic business. I’ve been writing about the problem of this dependency on TSMC for years. One of my first articles in 2013 was exhorting Intel. Well, I say you have to build a fab business.

还有太多东西要建。亚洲这些公司花了一段时间才收到信号。我认为总体上他们已经收到了，但「收到」之后，真正落地还要好几年。在算力极度短缺的情况下，你觉得这个过程要多久？有意思的是，这对英特尔和三星的逻辑代工业务意味着什么。我多年来一直在写「依赖台积电」这个问题。2013 年我最早的文章之一就是在劝英特尔：我说，你必须做代工生意。

You’re not going to be a designer anymore. There’s a huge business in manufacturing chips. I thought I was late writing it then. Their stock goes to the moon throughout the 2010s as they’re riding the sort of cloud wave. And it wasn’t until 2020 where they finally realize we fell behind. By the way, there’s this huge opportunity. We’re totally unprepared for it. We don’t have a customer service mindset or culture organization or all the IP building blocks and all these things that TSMC has. And they need a customer. They need customers to help them actually build a real foundry business. So I would write about this as a problem.

你不可能再只做设计，芯片制造是一门巨大的生意。我当时还以为自己写晚了。整个 2010 年代，受益于云计算的浪潮，他们的股价一飞冲天。直到 2020 年他们才终于意识到：我们落后了。顺便说，这里有个巨大的机会，而我们完全没准备好：我们没有客户服务的思维、没有相应的文化组织，也没有台积电拥有的那些 IP 积木块。而且他们需要客户——需要客户来帮他们真正把代工业务建起来。所以我把这写成一个问题。

And I write about the China issue. Like you’re dependent on a company that is 60 miles offshore of our greatest geopolitical opponent who thinks it’s theirs. So these are big problems. That’s where I came to appreciate this insurance issue. For a big tech company to go to Intel and say, Intel, you make our chip. And by the way, the biggest benefactor of this is going to be you because you’re going to learn how to work with a partner. And the biggest pain is going to be us because we’re going to have to figure out how to work with you. We could just go to TSMC.

我也在写中国的问题：你依赖的这家公司，距离我们最大的地缘政治对手只有 60 英里，而对方认为那是它的领土。所以这些都是大问题。正是在这里，我体会到「保险」这个角度。要让一家大型科技公司跑去对英特尔说：英特尔，你来给我们造芯片。而且顺便说，这件事最大的受益者会是你，因为你会学会怎么和伙伴合作；最大的痛苦在我们这边，因为我们得摸索怎么和你们配合。可我们完全可以直接找台积电。

They are awesome. They are so great to work with. We know they’re going to do a good job. It just never made rational sense for anyone to go work with Intel. That was their fundamental problem. In an unchanging world, TSMC would just win forever. But this is where TSMC, in some respects, made the same mistake as the money makers, made the same mistakes as Iran. If I can continue the analogy. Because they didn’t invest the last few years, the shortages are going to be so acute. Big 10 companies that were foregoing so much revenue and so many profits because we don’t have enough compute.

台积电太好了、太配合了，我们知道他们一定会做好。所以从理性上讲，任何人跑去和英特尔合作都不成立。这就是英特尔的根本问题。在一个不变的世界上，台积电会永远赢下去。但正是在这里，台积电在某些方面犯了和内存厂商一样的错误、和伊朗一样的错误——如果这个类比还能继续的话。因为他们过去几年没有投资，短缺会变得极其严重。

We will go through the pain of getting Intel up to speed, of getting Samsung’s logic up to speed. The scarcity is what ultimately saved Intel. I expect at some point that they’re going to announce some major partner for the first time. It’s going to be a big deal. But ultimately, TSMC brought it on themselves. It’s the cure for high prices is high prices thing. We’re going to route around them. There’s all these things as like an analyst sitting on the side. You can write these things. And it’s one of those things I’ve sort of warned. No one’s going to pay insurance that they don’t need to pay when that insurance expected value is negative.

那些因为算力不够而放弃了巨额收入和利润的「大十家」公司，会去承受把英特尔、把三星的逻辑代工拉到能用水平上的痛苦。是稀缺，最终救了英特尔。我预计某个时点他们会第一次宣布某个重要合作伙伴，那会是件大事。但归根结底，是台积电自己招来的。这就是「高价的解药就是高价」——我们会绕开他们。作为一个旁观的分析师，这些你都能写出来，而这也正是我一直在警告的那类事：当一份保险的期望值为负时，没有人会去买一份自己不必买的保险。

The way to solve the geopolitical problem of dependence on TSMC is to come up with a compute use case that is so massive that everyone is economically incentivized to bring other people up to speed. And then we get the sort of geopolitical insurance for free. If you think about the, let’s say, top 10 or 15 technology companies, which ones do you think have the most interesting setups today for their business? The answer is always Amazon. The reason Amazon is so compelling is the extent to which they build for them. They are their first best customer. They provide the scale to get basically anything off the ground, which they then sell to other people.

要解决「依赖台积电」这个地缘政治问题，办法是造出一个巨大到让所有人都有经济动机去把其他人拉上来的算力应用场景，这样我们就免费获得了那层地缘政治保险。如果看前 10 到 15 家科技公司，你觉得今天哪家的业务结构最有意思？答案永远是亚马逊。亚马逊之所以如此有吸引力，在于他们「先为自己构建」的程度：他们自己是自己的第一个、也是最好的客户。他们提供的规模几乎能让任何业务起步，然后再把它卖给其他人。

AWS is the most obvious example. AWS, contrary to sort of popular thought, was not spare Amazon capacity. Actually, it took a long time to get Amazon.com onto AWS. What it drove was the understanding that we can’t be having so many meetings. Like, we need to have just compute that you can plug in, purely API, surface. You don’t need to talk to anyone. It’s just there. And oh, by the way, if we do that for our internal retail teams, we could do that for anyone. Turns out the retail is so big, we have to start with everyone else. AWS actually started serving external customers before it served internal ones, but now it serves them all.

AWS 是最明显的例子。与流行看法相反，AWS 并不是亚马逊的闲置产能——实际上，把 Amazon.com 搬上 AWS 花了很长时间。它带来的是一个认识：我们不能开这么多会；我们需要一种插上就能用的算力，纯 API 暴露，不需要跟任何人沟通，它就在那儿。而顺便说，如果我们能为内部的零售团队做到这件事，我们就能为任何人做。结果发现零售业务太大了，反倒得先从外部客户开始。AWS 实际上先服务了外部客户，之后才服务内部，但现在两边都在用。

You got other products like, say, the logistics, where it was the opposite.

还有其他产品，比如物流，情况正好相反。

45 亚马逊的 Graviton、Trainium 与苹果的聚合者位置

Right now, we’re using external providers for logistics, UPS and FedEx and USPS. We need to build this up ourselves. And now they built it up themselves. They’re offering it to third parties. Other people can use their delivery services. You see this in market after market. They’re talking about some of their AI products or their chip products. What’s the beauty of the Graviton or the Trainium, particularly the early versions? The early versions were terrible. But if you’re on Amazon and you’re using some of their managed services, like, say, the Redshift database service, they don’t tell you what the processor is underneath that. You’re just buying a managed service.

我们现在的物流用的是外部服务商——UPS、FedEx、USPS。我们需要自己把这块建起来。现在亚马逊自己建起来了，而且开始向第三方提供，别人可以用他们的配送服务。这种故事在一个又一个市场里重复：他们谈自己的 AI 产品、自己的芯片产品。Graviton 和 Trainium 妙在哪里？尤其是最早的版本——早期的版本很糟糕。但如果你在亚马逊上、用着他们某个托管服务，比如 Redshift 数据库服务，他们不会告诉你底下是什么处理器：你买的只是一个托管服务。

So they can put all their crappy processors underneath the services they’re selling, and that gives them the volume and the capacity to iterate them and get better. And they get to the point where they can actually sell them externally. Because they were the first best customer for Graviton, Graviton got better. Because they were the first best customer for Trainium, Trainium got better. And now Trainium is obviously running Anthropic and AI products. We’ll see if any of them take off. They have call center software. Their call center or their customer experience is going through AI. By the way, it’s pretty good. I haven’t tried it. Moving back to America, I’ve been buying lots of stuff.

于是他们可以把自家那些还很粗糙的处理器塞在对外售卖的服务底下，而这给了他们足够的量和使用场景去迭代、去变好，直到有一天真的能对外卖。因为他们是 Graviton 的第一个、也是最好的客户，Graviton 变好了；因为他们是 Trainium 的第一个、也是最好的客户，Trainium 变好了。现在 Trainium 显然在跑 Anthropic 和各类 AI 产品。我们再看这些业务里有没有哪个能起来。他们有呼叫中心软件——他们的呼叫中心、客户体验正在被 AI 改造。顺便说，效果相当不错。我没试过。说回美国这边，我一直在买很多东西。

Every summer, I’d buy lots of stuff in a very brief amount of time. Sometime in like the last year or so, you can go on and you’re clearly talking to a chatbot, but the chatbot does a great job. And it actually does take care of the problem. So you can see that actually starting to work in that regard. But they’re building up these AI services for their own business that they’re going to make broadly available. And some of them will work, some of them won’t. It’s such an elegant approach, given they have so many investments in the real world. Their core business feels so impervious to AI for the model version of AI.

每年夏天，我都会在很短的时间里买一大堆东西。大概最近一年左右，你上网购物时明显是在跟一个聊天机器人打交道，但它干得非常好，而且真的把问题解决了。所以你能看到这件事开始真正跑起来了。他们在为自己的业务搭建这些 AI 服务，之后再把它们广泛地开放出去——有些会成，有些不会。考虑到他们在现实世界里投了那么多资产，这种打法非常优雅。他们的核心业务对「模型版的 AI」来说，显得几乎刀枪不入。

It will benefit from AI, but their moat feels deeper than anyone as far as their core business. And their ability to just sort of generate new business lines organically is very compelling. What about Apple? They’ve sat this whole thing out, it seems. It feels like it might be a situation of better be lucky than good, to a certain extent. Apple has their whole ecosystem. At the end of the day, they do own access to customers. So they can get suppliers. This is the classic aggregator play. If you own access to customers, suppliers come to you, not the other way around. So they can get suppliers for their AI as needed.

它会受益于 AI，但就核心业务而言，它的护城河感觉比任何人都深。而他们能凭一己之力不断长出新业务线，这一点非常吸引人。那苹果呢？看起来这整件事他们完全没参与。某种程度上，这有点像「运气比本事更重要」的局面。苹果有自己完整的生态。说到底，他们握着通往客户的入口，所以供应商会来找他们。这是经典的聚合者打法：如果你掌握了客户入口，供应商会来找你，而不是你去找供应商。所以需要 AI 的时候，他们能拿到供应商。

And by the way, to the extent it’s true that people don’t want to be productive, they just want to sort of a chatbot. Not only can they serve them a chatbot and finally getting a Siri that works, but you can see a future where this absolutely can work on device. And they actually don’t even need to pay for inference costs either because they’re using the customer’s electricity.

顺便说，如果「人们不想提高效率、只想要个聊天机器人」这件事成立，那他们不仅能给用户一个聊天机器人、终于让 Siri 变得能用，你还能看到这样一个未来：这件事完全可以在端侧跑通。而且他们甚至不需要为推理成本付钱——因为他们用的是客户自己的电。

48 苹果会不会掉进微软的陷阱；五家前沿模型公司

I don’t think we’re quite there. There’s a reason they’re using Google Cloud and NVIDIA chips, but you can certainly imagine a future where that’s the case. And they’re in physical goods. Actually making phones is hard. Having retail, having distribution for physical goods, they’re more insulated. The smartphone is so perfect. It’s small to fit in your pocket. It’s big enough to watch basically anything on it. You can run your whole life on it. All your entertainment is there. When we talk about customers who want to be entertained, the TV is now an accessory. It’s all on your phone. I don’t see anyone taking over the phone.

我不觉得我们已经到那一步了。他们现在用谷歌云和英伟达的芯片是有原因的，但你完全可以想象一个未来它不再需要。而且他们做的是实物生意。真的把手机造出来是很难的。有零售、有实物分销，他们更不容易被颠覆。智能手机太完美了：小到能塞进口袋，又大到几乎什么都能看；你的整个生活都能在它上面跑，所有娱乐都在上面。当我们说「想要娱乐的客户」时，电视如今只是配件，一切都在手机上。我看不出有谁能取代手机。

The question is, is the phone always going to be the center? Or is there a bit where, particularly in the home, this is where OpenAI’s efforts here are very interesting, where you want sort of an ambient AI, where you just talk to the AI and it tells you what you need. Apple is the best position to provide that, but can they provide that without having leading-edge models? Can they provide that if they’re so phone-centric? Or is it like a Microsoft situation? Microsoft didn’t miss mobile. They were very early to mobile. The problem is their mobile was a small PC. They assumed the PC would always be the center and their phones were going to be something that was off that.

问题在于：手机会永远是中心吗？还是说，尤其在家里，会出现另一种形态？这正是 OpenAI 在这方面努力很有意思的地方——你想要的是一种「环境式 AI」（ambient AI），你对着 AI 说话，它告诉你你需要什么。苹果是最有条件提供这件事的公司，但在没有前沿模型的情况下，他们能提供吗？在他们如此以手机为中心的前提下，他们能提供吗？还是说，这会是又一个微软式的处境？微软并没有错过移动时代，他们很早就进入移动了。问题在于他们的「移动」是一台小型 PC：他们假定 PC 永远是中心，而手机只是从它派生出来的东西。

Apple realized, no, we need to reset. The phone is not going to be accessory to the Mac. The phone is going to be the phone. The iPod helped them realize that and going with Windows and all that. But will they fall into a Microsoft-like trap? Assuming the phone’s so good, it’s always going to be the center. Or is this finally the time when actually ambient, the cloud, just in general, AI being everywhere, it can manifest through your phone, it can manifest through a device, it can manifest on your computer, is actually better and is actually disruptive to them? I think it’s possible. I also think it’s totally valid for Apple to double down on what they do.

苹果想明白了：不，我们要重置。手机不会是 Mac 的配件，手机就是手机本身。iPod 帮他们认识到这一点，再加上后来兼容 Windows 等等。但他们会掉进微软式的陷阱吗——认定手机太好了，所以它永远是中心？还是说，这一次终于轮到「环境式、云端、无处不在的 AI」——它可以出现在你手机里、出现在某个设备里、出现在你电脑上——反而更好、反而对他们有颠覆性？我认为这有可能。我也认为苹果把资源压在自己擅长的事上，完全说得通。

The other thing about the AI stuff is, on what basis should we expect Apple to be good at this? At the most crude level, AI is this probabilistic endeavor. Apple is the king of deterministic products. A physical product, you ship that iPhone, you ship it once, and it’s got to be good. If it’s bad, it costs you billions and billions and billions of dollars. Apple’s never had an iPhone recall. It’s amazing that care and decision-making and diligence and fierceness in terms of your supply chain and making hard decisions is very, very different than everything that goes into, like, making great AI. I generally prefer companies to do what they’re good at.

关于 AI 还有一点：我们凭什么期待苹果擅长这件事？最粗浅地说，AI 是一场概率性的工程，而苹果是确定性产品的王者。一台硬件，iPhone 出货一次，就必须是好的；如果它不好，代价是几十上百亿美元。苹果从没召回过 iPhone。而这份谨慎、这种决策方式、这种对供应链的严苛与较真，与「做出好 AI」所需要的所有东西，差别非常大。总体上，我更喜欢公司去做自己擅长的事。

So from my perspective, I’m fine with Apple not doing AI. I want them to keep making great devices. Of the five potential frontier AI winners, so OpenAI, Anthropic, Gemini, SpaceX AI, Grok, and Meta, which of those firms do you think has the most interesting setup? OpenAI and Anthropic obviously are the riskiest, but also have the biggest upside. Never discount, number one, the power of belief. They think they’re creating God. The most impactful things in history have usually been fueled by religion. The two religious organizations in Silicon Valley are OpenAI’s kind of, like, mainline. They go to church every Sunday. They’re sort of, like, evangelicals. That’s Anthropic.

所以在我看来，苹果不做 AI 挺好，我希望他们继续做好设备。「五家可能的前沿 AI 赢家」——OpenAI、Anthropic、Gemini、SpaceX AI、Grok 和 Meta——你觉得哪家的局最有意思？OpenAI 和 Anthropic 显然风险最大，但上行也最大。永远不要低估第一件事：信念的力量。他们认为自己在造神。历史上最有影响力的事情，通常都是被宗教驱动的。硅谷里的两个「宗教组织」：OpenAI 有点像主流教派，每周日都上教堂；而那种福音派式的存在，是 Anthropic。

They’re all in. It is core to their belief. That goes a long way. The fact you need to make a business work for you to survive goes a very long way. Google just needs search to not die too quickly. Meta has the huge advertising business. In a world where Meta was run by anyone other than Mark Zuckerberg, they would not be on the leading edge. That is one of the purest manifestations of founder energy, for better or for worse. Their business is so amazing. You see them just easily sort of doubling down on that. Google, there’s a bit where they had Google Cloud. They have TPUs.

他们是全押的。这属于他们信念的核心，这一点能走很远。而「你必须让生意跑通才能活下来」这件事，同样能走很远。谷歌只需要搜索别死得太快。Meta 有巨大的广告业务。如果 Meta 不是扎克伯格在管，他们根本不会待在前沿边缘上。这是创始人能量最纯粹的体现之一，无论好坏。他们的生意太好了，你能看到他们很轻松地继续加注。谷歌这边，他们有谷歌云，有 TPU。

They’ve been doing research in this. It makes sense why they’re pursuing this. Meta being, like, actually, we’re going to hire a completely new team. We’re going to start from scratch. This all, again, is pretty insane. Credit to Mark Zuckerberg in that regard. Again, you could decide whether that’s a good idea or not. And then SpaceX AI, data centers in space, the theory is there. Do they have to own their own model, though, to do that? They’d get better margins if they do. Then again, if we actually run out, whether through political opposition or power or whatever it might be, if we run out of data centers on Earth, they can run whatever model they want,

他们一直在做研究，所以他们追这件事是说得通的。而 Meta 是……真的，我们要招一个全新的团队，我们要从零开始。这一切，还是那句话，相当疯狂。这方面要给扎克伯格记一笔。当然，你也可以判断这是不是个好主意。然后是 SpaceX AI——太空数据中心，理论上是成立的。但要做这件事，他们非得有自己的模型吗？如果有，利润率会更好。但另一方面，如果有一天我们真的被卡住了——不管是来自政治反对、电力还是别的什么——如果地球上的数据中心不够用了，那他们想跑什么模型就跑什么模型。

as we’re seeing with selling their capacity to Anthropic right now. They’re all pretty interesting. Probably the case for SpaceX AI is probably the weakest because the data center in space play is so highly differentiated. If that plays out, I’m not sure to what extent they need to even have their own model. So why are you wasting billions and billions of dollars in the meantime? That’s a fair question. From a tactical perspective, I love the cursor acquisition. That makes so much sense for both companies. And so I’ve been intrigued to see what they do. Meta is probably the most interesting. You’ve written a lot about this recently.

就像现在他们把算力卖给 Anthropic 一样。这几家都挺有意思。SpaceX AI 的论证可能是最弱的，因为「太空数据中心」这张牌差异化太强了。如果这件事真跑起来，我不确定他们还需要在多大程度上拥有自己的模型——那你为什么还要在这期间烧掉几十上百亿美元？这是个公平的问题。从战术上看，我很喜欢收购 Cursor 这一步，对两家公司都非常合理。所以我一直很好奇他们会怎么做。Meta 可能是最有意思的那个。你最近写了很多这方面的东西。

I think there’s a very good case to make that it is more reckless to not be on the frontier if you’re a digital company. The counter to Meta is actually Microsoft. Microsoft is not on the frontier. The reason why Microsoft has $40 million of free cash flow last quarter, Microsoft paid a $10 billion dividend last quarter. There’s some money, but their play is, oh, we’re going to play all these off each other. We’re going to provide middleware. We’re going to provide the platform that enterprises will build on us, and we’re going to sort of disintermediate the models. I think it’s a rational play. It’s the IBM play of the 90s.

我认为有一个很有力的论证：如果你是一家数字公司，不待在前沿反而更不负责任。Meta 的反面恰恰是微软。微软不在前沿。微软上一季度之所以有 4000 万美元的自由现金流——顺便说，他们上一季度还发了 100 亿美元的股息。钱是有的，但他们的打法是：让这些模型互相博弈，我们提供中间件，我们提供企业会在上面构建的平台，然后我们对模型做「去中介化」。我认为这是理性的打法。这就是 90 年代 IBM 的打法。

History echoes. Everyone talks about Google. Google, like, following Microsoft. Microsoft follows IBM. And you can see that to an extent. What did IBM do? What’s the analogy? Well, so IBM had this dominant, we talked about it in the 70s. And then you fast forward to the 90s, and IBM is this very distressed asset. And the thought was IBM needed to break up into all these different pieces they had. So Lou Gerstner comes in and takes it over. Gerstner’s real key insight to IBM is we’re pretty mediocre at everything. It’s kind of like when I talked about Microsoft before. And that’s the price of monopoly. Once you’ve been a monopoly, you kind of lose your capacity to be good because you didn’t need to compete anymore.

历史会回响。所有人都在谈谷歌——谷歌像在跟随微软，微软跟随 IBM。某种程度上确实如此。IBM 做了什么？类比点在哪？IBM 曾经是绝对主导者，我们在 70 年代聊过。然后快进到 90 年代，IBM 成了一项非常困顿的资产。当时的想法是，IBM 需要拆成它手里那一堆不同业务。于是郭士纳（Lou Gerstner）进来接管。郭士纳对 IBM 真正的关键洞察是：我们在每一件事上都很平庸。这有点像我刚才谈微软时说的——这是垄断的代价。一旦你当过垄断者，你就某种程度上丧失了做好事情的能力，因为你不再需要竞争。

And I think a lot of tech incumbent companies have this problem. It didn’t matter what they did. They were going to rake in money. And if you don’t have the pressure, if you don’t have the incentive, if you don’t have the fear of death or the fear of God, as we talk about these model companies, then you don’t do your best work. And the problem is that once you lose that muscle, it’s gone. You’re just sort of fat and flabby. So what Gerstner realizes, actually, the worst thing IBM could do would be to break it up into component pieces because all those component pieces are actually not very good.

我认为很多科技在位公司都有这个问题。他们做什么都无所谓，钱照进。而如果你没有压力、没有激励、没有对死亡或对神的畏惧——就像我们谈这些模型公司时说的——你就不会做出最好的作品。问题在于，一旦这块肌肉没了，它就真没了。你只剩下虚胖和松弛。所以郭士纳意识到：IBM 最糟的选择恰恰是把它拆成一块块，因为那些零件其实都不怎么样。

Our biggest asset is that we’re big. It’s like, what? No, what does it mean we’re big? It’s the 90s. This internet thing is coming along. There’s all these companies that kind of know they have to figure out the internet and they don’t know what to do. They need someone who can come in, understand their business, and help them get online. That’s basically what IBM did. So they built out, and this is an echo of what’s happening now, huge consultant force. And they put all their time into building, basically it was middleware, where they would go in and they’d put this layer between a company’s old school mainframe, which all these companies had, and then modern web services.

我们最大的资产就是我们很大。听起来像什么？不，等等，「大」意味着什么？那是 90 年代。互联网这东西来了，一堆公司知道自己必须搞懂互联网，却不知道该怎么办。他们需要有人走进来，理解他们的生意，帮他们上线。这基本上就是 IBM 做的事。于是他们——这也是当下正在发生的事的回响——建起了一支庞大的顾问队伍，把全部精力投在搭一层「中间件」上：走进一家公司，在它那套老掉牙的大型机（这些公司都有）和现代的网络服务之间，插进一层东西。

On the other hand, they could have websites and e-commerce sites and all this sorts of things. And it gave IBM a 30-year lease on life. Yes, in theory, you could go get point solutions from all these hot Silicon Valley startups, but you don’t understand that. You don’t know how to do that. You know us.

这样一来，这些公司就能有网站、有电商站点，以及诸如此类的一切。这给了 IBM 三十年的生存期。理论上，你完全可以去那些火热的硅谷创业公司买一堆点解决方案，但你不懂那些，你不知道怎么做。而你认识我们。

55 微软的中间件剧本：90 年代的 IBM

We’ll come in. We’ll create all this middleware, build this big consulting force to help you implement it, and you’ll get online. And IBM basically brought all of corporate America online. That’s Microsoft’s playbook. Microsoft will help you figure out AI. It will help you figure out in a way where you’re not giving away the crown jewels to these companies. We’re going to build this platform, this harness, this sort of middle layer. We’re dependable. We’re stable. You know us. We have backwards compatibility to the 80s. You can build on us, and then we’ll manage all the changing models and what’s updating and do all those sorts of things.

我们会进场，帮你搭好所有这些中间件，建起一支庞大的咨询队伍帮你落地，然后你就上线了。IBM 基本上把整个美国企业界都带上了网。这就是微软的剧本。微软会帮你搞懂 AI——而且是在你不需要把王冠上的宝石交给那些公司的前提下来帮你。我们要搭的是平台、是 harness、是那层中间层。我们可靠、稳定、你认识我们、我们对 80 年代的东西都保持向后兼容。你可以在我们上面构建，然后所有模型的更替、所有更新换代，都交给我们来管。

And does that mean you’ll get the absolute best experience? No. Middleware saws off the sharp edges. You sort of get a lowest common denominator capacity. But if you value in this the oldest enterprise sales motion, how did Oracle go to market? Oracle went to market in the 1980s, Larry Ellison, with another technology taken from IBM, or just IBM didn’t want it, relational databases. And they’re like, you don’t want to be locked into IBM. Relational database, you could run anywhere. Come with us. The reason this is a joke is because Oracle locks you in more than anyone, right? But all of enterprise sales is companies whose long-term goal is to lock you in, getting you on board by trying to make you scared of being locked into somebody else.

那这是否意味着你会得到最好的体验？不会。中间件会把锋利边缘都锯掉，你拿到的是「最大公约数」式的能力。但如果在这个框架里，你看重的是最老的那套企业销售动作——甲骨文当年是怎么打市场的？80 年代，拉里·埃里森（Larry Ellison）带着一项同样从 IBM 拿来的技术（或者说 IBM 不要的技术）：关系型数据库。他们说的是：你不想被 IBM 锁死吧？关系型数据库，你在哪都能跑。跟我们走。这个笑话的地方在于，甲骨文锁死你的程度比谁都狠。可整个企业软件销售，就是一群长期目标是把你锁死的公司，靠让你害怕被别家锁死来让你上车。

All the cloud companies are like, oh, portability, whatever, you can do whatever. They’re like, oh, just use our service that only runs in our cloud, and now you’re locked in. That’s Microsoft’s playbook. It’s a very rational playbook, and I think it makes sense. That’s why they have extra money, because they’re not on the frontier. They are building massive data centers, but they’re building data centers for inference. They’re not building it for training. And their story about we’re investing in time in response to customer demand is more believable in that regard. They’re not having to tell a fungibility story where we’re building big data centers for training that will be used for inference down the road, maybe.

所有云公司都这样：哦，可移植性，随便你啦。然后他们说，哦，你就用我们这个只在自家云上跑的服务吧——于是你被锁死了。这就是微软的剧本。非常理性的剧本，我认为说得通。这也是他们为什么多出钱来：因为他们不在前沿。他们也在建巨大的数据中心，但他们建的是用于推理的数据中心，不是用于训练的。而他们那套「我们跟着客户需求、择时投入」的说法，在这一点上也更可信：他们不必去讲「可替代性」的故事——我们为训练建的大数据中心，将来或许能用于推理。

Go back to this notion that it’s reckless to not be in the front. The reason why that’s concerning, though, is at the end of the day, why are we using Microsoft products again? Because we did before. Like, to what extent does it actually make sense to have all these artifacts, all these documents, all these email inboxes? Can’t AI just do that? There’s a real threat here where, to Microsoft’s software business, the whole systems of record thing is funny, because one reason why systems of records are so powerful is, it’s so hard to move them to somewhere else, because it’s a very tedious, repetitive job.

回到「不待在前沿才是鲁莽」这个说法。它令人担心的地方在于：说到底，我们为什么还在用微软的产品？因为我们以前就用。所有这些文档、邮件收件箱这类「数字遗物」，到底在多大程度上真的有必要存在？AI 不能直接把这些干掉吗？对微软的软件业务来说，这是真实的威胁。而「系统记录」（systems of record）这件事很有意思：系统记录之所以强大，原因之一是把它搬到别处去非常困难，因为那是一件极其繁琐、重复的工作。

57 界面即腹地：Meta、TikTok 与注意力

Oh, AI is actually surprisingly good at that. I’m not sure how good the systems of record. Microsoft isn’t so much systems of record. They do have some of the dynamics business. It’s user interface. It’s like where you actually interact with the computer. That’s the part when you see codex, quad, co-worker, whatever, it is aimed like an arrow to the heart of what Microsoft has. In the long run, all digital companies are, but Microsoft is very much, their strategy is sound. It’s also desperate in an existential way, and also in a, they might pull it off because they’re desperate sort of way. Meta is not threatened immediately, but this is where my bullish view of AI comes in.

哦，AI 在这件事上其实出奇地擅长。我不确定「系统记录」这块会怎样。微软其实不太算系统记录，他们有一部分 Dynamics 业务。他们真正的东西是用户界面——也就是你实际与计算机交互的地方。这正是 Codex、Copilot、Cowork 之类的东西瞄准的地方，像一支箭射向微软的心脏。长期来看，所有数字公司都面临这个威胁，但微软尤其如此。他们的策略是扎实的，同时也带着一种存在意义上的绝望；而也正是在「绝望所以可能真能成」的意义上，他们有机会做成。Meta 不是立刻受威胁的，但这正是我对 AI 的看多观点所在。

I think all digital companies are threatened, and Meta is a digital company. They have software. Now, one worry is AI takes up more and more time. Time ultimately is Meta’s currency. We saw, OpenAI tried the Sora thing, didn’t really take off. Social networks are actually pretty hard. Also cost a lot of money. It’s kind of really interesting. So this came up with the creator payment stuff. So YouTube, very famously as paid creators kind of from the beginning. And that’s a much bigger drag on the business than people appreciate because YouTube has marginal costs to their content. Now, unlike a Netflix, they don’t have to pay that cost up front.

我认为所有数字公司都受到威胁，而 Meta 是一家数字公司。他们做的是软件。现在的担忧之一是：AI 会占掉越来越多的时间。时间归根结底是 Meta 的货币。我们看到 OpenAI 试过 Sora 那件事，没有真正起飞。社交网络其实相当难做，而且很烧钱。这一点很有意思。这跟创作者分成那件事有关。YouTube 从很早就开始给创作者付费，这一点非常有名。而这对生意的拖累比人们意识到的要大得多，因为 YouTube 的内容是有边际成本的。不过和 Netflix 不同，他们不需要预付这笔成本。

They’ll pay it after the fact. So their sharing revenue is a better model than a Netflix model. Netflix has to pay up front for content and then ideally make more money. YouTube pays along the way. But Facebook or Meta, they pay nothing. Instagram is this unbelievable product that generates all this money for which Facebook pays $0 for content. It’s unbelievable. It’s funny because you could see a world where for YouTube, AI-generated content could theoretically be a positive because the inference cost of generating content could be less than what they’re sharing with creators. For Meta, AI-generated content to the extent they’re the ones generating it is actually a worse margin profile than what they have today

他们事后付。所以他们的分成模式比 Netflix 的模式更好：Netflix 必须先为内容付钱，然后最好还能赚更多；YouTube 是一路付一路赚。但 Facebook、也就是 Meta，什么都不用付。Instagram 是一个不可思议的产品，为 Facebook 挣了这么多钱，而 Facebook 为内容付了 0 美元。简直难以置信。有意思的是，你可以想象一个世界：对 YouTube 来说，AI 生成的内容在理论上反而可能是正向的，因为生成内容的推理成本可能低于他们付给创作者的分成。而对 Meta 来说，如果他们自己生成 AI 内容，利润率反而比现在更差。

because what they have today is free. So they have attention. There is a bullish world where Meta is actually very well placed because in a world where we’re interacting with the AI all the time, the desire for a human connection becomes greater. And it’s sort of like a Meta going back to their roots. Meta, one of their biggest mistakes actually, Meta was always a social network company. They killed Snapchat or stopped Snapchat’s growth by realizing Snapchat is a great product. Let’s layer it onto our network. They brought their network to bear to kill Snapchat. The reason why TikTok was such a blind spot for them is TikTok is classified as a social network and it’s not a social network at all.

因为现在这个是免费的。所以他们拥有的是注意力。存在一个看多的世界：Meta 的位置其实非常好。因为在一个我们整天和 AI 交互的世界里，对「人的连接」的渴望会变得更强。这有点像 Meta 回到自己的本源。Meta 最大的错误之一其实是——Meta 一直是一家社交网络公司。他们曾经把 Snapchat 的成长掐死：他们意识到 Snapchat 是个好产品，那就把它叠加到我们的网络上来，用我们的网络去把它干掉。TikTok 之所以成为他们的巨大盲区，是因为 TikTok 被归类为社交网络，而它根本不是社交网络。

TikTok is an entertainment product. It doesn’t matter who you follow on TikTok. What you see on TikTok is a function of what you watched and you’re going to get more of the same. It’s a user-generated content network. And the insight from TikTok was the way to get the best content to limit it to your social network is an artificial constraint. We’re going to give you the best content from across the whole network. And the vast majority of content is going to be crap, but this is like the absolute question before. You don’t think about margins. You think about absolute numbers. The absolute amount of great content, even if the margin for great content is infinitesimal,

TikTok 是一个娱乐产品。你在 TikTok 上关注了谁并不重要：你看到什么，取决于你看了什么，然后你会看到更多同类。它是一个用户生成内容网络。TikTok 的洞察是：只把最好的内容局限在你的社交网络里，是一种人为约束。我们要把整个网络里最好的内容给你。绝大多数内容都很垃圾，但这就像前面说的「绝对值」问题：你不去想利润率，你想的是绝对数量。哪怕好内容的利润率薄到可以忽略不计。

if we have a ton of content, the absolute amount of great content is going to be very large. Then Meta is like, we’re a social network. Meta is serving you content from your network of people you know. And TikTok is serving you the best content from around the world. That’s why they took a huge chunk out of them. Meta had to shift. That’s what’s happened with Instagram and with Reels is it’s not really a social network. It is an entertainment product that pulls from the entire network. And social networking is like the group checked. It’s possible in AI. Actually, social network is important again because we actually want humans.

只要内容总量足够大，好内容的绝对数量就会非常大。而 Meta 的处境是：我们是社交网络，所以我们把「你认识的人的网络」里的内容推给你；TikTok 推给你的是全世界最好的内容。这就是为什么 TikTok 从他们身上咬下了一大块。Meta 不得不转向。Instagram 和 Reels 上发生的事就是这样：它其实不是社交网络，它是一个从全网取材的娱乐产品。而社交网络有点像那个被忽略的对照组。在 AI 时代，社交网络有可能重新变得重要，因为我们确实想要人。

We want to have some sort of connection to them. That will be interesting to see how that plays out. But the other thing with the models is they’re so impactful on advertising. The biggest impact of the models, the biggest monetization right now is probably not Anthropic or OpenAI. It’s the incremental gain that is happening for Google and Meta. Most of the stuff is pre-LLM. But we’re getting to LLMs, whether it be generating advertising content, what do we want? We want verifiable domains. How do you verify if a generated image is good for an ad? Does the ad sell or not? They actually can validate their image creation and their text creation in a way no one else

我们想要和他们有某种连接。这一点将来怎么演化会很有意思。但模型另一件事是：它们对广告的影响太大了。模型当下最大的影响、最大的变现，很可能不是 Anthropic 或 OpenAI，而是谷歌和 Meta 获得的边际增益。现在大部分东西还是前 LLM 时代的。但我们正在走向 LLM：无论是生成广告内容——而这里我们要什么？我们要可验证的领域。你怎么验证一张生成的图对广告是不是好？广告卖出去了没有？他们验证自己生成的图像和文案的能力，是别人无法企及的。

can. And their validation is the ad marketplace. Running a gazillion A-B tests on all these different things, see what works, see what doesn’t. Most ads are a throwaway. It’s fine. The vast majority of ads don’t convert. They have this massive advantage, this huge liquid market that is a verification machine where the verifiers are humans deciding whether they click on that ad and make a purchase or not. But they’re doing it at global scale. That can actually have a feedback loop to make their products better. You’re also going to get a world where ad matching is actually still fairly crude. Here’s the qualities of the person.

而他们的验证器就是广告市场本身：跑海量的 A/B 测试，看什么有效、什么无效。大多数广告是一次性的消耗品，没关系，绝大多数广告本来就不转化。他们拥有这个巨大的优势——一个体量庞大、流动性极高的市场，本身就是一台验证机器，而验证者就是一个个决定要不要点击、要不要下单的真实人类。而且这一切是在全球尺度上发生的。这能形成反馈回路，让他们的产品变得更好。你还会进入一个广告匹配仍然相当粗糙的世界：这边是这个人的特征。

Here’s the qualities of the ad. And it’s like you create an embedding, like a vector calculation and see what numbers match. And then you sort of match an ad to the person. What do LLMs do? LLMs predict. We’re going to move to this world where Meta is going to look at people and say, this person probably wants to see this next. And they’re going to go find that thing and show it to them. The potential upside in terms of showing people better ads that are more relevant to them, they only need to increase a few percentage points for the returns to be billions and billions

这边是这条广告的特征。做法是生成一个 embedding，像向量计算一样看哪些数字匹配，然后把广告匹配给人。而 LLM 做什么？LLM 做预测。我们会走向这样一个世界：Meta 看着一个人说，这个人接下来大概想看这个，然后去找出那个东西、展现给他。就「给人看更相关、更好的广告」这件事而言，只要提升几个百分点，回报就是几十上百亿美元。

of dollars. This alone is worth them investing in being on the leading edge, in having these amazing models. I think a big problem Meta has is they don’t tell this story. It’s weird, but Mark Zuckerberg has the same problem Sam Altman does. He doesn’t love ads. They have the best ad business in the world. They have an ad business that I think is a societal positive. You and I have set up these little content businesses that make great money. Content is you get a ride on social media. I grew up on Twitter, people sharing my links. It was amazing. If you’re selling some product, the beauty of the internet is there is a niche out there

光这一件事，就值得他们投入去待在前沿、去拥有这些了不起的模型。我认为 Meta 的一个大问题是：他们不讲这个故事。很奇怪，但扎克伯格和 Sam Altman 有同样的问题——他并不热爱广告。他们拥有世界上最好的广告业务。我甚至认为他们的广告业务对社会是正向的。你和我也做了这些小小的内容生意，挣得还不错。做内容就是要搭上社交媒体的顺风车。我是在 Twitter 上成长起来的，大家分享我的链接，那太棒了。如果你在卖某个产品，互联网的美妙之处在于，某个角落里就是有一群人想要那个产品。

that wants that product. The question is, how do you find the niche? Facebook advertising. That’s what it does. It helps products find the people who didn’t even know they wanted that product, but when they get it, they’re so happy they got it. That’s a huge societal positive. You have new business from a new entrepreneur making a new product. You have customers who are happy they got something that they didn’t know they would get otherwise. Those customers, by the way, got lots of free entertainment and they’d have to pay for it along the way and Meta made a bunch of money for themselves and their shareholders, which is basically

问题是怎么找到那个角落。Facebook 广告。这就是它做的事：它帮产品找到那些甚至不知道自己想要这个产品的人；而当他们拿到手时，会非常高兴自己得到了它。这是巨大的社会正向。你会看到新创业者做出新产品带来的新生意；你会看到顾客因为得到了本来不知道会得到的东西而开心。顺便说，这些顾客还获得了大量免费的娱乐，本来他们得为此付费，而 Meta 为自己和股东赚了一大笔钱，基本上就是——

01:04 广告的社会价值、ATT，与「智能会不会变成大宗商品」

everywhere in the world. This is why advertising is great and Meta’s advertising in particular is awesome. And I get frustrated that Meta doesn’t talk about that. Mark Zuckerberg has never really talked about the societal benefits of advertising except in passing in 20 years. He’s handed it off to other people to take care of. And maybe there’s a bit where him not paying attention is why there is a certain grit and grind that goes into building an advertising business. People get frustrated or have questions about it as far as data and all those sorts of things. And maybe there was a bit where he didn’t want to be involved in it and wipe his hands

——在世界各地都是这样。这就是为什么广告是件好事，而 Meta 的广告尤其棒。让我沮丧的是，Meta 自己不讲这一点。20 年来，扎克伯格从来没有真正谈过广告的社会价值，只有零星的提及。他把这件事交给别人去打理。也许正是因为他不上心，做广告这门生意才有了那种必须的韧劲和苦功。人们对数据之类的事情会有疑虑、会不满。也可能他确实不想掺和这些，只想把手洗干净。

of it. But you saw this when Apple passed ATT. App tracking transparency was one of the worst antitrust violations in the history of technology. Apple unilaterally obliterating all these business models while they’re simultaneously building their own as far as advertising goes and doing all this tracking. Why? Trust us. And meanwhile, they’re running these advertisements. I remember that advertisement of people on the bus, like overhearing everyone around them, what they’re saying. That was such a dishonest representation of how advertising works on the Internet. You had Tim Cook in Congress talking about companies selling data. Facebook’s not selling your data. That’s value to them. Why would they sell the data?

苹果推出 ATT 的时候你就看到了这一点。App Tracking Transparency 是科技史上最恶劣的垄断行为之一。苹果一边单方面毁掉所有这些商业模式，一边同时在建设自己的广告业务、做着自己的追踪。凭什么？相信我们。与此同时，他们自己也在投放那些广告。我记得那条广告：公交车上的乘客，能听到周围所有人说的话——那是对互联网广告运作方式极其不诚实的呈现。还有 Tim Cook 在国会里谈「公司卖你的数据」。Facebook 不卖你的数据。数据对他们来说是价值，他们为什么要卖？

Meta was not prepared to respond. And I think you got this to Sheryl Sandberg back in the day. She, in every call, would talk about advertising, how great it is, and have a bunch of case studies. People who are benefiting from advertising and these new entrepreneurs. And then she left, and it’s kind of like that hole never got filled. It feels like it’s a company that’s kind of like embarrassed. We make a lot of money from ads, but we got glasses, and we’re doing AI. It’s like, you have ads, and ads are awesome. I think if they had communicated that more consistently, they would be in a better place generally

Meta 当时完全没有准备好在这一点上回应。我觉得这要归到当年 Sheryl Sandberg 身上。她每次电话会都会谈广告，谈它有多好，摆出一堆案例：谁因此受益，哪些新创业者因此起来。然后她离开了，那个缺口似乎就一直没被填上。感觉这是一家有点「不好意思」的公司：我们从广告里赚了很多钱，但我们还有眼镜，我们在做 AI。可事实是：你拥有广告，而广告是件了不起的事。我认为如果他们一直更一致地讲这件事，他们的整体处境会更好。

from a PR perspective. They would have been in a better place relative to Apple. And I think they would have an easier time right now convincing Wall Street that let us invest. The other problem is they’ve spent cumulative hundred some billion dollars on Oculus, which I dated all along. And so there’s a bit where, why should we let you spend money again? The one major player and company that we haven’t talked about much is Jensen and NVIDIA. And I’m curious how you would tie this back to the notion of not understanding commodity markets in Silicon Valley, whether or not you think compute ultimately is a commodity.

从公关角度会更好，相对苹果的位置也会更好。而且他们现在说服华尔街「让我们继续投资」也会更容易。另一个问题是：他们在 Oculus 上累计烧掉了一千多亿美元，这一点我从来都不看好。所以会有这样一种情绪：凭什么再让你花一次钱？我们还没怎么谈到的一个主要玩家是黄仁勋和英伟达。我很好奇你会怎么把它联系回「硅谷不懂大宗商品市场」这件事——你是否认为算力最终会变成大宗商品？

I’m curious whether or not you think intelligence will ultimately be a commodity. It’s interesting that intelligence and compute, which seem to be by far the most interesting and important topics in tech, both might be commodities and less differentiated than the most interesting thing about the Internet is free distribution. Bandwidth is a commodity. The fact that I can pull out my phone right now and connect to any information source in the world for free, free on a marginal cost basis, is because it’s a commodity. It changed the world. Commodities change the world. There is a aspect of differentiated products by definition have lower TAMs because there’s

我很好奇你是否认为智能最终会变成大宗商品。有意思的是，智能和算力——这两个看起来是科技界最有趣、最重要的题目——可能都是大宗商品，都比不上互联网最有趣的那一点：免费分发。带宽是大宗商品。我此刻能掏出手机、免费连接到世界上任何信息源——按边际成本算免费——就是因为它是大宗商品。它改变了世界。大宗商品改变世界。差异化产品按定义 TAM 更低，因为存在弹性问题。

an elasticity aspect to it. Not everyone can afford to pay for it. People’s willingness to pay is going to differ. Your market is going to be constrained. Apple’s never going to serve the whole world by selling a device, whereas a Google can because it’s free. That matters. You’re paying for a commodity, but to the extent it is available to everyone is the extent it is impactful. The Internet is a commodity. It changed the world. So I don’t think it’d be weird that intelligence ends up a commodity and changes the world. Commodities often are not thought of as as good of businesses as these differentiated harm reduction products.

不是每个人都付得起，人们的支付意愿各不相同，你的市场必然受限。苹果永远不可能靠卖硬件服务全世界，而谷歌可以，因为它是免费的。这很重要。你为一样大宗商品付钱，但它对所有人可得的那一部分，正是它产生影响的那一部分。互联网是大宗商品，它改变了世界。所以我并不觉得「智能最终变成大宗商品并改变世界」有什么奇怪。大宗商品常常不被看作好生意，比不上那些差异化的、能减少伤害的产品。

Curious for your thoughts on Jensen and NVIDIA specifically. NVIDIA’s position is, I think, definitely unnatural. You look at NVIDIA, they’ve maintained all their margins. Isn’t that amazing? It’s 2026 and everyone’s coming for them and they’re still charging however much money for a chip. But they’re actually not maintaining their margins because this whole question of circular financing is people talk about Lucent and things like that. And, you know, this whole deal and NVIDIA is providing a 25 percent backstop. But if you actually ascribe a value to that, to NVIDIA’s taking equity in the NeoClouds or whatever, they guarantee they’re going to buy all their computers in 2030.

我很好奇你对黄仁勋和英伟达的具体看法。我认为英伟达的地位绝对是不自然的。你看英伟达，他们维持住了所有利润率。这不神奇吗？都 2026 年了，所有人都在围攻他们，他们一片芯片还是敢收那么多钱。但实际上，他们并没有真正维持住利润率，因为这整个循环融资的问题——人们会谈朗讯之类的事。你懂的，这一整套交易里，英伟达提供了 25% 的兜底。但如果你真给它估个价：英伟达入股这些 NeoCloud，或者诸如此类，他们保证 2030 年会买下他们所有的机器。

Why do they do that? So that the entity in question can get a lower cost of capital so they can buy over GPUs, et cetera. But implicit in that, why do they get a lower cost of capital? They get a lower cost of capital because NVIDIA assumed risk. This is my point before. Risk never disappears. It just appears somewhere else. Taking on risk has a price. There is a world where AI takes off, it never stops, and everything is fine. And NVIDIA captured all the upside of their risk. But there’s also a world where, say, there’s this NeoCloud they backed up, a ton of compute

他们为什么这么做？为了让相关实体拿到更低的资金成本，好去买 GPU 等等。但隐含的含义是：他们凭什么拿到更低的资金成本？因为他们拿到更低资金成本，是因为英伟达承担了风险。这就是我前面的观点——风险从不消失，它只是出现在别的地方。承担风险是有价格的。存在这样一个世界：AI 起飞了，一路不停，一切都好，英伟达把所有风险的上行都收走了。但也存在另一个世界：比如他们支持的那家 NeoCloud，一大堆算力涌向市场。

comes to market. The hyperscalers have plenty of compute. They don’t have enough compute. NVIDIA is paying for a compute that no one wants. They just lost a bunch of money. If you think about it, there’s an expected value of that investment. That expected value, it’s not zero. It’s not 100%. It’s somewhere in the middle. But that is a diminution of NVIDIA’s profitability. If you actually look at their business holistically, what that is, is a price cut. Now, the price cut didn’t show up in margins, it didn’t show up in what they’re offering. But a lot of what NVIDIA is doing is, how can we maintain our margins, even if the wide

超大规模云厂商有充足的算力，可他们算力不够。英伟达在为没人想要的算力付钱。他们刚刚亏了一大笔钱。如果你仔细想，那笔投资有一个期望值。那个期望值不是零，也不是 100%，在两者之间。但那就是英伟达盈利能力的一种减损。如果你整体看他们的生意，那本质上是一次降价。这次降价没有体现在利润率上，没有体现在他们的报价上。但英伟达做的很多事情，就是「我们怎么维持住利润率，哪怕从更宽的——

view, sort of discounted cash flow, expected value, holistic view of our company. People do discounted cash flows, but are you actually considering all these pieces? The reality is, is that moving stuff off the balance sheet, by and large, works. But they’re doing all these deals to maintain what feels somewhat unnatural. We have seen price cuts. They’re just manifesting in these very bizarre sort of ways. Now, in the long run, I think the challenge is, their ultimate competitors are the hyperscalers, particularly Google and Amazon. So Google and Amazon aren’t just building their own chips, but they’re also looking to sell those chips externally. Google already made a deal to sell like 20% of their TPUs to Anthropic.

——角度、从贴现现金流、从期望值、从公司整体来看并不成立」。人们会做 DCF，但你真的把所有部分都算进去了吗？现实是，把东西挪出资产负债表，大体上是有效的。但他们做这些交易，是为了维持一种感觉上不太自然的东西。我们其实已经看到降价了，只是它以这些非常古怪的方式表现出来。长期看，我认为挑战在于：他们最终的竞争对手是超大规模云厂商，尤其是谷歌和亚马逊。谷歌和亚马逊不只是造自己的芯片，他们还想把这些芯片卖到外面去。谷歌已经达成协议，把大约 20% 的 TPU 卖给 Anthropic。

On the last earnings call, Andy Jassy practically confirmed that they’ll be selling Trainium 3 or maybe Trainium 4 or they’ll be selling Trainium chips sort of eventually externally, which makes sense. That gives them a long-term buy-in to these companies. There’s a huge amount of R&D that goes into developing chips. They get more leverage on their spend. It all makes sense. And by the way, they’re not selling their chips on differentiation. They’re selling their chips as commodities. NVIDIA is the one selling differentiation. People aren’t going to Amazon to use Trainium. So they’re not cannibalizing the attractiveness of their cloud by selling Trainium outside. So they’re NVIDIA’s biggest problem.

上一次财报电话会上，Andy Jassy 几乎确认了他们会把 Trainium 3，或者也许是 Trainium 4 卖出去——他们最终会对外销售 Trainium 芯片，这完全说得通。这给了他们对这些公司的长期绑定。芯片研发需要巨额投入，这样他们对这笔开支有更多杠杆。一切都很合理。顺便说，他们卖芯片不是靠差异化，是把芯片当大宗商品卖。英伟达才是卖差异化的那一家。人们去亚马逊不是为了用 Trainium。所以把 Trainium 卖到外面，并不会削弱他们云业务的吸引力。所以他们是英伟达最大的麻烦。

Because what’s the number one advantage that the hyperscalers have? Scale. Lower cost of capital. It’s a capital fight. They have a lower cost of capital than the NeoClouds do. The NeoClouds are, they’ll buy NVIDIA left, right, and center. And by the way, it also makes total sense that why SpaceX, I like Elon’s out there, we will always buy NVIDIA because they’re the best. No, you’ll buy NVIDIA because they’re the most fungible. NVIDIA is true. It is the most fungible. CUDA’s moat is dramatically diminished because the models don’t care what they run on. And that’s what actually matters, what’s built on top of the models. But it still matters.

因为超大规模云厂商的头号优势是什么？规模。更低的资金成本。这是一场资本战。他们的资金成本比 NeoCloud 更低。而 NeoCloud 会拼命买英伟达。顺便说，这也完全说得通：为什么 SpaceX……我喜欢 Elon 说的「我们永远买英伟达，因为他们最好」。不，你们买英伟达是因为它最「通用可替换」。英伟达确实如此，它是最通用可替换的。CUDA 的护城河已经大幅削弱，因为模型不在乎自己跑在什么上面。真正重要的是在模型之上构建了什么。但它仍然重要。

01:10 英伟达、电力约束，与「我们希望泡沫留下什么」

It’s still something of a moat. If you want to play the game SpaceX is doing, where we’re going to build a lot and rent it out, but reserve the right to pull it back, of course you’re going to be on NVIDIA because the easiest way to rent it out is to be on NVIDIA. You saw this very early, by the way. You go back to 2024, 2023. NVIDIA starts talking about all these sovereign clouds. They start talking about, they tried to call these Neutron models. They had this thing in 2024, I remember. It was the first one where it was like the rock star GTC at San Jose and like the huge

它仍然算某种护城河。如果你想玩 SpaceX 那套游戏——我们要建很多、然后租出去，但保留随时收回的权利——那你当然会选英伟达，因为最容易出租的方式就是建立在英伟达上。顺便说，这一点你很早就看得到了。回到 2024、2023 年，英伟达开始谈这些「主权云」。他们开始讲，还试图把它们叫做 Neutron 模型之类的。我记得 2024 年有那么一次——那是第一届像摇滚明星一样的圣何塞 GTC，巨大的——

Coliseum and just no one comes out. It was a very boring GTC. The old ones used to be NVIDIA demonstrating like 50 gazillion things as they’re throwing stuff at the wall. They knew they had something with GPUs and they’re trying to like find the use case. Once LLMs showed up, it’s like, oh, we have the use case. But they were coming up with all these enterprise offerings. I can’t remember what they were called. They’re like these modules, basically, that, of course, they were free, but they only ran on NVIDIA. And you could see what they were doing is they were trying to lock people in.

——竞技场那么大，结果什么都没端出来。那是一届非常无聊的 GTC。以前的 GTC 是英伟达把五十亿样东西摆出来演示，像是在往墙上扔东西看看哪样粘得住。他们知道自己在 GPU 上有东西了，正在找用例。LLM 一出现，就成了：哦，用例有了。但在那之前，他们在搞一堆企业级的产品。我记不清叫什么了。基本上就是一些模块，当然是免费送的，但只能在英伟达上跑。你能看出他们在做什么：他们想把人锁住。

Intel is a good example here. AMD cleaned them out in hyperscaler sales because the hyperscalers were put in the effort to get stuff working on AMD versus Intel. There are still small differences, even though they’re x86, because they’re buying at such scale. The investment to do it is worth it to get a better chip or lower price or whatever it might be. The part of Intel’s business that never floundered was selling to government and selling to enterprises. They don’t have the resources of a hyperscaler. They’re not buying at that scale. They’re just going to keep buying what they had before. That’s why NVIDIA talks about selling to sovereign clouds.

英特尔是个好例子。在超大规模客户的销售里，AMD 把英特尔清了出去，因为超大规模厂商愿意花功夫让东西能在 AMD（而不是英特尔）上跑起来。哪怕都是 x86，差异仍然存在，因为他们采购的量太大，为拿到更好的芯片或更低的价格，这点投入是值得的。英特尔那块从来没有垮掉的业务，是卖给政府和卖给企业。那些客户没有超大规模厂商的资源，采购量没到那个量级，他们只会继续买以前买的东西。这就是为什么英伟达要谈「卖给主权云」。

That’s why they talk about selling to enterprises, because they want to get in these markets where they’re not going to be balancing this chip versus that chip. The hyperscalers have always been the threat to NVIDIA for that reason. They’re actually bigger. So you have this issue where the hyperscalers are the threat. The hyperscalers have a better cost of capital than the other companies NVIDIA wants to buy them. That’s how you get this deal this week. I see this deal as a response. That’s why it goes with the Google deal. Google can just issue equity. Shareholders don’t love it, but their monetization capacity is much higher than NVIDIA or NVIDIA’s

这就是为什么他们要谈「卖给企业」——他们想进入那些不会拿这颗芯片和那颗芯片反复权衡的市场。正因如此，超大规模厂商一直是英伟达的威胁：他们实际上更大。所以你的处境是：超大规模厂商是威胁，而且他们的资金成本比英伟达想卖货的那些公司更好。这周那笔交易就是这么来的。我把这笔交易看作一种回应。这也是它和谷歌那笔交易连在一起的原因。谷歌可以直接增发股票，股东不喜欢，但他们的变现能力比英伟达——或者英伟达的——

customers are. I think what NVIDIA is hoping for, maybe they wouldn’t say this in so many words, but if we get to a world where we actually run out of power, that’s probably good for NVIDIA because in a world where we’re totally constrained on power, we have to get the best efficiency, the best token efficiency. And I think NVIDIA is still the most token efficient. So that is a good world for them. Probably the biggest problem for NVIDIA over the last couple of years is I think the U.S. has actually brought a lot more power online than expected. It surprised me. Whether it be what Elon did sort of behind the meter, which has been replicated, or West Texas

——客户高得多。我认为英伟达期待的是（也许他们不会把话说这么直白）：如果我们进入一个电力真的不够用的世界，那对英伟达大概是好事，因为在电力被完全约束的世界里，我们必须拿到最高效率、最高的 token 效率。而我认为英伟达仍然是最省 token 的。所以那对他们是个好世界。过去几年英伟达最大的麻烦，我认为是美国实际上把比预期多得多的电力接入了电网。这让我意外。不管是 Elon 那种「表后」自建（后来被广泛复制），还是西得克萨斯——

is a natural gas, but even like restarting nuclear plants. We love how the U.S. responds to these things. It’s awesome. It’s actually one of the biggest encouraging signals about the U.S. is I was writing early on, assume this is a bubble. You want there to be a long-term payout. The dot-com, we got fiber in the ground. And by the way, Google’s played this game before. Google built its business by buying up dark fiber. They had the killer search engine, but so much of the power of what they do is because they bought up all this dark fiber that was basically free after the dot-com era.

——的天然气，甚至包括重启核电站。我们真的很喜欢看美国应对这类事情的方式，太棒了。这实际上是关于美国最令人鼓舞的信号之一：我早就在写，假设这是一场泡沫，你希望它留下长期的回报。互联网泡沫里，我们得到了埋在地下的光纤。顺便说，谷歌玩过这个游戏。谷歌是靠买暗光纤起家的：他们有一个杀手级的搜索引擎，但他们做事的很大一部分力量，来自在互联网泡沫之后几乎免费地买下所有这些暗光纤。

Our core internet still runs on WorldCom fiber. That was a lasting benefit. The railroads, BNSF is throwing off money that’s going to Google from Northern Pacific and Jay Cooke selling bonds to retail investors. You want a bubble that produces something that lasts. And very long, it’s like, what’s going to last from AI? The GPUs don’t last that long. Data centers, okay, fine. But what is it going to be? Power. It has to be power. If we’re in a world where this all blows up and we have way too much power, that is an amazing world to be. We’ve always been energy constrained. Energy undergirds everything.

我们的核心互联网至今还跑在 WorldCom 的光纤上。那是一份持久的收益。铁路也是，BNSF 如今还在给谷歌分红，那份资产一路追溯到北太平洋铁路、追溯到 Jay Cooke 向散户卖债券。你希望一场泡沫能留下一些持久的东西。那 AI 会留下什么？GPU 撑不了那么久。数据中心，还行吧。但会是什么？是电力。只能是电力。如果我们处在一个「这一切都崩了、我们电力严重过剩」的世界里，那会是一个美妙的世界。我们一直受制于能源。能源支撑着一切。

What would it be like to live in a world of energy abundance? It’s hard to even imagine because our minds are so constrained by the fact we’ve actually always been in energy scarcity. I think we’ve done an unbelievable job. Power for sure is a constraint. It’s going to be a constraint. But I think it has taken longer to become a constraint than anyone expected. And I wouldn’t be surprised if that includes Jensen Huang. I think he thought insufficient power was going to be NVIDIA’s moat sooner than it happened. It turns out that the longer we have enough power, the more time Amazon has to make Trainium better.

生活在一个能源丰裕的世界里会是什么样？甚至很难想象，因为我们的思维被「我们一直处于能源稀缺」这个事实牢牢限定住了。我认为我们做得非常了不起。电力当然是一个约束，它会是一个约束。但我认为它变成约束所花的时间，比任何人预期的都长。而且我猜这「任何人」里包括黄仁勋。我认为他以为电力不足会更早成为英伟达的护城河。结果却是：电力够用的时间越长，亚马逊就有越多时间去把 Trainium 做得更好。

The more time Google has to make TPUs competitive from an efficiency standpoint. And if we get in a world where these are in a world where those margins seem very hard to sustain. I love hearing your takes on just everything going on. It’s the most interesting time I’ve ever observed in this world that you love so much. So thank you so much for your time. Thank you very much.

谷歌就有越多时间让 TPU 在效率上变得有竞争力。而如果我们进入那种状态——那些利润率会显得非常难以维持。我很喜欢听你聊这一切。这是我在你如此热爱的这个世界里，观察到的最有意思的时代。非常感谢你抽时间。非常感谢。