Dylan Patel - The Infinite Demand for Tokens, Claude Mythos, and Supply Constraints - [Invest Like the Best, EP.469]

频道: Invest Like the Best with Patrick O'Shaughnessy
视频: https://colossus.com/episode/supply-demand-of-tokens/
原文语言: en
统计: 共 72 轮 · Patrick O'Shaughnessy 8 · Dylan Patel 61

[0:00]

Ramp is the only platform built to make your finance team leaner, faster, and better,saving businesses 5% annually on average so you can stay focused on growth.Ramp customers grew revenue 3.2 times faster than the average American business.Visa, Vercel, Cursor, Stripe, Notion, 11Lab, Shopify, and 70,000 other businesses all runon Ramp.Mine does too, and so should yours.Learn more at ramp.com slash invest.OpenAI, Cursor, Anthropic, Perplexity, and Vercel all have something in common.They all use WorkOS.To achieve enterprise adoption at scale, you have to deliver on core capabilities like SSO,SCIM, RBAC, and Audit Logs.Instead of spending months building these mission-critical capabilities yourself, you can just use WorkOSAPIs to gain all of them on day zero.That's why so many of the top AI teams you hear about already run on WorkOS.WorkOS is the fastest way to become enterprise-ready and stay focused on what matters most,your product.Visit WorkOS.com to get started.Felix by Rogo is a personal finance agent that turns a single prompt into finished,client-ready work using your firm's own templates, context, and standards.Send Felix an email like,take these comments and turn them for me,

Ramp 是唯一一个专为让你的财务团队更精简、更快、更好而打造的平台，平均每年为企业节省 5% 的开支，让你能专注在增长上。Ramp 的客户收入增速是美国企业平均水平的 3.2 倍。Visa、Vercel、Cursor、Stripe、Notion、ElevenLabs、Shopify，以及另外 7 万家企业都跑在 Ramp 上。我的公司也是，你的也该是。详情见 ramp.com/invest。OpenAI、Cursor、Anthropic、Perplexity 和 Vercel 有一个共同点：它们都用 WorkOS。要想在企业级市场规模化落地，你必须交付这些核心能力：SSO（单点登录）、SCIM（账号自动开通与同步）、RBAC（基于角色的权限控制）和审计日志（Audit Logs）。与其花几个月自己造这些关键能力，不如直接用 WorkOS 的 API，第零天就全部拥有。这就是为什么你听说过的这么多顶尖 AI 团队都已经跑在 WorkOS 上。WorkOS 是让你达到企业级就绪、并持续专注于最重要的事——你的产品——的最快路径。访问 WorkOS.com 开始使用。Rogo 出品的 Felix 是一个个人金融智能体（agent），它用你所在机构自己的模板、上下文和标准，把一句提示词变成可以直接交付客户的成品。你给 Felix 发一封邮件，比如「把这些意见帮我改进去」，

[1:11] Patrick O'Shaughnessy

or update my tracker with the context of these emails.And Felix sends back finished PowerPoint decks, Excel models, and sourced research.Felix works the way your team already does,delivering work quickly and accurately around the clock.Learn more at rogo.ai slash Felix.Hello and welcome, everyone.

或者「用这几封邮件里的信息更新一下我的跟踪表」。Felix 就会回给你做好的 PowerPoint 演示稿、Excel 模型和带出处的研究材料。Felix 按你团队本来的工作方式干活，全天候快速而准确地交付成果。详情见 rogo.ai/felix。大家好，欢迎收听。

[1:30] Patrick O'Shaughnessy

I'm Patrick O'Shaughnessy, and this is Invest Like the Best.This show is an open-ended exploration of markets, ideas, stories, and strategiesthat will help you better invest both your time and your money.If you enjoy these conversations and want to go deeper,check out Colossus, our quarterly publication with in-depth profilesof the people shaping business and investing.You can find Colossus along with all of our podcasts at Colossus.com.Patrick O'Shaughnessy is the CEO of Positive Sum.All opinions expressed by Patrick and podcast guests are solely their own opinionsand do not reflect the opinion of Positive Sum.This podcast is for informational purposes onlyand should not be relied upon as a basis for investment decisions.Clients of Positive Sum may maintain positions in the securities discussed in this podcast.To learn more, visit PSUM.vc.This is my second conversation with Dylan Patel.Dylan is the founder and CEO of SemiAnalysis,where he tracks the semiconductor supply chain and AI infrastructure build-out.This conversation is about the supply and demand of tokens.On demand, Dylan describes something completely explosive.He explains why the Frontier model is the only model anyone wants

我是 Patrick O'Shaughnessy，这里是《Invest Like the Best》。本节目是对市场、思想、故事与策略的开放式探索，帮助你更好地投资你的时间和你的钱。如果你喜欢这些对话、想更深入，可以看看 Colossus——我们的季刊，深度刻画那些正在塑造商业与投资的人。Colossus 和我们所有播客都能在 Colossus.com 上找到。Patrick O'Shaughnessy 是 Positive Sum 的 CEO。Patrick 与播客嘉宾表达的所有观点均为其个人观点，不代表 Positive Sum 的立场。本播客仅供参考，不应作为投资决策的依据。Positive Sum 的客户可能持有本播客中讨论的证券。了解更多请访问 PSUM.vc。这是我第二次和 Dylan Patel 对话。Dylan 是 SemiAnalysis 的创始人兼 CEO，在那里他追踪半导体供应链和 AI 基础设施的建设。这次对话讲的是 token 的供给与需求。在需求侧，Dylan 描述的东西完全是爆炸性的。他解释了为什么前沿模型（Frontier model，即当下能力最强的那一档模型）是唯一有人想要的模型，

[2:37] Patrick O'Shaughnessy

and willingness to pay for it is nearly unbounded.His own firm has gone from tens of thousands of dollars in AI spend last yearto 7 million run rate this year.On supply, we walk through the bottlenecks across memory, logic, and fab equipmentthat will determine how fast any of this can scale.We also cover mythos and what the leading labs need to doto fix their growing perception problem.Please enjoy my conversation with Dylan Patel.You told me this incredible story about how your own team's use of tokens

以及为什么人们为它付费的意愿几乎没有上限。他自己的公司在 AI 上的花费，从去年的几万美元涨到了今年 700 万美元的年化运行速度。在供给侧，我们会逐一走过存储（memory）、逻辑芯片（logic）和晶圆厂设备这几处瓶颈——它们决定了这一切到底能扩张得多快。我们还聊了 Mythos，以及领先的实验室们要做什么，才能修好自己日益严重的公众观感问题。请欣赏我与 Dylan Patel 的对话。你跟我讲过一个特别惊人的故事，关于你们自己团队的 token 用量

[3:06] Dylan Patel

has changed dramatically this year.Yeah.Can you retell that story and what it is teaching you about what's going on in the world?Last year, we thought we were heavy users of AI.Everyone's using ChatGPT.Everyone's using Claude, providing whatever subscriptions anyone wants.On the order of spend of tens of thousands of dollars for our firm.This year, the spend has just skyrocketed.And it really started in late December with Opus.That included Doug O'Loughlin, who's president.He's very much leading the charge in the sense of non-technical people using AI for coding.And so he's basically pilled the whole firm slowly over time.I think he's been the leader in doing that.Obviously, the engineers were using it anyways.But spend in January just started to inflect and rocket and rocket and rocket and rocket.We signed an enterprise contract with Anthropic.And it's gone to the point where now, I think when I last talked to you,it was 5 million spend rate.It's actually 7 million spend rate now.So we're spending 7 million.That was last week, by the way.And a lot of that is just the usage.People who have never coded before are using Claude Code and spending thousands of dollars,sometimes a day.And it also lets some people spend thousands of dollars one day,

今年发生的剧变。是的。你能把那个故事再讲一遍吗？以及它让你看到了世界正在发生什么？去年我们以为自己已经是 AI 的重度用户了。人人都在用 ChatGPT，人人都在用 Claude，谁想要什么订阅我们都给。整个公司的量级大概是几万美元的开销。今年这个开销直接冲天了。真正的起点是去年 12 月底的 Opus。带头的人里包括我们的总裁 Doug O'Laughlin。他非常主动地在推动一件事：让非技术背景的人也用 AI 写代码。他基本上是慢慢地把全公司都「点化」了，我觉得他是这件事上的领头人。工程师们本来就在用。但 1 月份开销就开始拐头，一路火箭般往上冲、往上冲、往上冲。我们跟 Anthropic 签了企业合同。到现在——我记得上次跟你聊的时候是 500 万美元的年化开销——现在实际上是 700 万了。所以我们在花 700 万。顺便说一句，这还是上周的数字。其中很大一部分就是使用量。那些以前从没写过代码的人在用 Claude Code，一天能花掉好几千美元。也可能有人今天花好几千，

[4:15] Dylan Patel

or spend a couple hundred dollars a couple days, and then they go back to $1,000.It's very variable across each individual user.But across a firm, we're spending 7 million dollars a year now on Claude Code at the current rate,versus our salary expense being in the neighborhood of $25 million.So we're north of 25% of spend on Claude Code as a percentage of salary.And if this trajectory continues, then we'll spend more than 100% by the end of the year,which is a bit terrifying.Thankfully, I don't have to decide between people and AI, because our company's growing so fast.It's more so like, okay, well, I don't have to hire nearly as fast,and I can spend a lot more on AI, and it works, and we just grow faster.But I think other folks will start to reckon with the fact that,huh, if this person can do the work of 5 to 10 to 15 people using Claude Code,then all of a sudden, I should probably cut people.And the use cases are so broad.Give a couple examples.Okay, so for example, one thing is we have a reverse engineering lab in Oregonthat we've been building for a year and a half.We have a bunch of fancy microscopes, scanning electron microscopes.The whole purpose of this is you reverse engineer chips.

接下来几天每天花几百，然后又回到 1000 美元。每个人之间波动很大。但整个公司算下来，我们现在按当前速度一年在 Claude Code 上花 700 万美元，而我们的工资支出大约在 2500 万美元这个量级。所以 Claude Code 的开销占工资的比例已经超过 25%。如果这条曲线继续走下去，到年底我们会花掉超过 100%，这有点吓人。谢天谢地，我不用在人和 AI 之间做选择，因为我们公司增长太快了。更像是：好吧，那我不用招人招得那么快了，我可以在 AI 上多花很多钱，而且它管用，我们还长得更快。但我觉得其他人会开始面对这么一个现实：嗯，如果这个人用 Claude Code 能干 5 到 10 到 15 个人的活，那我大概是该裁人了。而且用例的范围广得离谱。举几个例子。好，比如说，我们在俄勒冈有一个逆向工程实验室，已经建了一年半。我们有一堆高级显微镜、扫描电子显微镜。它的全部目的就是逆向工程芯片。

[5:20] Dylan Patel

You get the architecture out of it.You get the materials that they're using to manufacture,and this is some of the data we sell.This is a very slow process of analyzing that data.Instead, one person on the team,they've been able to spend, with a couple thousand dollars of Claude tokens,they've been able to create this application that is GPU accelerated,runs on a server that we have at CoreWeave,and anytime we send it an image,it will take the picture of the chip and overlay where every single material is.Oh, this part is copper.Oh, this part of the gate is tantalum.This part of the gate is germanium.This part of the gate is cobalt.And so you can do a finite element analysis of the entire stack up of the chip.Very, very quickly, visual with a dashboard, GUI, it's everything.A few thousand dollars of Toclaude.The person previously worked at Intel,and he said that was an entire team's job to build that and maintain that.I'll rack that up across the entire firm.It's insane.Another example that I think is super fun is Malcolm.He's an economist at a major bank before.Their economist department was like 100 or 200 people.What he built was the most incredible thing ever.He piped all of this different data,

你从芯片里把架构扒出来，把它制造时用的材料扒出来，这些就是我们卖的一部分数据。分析这些数据是一个非常慢的过程。结果团队里有一个人，他花了几千美元的 Claude token，就做出了一个应用：GPU 加速，跑在我们放在 CoreWeave 的一台服务器上，任何时候我们发一张图给它，它就会把芯片的照片拿过来，把每一种材料的位置叠加标出来。哦，这块是铜。哦，栅极的这部分是钽。这部分是锗。这部分是钴。于是你就能对整颗芯片的叠层做有限元分析（finite element analysis，一种把结构切成小块、逐块算物理特性的仿真方法）。非常非常快，有可视化、有仪表盘、有图形界面，什么都有。就几千美元的 Claude token 换来的。这个人以前在英特尔工作，他说这在英特尔是一整个团队的活——搭它、维护它。把这种事在全公司乘一遍，简直疯了。另一个我觉得特别好玩的例子是 Malcolm。他以前是一家大银行的经济学家，他们那个经济研究部门有一两百人。他做出来的东西是我见过最不可思议的。他把各种不同的数据全都接了进来，

[6:22] Dylan Patel

FRED data and all these other data, employment reports,and all these other things from various APIs.We signed a couple of contracts with folks to get API access to data,pulled it all in, started running regressions,started looking at the impact of various economic revolutions on the economy.From a deflationary, inflationary perspective,the Bureau of Labor Statistics has this entire set of 2,000 tasks.And so he did that with AI.Which ones can be done by AI?

FRED 数据（美联储经济数据库）以及其他各种数据、就业报告，还有各种 API 来的东西。我们跟几家签了合同，拿到数据的 API 权限，全都拉进来，开始跑回归，开始看各种经济变革对经济的影响，从通缩和通胀的角度去看。美国劳工统计局（Bureau of Labor Statistics）有一整套 2000 项任务的清单，他就用 AI 把这件事做了：哪些能被 AI 干？

[6:46] Dylan Patel

Which ones cannot? And grading them across a rubric.About 3% are doable now with AI.And so he's created this metric so that you can measure things that can be done by AI,what the cost of being able to do those with AI,and therefore the deflationary aspect of it.Phantom GDP is what he's called it.Output can go up, but because cost falls so much,actually GDP theoretically shrink.So he created this whole analysis and a brand new benchmark of language models,a set of evals across 2,000 different evals.He did this all by himself.This is all by himself, yeah.And he's like, dude, this would have taken the team of 200 economists a year.He's like completely cracked out on Claude.He's like, everything has changed.How do you think about it as a business owner,going from close to zero to 25% accelerating towards whatever percent of total spend?

哪些不能？并且按一套评分标准给它们逐项打分。结论是现在大约 3% 是 AI 能做的。于是他造出了这么一个指标，让你可以衡量哪些事能被 AI 做、用 AI 做这些事的成本是多少，进而算出它的通缩效应。他管这叫「幽灵 GDP」（Phantom GDP）：产出可以上升，但因为成本降得太多，理论上 GDP 反而会缩水。所以他做出了一整套分析，还做了一个全新的语言模型基准（benchmark）——一套包含 2000 项不同评测的评测集。这全是他一个人做的。全是他一个人，对。他说，兄弟，这活儿放以前得 200 个经济学家干一年。他完全被 Claude 上头了。他说，一切都变了。作为一个企业主，你怎么看这件事——从接近于零，涨到占总支出的 25%，还在往某个更高的比例加速？

[7:29] Dylan Patel

At what point are you like, whoa, I need to put the brakes on thisand be careful how much we're spending.Maybe we don't need to spend on the most cutting it on Opus 4.7,which came out today.Maybe I can throw it back to something that's a little bit cheaper.I think I'm in the information business.We sell analysis.We do consulting.We create data sets.I don't see why this wouldn't be completely commoditizedon a pretty rapid basis if I'm not constantly improving.My first product that I was selling as a data set,there's more people trying to do it.Now we've made it constantly better and better and better and more detailed.And so therefore it sells a market.But the way we were doing it in 2023 is not terribly different.It's basically what everyone else is doing now.If I don't move up the bar, then I will be commoditized.If I don't move fast enough, I will also lose my edge.So the question is, yes, AI commoditizes things,just like it commoditizes software.Those who can move fast and keep control of their customersand keep providing them an awesome serviceand keep improving the service won't shrink.They'll grow faster.Those who are incumbent and not doing anything, they're going to lose.And so it's a bit of an existential.

到什么程度你会觉得，哇，我得踩个刹车了、得小心花钱了？也许我们不需要用最尖端的 Opus 4.7（今天刚发布的那个），也许我可以退回到便宜一点的东西。我觉得我做的是信息生意。我们卖分析、做咨询、做数据集。如果我不持续变强，我看不出这门生意为什么不会很快被彻底商品化。我最早卖的那个数据集产品，现在想做的人更多了。我们把它不断做得更好更好更细，所以它还能卖出市场价。但我们 2023 年做它的方式，跟现在别人做的方式其实差别不大。如果我不把标准往上抬，我就会被商品化；如果我动得不够快，我也会失去优势。所以问题是：是的，AI 会把东西商品化，就像它把软件商品化一样。那些能动得快、能守住客户、能持续提供出色服务并不断改进服务的人，不会萎缩，他们会长得更快。那些占着在位者位置却什么都不做的，会输。所以这有点像一个生存问题。

[8:33] Dylan Patel

If I don't adopt AI, someone else will, and they will beat me.Another easy example is the energy space.So we've had a few energy analysts for like a year now.We've been trying to build out this energy model.It's very complex.Energy's data services market is something like $900 million.So obviously a huge market for me to try and break into.And we've been like slowly grinding at it.And it's been helpful for our data services business.We really hadn't broken into the energy data services businessdespite a year of having multiple people on the team.Then Cloud Code, Psychosis hits one of the peoplewho leads the data center in energy and industrial business at SemiAnalysis.This Jeremy hits him.And now all of a sudden in three weeks, he spent a lot.He was spending like $6,000 a day.It was an insane amount.But he scraped every single power plant in the U.S.,every single transmission line above a certain voltageand created this entire mapping of the entire U.S. grid,as well as a lot of demand sources,all from various public sources of data.And it's got like this dashboard where you can view and check.You can see all the micro regions of the U.S.where there's power deficits and surpluses.All of these details built in a handful of weeks.

如果我不采用 AI，别人会，而他们会打败我。另一个很好懂的例子是能源领域。我们有几个能源分析师，大概一年了。我们一直想搭一个能源模型，非常复杂。能源的数据服务市场大概有 9 亿美元，显然对我来说是个很大的、值得切进去的市场。我们一直在慢慢磨。它对我们的数据服务业务有帮助，但尽管团队里放了好几个人磨了一年，我们其实一直没真正切进能源数据服务这一块。然后 Claude Code 的「精神病发作」（psychosis，指人被 AI 用上头、疯狂沉迷的那种状态）击中了负责 SemiAnalysis 数据中心、能源与工业业务的一个人——Jeremy。突然之间，三周内，他花了很多钱，一天能花 6000 美元，是个疯狂的数字。但他把美国每一座电厂、每一条超过某个电压等级的输电线路全都爬下来了，做出了整张美国电网的地图，还有大量需求侧的来源，全部来自各种公开数据。而且还带一个仪表盘，你能查能看，能看到美国每一个微区域的电力缺口和盈余。所有这些细节，几周之内做出来。

[9:35] Dylan Patel

We started showing some of our customerswho buy our data center data set but are energy traders.We showed some of them and they're like,wow, how long did this take you?This is really good.This is better than XYZ Company.And then we like dig deeper.XYZ Company has 100 peopleand have been working on this for a decade.Obviously, our thing is not fully as robust,but in some ways it is better.I'm going to commoditize these energy services companies,data services company.Hey, who's going to come commoditize me if I don't move faster?

我们开始给一些客户看——那些买我们数据中心数据集、但本身是能源交易员的客户。给其中一些人看了，他们说：哇，这个你们做了多久？真的很好。这比某某公司的还好。然后我们再往下挖，某某公司有 100 个人，干这件事干了十年。当然我们的东西还没有他们那么完备，但在某些方面确实更好。我要去把这些能源服务公司、数据服务公司商品化掉。那问题是——如果我不动得更快，谁会来把我商品化掉？

[10:00] Patrick O'Shaughnessy

And so the question from a business owner's perspective is,yeah, I'm spending a lot,but what does that spend getting me?Is it getting more revenue?Are you worried that in the limit,the people that control capital and are investing capitalwho are often hiring you for what you dowill just say, well, we have analysts toowho are really smart about this.Like, we'll just build this ourselves.If it's getting that easy,at what point does it just all poolinto the investment firms that stand to gain the mostbecause they have the most leverage on top of the dataor the insights that they glean?

所以从企业主的角度，问题是：对，我在花很多钱，但这些钱给我换来了什么？有没有换来更多收入？你会不会担心，在极端情况下，那些掌握资本、在配置资本、而且经常是雇你干活的人，会说：我们也有分析师，也很懂这个，我们自己搭就行了。如果这事变得这么容易，到什么时候，一切就都汇聚到那些投资机构手里了——因为它们在数据或洞见之上有最大的杠杆、最能从中获益？

[10:29] Dylan Patel

First of all, any information services business,obviously, I don't generate as much valueas my customer does from said information.Because if I sell you information for a dollar,you're only buying it for a dollarbecause you know that information helps you make a decisionthat lets you make more than $1.And so therefore, you have made more money off of methan I did from the information myself.These investment funds all have their own information services,you know, especially like the super,the Jane Streets of the world and the Citadels.They're really detailed on their data.And yet, these sort of folks also purchase data from usand continue to do so and continue to grow with usbecause I think there's just some it factor, right?

首先，任何信息服务生意，显然，我从这条信息里产生的价值，都比不上我的客户从中获得的价值。因为如果我把一条信息一美元卖给你，你之所以愿意花一美元买，是因为你知道这条信息能帮你做一个决定，让你赚到超过一美元。所以你从我这儿赚到的钱，比我从这条信息本身赚到的还多。这些投资基金全都有自己的信息服务，尤其是那些超级机构，比如 Jane Street 和 Citadel 这一类，他们的数据做得非常细。但这类人也照样从我们这儿买数据，而且持续买、跟着我们一起增长——我觉得就是有某种「说不清的东西」在。

[11:04] Dylan Patel

We move faster.We're more nimble.You're at the edge.We're a smaller team that's focusedon just one specific thing,AI infrastructure and the huge revolution that causesin AI and tokenomics and all these things.And we see where it's headed.And so we're moving faster and building faster.I think investment professionals,yes, they'll try and build some of the stuff we do.And more likely, they'll just buy the data from usand it's cheaper for them to buy the data from usand then build on top of itthan it is to build it themselves.I feel like every conversation I have with you,what I'm always getting atis just supply and demand of tokens.That's the thing that's interesting to mein the world right now.What has this experience taught you about the demand?

我们更快、更灵活，我们在最前沿，我们是一支只聚焦在一件具体事情上的小团队：AI 基础设施，以及它在 AI、token 经济学（tokenomics）等等方面引发的巨大变革。而且我们能看到它要往哪儿走。所以我们跑得更快、造得更快。我认为投资专业人士，是的，他们会试着自己做一部分我们做的事；但更可能的是，他们直接从我们这儿买数据，因为从我们这儿买、再在上面做加工，比自己从头造要便宜。我感觉每次跟你聊，我真正想聊的都是 token 的供给和需求。这是我眼下觉得世界上最有意思的事。这段经历让你学到了什么关于需求的事？

[11:43] Dylan Patel

Has it changed your view on the demand side of that equation?Just feeling it viscerally yourself?If we take a step back and look at the macro lens, right?Anthropic has gone from 9 billion revenueto what, they're at 35, 40 billion?

它有没有改变你对这个等式需求侧的看法？就是你自己切身感受到的这一面。如果我们退一步，用宏观视角看：Anthropic 的收入已经从 90 亿美元涨到了多少，350 亿、400 亿？

[11:55] Dylan Patel

Now, probably by the time this airs,40, 45 billion, who does?ARR.Their compute has not grown to the same degree.And if you do the calculationsand you assume they didn't decreasetheir research and development compute,they clearly didn't.They're released, they have mythos.They have Office 4.7.So they clearly didn't decreasetheir research compute spend.So ultimately what they've done,even if you assume all incremental computethey've gotten has gone towards inference,their margins are at a floor of 72%.In reality, some of that incremental computethey've got probably went to research and development.It may be higher than 72% gross margins.To be clear, at the start of the year,there was a leak from their funding around docs.Someone leaked it.30-something percent gross margins.Where on earth does a business like thisgrow margins like that?

等这期播出的时候，大概是 400 亿、450 亿了，谁知道呢。是 ARR（年化经常性收入）。他们的算力并没有以同样的幅度增长。如果你去算一算，并且假设他们没有削减研发用的算力——他们显然没削减，他们在发模型，他们有 Mythos，有 Opus 4.7，所以他们显然没有削减研发算力的投入——那么最终，即便你假设他们拿到的所有增量算力都投进了推理，他们的毛利率下限也在 72%。而实际上，那些增量算力里有一部分很可能去了研发，所以毛利率可能比 72% 还高。说清楚一点：今年年初，他们融资轮的文件泄露过一次，有人把它捅了出去，当时是 30 几个百分点的毛利率。这世上还有什么生意能把毛利率涨成这样？

[12:37] Dylan Patel

It's in principle, right?Their demand is so high.They're able to cut back on usage limits,rate limits, all these things.What really matters is having an Anthropic repand having an enterprise contract with themand getting the rate limit increases that you needbecause otherwise tokens are ultimatelysuper, super in demand.Whoever can pay for them,Anthropic has the same problem.I mean, not problem.It's just the reality of how capitalism works.Yes, people are sending them $40 billion ARR in tokens,but those tokens are generating way morethan $40 billion in value.Various businesses will have differentvalue generation per token,but as we get more and more intelligent,what really matters is accessto these most intelligent tokensand leveraging them at things.You as a person decidingwhat is the best way to leverage these tokensto grow business and generate valuebecause a lot of folks will want tokensand generate tokens,but the shitty SaaS startup in SFwho is using Claudeto generate their software productis not necessarily actually creating a ton of valueand therefore they're going to get pricedout of tokens soon enough.Vanta automates security and compliancefor over 16,000 fast-moving companies

原理上说，是因为他们的需求太高了。他们可以砍使用额度、砍速率限制（rate limit，即单位时间内你最多能用多少）等等。真正重要的是有一个 Anthropic 的客户经理、跟他们签一份企业合同、拿到你需要的速率上限提升，否则 token 归根到底就是极度、极度供不应求。谁付得起谁拿走。Anthropic 也有同样的问题——不，也不算问题，这就是资本主义运转的现实。是的，人们正在给他们送去 400 亿美元 ARR 的 token 消费，但这些 token 产生的价值远超 400 亿美元。不同生意每个 token 能创造的价值不一样，但随着模型越来越聪明，真正重要的是——能不能拿到这些最聪明的 token，以及怎么把它们用在事情上。你作为一个人，要决定怎么用这些 token 才是最好的方式，去把生意做大、创造价值。因为很多人都会想要 token、都会生成 token，但旧金山那家烂 SaaS 创业公司拿 Claude 生成他们的软件产品，其实并没有创造多少价值，因此他们很快就会被 token 的价格挤出去。Vanta 为超过 16000 家快速成长的公司自动化安全与合规，

[13:44]

like Ramp, Cursor, and Harvey,keeping them audit-ready around the clock.It's the number one agentic trust platform,and it now helps companies like yourswatch for the risks that show up between auditsacross your vendors, your AI tools,and your whole environment.Every new tool your team signs up for,every vendor that turns on AI features,is an opportunity for something to go wrong,and most security programsweren't built for AI's pace of growth.The Vanta agent works like a 24-7 GRC engineerin the background,finding issues, drafting fixes for you,and cutting vendor assessment time by up to 50%.Whether you're a fast-growing startupor a global enterprise,Vanta helps you earn and prove trust.Invest like the best listenersget a special offer for $1,000 offat vanta.com slash invest.Ridgeline is the first end-to-end system of recordwith embedded AI for investment management firms,running portfolio accounting, reconciliation,reporting, trading, and complianceon one unified platform.Firms are moving off legacy technologyand onto Ridgelinebecause of how far ahead Ridgeline's AI features arecompared to anything elsein investment management software.Which is why I believe that firmsthat come out ahead in the AI era

比如 Ramp、Cursor 和 Harvey，让它们全天候保持随时可审计的状态。它是排名第一的智能体信任平台，现在还能帮你这样的公司盯住两次审计之间冒出来的风险——覆盖你的供应商、你的 AI 工具和你的整个环境。你团队新注册的每一个工具、每一个开启了 AI 功能的供应商，都是出事的机会，而大多数安全方案压根不是为 AI 这种增长速度设计的。Vanta 的智能体像一个 7×24 小时在后台工作的 GRC（治理、风险与合规）工程师，帮你发现问题、起草修复方案，把供应商评估的时间最多缩短 50%。无论你是高速成长的初创公司还是全球性大企业，Vanta 都能帮你赢得并证明信任。《Invest Like the Best》的听众可在 vanta.com/invest 获得 1000 美元优惠。Ridgeline 是第一个面向投资管理机构、内嵌 AI 的端到端记录系统，把组合会计、对账、报告、交易和合规跑在一个统一平台上。各家机构正在从老旧技术迁移到 Ridgeline，因为 Ridgeline 的 AI 功能领先于投资管理软件里的任何其他东西。这也是为什么我相信，在 AI 时代跑出来的机构

[14:53] Patrick O'Shaughnessy

will be the ones runningon Ridgeline's unified platform.If you're serious about your firm's AI strategy,Ridgeline should be part of that conversation.You can request a demo at ridgeline.ai.I had this experience just today

会是那些跑在 Ridgeline 统一平台上的机构。如果你认真对待你所在机构的 AI 战略，Ridgeline 应该出现在这个讨论里。你可以在 ridgeline.ai 申请演示。我今天就刚好有过这么一次体验，

[15:09] Patrick O'Shaughnessy

where on the flight here,I got really limited out on something.I saw 4.7 came outand what I immediately wantedwas to be on 4.7 that second.I couldn't think about using 4.6 anymoreand not this 4.7 is out.I was perfectly happy with 4.6for the last many weeks.It's amazing.Are you surprised that people are so insistenton going to the most expensiveleading edge thing to the degree they are?

在来这儿的飞机上，我被某个东西严重限流了。我看到 4.7 发布了，我当下唯一想要的就是立刻用上 4.7。我一秒都无法再忍受用 4.6，就因为 4.7 已经出来了。而在过去好几周里我用 4.6 一直很满意。这太神奇了。人们对「必须用最贵、最前沿的那个」执着到这个程度，你觉得意外吗？

[15:34] Dylan Patel

Without a doubt.I think one of my funniest memoriesin the past month and a halfis myself and a buddy of mine, Leopold,being on our kneesin front of an Anthropic co-founderbegging him for access to Mythosand then pretending it doesn't exist.Because we knew it existedand we're like, please give us access.And he's like, I don't know what you're talking about.What was your reaction to that rate cardor that eval card coming out?

毫不意外。我过去一个半月里最好笑的记忆之一，就是我和我一个哥们儿 Leopold 跪在一位 Anthropic 联合创始人面前，求他给我们 Mythos 的访问权限，而他还假装它不存在。因为我们知道它存在，我们就说，求你了给我们开权限。他说，我不知道你在说什么。那份价目表——或者说那份评测卡——出来的时候，你的反应是什么？

[16:00] Dylan Patel

It was rumored in the Bay Area.We knew it was supposed to be really good.But if you just look at the benchmarks,obviously benchmarks change over time.Mythos is potentially the biggest step upin model capabilities in two years.I think that's really, really an important detailthat it's so good that they're like,don't want to release it,even though they already announced the priceto their people that they dida selective release for Cyber forand it's 5 or 10x the token cost.They just don't want to release itbecause they're worried about the impact on the world.And they're releasing a worse version,Opus 4.7, to us.And they explicitly said in the model card,hey, we actually preferentially made it worse at Cyber.I don't know if you read that.Whoever you are, if you have enough capital,you should get a freaking Enterprise Anthropic subscriptionwhere you pay per token,not with these subscriptions,because then you won't get rate limited much.And then you need to figure outhow to leverage those tokensto the highest value taskand make money off of it.Because ultimately what you're doing,maybe like a year from now or two years from now,the business is actually just arbitrageing tokens, right?

湾区早就有传闻了，我们知道它应该非常好。但如果你就看基准分数——当然基准会随时间变化——Mythos 可能是两年来模型能力上最大的一次跃升。我觉得有一个细节非常非常重要：它好到他们不想发布，尽管他们已经把价格告知了那些做定向发布、面向网络安全（cyber）的客户，而且价格是 token 成本的 5 倍或 10 倍。他们就是不想发，因为他们担心它对世界的影响。他们发给我们的是一个更差的版本，Opus 4.7。而且他们在模型卡（model card）里明说了：嘿，我们特意把它在网络安全方面做弱了。不知道你有没有读到。不管你是谁，只要你有足够的资本，你就该去搞一份 Anthropic 的企业订阅、按 token 付费，而不是用那种包月订阅，因为那样你就不太会被限流。然后你要想清楚怎么把这些 token 用在价值最高的任务上，并且从中赚到钱。因为归根到底你在做的事，可能一年后或两年后，这门生意其实就是在做 token 的套利。

[16:58] Dylan Patel

The tokens are amazing,but let's figure out what direction to point them in.And then three or four years from now,the model will know what to do with the tokensand how to make the most value.You need to look at this retroactively,pick any benchmark.The cost to hit a certain capability tierused to cost X.And now it costs 1 100th or 1 1,000th of that.DeepSeq, for example, on GPT-A4was 1 600th the cost.And since then,the costs have fallen furtherfor GPT-4 class models.Of course, no one gives a crapabout GPT-4 class models.They want the Frontierbecause the Frontier lets themcreate the economically valuable things.But GPT-4 class modelscan still be used in stuff.And so people are using themin some like tiny use cases.It's just the cost have fallen so fast.It's not really what's driving the demand.What's driving demandis all these new use cases.Yeah, current 4.6 Opusor 4.7 Opus tier modelsa year from now,my spend for the same exact qualityof the modelwould probably be like 70k.I bet you it'll be 100 times cheaper.Irrelevant because I'm going to be usinga way, way, way better modelwhich can do way, way better things.Anthropic Mythos is more expensive as a model,but it spends a lot less tokens

token 很棒，但我们得想清楚该把它们指向哪个方向。再过三四年，模型自己就知道该拿这些 token 干什么、怎么创造最大价值了。你得回头看：随便挑一个基准。达到某个能力档位过去要花 X，现在只要百分之一或千分之一。比如 DeepSeek，相对于 GPT-4，成本是六百分之一。从那以后，GPT-4 这一档模型的成本还在继续下降。当然，没人再在乎 GPT-4 这一档的模型了。他们要的是前沿，因为只有前沿才能让他们创造出有经济价值的东西。但 GPT-4 档的模型还是能用在一些地方，所以人们把它们用在一些很小的场景里。只是成本掉得太快了，这并不是真正在驱动需求的东西。驱动需求的是所有这些新用例。是的，现在 4.6 Opus 或 4.7 Opus 这一档的模型，一年后，我要买到完全相同的模型质量，我的开销可能只有 7 万美元。我打赌会便宜 100 倍。但这不重要，因为我到时候会用一个好得多得多得多的模型，它能干好得多的事情。Anthropic 的 Mythos 单价更贵，但它完成同一件事花的 token 少得多，

[18:07] Dylan Patel

to do the thing.And therefore, it is actually cheaperin most tasks than 4.6 Opusbecause it's just way more efficient,even though each individual token is smarter.So, yeah, there's crazy geniusescreating huge cost efficiency improvementsevery day.They work at the labsand they're making the modelsway more efficient.You see it every generation.A GPT, what was it,5 Nano or whateverwas better than GPT-4.Or 5 Mini was better than GPT-4and it was like 1 100th the cost.This just happens.And we accept it at its face value,but ultimately you keep making things cheaperand then you keep scaling them upand you keep getting humongous improvements.When I last saw you,Mythos had just come out,maybe the day before or something,or the card had just come out.And you said something like,it actually made you feel like a little scared.It was so good.What did you mean by that?

所以在大多数任务上它其实比 4.6 Opus 更便宜，因为它效率高太多了，尽管每一个 token 更贵、更聪明。所以，是的，有一群疯狂的天才每天都在创造巨大的成本效率改进。他们在实验室里工作，把模型做得效率高得多。每一代你都能看到。GPT——那个叫什么来着——5 Nano 之类的，比 GPT-4 还强；或者 5 Mini 比 GPT-4 强，而成本只有百分之一。这种事就是会发生。我们把它当成理所当然接受了，但归根到底就是：你不停把东西做便宜，然后不停把规模加上去，然后你就不停拿到巨大的能力提升。我上次见你的时候，Mythos 刚出来，可能就是前一天，或者是它的卡刚出来。你当时说了一句话，说它让你有点害怕，因为它太好了。你那话是什么意思？

[18:54] Dylan Patel

Anthropic's whole goal in 2025and even a lot of 2024,they're like,hey, by the end of 2025,we need an L4 software engineerin our model.And they by and large achieved thatwith 4.6 Opus.What they didn't say is that,and if you look at Mythos,if you compare benchmarks,it's like an L6 engineer.So L4 is like pretty new.L6 is like quite well experienced.I think Anthropic said that the model internallywas available in February.So in two months,they've gone from L4 engineerto L6 engineer.What's next?

Anthropic 在 2025 年、乃至 2024 年很大一部分时间里的整个目标，就是：嘿，到 2025 年底，我们的模型里得有一个 L4 级软件工程师。他们靠 4.6 Opus 基本上做到了。他们没说的是——如果你去看 Mythos、去比基准——它像一个 L6 级工程师。L4 大概相当于比较新的工程师，L6 就是相当有经验的了。我记得 Anthropic 说这个模型在内部 2 月就可用了。所以两个月里，他们从 L4 工程师走到了 L6 工程师。那接下来是什么？

[19:26] Dylan Patel

When you think about the model progress,it's only accelerated.Anthropic's release cadence has compressed.OpenAI's release cadence has compressed.Why?Because generally to make a better model,you need a few things, right?

你去想模型的进展，它只有加速。Anthropic 的发布节奏被压缩了，OpenAI 的发布节奏也被压缩了。为什么？因为一般来说，要做出一个更好的模型，你需要几样东西。

[19:36] Dylan Patel

You need amazing compute.Compute is very expensiveand it has a timescale that we trackand it's like,it's growing,but it's set in stonefor the next short term.It's like set in stonewhat you've already signed.And there will be delays and shiftsand somehow you can find a little more,but it's generally pretty set in stone.There's amazing researchersthat people are payingtens of millions of dollars for.And then lastly,there's implementation.And implementation historicallyhas been very difficult.If I have an idea,now I have to implement it.Implementing is hard.Now ideas are there.Implementation is very easy.It's expensive,but it's very easy.So how does one decidewhat ideas to implement?

你需要非常好的算力。算力非常贵，而且它有一个我们在追踪的时间尺度——它在增长，但短期内基本是板上钉钉的，你已经签下的合同就是定死的。会有延迟、有挪动，你也许能多挤出一点，但整体上相当固定。还有那些非常优秀的研究员，人们花几千万美元去请。最后一样是实现（implementation）。而实现在历史上一直非常困难：我有一个想法，然后我得把它实现出来，实现很难。现在想法有的是，实现变得非常容易了——它贵，但很容易。那么问题就变成：一个人该怎么决定去实现哪些想法？

[20:11] Dylan Patel

And it turns outif your implementationis just so much easier,now you can just implement more ideasand move on the treadmillfaster and faster and faster,whether that is AI model research.And so now your model release cadenceis shrunk down to two monthsfrom where it was six months before.Or I want to take every power plantin the US and every transmission lineand model it and run regressionsand see the micro supply and demand.I can also do that.The idea is cheap.Which idea makes sense?

结果是，如果你的实现变得容易这么多，你现在就可以实现更多想法，在跑步机上越跑越快、越跑越快——不管那是 AI 模型研究，于是你的模型发布节奏从原来的六个月缩到了两个月；还是「我想把美国每一座电厂和每一条输电线路都建模、跑回归、看微观供需」，我也能做。想法是便宜的。哪个想法说得通？

[20:35] Dylan Patel

Which idea is worth the capitalthat you have to spend on the tokensbecause the implementation is there?That's the key learning.And if implementation costscontinue to tank,which they are,we don't even have Mythos yet.It's only been a handful of hourssince Opus 4.7 launched,but my team is pretty excitedabout it internally.What now comes to the world?

哪个想法值得你为它花掉买 token 的资本？因为实现这一环已经具备了。这就是关键的学习。而如果实现成本继续暴跌——它确实在暴跌——我们甚至还没拿到 Mythos，Opus 4.7 发布也才几个小时，但我团队内部已经相当兴奋了。这个世界接下来还会出现什么？

[20:56] Dylan Patel

It's a complete reorderingof how economies work.What used to matter a lotwas executionwas very, very fucking difficultand ideas were cheap.Now ideas are cheapand plentiful,but execution is very easy.So really only the good ideasare the ones that can justifythe spend on super cheap implementation.So are you actually scaredor does it just introducean uncertaintythat's hard to grapple with?

这是对经济运转方式的一次彻底重排。过去非常重要的是执行——执行他妈的太难了，而想法很廉价。现在想法便宜又充裕，但执行非常容易。所以真正站得住的只有好想法——好到能证明这笔超便宜的实现开销花得值。那你到底是真的害怕，还是它只是引入了一种很难把握的不确定性？

[21:20] Dylan Patel

Uncertainty is there,but I do think that causessome fear in terms ofhow does society reform itself?How does one exist in a worldwhere actually your abilityto implement somethingis not actually that important?

不确定性是有的，但我确实认为它带来了某种恐惧：社会该怎么重新组织自己？在一个「你把东西实现出来的能力其实没那么重要」的世界里，一个人该怎么存在？

[21:37] Dylan Patel

Your ability to choosethe correct ideafor AI to implementand then your abilityto sell that ideaor sell what the AIhas implementedis what matters.Your ability to garner capitaltowards that is what matters.And going back to the point ofit's very importantto have the newest model always,who's going to have accessto the newest model?

你选对该让 AI 去实现哪个想法的能力，以及你把这个想法卖出去、或者把 AI 实现出来的东西卖出去的能力，才是重要的。你为它筹集资本的能力才是重要的。回到「永远拥有最新模型非常重要」这一点——谁会拿到最新的模型？

[21:55] Dylan Patel

Anthropics Project,I know it's not called Earwig,but I troll Anthropic peopleby calling it Earwig.Glasswig.Anthropic Earwig,where they only releaseMythos to certain companiesfor cyber,that's just going to besomething that continues.Models will haveless broadand less broad deployment.I know OpenAIand Anthropicand all these people are like,we want to havegreat AI for everyone.AI is very fucking expensive.Who's going to payfor the trillion dollarsof infrastructure?

Anthropic 那个项目，我知道它不叫 Earwig（蠼螋），但我就爱这么叫来气 Anthropic 的人。（是 Glasswing。）Anthropic 的「Earwig」项目——他们只把 Mythos 发给特定几家公司做网络安全。这种事只会继续下去。模型的部署范围会越来越窄、越来越窄。我知道 OpenAI、Anthropic 这些人都说，我们想让所有人都用上很棒的 AI。可 AI 他妈的太贵了。那上万亿美元的基础设施谁来付？

[22:21] Dylan Patel

People who have moneyand we can builduseful things with AI.And then you don't wantpeople to distill your modelsso you don't release them broadly.You release them to fewerand fewer set of customers.Those customersare also now wrestlingover the tokensunless Anthropic jacks them.They could double their pricingon Opusand I would continue to payand I bet most userswould continue to pay.I bet that wouldn't solvetheir humongouscapacity problemthat they have.So then the question becomes,where does this cycle endwhere token usageand thereforethe benefits of those tokens,the additional valuegenerated on topof those tokensaggregates among fewerand fewerand fewer companies?

有钱的人，以及那些能用 AI 造出有用东西的人。而且你不希望别人蒸馏（distill，指用你的模型输出去训练一个便宜的仿制模型）你的模型，所以你不会广泛发布它，你会发给越来越少的一批客户。而这些客户现在也在为 token 互相角力——除非 Anthropic 大幅提价。他们可以把 Opus 的价格翻一倍，我照付，我打赌大多数用户也照付。我打赌那也解决不了他们那个巨大的产能问题。那问题就变成：这个循环会在哪儿结束——token 用量、以及随之而来的 token 带来的好处、在这些 token 之上创造出的额外价值，会聚集到越来越少、越来越少、越来越少的公司手里？

[22:55] Dylan Patel

I don't have mythos.You know who has mythos?Top freaking banks.Now they're only using itfor cybersecuritybut at some pointI can envision a worldwhere, hey maybe I,because I have anenterprise Anthropic contractand because Anthropic peoplekind of like me,they're willing to give usslightly earlier accessor slightly higher rate limitsor something for a model.I hope that's what happens.And then my competitor,whoever that is,doesn't have thatand I'm able tofucking crush them.There are peoplelike Ken Griffin of Citadelis super well connectedand super rich.He goes and signs a dealwith Open Air Anthropicthat's like,yeah, I'm going to getaccess to your modelsand I'll buy the first$10 billion worthof tokens each yearso whenever you releasethe model,I'll spend the first$10 billion tokensand then everyone elsecan get the modelafter that.And it's like,okay, well now what does that do?

我没有 Mythos。你知道谁有 Mythos 吗？那些顶级银行。现在他们只拿它做网络安全，但我能想象某个时点会变成这样：嘿，也许我，因为我有 Anthropic 的企业合同、因为 Anthropic 的人还挺喜欢我，他们愿意给我们稍微早一点的访问权、或者稍微高一点的速率上限。我希望是这样。然后我的竞争对手，不管是谁，没有这个，我就能把他们碾碎。有些人，比如 Citadel 的 Ken Griffin，人脉超广、超有钱。他去跟 OpenAI 或 Anthropic 签一个协议：好，我要拿到你们模型的访问权，而且每年我先买 100 亿美元的 token——所以你每次发布模型，前 100 亿美元的 token 我来花，之后别人才能拿到这个模型。那么，好吧，这会带来什么？

[23:39] Dylan Patel

Now he's going to crusheveryone in the markets.That's just an example.It could be any number of things.It could be cyber,like Anthropic is worried about,oh, now I can hack people.It could be information servicesbusiness like myselfwhere I crush someone else.I think it's such a broad base.We don't knowwhat these models can do.Anthropic doesn't knowwhat these models can do.No one knowswhat these models can do.It's up to the end userto figure outwhere they can leveragethe tokens to seewhat they can buildand imagine,which is tremendously productiveand uplifting for humanity.But then what happensto the concentrationof resourcesand usage of it?

现在他要在市场上把所有人碾碎。这只是一个例子，可以是任何事情。可以是网络安全，就像 Anthropic 担心的那样——哦，现在我能黑别人了。也可以是像我这样的信息服务生意，我把别人碾碎。我觉得这个面非常宽。我们不知道这些模型能做什么。Anthropic 不知道这些模型能做什么。没人知道这些模型能做什么。这要靠最终用户自己去发现，他们能把 token 用在哪儿、能造出和想象出什么——这一点极其有生产力，也让人振奋。但接下来，资源和使用的集中化又会变成什么样？

[24:05] Dylan Patel

Presumably right now,robotics or robotsconsume relatively zero tokensversus everything else.What's your view of that?If that's like a seconddemand curvethat could start to ratchet,there's a new startupevery single daywithin a mile of heretrying to buildsomething interestingin robotics,and so there's this conceptof software-only singularity,which is thatthe world has AI singularitybut only in softwareand now what aboutthe rest of the world?

大概现在，机器人相比其他一切，消耗的 token 几乎是零。你怎么看这件事？如果那是一条可能开始往上爬的第二需求曲线的话。这里方圆一英里之内每天都有新的创业公司，想在机器人上做出点有意思的东西。所以有一个概念叫「纯软件奇点」（software-only singularity）——世界上出现了 AI 奇点，但只发生在软件里，那世界的其余部分怎么办？

[24:31] Dylan Patel

The vast majorityof the world is physical.You can see the worldorient around hardwarenot software.That's actually whyI think software-only singularityis like just a blipand not like we doget everything elsebecause once softwareis super easy,what makes robotsreally hard?

世界的绝大部分是物理的。你能看到世界正在围绕硬件而不是软件重新组织。这其实就是为什么我觉得纯软件奇点只是一个小插曲，而不是说我们拿不到其他东西——因为一旦软件变得超级容易，那让机器人真正难的又是什么？

[24:46] Dylan Patel

It's programming microcontrollersand actuatorsand controlling all this stuffis very difficultand right now,the interesting thingabout models,AI models,is they're actuallyreally inefficientin learning.It's just we're ableto give them so much datathat they're able to learnand pass us in certain ways.Currently,the robot models,VLA's,vision language action models,which is very popularright now,is probablynot going to bethe thing that ultimatelyscales beyond.They are inefficientin dataand we can't scalethe data for themfast enough.There is going to besome way tolarge-scale pre-trainedrobot modelswhere just like humanssee all this datathroughout their livesand what's interestingis humans,the reason why we're so goodis we're sample efficient.One example,two example,we're good.So applying that to robotics,so once you have thissoftware-only singularityimplementation is super cheap,anyone can start to buildthese models that nowrobots are actually usefuland so I thinkin the next 6 to 18 months,we'll start seeingreal breakthroughs in robotics,that enablefew-shot learning,i.e. there's a pre-trainedrobot modeland now there's a robotthat you have hiredor bought or whatever.You show it a few examples

是给微控制器和执行器（actuator，把电信号变成实际动作的部件）编程、控制这一大堆东西，这非常难。而现在关于模型有意思的一点是：AI 模型在学习上其实非常低效，只不过我们能喂给它们海量数据，所以它们才学得会、甚至在某些方面超过我们。目前的机器人模型，也就是现在很流行的 VLA（vision language action models，视觉-语言-动作模型），大概率不是最终能规模化下去的那个东西。它们在数据上很低效，而我们没法把数据规模拉起来得够快。将来一定会出现某种大规模预训练的机器人模型——就像人类一生中看到了海量数据一样。有意思的是，人类之所以这么强，是因为我们是样本高效（sample efficient）的：一个例子、两个例子，我们就会了。把这个应用到机器人上——一旦你有了纯软件奇点、实现变得超级便宜，任何人都能开始造这类模型，机器人就真的有用了。所以我认为未来 6 到 18 个月，我们会开始看到机器人领域真正的突破，能实现少样本学习（few-shot learning，只看几个示范就学会）——也就是有一个预训练好的机器人模型，然后你雇了或买了一台机器人，你给它演示几遍，

[25:54] Dylan Patel

and it's able to do it.Right now,there's a lot of companiesdoing robots for advertisementor robots for simple stufflike that,but it'll be like,oh, folding clothes,sure, sure, sure,no, but it's going to getreally niche.Robots just for cleaningchalkboardsand it's a rental serviceor it'll be a model packagethat you downloadonto your standard robotthat then does that.And anyways,there'll be a huge explosionand physical good accelerationand deflationary effects there,but that's ultimatelygoing to keeptoken demand going crazy.I don't think token demandslows down personally.Did you learn anything elseabout the worldbased on Mythos' resultsand how it was built?

它就能做了。现在有很多公司在做机器人打广告，或者做类似的简单东西，但将来会是——哦，叠衣服，好好好，不，它会变得非常细分。专门擦黑板的机器人，而且是一种租赁服务；或者是一个模型包，你下载到你的标准机器人上，它就会做那件事。总之，会有一次巨大的爆发，实体商品加速、通缩效应出现，但这归根到底还是会让 token 需求继续疯涨。我个人不认为 token 需求会慢下来。基于 Mythos 的结果和它的构建方式，你还了解到了关于这个世界的其他什么事吗？

[26:28] Dylan Patel

This is my way of askingif you break downthe componentsof the scaling laws.Mythos is a materiallylarger modelthan prior models.And so, yes,it is a much larger model.What chip it's trained onis not really relevant.It's the scale.And obviously,to 100,000 blackwellsis a coolantto hundreds of thousandsof prior generation chips.TPUs and Traniumhave their differentrelease cadence,so it's not exactlylike mirrored one-to-one.But ultimately, yes,Mythos is a significantlylarger model.It's proof that the scalinglaws still work.Everything about itshows the trendlinecontinues of models.More compute into modelmakes model better.And along the whole way,it's not just more computeinto model makes model better.Along the whole way,we're also gettingthese compute efficiency wins,which are,as all this research computethat the labs are spendingis actually turning into,if I want X capabilityto your model,every six months,that cost,or every two months,that cost is dramaticallydecreasing.But then if I scale it upmassively,I get a humongouscapability drop as well.And so, yes,it's proof that thisis still happening.Google and Anthropicare not heavy,heavy users of GPUson the training side,but OpenAI,they'll start having

这是我在问你能不能拆解一下缩放定律（scaling laws，即模型能力随算力、数据、参数增加而可预测提升的规律）的各个组成部分。Mythos 是一个比之前的模型实质上更大的模型。所以，是的，它是一个大得多的模型。它在什么芯片上训练其实不重要，重要的是规模。显然，10 万张 Blackwell 大致相当于几十万张上一代芯片。TPU 和 Trainium 有各自不同的发布节奏，所以并不是完全一一对应的。但归根到底，是的，Mythos 是一个显著更大的模型。它证明了缩放定律仍然有效。关于它的一切都表明模型的那条趋势线在继续：往模型里投更多算力，模型就更好。而且一路上还不只是「更多算力让模型更好」；一路上我们同时也在拿到这些算力效率的提升——实验室们花的所有研发算力，实际上正在变成：如果我想要模型有 X 能力，每六个月、甚至每两个月，这个成本就在大幅下降。而如果我再把规模拉大，我还会拿到一次巨大的能力跃升。所以，是的，这证明了这件事仍在发生。Google 和 Anthropic 在训练侧并不是 GPU 的重度用户，但 OpenAI 会开始有

[27:32] Dylan Patel

their new class of models.I think they're takinga more sensible,principled approachto scaling in small steps.Anthropic really wentfor a huge jump.We'll see better and bettermodels throughout the year,and the release cadenceis only going to get faster.We've gone a long wayin the conversationwith saying almost nothingabout OpenAI,which would have beenso strange 12 months ago.So this is the interesting thing.Everyone's like,okay, so Anthropic'sjust won, right?

他们新一类的模型。我觉得他们在扩张上采取的是更稳妥、更有原则的小步走法，而 Anthropic 是直接来了个巨大的跳跃。今年我们会看到越来越好的模型，发布节奏只会更快。我们这场对话已经聊了很久，却几乎没提 OpenAI——这在 12 个月前会显得非常奇怪。所以有意思的地方在这儿。大家都说：好了，那就是 Anthropic 赢了呗？

[27:52] Dylan Patel

They had Mythos in February.They never even released itbecause they didn't feelthe need to.They're already sold out.Their revenue's alreadyadding $10 billion a month.And then you've gotOpus 4.7 today,all before OpenAI'salleged spud release,which media such asThe Informationand others have posted about.So clearly,Anthropic is in the lead,and OpenAI is cooked.What's interesting isbecause Anthropichas such bounds on compute,and they can only grow itso fast,and to the point of,Dario used to gloat abouthow OpenAIwas being too aggressiveon compute,and Anthropicwas more sensiblein their scaling,and now Anthropicis like,fuck, I wish we hada lot more compute.OpenAI is able to paythe bills perfectly fine.In fact,they've raised a ton of moneyto get incremental computein addition to theirresponsible levelsof computethat they were buyingfrom Oracle and Core,and SoftBank,and all these people,and Microsoft,such as Tranium.Now they're getting Traniumas well from Amazon.They've done thisinsane thing on compute.They also knowthey need more,but what's interesting isif you were to sayOpus 4.6,let's ignore modelsgetting better over time.Let's just take diffusionof this technology.You and I may jump on the model

他们 2 月就有了 Mythos，一直没发，因为他们觉得没必要发——他们本来就已经卖光了。他们的收入已经在每个月增加 100 亿美元。然后今天又来了 Opus 4.7，全都发生在 OpenAI 那个据称要发的新模型（The Information 等媒体报道过）之前。所以显然 Anthropic 领先，OpenAI 完蛋了。有意思的是，因为 Anthropic 在算力上有这么强的约束，他们只能长这么快。说到这个，Dario 以前还得意地说 OpenAI 在算力上太激进，而 Anthropic 在扩张上更稳妥；现在 Anthropic 心想：操，真希望我们有多得多的算力。OpenAI 完全付得起账单。事实上，他们已经融了一大笔钱去拿增量算力，而这是在他们本来从 Oracle、CoreWeave、软银、微软等等那里买的「负责任规模」的算力之上的。比如 Trainium，他们现在也从亚马逊那儿拿 Trainium。他们在算力上做了这件疯狂的事。他们也知道自己还需要更多。但有意思的是，如果你说 Opus 4.6——先不管模型会随时间变好这件事，就只看这项技术的扩散：你和我可能第一天就跳上新模型，

[28:58] Dylan Patel

immediately day one,but other businessestake time,and it takes timefor people to learn,and the spark of,oh shit,quad psychosis momentdoesn't hit everyoneat the same time.And so,by the end of the year,let's say a 4.6 Opustier model the economywould spend $100 billion on.I don't think that's unreasonable.It's spending $40 billionright now.That's like a linearextrapolation.It's a linear extrapolation,not an exponential.To get the exponential,you need the better models.Anthropic won't haveenough compute to do that,and presumably OpenAIand Googlewill hit that tiersoon enough.Whoever hits that tier next,sure,Anthropic may get to charge70 plus percentgross margins,but if OpenAIhits it next,they charge 50 percentin gross margins,they still get all of thisincremental demand,and probably they alsowon't have enough computeto serve all the users.Sure,maybe Mythosis a modelwhere if the worldhad enough compute,it'd be $500 billionof revenueor something crazy.There is such demandfor these tokensand such limitationson compute.We see this withH100 prices skyrocketingand all these other things.The useful life of these GPUscontinue to extend.It's pretty cleareven the tier 2 labis going to be sold out

但其他企业需要时间，人们学会也需要时间，那种「我操」的 Claude 上头时刻不会同时击中所有人。所以，到今年年底，假设一个 4.6 Opus 档位的模型，整个经济会在它上面花 1000 亿美元。我不觉得这个数字不合理。现在花的是 400 亿。这只是线性外推。是线性外推，不是指数。要拿到指数，你需要更好的模型。Anthropic 不会有足够的算力去做到；而 OpenAI 和 Google 大概很快也会摸到那个档位。谁下一个摸到那个档位——好，Anthropic 也许能收 70% 以上的毛利率，但如果 OpenAI 下一个摸到，他们收 50% 的毛利率，他们照样吃到所有这些增量需求，而且大概率他们也没有足够算力服务所有用户。当然，也许 Mythos 是这样一个模型：如果世界上有足够算力，它能做到 5000 亿美元的收入，或者某个疯狂的数字。对这些 token 的需求就是这么大，而算力就是这么受限。我们从 H100 价格飞涨和其他种种迹象都能看到这一点。这些 GPU 的使用寿命还在不断延长。很明显，连第二梯队的实验室都会卖光

[30:04] Dylan Patel

of tokens,let alone the tier 1 lab.The tier 1 labwill have better margins,but the tier 2 labwill be sold out,and probably the tier 3 labwill also be closeto sold out.Economic valuethat the best modelcan deliveris growing fasterthan our abilityto actually servethose tokens to peoplevia the infrastructure.And so this gapwill continue to growand the model labswill continue to haveexpanding marginsuntil peoplein the hardware supply chainand infrastructure supply chainare like,wait, no,why don't I justjack up my margins?

token，更别说第一梯队了。第一梯队的实验室毛利更好，但第二梯队的会卖光，而且第三梯队大概也会接近卖光。最好的模型能交付的经济价值，增长得比我们通过基础设施把这些 token 真正服务给人的能力还快。所以这个缺口会继续扩大，模型实验室的利润率会继续扩张——直到硬件供应链和基础设施供应链里的人说：等等，不对，我为什么不把我的利润率也拉上去？

[30:29] Patrick O'Shaughnessy

So suffice to say,I think the assessment todayor your assessmentof the demand sideis completely explosivein your own particular examplehere at Semi-Analysis,but just more broadlythat you call itAI psychosisis people fallinto this experienceof what they can do,the implementation difficultygoing completely away.I've certainly felt that.My own token spendis just throughthe absolute roofjust in a matter of weeks.So that feels likea pretty good assessment.Anything we're missingon the demand side?

所以可以这么说，我觉得今天的判断——或者说你对需求侧的判断——是完全爆炸性的：既体现在你自己在 SemiAnalysis 这个具体案例上，也更普遍地体现在你说的那个「AI 精神病」——人们掉进那种体验里，实现的难度完全消失了。我肯定也感受到了。我自己的 token 支出在短短几周内就冲到了绝对高位。所以这听起来是个相当准确的判断。需求侧还有什么我们漏掉的吗？

[30:54] Dylan Patel

If you don't use more tokens,you'll never escapethe permanent underclass.Either you use more tokensand you generateeconomic value,outsize economic valuefor the use of those tokens.A lot of peopleare doing itthe boring, lazy way.Oh, I guess I'll just workone hour a dayinstead of eight hours a dayand I'll have AIdo most of my job.That's the boring way.The cool way isI'll still work eight hours a dayand I'll do 8x the workand maybe I'll make5x the money.You can't do thiswith a job, obviously.There's peoplewho have multiple jobsand there's peoplewho start companiesand start selling stuff.There's people who are hustling,which is what I viewlike you and I as doingas we're mostly hustling.Get that economic valueon this AIbefore everyone is usingit in its table stakesbecause it's stillnot table stakes.So, if you don't usemore tokensand generate the valuefrom themand capture that value,there's three different problemshere.Using more tokens,generating valuefrom those tokensand capturing valuefrom the valuethat you createdfrom the tokens.If you don't dothese three things,you'll never escapethe permanent underclass,i.e. as modelscontinue to skyrocketcapability and theconcentration of resources

如果你不用更多 token，你永远逃不出永久底层阶级。要么你用更多 token 并且创造出经济价值——相对于你用掉的这些 token 来说是超额的经济价值。很多人在用无聊、偷懒的方式做这件事：哦，那我一天就工作一小时而不是八小时吧，让 AI 干我大部分的活。那是无聊的方式。酷的方式是：我还是一天工作八小时，但我干八倍的活，也许我能多赚五倍的钱。当然，你打一份工是做不到这件事的。有些人打好几份工，有些人开公司卖东西，有些人在拼——我觉得你和我基本上就是在拼。在所有人都在用、它变成标配之前，先把这波 AI 的经济价值拿到手，因为它现在还不是标配。所以，如果你不用更多 token、不从中产生价值、不把价值捕获下来——这里其实是三个不同的问题：用更多 token、从这些 token 里产生价值、以及从你用 token 创造出来的价值里捕获价值。如果你不做这三件事，你永远逃不出永久底层阶级——也就是说，随着模型能力继续飞涨、资源的集中化

[31:51] Dylan Patel

potentially happens.All right,let's talk about supply.What is changingat the frontierof supplyingthe entire stackthat's requiredto serve all these tokensas the demand curve explodes?As demand skyrockets,prices are going upfor everythingon the supply side,whether it bethe end GPUs,their prices are going up.In addition,their useful lifeis extending.H100 prices look like this.Yeah, exactly.There's people who have arguedGPUs full livesare less than five years.Complete nonsense.Their cluster is nowre-signing.Three or four-year-oldhopper clustersare re-signingfor three or four more years.There's A100 clustersthat are re-signingfor another couple years.So the useful lifeis clearly not five years.It's maybe evenseven or eight years,arguably.We don't know yet.We'll see when hopper gets there,but it's clearly not five years.So the useful lifeis extendingand the prices are going upon that renewal.In effect,the gross marginwas not 35% on a cluster.It's beyond that.So margins are expandingin the cloud layer.Margins are extremely healthyon the hardware layerwith NVIDIA still charging 75%or whatever percentgross margin.As we move downthe stack memory,obviously marginshave skyrocketed there.Places like Optics

可能会发生。好，我们来聊供给。在服务这些 token 所需要的整个技术栈的前沿，随着需求曲线爆炸，什么正在变化？随着需求飞涨，供给侧的一切价格都在涨，不管是终端 GPU——它们的价格在涨；此外，它们的使用寿命还在延长。H100 的价格是这样的走势。对，正是。有人论证过 GPU 的完整寿命不到五年，完全是胡说八道。他们的集群现在正在续签。三四年前的 Hopper 集群正在续签三到四年。还有 A100 的集群在续签好几年。所以使用寿命显然不是五年，甚至可以说可能是七八年。我们还不知道，等 Hopper 走到那儿我们就知道了，但显然不是五年。所以使用寿命在延长，而续签时价格还在涨。实际效果就是，一个集群的毛利率不是 35%，比那高得多。所以云这一层的利润率在扩张。硬件层的利润率极其健康，NVIDIA 还在收 75% 之类的毛利率。往下走到存储，显然那儿的利润率已经飞涨了。像光学器件

[32:59] Dylan Patel

and Logic,there are large prepaymentsand margins are growing slowly.More so,the companiesthat are making chipslike NVIDIAare paying huge prepayments.So in effect,the cast of capitalor timing of cashor the returnon invested capitalis going upeven if the gross margin isn't.And you see thisacross the whole supply chain.You see ASMLis completely sold outand they need Carl Zeissto expand faster.Everyone's either sold outand margins are going upor they're getting prepaymentswhich increases the returnon invested capitalbecause the invested capitalis lower.And so this is a consistent trendacross any part.It's even liketo make a PCBrequires copper foil.And that copper foilis sold outand people are makingprepayments for it.Anything and everythingthat has a pulseand is sold out,people are jumpingto get more incremental supplyand fighting overthe supplyfor the years after.What do you thinkare the most importantbottlenecks?

和逻辑芯片这些环节，有大额预付款，利润率在缓慢增长。更重要的是，像 NVIDIA 这样造芯片的公司在支付巨额预付款。所以实际上，资本的成本、或者说现金的时点、或者说投入资本回报率（ROIC）在上升，即使毛利率没上升。你在整条供应链上都能看到这一点。你看到 ASML 完全卖光了，他们需要蔡司（Carl Zeiss）扩产更快。每个人要么卖光了、利润率在涨，要么在拿预付款——预付款会提高投入资本回报率，因为投入的资本更少了。所以这是任何环节都一致的趋势。甚至像做一块 PCB（印刷电路板）需要铜箔，而铜箔也卖光了，人们在为它付预付款。任何有点用处的东西只要卖光了，人们就抢着去拿增量供给，并为之后几年的供给互相争夺。你觉得最重要的瓶颈是什么？

[33:49] Patrick O'Shaughnessy

Typically in economic historywhen there's this kind of demand,supply reorientsand rises very,very quicklyto meet the demand.It seems likeit's almost impossiblefor supplyright now in this momentto keep up.Famous last words,every shortageis followed by a glut historically.But what are themost interesting bottlenecksto youacross the supply side?

在经济史上，通常出现这种量级的需求时，供给会重新调配、并且非常非常快地跟上。而现在这个时刻，供给似乎几乎不可能跟得上。当然，历史上「每一次短缺之后都跟着一次过剩」这句话很容易变成著名的临终遗言。但在供给侧，你觉得最有意思的瓶颈是哪些？

[34:10] Dylan Patel

Supply chainsare usually very fast to react.One unique thingis that our supply chainsnow are more complexthan everand that things we're buildingare more complex than everand therefore the lead timesare longer.And it's not likewe haven't seen18 month long lead timesin other industries.It's justbuilding incremental supplydidn't take years.And this is the casewith memory.Memory can only grow capacitylow double digit percentagesa year, right?

供应链通常反应非常快。一个独特之处在于，我们现在的供应链比以往任何时候都复杂，我们在造的东西也比以往任何时候都复杂，因此交期更长。倒不是说别的行业没出现过 18 个月的交期，只是「建增量产能」以前不需要好几年。而存储就是这种情况。存储的产能每年只能增长低两位数的百分比，对吧？

[34:33] Dylan Patel

20, 30% a year.Even less for NAND,a little bit higher for DRAM,but whatever.Even though the demand signalwas very strongat the end of 2025,the memory companiesimmediately started reacting.None of that incrementalcapacity really gets hereuntil the secondthat they've decided to doin addition to thetypical 20 to 30%.They can stretch a little bit,but really,the true incremental supplydoesn't come until 28,which is a very unique thing.Even if they wantedto build as fast as possible,it doesn't come until 28.Late 27 at best.So,the result ismemory prices havegone through the roofand guess what?

一年 20%、30%。NAND 还更低一点，DRAM 高一点，但差不多。即便 2025 年底需求信号已经非常强、存储厂商立刻就开始反应，那些增量产能——在他们决定要做的那一刻之后——真正到位，也要等到 2028 年。这是在通常的 20% 到 30% 之外的部分。他们能稍微拉伸一点，但真正的增量供给要到 28 年才来，这是一个非常独特的情况。就算他们想以最快速度建，也得等到 28 年，最早也是 27 年底。所以结果就是存储价格已经冲上天了，而且你猜怎么着？

[35:05] Dylan Patel

They're going to doubleand triple again.At least on DRAM,especially.People are like,oh, the memory storyis overplayed,everyone gets it.It's like,no, no, no,you don't get it.DRAM will doubleor triple from here stillbecause that's how muchcapacity is requiredand they have to stealcapacity from somewhere elseand the only wayto steal capacityfrom somewhere elsein a capitalist economyis demand destructionvia higher pricing.We're not rationing stuff hereand so ultimatelythat's what's going to happenand so marginscontinue to go up.I think Logicalso has humongouscapacity problems.TSMC just had their earnings,they keep upping CapEx,ultimately takes themquite some timeto build fabsand they're tryingto do everythingthey can to squeezeevery little outputout of every fabthat they havebut ultimatelythey're not raisingprices fastbecause they're good peopleit seems like.Single digit price increasesinstead of triple digitprice increaseslike the memory guyshave had.So you ultimatelyhave this marketwhere yeah,TSMC is a great companybut are they actuallygoing to extractall the value?

它们还会再翻一倍、翻两倍。至少 DRAM 尤其如此。人们说，哦，存储这个故事被讲过头了，大家都懂了。不不不，你没懂。DRAM 还会从这儿再翻一倍或两倍，因为需要的产能就是这么多，而他们必须从别的地方把产能抢过来；而在资本主义经济里，从别处抢产能的唯一办法，就是靠涨价来做需求毁灭（demand destruction，即价格高到把一部分需求逼走）。我们这儿不搞配给制，所以最终就会这样，所以利润率会继续上涨。我认为逻辑芯片也有巨大的产能问题。台积电刚出了财报，他们不停上调资本开支（capex），但建晶圆厂终究要花不少时间，他们在想尽一切办法从现有每一座厂里挤出每一点产出。但归根到底他们不快速涨价，看起来是因为他们是好人——只涨个位数的价，而不像存储那帮人涨三位数。所以你最终会得到这样一个市场：是的，台积电是一家伟大的公司，但它真的会把全部价值提取走吗？

[36:00] Dylan Patel

I mentioned thingslike copper foil,glass fibers for PCBs,lasers,these are thingsthat are well understoodin niche supply chainsbut they're very,very tightand ultimately upstreamthe semiconductorwafer fabricationequipment supply chainis one that's gone upa lot but it's stillvery underappreciated.TSMC,CapEx this yearthey say 56.We've had 57.4 billionsince Januaryand we may up itslightly morejust because we seesome ways that theycan get incrementalCapEx but what peoplearen't focusing onis what does thatmean next yearand what does thatmean the year afterand it turns outthree years from nowTSMC is going tospend $100 billionon CapEx.Maybe two years from nowit might be 28.But sincerelythey may spend$100 billionon CapEx in 2028and people justcan't fathom thatbut what does thatmean for theirdownstream supply chains?

我提到过铜箔、PCB 用的玻纤、激光器这类东西——这些在小众供应链里是被充分理解的，但都非常非常紧张。而再往上游，半导体晶圆制造设备这条供应链已经涨了很多，但仍然被严重低估。台积电今年的资本开支，他们说是 560 亿。我们从 1 月起给的数是 574 亿，而且我们可能还会再往上调一点，就因为我们看到他们还有一些办法能挤出增量资本开支。但人们没在关注的是：这对明年意味着什么？对后年又意味着什么？结果是三年后台积电的资本开支会到 1000 亿美元。也许两年后就到了，可能是 2028 年。但说真的，他们可能在 2028 年花 1000 亿美元的资本开支，人们就是想象不到这个数字。可这对他们的下游供应链意味着什么？

[36:46] Dylan Patel

Companies likeLamb Researchor Applied Materialsor ASMLor their furtherdownstream supply chainslike MKSIand all these othercompanies.The tail whipjust gets whippedharder and harderand harderand that's a shortageif TSMC wants tospend $100 billionin 2028which is a realpossibility.I think peoplewould think that'sinsane but that'sa real, real possibility.What about other partsof the chip ecosystemwhere GPUs have beencompletely dominant?

像 Lam Research（泛林）、Applied Materials（应用材料）、ASML，或者更下游的供应链，比如 MKS Instruments 之类的公司。这条鞭子的尾巴只会被甩得越来越狠、越来越狠、越来越狠——如果台积电真的想在 2028 年花 1000 亿美元，那就是一场短缺，而这是真实存在的可能性。我觉得人们会觉得这很疯狂，但这真的、真的有可能。芯片生态里其他部分呢——GPU 一直是完全占据主导的，

[37:10] Dylan Patel

What about like CPUsor ASICs or thingsthat start to pop outas both opportunitiesand bottlenecksbeyond justNVIDIA's GPU dominance?Yeah, I mean ASICsare obviously taking offbut I'll pivot awayfrom AI chipsto talk aboutthese other things.There's a projectwe did on FPGAsand it turns outthere's 120 FPGAsper next generationRAC, AI RACand then what aboutall the FPGA names?

那 CPU、ASIC（专用集成电路，为某个特定任务定制的芯片），或者其他在 NVIDIA 的 GPU 统治之外冒出来的、既是机会也是瓶颈的东西呢？对，ASIC 显然在起飞，但我要从 AI 芯片岔开去聊点别的。我们做过一个关于 FPGA（现场可编程门阵列，一种出厂后还能改电路逻辑的芯片）的项目，结果发现下一代 AI 机架里每一台要用 120 颗 FPGA。那 FPGA 那些公司呢？

[37:32] Dylan Patel

CPUIs,all these reinforcementlearning environmentsplus all the slop codeyou and I are generatingthat is now runningon someVercel instanceor whatever it isor some AWS instanceor some bucketthat we've spun upall of that requires CPUand so CPUsare completely sold outand demandis skyrocketing there.Yeah, how peopleunderstand the rolethat CPU playsand everything?

CPU 也是——所有这些强化学习环境，加上你我在生成的所有那些垃圾代码，现在跑在某个 Vercel 实例上、或者某个 AWS 实例上、或者我们开的某个存储桶上，所有这些都需要 CPU。所以 CPU 完全卖光了，需求在那儿飞涨。是啊，人们对 CPU 在这一切中扮演的角色理解得怎么样？

[37:51] Dylan Patel

There's two main reasonswhy you needtons of CPUs.One is when you'redoing reinforcement learningthe CPU isvery critical to that.So before you wouldthrow all the internet'sdata into the modeltrain itand it spitssome stuff out.Now you trainall the world'sinternet'syou put all theinternet datainto the modelthen you put itin this environment.This environmentis like,hey modeltry this outand it tries stuff outtries a bunchof different thingsand in the endthere is an environmentwhich scoreswhether or notwhat it tried outis successfuland it grades it.These environmentscan be anything.It can be,hey,check if the textwas outputtedin the right waystructured outputs.It can be very simple stuffor it can bevery complex stuffand people are startingto get intovery complex thingslike,hey,I want you toopen this filechange it,edit it,update it,submit it to this website.I want you to open upthis physics simulationfrom Siemensand edit this CAD modelso the environmentscan get more and more complexand those environmentsrun on CPUs.They don't run on GPUs.They don't run on ASICs.The ASICs run the modelthat takes the input datafrom the environment,runs it through the model.The model creates outputsof various different trajectories,

你需要大量 CPU 主要有两个原因。第一，做强化学习（reinforcement learning）的时候，CPU 非常关键。以前你是把全互联网的数据扔进模型里训练，它吐出一些东西。现在你先把全互联网的数据放进模型，然后把它放进一个「环境」里。这个环境会说：嘿，模型，你试试这个。它就去试，试一堆不同的做法，最后环境会判定它试出来的东西成功与否，并给它打分。这些环境可以是任何东西：可以是「检查这段文本有没有按正确格式输出」这种结构化输出的检查，可以是很简单的东西，也可以是很复杂的东西。人们现在开始做非常复杂的：嘿，我要你打开这个文件、改它、编辑它、更新它、提交到这个网站；我要你打开西门子（Siemens）的这个物理仿真、编辑这个 CAD 模型。所以环境可以变得越来越复杂，而这些环境跑在 CPU 上。它们不跑在 GPU 上，不跑在 ASIC 上。ASIC 跑的是模型——把环境给的输入数据拿进模型跑一遍，模型产出各种不同的轨迹（trajectory），

[38:57] Dylan Patel

ways that it thinkit could solve itin different instances.Those trajectoriesare graded slash scoredand the onesthat are successfulyou train onand you updateand you iterate,iterate, iterate.And so CPUsare very usefulfor that oneand then once you havethese great modelsand you're deploying them,those modelsare generating code.They're generatinguseful output.That useful output,it doesn't gofrom a GPUstraight to the human brain.It goes from a GPUor an ASICthrough toa deployed appthat you're deployingsomewhere that actuallyjust runs on CPUs.So that's another areawhere there's a lot of demandand things are sold outin a large, large way.As you continueto assessand try to bethe world's best informed personon both the trajectoryof supply and demand,what are thingsthat you wish you knewto make that understandingthat you don't know?

也就是它认为在不同情形下可以怎么解决问题的各种路径。这些轨迹会被评判、打分，成功的那些你就拿来训练、更新、迭代、迭代、迭代。所以 CPU 在这一块非常有用。然后，当你有了这些很棒的模型并把它们部署出去，这些模型在生成代码、生成有用的输出。而这些有用的输出，并不是从 GPU 直接进到人脑里的，它是从 GPU 或 ASIC 出来，进到你部署在某处的一个应用里——那个应用其实就是跑在 CPU 上的。所以这是另一个需求很大、而且大面积卖光的领域。你一直在评估、并努力成为这个世界上对供给和需求走向了解得最清楚的人，那有什么是你希望自己知道、但目前还不知道的？

[39:43] Dylan Patel

I think the hardest areafor usand for everyoneis understandingtokenomics,economics of tokens.I think we havea really tremendouslygood insightinto how much it coststo run infrastructure,what the cost of tokens are,what the cost of models are,what the marginsof these labs are.But the usageand adoptionis what's really difficultto model.Continuously, right?

我觉得对我们、也对所有人来说，最难的领域是理解 tokenomics（token 经济学，即 token 的成本、定价与价值创造这本账）。我认为我们对「跑基础设施的成本是多少、token 的成本是多少、模型的成本是多少、这些实验室的利润率是多少」有非常好的洞察。但使用量和采用率才是真正难以建模的。而且是持续地难，对吧？

[40:05] Dylan Patel

January,we had crazy estimatesfor February.Anthropics smashed them.How do we calibratethis model?What are the data sourcesfor this?February,we had crazy assumptionsfor March.I know people were like,you're crazy, Dylan.And then they smashed them.Everyone sees the numberof 10 billionand they're like,what the fuck?

1 月的时候，我们给 2 月做了疯狂的估计，Anthropic 把它们全砸穿了。我们该怎么校准这个模型？数据源在哪儿？2 月的时候我们给 3 月做了疯狂的假设，我知道有人说，Dylan 你疯了。然后他们又砸穿了。所有人看到 100 亿这个数字都会想：什么鬼？

[40:18] Dylan Patel

How do they add10 billion of revenue?Who's using all these tokens?Why are they using them?What are they buildingwith them?And then more importantly,with what they're buildingwith these tokens,how is that actuallydiffusing into the economyand what valueis that generating?

他们怎么可能加上 100 亿的收入？谁在用掉这些 token？他们为什么要用？他们拿这些 token 在造什么？以及更重要的是，他们用这些 token 造出来的东西，是怎么扩散进经济的？创造了什么价值？

[40:30] Dylan Patel

Because it's not reallysomething that you can capturein any GDP statistic.All of the valueof the tokens that I useget transformedinto better information,which I then sellat a discountto what people usedto sell informationfor relatively.And therefore,that informationis now making its waythroughout the economyand people are makingbetter investment decisionsor better competitive decisionsif they're a semi-carter companyor a data center companyor a hyperscaler.What is the valueof this?

因为这其实不是任何一个 GDP 统计口径能捕捉到的东西。我用掉的 token 的全部价值，被转化成了更好的信息，而我又以相对过去信息售价打了折的价格把它卖出去。因此这些信息现在正在经济里流转，人们据此做出更好的投资决策，或者——如果他们是半导体公司、数据中心公司、超大规模云厂商——做出更好的竞争决策。这东西的价值是多少？

[40:56] Dylan Patel

What has that doneto the economy?It's clearly,by every subjective metric,amazing.But where is the phantom GDP?What is the phantom GDP?How do we trackthe real economic value?Because the GDP metricsare not accurateif you were to say,what is the GDPthat Dylan Patel is making?

它对经济做了什么？从每一个主观指标看，它显然是惊人的。但幽灵 GDP 在哪儿？幽灵 GDP 是什么？我们怎么追踪真实的经济价值？因为 GDP 指标是不准的。如果你要问「Dylan Patel 创造的 GDP 是多少」，

[41:13] Dylan Patel

It's tiny comparedto the valuethat I think is being created.And I think you would saythe same for Patrick.What is the valuebeing created by these tokens?Not on a basis of simple,what is the knock-on effectof all the thingsthat these things are doing?

跟我认为正在被创造出来的价值相比，那个数字小得可怜。我想 Patrick 你也一样。这些 token 创造的价值到底是多少？不是按简单口径算，而是这些东西所做的一切带来的连锁效应是什么？

[41:25] Dylan Patel

And I think that's the realquestion and challenge.It's hard to measure.I think we've gota tremendous readingon the supply side of things.I think we've gota tremendous readingon even a lotof the demand side signals,but it's what is the valuethese tokens are generating?

我觉得这才是真正的问题和挑战：它很难度量。我觉得我们对供给侧有非常好的把握，我也觉得我们对需求侧的很多信号有非常好的把握，但这些 token 到底在创造什么价值——

[41:38] Dylan Patel

That's hard to quantifyand measure.I hope we get a chanceto do this every three monthsbecause this changesso quickly.What do you thinkis going to happen next?When I come backthree months from nowand we're in San Franciscotogether again,what do you expect?

那件事很难量化和度量。我希望我们能每三个月做一次这个对话，因为它变化得太快了。你觉得接下来会发生什么？三个月后我回来，我们又在旧金山碰面，你预期会是什么样？

[41:50] Dylan Patel

Large-scale protests.Really?Yeah, I think there will bea large-scale protestagainst Anthropicand up at AI.People hate AI.AI is less popularthan ICE,less popularthan politicians.With Anthropicadding so much revenue,that's going to startcausing business changesdownstream.People are going to getmore and more scaredof AI.They'll start blamingmore and moreof their own problemsand things that are global,have been deep-seatedproblems for a long time.Those will bubble upand be blamed on AI.Probably some politicianor some influencerwill be able to starttaking and weaponizingAI against people.You look at the commentsof news articleswhere Sam Altmanhad a Molotov cocktailthrown at his housetwice in two weeks,people are cheering it on.And this is justthe beginning,so I think we'll seelarge-scale protestsagainst AIin three months.What is the counterweightto that?

大规模抗议。真的？是的，我认为会出现针对 Anthropic 和 OpenAI 的大规模抗议。人们讨厌 AI。AI 的受欢迎程度比 ICE（美国移民与海关执法局）还低，比政客还低。随着 Anthropic 增加这么多收入，它会开始在下游引发商业上的变化。人们会越来越怕 AI。他们会开始把越来越多自己的问题、以及那些全球性的、长期存在的深层问题，都算到 AI 头上。这些会浮上来，被归咎于 AI。大概率会有某个政客或某个网红能够把 AI 拿来当武器去对付人。你去看新闻文章下面的评论——Sam Altman 家两周内被扔了两次燃烧瓶，人们在下面叫好。而这只是开始。所以我认为三个月后我们会看到针对 AI 的大规模抗议。那对冲这件事的力量是什么？

[42:40] Dylan Patel

How should the AI industryhead that off?First of all,Sam Altman and Dariohave to stop gettingon interviews.They're so uncharismatic.I don't knowwhat they're doing.Every interview they dois like,wow, normal peopleare going to hate youeven more.Sam being on Tucker Carlsonprobably made all Republicanshate open AI.I'm just guessing.Same with Dario.I think that's first.Two,they need to startshowing uplifting thingsthat can be done with AI.Three,they need to stop talkingabout how the capabilitiesare going to changethe whole world constantlybecause then peopleare going to get fearof that capabilitybecause they have no connection.They don't know how to use it, yeah.There's no connectionto it either.The average persondoesn't knowan anthropic employee.The average persondoesn't knowan open AI employee.The average persondoesn't knowwho these people are,what their goals are,and they just view themas a sneaky cabalof 5,000 peopleat this companythat are going tochange the worldand automate all the jobsand destroy society.That's what they view it asand as peoplewho are fundingthe buildingof all these data centersand power plantsthat are going topollute the world.They don't quite understandwhat's happening.

AI 行业该怎么提前化解？第一，Sam Altman 和 Dario 必须别再上访谈了。他们太没有个人魅力了，我不知道他们在干嘛。他们做的每一场访谈都像是：哇，普通人只会更讨厌你。Sam 上 Tucker Carlson 的节目大概让所有共和党人都讨厌 OpenAI 了，我只是猜。Dario 也一样。我觉得这是第一点。第二，他们需要开始展示 AI 能做的那些鼓舞人心的事。第三，他们需要停止不停地讲这些能力将怎样改变整个世界，因为那样人们只会对那种能力产生恐惧——因为他们跟它没有连接。他们不知道怎么用它，对。跟它也没有任何连接。普通人不认识任何一个 Anthropic 的员工，普通人不认识任何一个 OpenAI 的员工。普通人不知道这些人是谁、他们的目标是什么，他们只把这些人看成一个鬼鬼祟祟的、5000 人的密谋集团，要去改变世界、自动化掉所有工作、摧毁社会。这就是他们的看法。而且在他们眼里，这些人还在出钱建那些会污染世界的数据中心和电厂。他们并不真正理解正在发生什么。

[43:38] Dylan Patel

So they have to stoptalking about the futurething that's going to happenand only talk aboutpresent how uplifting AI is.I think it's a hugereorg and rebrandingthat needs to be done.This is so much fun.I love doing this with you.Thanks for your time.Awesome, thanks.If you enjoyed this episode,visit Colossus.com.You'll find every episodeof this podcastcomplete withhand-edited transcripts.You can also subscribeto Colossus,our quarterly print,digital,and private audio publicationfeaturing in-depth profilesof the founders,investors,and companiesthat we admire most.Learn more atColossus.comslash subscribe.You know how smalladvantages compound over time?

所以他们必须停止谈论未来将要发生的那件事，只谈现在——AI 现在有多么鼓舞人心。我觉得这需要一次巨大的重组和品牌重塑。这太好玩了，我很喜欢跟你做这个。谢谢你的时间。太好了，谢谢。如果你喜欢这一期，请访问 Colossus.com。你会在那里找到本播客的每一期，都配有人工编辑的文字稿。你也可以订阅 Colossus——我们的季度印刷、数字与私人音频出版物，深度刻画我们最欣赏的创始人、投资人和公司。了解更多请访问 Colossus.com/subscribe。你知道小小的优势是怎么随时间复利的吗？

[44:33]

That's true in investingand just as truein how you run your company.Your spending systemis your capital allocation strategy.Ramp makes it smarterby default.Better data,better decisions,better economics over time.See how atramp.comslash invest.As your business grows,Vanta scales with you,automating complianceand giving you a singlesource of truthfor security and risk.Learn more atvanta.comslash invest.The best AIand software companiesfrom OpenAIto Cursorto Perplexityuse WorkOSto become enterprise-readyovernight,not in months.Visit workos.comto skip the unglamorousinfrastructure workand focus on your product.Ridgeline is redefiningasset management technologyas a true partner,not just a software vendor.They've helped firms5x in scale,enabling faster growth,smarter operations,and a competitive edge.Visit ridgelineapps.comto see what they can unlockfor your firm.Every investment firmis uniqueand generic AIdoesn't understandyour process.Rogo does.It's an AI platformbuilt specificallyfor Wall Street,connected to your data,understanding your process,and producing real outputs.Check them outat rogo.aislash invest.T shirt

在投资上是这样，在你怎么经营公司上同样如此。你的花钱系统就是你的资本配置策略。Ramp 让它默认变得更聪明：更好的数据、更好的决策、随时间更好的经济效益。详见 ramp.com/invest。随着你的业务增长，Vanta 会随你一起扩展，自动化合规、为安全与风险提供单一事实来源。详见 vanta.com/invest。从 OpenAI 到 Cursor 到 Perplexity，最好的 AI 与软件公司都用 WorkOS 一夜之间达到企业级就绪，而不是花几个月。访问 workos.com，跳过那些不体面的基础设施苦活，专注于你的产品。Ridgeline 正在重新定义资产管理技术——作为一个真正的合作伙伴，而不只是一个软件供应商。他们已帮助多家机构把规模做大 5 倍，实现更快的增长、更聪明的运营和竞争优势。访问 ridgelineapps.com，看看他们能为你的机构释放什么。每一家投资机构都是独特的，而通用 AI 并不理解你的流程。Rogo 理解。它是一个专为华尔街打造的 AI 平台，连接你的数据、理解你的流程、产出真实的成果。访问 rogo.ai/invest 了解。