章节 1：Nvidia 收购 Groq 与推理芯片的战略博弈

📝 本节摘要：

本章从 Nvidia 收购 Groq 这一热点事件切入。Dylan 指出，Nvidia 过去通过“通用 GPU”赢得了第一波 AI 浪潮，但在推理工作负载（Inference）日益庞大的今天，市场对“专用芯片”的需求正在激增（如专注于解码速度的 Groq 或专注于 KV Cache 的芯片）。

>

核心观点在于：Jensen Huang 极度信奉“只有偏执狂才能生存”。面对可能在成本和性能上击败 Nvidia 的单点解决方案（Point Solutions），Nvidia 选择通过收购（如 Groq）来获取稀缺的芯片工程人才，从而快速构建针对不同细分市场的专用产品线，以维持其高昂的利润率和霸主地位。

[原文] [Intro Clip]: this is the biggest change in human history maybe ever what's about to happen with AI this is the biggest revolution bigger than industrial revolution jensen is very paranoid about losing if he just kept making his mainline chip people crush him on cost and performance acquiring Grock is how you get those resources to make more solutions for different parts of the market to stay king at the end of the day this is an economic war if the US and the West win in AI China will not rise to be the global hedgeimony but without AI China definitely will rise they're just going to outrun America hi I'm Matt Turk welcome back to the Matt podcast today I'm joined by the one person Wall Street and Silicon Valley turn to when they need to cut through the hardware hype Dylan Patel of Semi analysis we dove into many of the most important topics of today nvidia's massive move to acquire Grock the truth about the capex bubble whether the US power grid can actually handle the AI boom and the geopolitical

[译文] [前言片段]: AI 即将带来也许是人类历史上最大的变革，这是一场比工业革命更宏大的革命。Jensen（黄仁勋）对于失败非常偏执（paranoid）。如果他只坚持制造主线芯片，人们会在成本和性能上击垮他。收购 Groq 是获取资源的方式，以便为市场的不同部分制造更多解决方案，从而保持王者地位。归根结底，这是一场经济战。如果美国和西方在 AI 领域获胜，中国将不会崛起成为全球霸主；但如果没有 AI，中国绝对会崛起，他们会直接超越美国。嗨，我是 Matt Turk，欢迎回到 Matt 播客。今天加入我的是华尔街和硅谷在需要看穿硬件炒作时都会求助的那个人——SemiAnalysis 的 Dylan Patel。我们深入探讨了当今许多最重要的话题：Nvidia 收购 Groq 的大动作、资本支出（Capex）泡沫的真相、美国电网是否真能承受 AI 繁荣，以及地缘政治。

[原文] [Intro Clip]: chess match playing out between the US and China but I have to warn you this conversation went off the rails in the best possible way and we ended up going into all sorts of fun tangents like the strange phenomenon of Chinese romance dramas set inside semiconductor factories and what's really like when three AI famous roommates live together in SF please enjoy this fantastic conversation with Dylan hey Dylan welcome

[译文] [前言片段]: 美中之间正在上演的棋局。但我必须警告你，这次谈话以最好的方式“跑题”了，我们最终聊到了各种有趣的题外话，比如以半导体工厂为背景的中国浪漫剧这一奇怪现象，以及三位 AI 界的知名室友在旧金山合租的真实生活。请享受这段与 Dylan 的精彩对话。嘿，Dylan，欢迎。

[原文] [Dylan Patel]: hello how are you

[译文] [Dylan Patel]: 哈喽，你好吗？

[原文] [Matt Turck]: i'm great i'd love to start with Grock and Nvidia since it's still fresh so not so long ago Nvidia was saying that uh one GPU could do it all and now they're doing this acquisition non-exclusive deal with Grock what does that mean from your perspective

[译文] [Matt Turck]: 我很好。我想从 Groq 和 Nvidia 开始聊，因为这事儿还很新鲜。不久前，Nvidia 还在说一个 GPU 就能搞定一切，而现在他们正在进行这项收购——与 Groq 的非独家交易。从你的角度来看，这意味着什么？

[原文] [Dylan Patel]: it's very clear we're not sure where AI models are headed in terms of you know over the next few years what happens to the architecture but you know the thing that I think everyone is sort of like agreed on is models are pretty auto reggressive right next token generation is like the thing but beyond that right attention mechanisms changed the how how it works everything changes right could could change

[译文] [Dylan Patel]: 很明显，我们并不确定 AI 模型在架构方面未来几年会走向何方。但我认为大家某种程度上达成共识的是，模型是非常自回归（auto-reggressive）的，对吧？“下一个 Token 生成”（next token generation）是核心。但除此之外，注意力机制（attention mechanisms）改变了它的工作方式，一切都在变化，对吧，一切都可能改变。

[原文] [Dylan Patel]: and so what's interesting is the reason Nvidia one is because they just took like the widest surface area bet and then people kept developing models on that and that kind of shape worked but now the workload is so large that there is room for specialization that will give you 10x increases in certain domains right

[译文] [Dylan Patel]: 有趣的是，Nvidia 之前之所以赢了，是因为他们押注了“最广泛的覆盖面”（widest surface area bet），然后人们不断在该基础上开发模型，这种形态奏效了。但现在工作负载如此之大，以至于有了“专用化”（specialization）的空间，这能在某些特定领域给你带来 10 倍的提升，对吧？

[原文] [Dylan Patel]: in a general purpose workload grock doesn't work right you know it can't train it can't you know it can't inference really really large models um cost efficiently right you can't serve many many many users but what it can do is it can go bl screamingly fast right same with the cerebrous open AI deal but that's like one workload right uh very decode focused right gener doing auto reggressive tokens in a in a single stream super fast

[译文] [Dylan Patel]: 在通用工作负载下，Groq 是行不通的，对吧。你知道它不能训练，也不能真正具有成本效益地推理（inference）非常非常大的模型，你无法服务海量的用户。但它能做的是快到尖叫（screamingly fast），对吧？Cerebras 与 OpenAI 的交易也是同理。但这就像是针对单一工作负载，非常专注于解码（decode focused），在单流中极速地进行自回归 Token 生成。

[原文] [Dylan Patel]: another direction AI models could head right we don't know are models going to think in one token stream or is it actually they're constantly context switching right and they're going from they have this humongous humongous context and they're generating in multiple parallel streams right and so Google and openi have both released mechanisms of this with their pro models where the model actually doesn't just have one single chain of thought for reasoning it has multiple right

[译文] [Dylan Patel]: AI 模型可能朝向的另一个方向——我们不知道模型是会以单一 Token 流进行思考，还是实际上它们在不断地进行上下文切换（context switching）？它们可能拥有巨大无比的上下文，并且在多个并行流中生成。Google 和 OpenAI 都在其 Pro 模型中发布了这种机制，模型实际上不只有一条单一的推理思维链（chain of thought），而是有多条，对吧？

[原文] [Dylan Patel]: and then I don't exactly like you know and and and how they choose which one and what the final answer to you delivers is is an area of research um but there there is room for that kind of chip right something that works on very parallel a lot lot of streams of chain of thought and maybe the latency requirements are not as crazy right maybe you don't want to go blindingly fast right maybe you're okay with it being you know because I can spin up 100 parallel you know streams of thought or agents or whatever you want to call them maybe I I care a lot about cost there and because it's 100 in parallel instead of one going super super fast it's not as deep right the tree search or the depth of the inference is not as deep but it is much wider

[译文] [Dylan Patel]: 至于它们如何选择哪一条路径以及最终给你提供什么答案，这是一个研究领域。但这种芯片是有生存空间的，即那种能处理高度并行、大量思维链流的芯片。也许（这种场景下）延迟要求没那么疯狂，对吧？也许你不需要快到瞎眼。也许因为我可以启动 100 个并行的思维流或智能体（Agents）——随你怎么称呼——我可能更在乎成本。而且因为是 100 个并行而不是一个极速运行，它的深度没那么深，树搜索或推理的深度没那么深，但它要宽得多。

[原文] [Dylan Patel]: you know there's other parts of inference hey process do creating the KV cache so Nvidia has a chip for that right that's the CPX so they they've made the CPX they bought Grock for decode and then they still have their general purpose GPU so they've they're kind of trying to cover their bases because unlike the first wave of AI chip companies where they sort of just made chips and then tried to figure out where it would work right they had a thesis Grock and Cerebrus both as well as Samanova right which was put a lot of memory on the chip and not necessarily in the case of Cerebrus and Grock no memory off chip and in the case of Samanova less memory offchip or slower memory offchip with higher capacity you know they they sort of all made similar bets in that direction and it didn't work for a while until it kind of did right um because there's a workload that now necessitates it

[译文] [Dylan Patel]: 推理还有其他部分，比如创建 KV Cache（键值缓存）。Nvidia 为此专门有一款芯片，那就是 CPX。所以他们制造了 CPX，他们买下 Groq 用于解码（Decode），然后他们仍然拥有通用 GPU。他们是在试图全面布局（cover their bases）。因为这不像第一波 AI 芯片公司，那时候大家只是造出芯片然后试图搞清楚它能用在哪。Groq、Cerebras 以及 SambaNova 当时都有一个论点，就是在芯片上放大量内存。Cerebras 和 Groq 是片外无内存（no memory off chip），SambaNova 是片外内存较少或较慢但容量更大。他们都在那个方向上做了类似的押注，这在很长一段时间内都没奏效，直到现在某种程度上奏效了，因为现在的某种工作负载需要它。

[原文] [Dylan Patel]: nvidia recognizes they're they're the leader they're at the tent pole hey in one respect they can just run faster than everyone but it's kind of hard to be 2x better than Google or or OpenAI or whoever else's internal chip right to justify their you know 75% plus margins right and then they have to be 2x to 4x better to justify 4x better to justify their margins because that's what they're charging above COGS you know the question is what what architecture will deliver that

[译文] [Dylan Patel]: Nvidia 意识到他们是领导者，他们处于顶梁柱的位置。从一方面看，他们可以跑得比谁都快，但想要比 Google 或 OpenAI 或其他任何人的内部自研芯片好上 2 倍是很难的，尤其是为了证明他们 75% 以上的利润率（margins）是合理的，对吧？他们甚至需要好上 2 到 4 倍，甚至 4 倍，才能证明其利润率的合理性，因为那是他们在销货成本（COGS）之上收取的溢价。问题在于，什么样的架构能实现这一点？

[原文] [Dylan Patel]: well yes keep the programmability of their GPUs is great for training and for a lot of workloads but you know guess what i think I think a lot of people will just be downloading an open source model downloading an inference framework and pressing go right a little bit more complicated than that but that's that's going to be the consumption method for a lot of enterprises a lot of uh startups a lot of tech companies is they're just going to do that or they're going to rent the G GPUs or or rent the chips and then download an open source framework and model and go right

[译文] [Dylan Patel]: 是的，保留 GPU 的可编程性对训练和许多工作负载来说都很棒。但你知道吗，我认为很多人只会下载一个开源模型，下载一个推理框架，然后点击“运行”。当然比这稍微复杂一点，但这将是许多企业、初创公司和科技公司的消费方式。他们就是会这样做，或者租用 GPU 或芯片，然后下载开源框架和模型并开始运行。

[原文] [Dylan Patel]: and Nvidia recognizes this and hey there is room for products that aren't general purpose right the general purpose GPU will still probably be the main line for training and for a lot of inference and for costefficient inference but maybe blindingly fast or workloads that have a ton of prefill i.e creating the the KV cache maybe that those workloads could be different chips right and the CPX chip they announced right they say it's for the context processing creating the KV cache it's also really useful for video models because video models don't care about memory bandwidth and so you know why pay for the expensive memory that the general purpose chip has or why do what Grock is doing which is tying hundreds or thousands of chips together and not having memory but keeping the entire model on chip

[译文] [Dylan Patel]: Nvidia 意识到了这一点，即“非通用”产品是有生存空间的。通用 GPU 可能仍将是训练、许多推理任务以及高成本效益推理的主线。但也许对于极速任务，或者有大量预填充（Prefill，即创建 KV Cache）的工作负载，这些可能需要不同的芯片。他们发布的 CPX 芯片，他们说是用于上下文处理（Context Processing）、创建 KV Cache 的。它对视频模型也非常有用，因为视频模型不太在乎内存带宽。所以，为什么要为通用芯片上昂贵的内存买单？或者为什么要像 Groq 那样把成百上千个芯片连在一起，没有片外内存，而是把整个模型放在芯片上？

[原文] [Dylan Patel]: the trade-off for that of course is you need thousands of chips and you have less compute per chip and so like Nvidia's trying to capture the whole surface area because again you don't know where models are headed and it's hard to say where the research is headed

[译文] [Dylan Patel]: 当然，那样做（Groq 模式）的权衡是你需要数千个芯片，而且每个芯片的算力较低。所以 Nvidia 正试图捕捉整个“覆盖面”（surface area），因为同样地，你不知道模型会走向何方，也很难说研究会走向何方。

[原文] [Matt Turck]: and do you think it's a good thing for the market yet another one of those deals that's structured as a as a license but really an acquisition i certainly think it's not good from an anti-competitive sense right i don't think people should just be able to buy companies without like any antitrust like process at all now in the case of like a large company buying a startup I'm completely fine with it the flip side is like hey we know the deal is happening right uh this happened for a company I was an adviser for Nvidia acquired in fabrica just maybe a few months before they did Grock and similar style of deal right if someone wanted to strike it down that's the biggest limbo right we've seen this happen in venture and you probably know more stories of this but like a company trying to get acquired they get stuck in limbo for like a year and then it falls apart stories yeah it falls apart the deal did because some regulatory BS and now the company was and the founders were focused on getting the deal done instead of like making the product better for a year and now they're like behind or you know they they they weren't focused on growth as much right you know you only have so much time as a founder so in that sense I like the license deals right

[译文] [Matt Turck]: 你认为这对市场是件好事吗？这又是一笔名义上是“授权”但实际上是“收购”的交易。

[Dylan Patel]: 我当然认为从反竞争的角度来看这并不好，对吧？我不认为人们应该能够完全不经过任何反垄断流程就买下公司。如果是大公司收购初创公司，我完全没意见。但这事的另一面是，嘿，我们知道交易正在发生。这种事发生在我担任顾问的一家公司身上——Nvidia 收购了 Infabrica，就在他们搞定 Groq 之前的几个月，也是类似的交易风格。如果有人想否决它，那就会陷入最大的僵局（limbo）。我们在风投圈见过这种情况，你可能知道更多这类故事：一家公司试图被收购，结果陷入了一年的僵局，然后交易黄了。是的，交易黄了，因为一些监管的破事（regulatory BS）。而现在这家公司和创始人在这一年里专注于搞定交易而不是优化产品，现在他们落后了，或者他们没那么专注于增长了。作为创始人你的时间是有限的。所以从这个意义上说，我喜欢这种“授权”式的交易。

[原文] [Matt Turck]: so now is uh Nvidia also dominating the the inference market is there any world where Nvidia is no longer the king or they seem to be getting stronger

[译文] [Matt Turck]: 那么现在 Nvidia 是不是也主宰了推理市场？有没有这种可能，即 Nvidia 不再是王者，还是说他们看起来越来越强了？

[原文] [Dylan Patel]: I think the thing about Nvidia is they take the Andy Grove mentality like more serious than anyone else right like okay fine Google like implemented OKRs because Intel did it but that's like you know management stuff right only the paranoid survive right this is like core to the Bay Area um core to Nvidia um Jensen is very paranoid about losing right these specializations if he just kept making his mainline chip would mean people could you know point point solutions for specific parts of the market would crush him on cost and performance then he can't justify his margin that's a threat to Nvidia's business model as a whole

[译文] [Dylan Patel]: 我认为 Nvidia 的特点是，他们比任何人都更认真地贯彻安迪·格鲁夫（Andy Grove）的心态。好吧，Google 实施 OKR 是因为 Intel 做了，但那是管理层面的东西。“只有偏执狂才能生存”（Only the paranoid survive），这就像是湾区的核心，也是 Nvidia 的核心。Jensen 对于失败非常偏执（paranoid）。如果他只坚持制造主线芯片，这些“专用化”趋势意味着人们可以用针对市场特定部分的“单点解决方案”（point solutions），在成本和性能上击垮他。那样他就无法证明他的高利润率是合理的，这对 Nvidia 的整个商业模式都是威胁。

[原文] [Dylan Patel]: especially if the best model only changes every 3 months or the model you want to roll out okay well then you're going to have three months to figure out how to make a model work on one chip architecture for that point solution and you know it's fine software software advantage of Nvidia is not that important then Jensen's super paranoid about losing and frankly it's really hard to hire enough talented chip people when you look across the market there is only a few companies who have successfully created a chip architecture software to run the models accurately run the run the models accurately right

[译文] [Dylan Patel]: 特别是如果最好的模型每 3 个月才变一次，或者你要推出的模型确定了，那你就有三个月的时间去研究如何让模型在那个针对特定点解决方案的芯片架构上运行。这其实没问题，那时 Nvidia 的软件优势就没那么重要了。Jensen 超级偏执于失败。坦率地说，很难招到足够多的天才芯片人才。环顾市场，只有寥寥几家公司成功创造了芯片架构和软件来准确地运行模型，是准确地运行模型，对吧？

[原文] [Dylan Patel]: like cuz you can look at random APIs of say an Alibaba Quen model and different people are doing all sorts of tricks like quantizing it but also many other tricks which then end up like making the model quality lower you know building a rack scale solution networking thousands of chips together and then deploying an API and Grock did the whole thing with frankly not that many people

[译文] [Dylan Patel]: 因为你可以看看比如阿里巴巴 Qwen 模型的各种随机 API，不同的人在搞各种花样，比如量化（quantizing），但也还有很多其他的花样，结果导致模型质量下降。你知道，构建一个机架级（rack scale）解决方案，把成千上万个芯片联网在一起，然后部署一个 API，Groq 用坦白说并不多的人手就做到了这一切。

[原文] [Dylan Patel]: so now it's like okay well I'm Nvidia I want to make four different chip architectures and actually four different point solutions maybe the general purpose and then one here one here one here and in addition my general purpose thing is actually not just like a GPU chip it's like GPU chips CPU chips networking chips NV switch nicks like you know there's many many chips and each of those chips has many chiplets you don't have enough engineering resources right and so like acquiring gro is like how you get those resources to make more solutions for different parts of the market

[译文] [Dylan Patel]: 所以现在情况是：好吧，我是 Nvidia，我想制造四种不同的芯片架构，实际上是四种不同的单点解决方案。也许是通用型，然后这儿一个，那儿一个，这儿再一个。而且我的通用型产品不仅仅是一个 GPU 芯片，它包括 GPU 芯片、CPU 芯片、网络芯片、NV Switch、网卡等等。你知道有非常多的芯片，每个芯片又有许多小芯片（chiplets）。你根本没有足够的工程资源，对吧？所以，收购 Groq 就像是你获取这些资源（人才）的方式，以便为市场的不同部分制造更多的解决方案。

紧接上一章关于 Nvidia 战略转型的讨论，本章 Dylan 首先简要点评了 Nvidia 面临的竞争威胁（初创公司与巨头自研芯片），随即深入探讨了核心话题：CUDA 护城河是否依然坚固？

Dylan 提出了一个反直觉的观点：传统的“CUDA 护城河”（即编写底层内核代码的能力）正在消失，因为大多数人不再手写 CUDA。取而代之的是，Nvidia 正通过解决推理成本（Inference Cost）和KV Cache 管理等新一代复杂工程问题，构建起一层更隐蔽、更难以逾越的“新软件护城河”。

章节 2：CUDA 护城河的演变：从内核代码到推理软件

📝 本节摘要：

本章首先触及了 Nvidia 面临的围剿——从 Etched 等初创公司到 Google、Amazon 等巨头的自研芯片。

>

随后对话转向核心议题：CUDA 还是无敌的吗？ Dylan 指出，随着开源推理引擎（如 VLM）的普及，开发者不再需要手写 CUDA 代码，这意味着传统的编程壁垒正在降低。然而，Nvidia 正在构建新的防线。

>

在 AI 代理（Agents）和代码生成（如 Cursor、Claude Code）等需要频繁上下文切换（Context Switching）的场景中，最大的瓶颈变成了KV Cache（键值缓存）的管理与预填充成本。Nvidia 正通过开发复杂的软件（如 KV Cache 管理器）将存储、内存与计算打通，这种系统级的优化能力构成了其在“后 CUDA 时代”的新护城河。

[原文] [Dylan Patel]: as far as like are they threatened like I think I think like obviously There's some cool startups out there right that are raising a lot right currently or have raised such as Etched Maddx uh positron these new age of AI companies there's also the prior age of like Cerebrris is is out there still right you know Tenstor etc and there's so there's a lot of AI chip companies on the startup side but then there's also you know Google's TPU AMD GPUs uh Amazon Tranium uh who are all really credible competitors and then you know Meta's MTIA is somewhat credible and then you know Microsoft Somaya is not credible but like you know maybe it will be one day right so you sort of have like a lot of competition they've got to hold the gates back and so I think is there a risk to them being I mean like there's there's risk from all of those companies that I mentioned and and you know effectively California/ Seattle right only two two places there's there's also chips from other parts of the world right obviously China has a number of different AI chip companies that are doing cool things anyone would have told you Grock was you know their business revenue their revenue was not like stellar right in fact they missed revenue last year significantly and yet they got bought right because the value of the IP was there and the value of the team anyone else would have been like well why the heck would I buy this right uh makes no sense there's definitely a credible threat yeah

[译文] [Dylan Patel]: 至于他们是否受到威胁，我认为，显然有一些很酷的初创公司正在融资或已经融了很多钱，比如 Etched、MatX、Positron，这些新时代的 AI 公司；还有像 Cerebras 这样的上一代公司依然存在，还有 Tenstor 等等。所以在初创公司方面有很多 AI 芯片公司。但同时还有 Google 的 TPU、AMD 的 GPU、Amazon 的 Trainium，这些都是非常可信的竞争对手。然后 Meta 的 MTIA 也算有点竞争力，至于 Microsoft 的 Maia 目前还不可信，但也可能有一天会变强，对吧？所以你面临着大量的竞争，他们必须守住大门。所以我认为他们有风险吗？也就是我提到的所有那些公司带来的风险，实际上主要集中在加州和西雅图这两个地方。当然世界其他地方也有芯片，显然中国有一些 AI 芯片公司在做很酷的事情。任何人都会告诉你 Groq 的业务收入其实并不亮眼，实际上他们去年严重未达营收预期，但他们还是被收购了，因为 IP（知识产权）的价值在那里，团队的价值在那里。换做别人可能会问：“我到底为什么要买这个？”这看起来没道理。但确实存在可信的威胁。

[原文] [Matt Turck]: and do you think uh CUDA is going to remain that mode I guess a combination of CUDA and whatever came out of the Melanox acquisition like do do those persist as long-lasting advantages

[译文] [Matt Turck]: 那么你认为 CUDA 还会继续成为那个护城河（Moat）吗？我猜是 CUDA 加上收购 Mellanox 后带来的那些技术的结合，这些会作为长期的优势持续存在吗？

[原文] [Dylan Patel]: I think they do I think networking is super important I think uh the CUDA software mode is very important but it's also like changing rapidly right It's an incredible amount of the software that Nvidia GPUs run on is not from Nvidia it's it's the developer ecosystem that's open sourcing it when you look at for example VLM and SGLANG right these support AMD GPUs almost as first class citizens now and VLM is getting significant support for TPUs for tranium and there will be other chips coming out from startups that also support VLMs SGLang now like how difficult is it you know the the the reason why CUDA is so important is like okay I can do whatever I need to do right programming a GPU

[译文] [Dylan Patel]: 我认为会。我觉得网络连接（Networking）超级重要，CUDA 软件护城河也非常重要，但它也在迅速变化，对吧？Nvidia GPU 上运行的软件有惊人的数量并非来自 Nvidia，而是来自开源的开发者生态系统。例如你看 vLLM 和 SGLang，它们现在支持 AMD GPU 几乎就像支持一等公民一样，而且 vLLM 也正在获得对 TPU 和 Trainium 的重要支持，未来初创公司推出的其他芯片也会支持 vLLM 和 SGLang。现在的问题是这有多难？你知道，CUDA 之所以如此重要，是因为“好吧，我可以做任何我需要做的事，我可以给 GPU 编程”。

[原文] [Dylan Patel]: i think most AI chips will not be consumed by people programming anything for it they will download an open source inference engine and they will download an open source model and then they will put it on the and it's really simple to download VLM and like make it work like it's not that hard to set up uh you know a server and Nvidia's putting out a lot of open source software like Triton inference server and and uh Dynamo and all these things to to make it easy because that is the consumption model ultimately for the majority of AI right is and it might be like oh it's my own inference engine but most servers will not run code besides the inference engine and the model it's like not like people are actually like researchers are like writing code for GPUs to see ideas if they'll work and train models and all these things or just mess around with them to figure out you know infra performance or whatever it is but most of it won't be there

[译文] [Dylan Patel]: 我认为大多数 AI 芯片的消费方式，并不是人们去为它编写任何程序。他们会下载一个开源推理引擎，下载一个开源模型，然后把它跑起来。下载 vLLM 并让它运行真的很简单，设置服务器没那么难。Nvidia 也推出了很多开源软件，比如 Triton 推理服务器、Dynamo 等等，来让这一切变得简单。因为这最终将是大多数 AI 的消费模式。虽然可能有人会说“哦，这是我自研的推理引擎”，但大多数服务器除了推理引擎和模型之外不会运行别的代码。这不像研究人员为了验证想法、训练模型或搞清楚基础设施性能而真的在为 GPU 写代码。绝大多数情况都不是那样。

[原文] [Dylan Patel]: and so CUDA as a mode CUDA language is like you know like it's like fine right like you know no one actually writes CUDA right like most people write PyTorch and then like torch compile and then they just run it on the GPU they don't write CUDA but a lot of this CUDA mode is like how does PyTorch translate into high performance GPUs and that surface area from when people were just writing like hardcore when people are hardcore writing CUDA kernels to like hey they're writing PyTorch and then it's compiling down to GPUs versus oh I'm just downloading VLM is it is a it is a curve of like not a ton of people that can do CUDA kernels a whole lot more people can do PyTorch right random you know PhDs and random people it's very simple right a crapload of people can do VLM download it run it on a server well if it now supports other chips what is the CUDA mode's recognized this and they've been building software that is not necessarily the CUDA remote and I I can give some examples

[译文] [Dylan Patel]: 所以，CUDA 作为一种护城河，或者说 CUDA 语言，你知道，它还行。但没人真的去写 CUDA 了，对吧？大多数人写 PyTorch，然后用 Torch Compile，接着在 GPU 上运行。他们不写 CUDA。但这个“CUDA 护城河”的很大一部分在于 PyTorch 如何转化为 GPU 上的高性能表现。这个领域已经发生变化：从人们硬核地手写 CUDA 内核，变成他们写 PyTorch 然后编译到 GPU，再变成“哦，我只是下载 vLLM”。这是一条曲线：能写 CUDA 内核的人不多；能写 PyTorch 的人多得多（随便哪个博士或普通人都能写，很简单）；而能做“下载 vLLM 并在服务器上运行”的人简直多如牛毛。如果 vLLM 现在支持其他芯片了，那 CUDA 护城河还剩什么？Nvidia 意识到了这一点，他们一直在构建不一定是传统 CUDA 护城河的软件，我可以举一些例子。

[原文] [Dylan Patel]: All right so the name of the game is fast tokens and lowest cost tokens right and lowest cost tokens happens by your chip being fast but there's also tricks right one example right like I mentioned with you know the CPX versus Grock right is processing your prefill contacts right super cheap CPX right if I'm if I'm care a lot about speed then Grock these are optimizations on the hardware side there's optimizations on the software side as well right

[译文] [Dylan Patel]: 好的，现在游戏的核心是“快速 Token”和“最低成本 Token”。最低成本可以通过芯片速度快来实现，但也有技巧。一个例子，就像我提到的 CPX 对比 Groq，处理预填充（Prefill）上下文，CPX 超级便宜；如果我非常在乎速度，那就选 Groq。这些是硬件层面的优化，但在软件层面也有优化。

[原文] [Dylan Patel]: and so one example is when I'm doing for example if I look at a cloud code or a cursor type application right the workload is like it takes your repo takes the relevant parts of your repo puts it in the context of the LLM it prompts it generates right and if it's an agent mode it it it circulates the context a couple times it'll collapse put things off to the side access different contexts but what's you know especially when you think about an agent for software and you can see this in codeex you know Codex Codex actually not as good as cloud code but it can do work on time horizons of like 9 10 hours um and do like a big refactor better than cloud code can even though most of the times cloud code is better

[译文] [Dylan Patel]: 举个例子，比如我看像 Cloud Code（可能是指 Claude Code 或类似工具）或 Cursor 类型的应用程序。它的工作负载是：抓取你的代码仓库，提取相关部分，放入大模型（LLM）的上下文中，提示它，然后生成。如果是代理（Agent）模式，它会循环处理上下文几次，折叠内容，把东西放一边，访问不同的上下文。特别是当你想到软件代理时——你可以在 Codex 中看到这一点。Codex 实际上不如 Claude Code 好，但它能在 9 到 10 小时的时间跨度上工作，做一个大的重构，甚至比 Claude Code 做得更好，尽管大多数时候 Claude Code 更强。

[原文] [Dylan Patel]: and and what's interesting about Codeex does is it'll like take your repo it'll identify parts if you're asking it to refactor it identify parts write stuff you know make like these notes for itself everywhere collapse the context switch from this part of the repo to that part of the repo to this part of the repo but when you think about it it's like oh if this thing is just generating tokens all the time plus it's switching what my context is constantly that's really expensive right if you think about like what's the cost of inference um I want to say it's like it's it's $10 per million tokens of output and or and $3 for decode or 10 for decode and three for prefill

[译文] [Dylan Patel]: Codex 有趣的地方在于，它会获取你的代码库，识别各个部分。如果你让它重构，它会识别部分、写入内容、到处给自己做笔记、折叠上下文，从代码库的这部分切换到那部分，再到这部分。但你想想，如果这东西一直在生成 Token，加上它不断地切换上下文，那是真的很昂贵的，对吧？如果你考虑推理的成本——我想说大概是每百万输出 Token 10 美元，解码（Decode）是 10 美元，预填充（Prefill）是 3 美元。

[原文] [Dylan Patel]: uh and so if you think about oh it just worked for nine hours on one task one refactor huge value but if it changed context a ton of times and your context is like 30k usually or 50k or you know heading to hundreds of thousands you know how long your how big your repository is and how much context switch now you're spending all this money on on prefill right not the decode tokens but actually why am I like regenerating the KV cache I can actually just like store the KV cache elsewhere and then when I need it again I can pull it and and plop it into CPU memory or into GPU memory

[译文] [Dylan Patel]: 所以如果你想，哦，它在一个任务、一个重构上工作了 9 个小时，价值巨大。但如果它切换了无数次上下文，而你的上下文通常是 30k 或 50k，甚至迈向几十万——你知道你的代码库有多大，上下文切换有多少——那你现在把钱全花在预填充（Prefill）上了，而不是解码 Token。实际上，我为什么要重新生成 KV Cache（键值缓存）呢？我其实可以把 KV Cache 存储在别处，当通过我再次需要它时，把它拉出来，直接丢进 CPU 内存或 GPU 显存里。

[原文] [Dylan Patel]: and so Nvidia's got this like KV cache manager and they've been working really hard on like making it so they can interface SSDs and stick the KV cache on there and pull it out whenever they want so for this kind of workload and then if you do this and you look at like coding as an application and you like look at these coding companies and how much they're paying for prefill versus decode actually majority of their cost is pre-fill tokens not decode tokens because their context is just so large and it's switching all the time even in agent modes you know if you can now not have to do the pre-fill your costs go down dramatically but that's a very complicated thing to do from a software perspective

[译文] [Dylan Patel]: 所以 Nvidia 有了这个“KV Cache 管理器”，他们一直在努力开发，让它可以连接 SSD，把 KV Cache 存放在那里，随时取用。对于这种工作负载，比如编程应用，你看看这些编程公司在预填充和解码上花了多少钱，实际上他们的大部分成本是预填充 Token，而不是解码 Token。因为他们的上下文太大了，而且一直在切换，即使在代理模式下。如果你现在可以不必做预填充，你的成本会急剧下降。但从软件角度来看，这是一件非常复杂的事情。

[原文] [Dylan Patel]: you know companies like Enthropic Google OpenAI have already done it but what about the wide world right and so Nvidia is trying to make the open source software for this and that's like CUDA mode but it's like actually no none of this is CUDA right like it's like memory management and like you know storage management and when do you call what and how do you transfer it and how do you like spread the KV cache across a bunch of different storage nodes and what happens when you read it and the network congestion just like all these things yeah it's like Nvidia's wheelhouse but it's not CUDA and I think like the easy way to say it is it is the CUDA mode right and so things like this KV cache manager and many other things they're trying to do to reduce the cost of inference like is how they build the new CUDA mode because again today it's it's you know it is quite I mean AMD is like not fully there yet and TPU is being added right now and tranium is being added soon as well to VLM but all of them will have a very good UX for download model run model on VLM by the middle of the year I think right certainly AMD is already there

[译文] [Dylan Patel]: 你知道，像 Anthropic、Google、OpenAI 这样的公司已经做到了这一点，但广大的外部世界呢？所以 Nvidia 正试图为此制作开源软件。这就像是“CUDA 护城河”，但实际上这些都不是 CUDA，对吧？这更像是内存管理、存储管理、何时调用什么、如何传输、如何将 KV Cache 分散在一堆不同的存储节点上、读取时会发生什么、网络拥塞等等。是的，这是 Nvidia 的强项，但这不是 CUDA。我觉得简单的说法是“这就是 CUDA 护城河”。所以像这个 KV Cache 管理器以及他们试图做的许多其他降低推理成本的事情，就是他们构建新 CUDA 护城河的方式。因为再说一次，今天 AMD 还完全没到位，TPU 正在被添加进 vLLM，Trainium 也快了，但到今年年中，所有这些芯片在 vLLM 上“下载模型、运行模型”的用户体验（UX）都会变得非常好。当然 AMD 现在已经做到了。

[原文] [Dylan Patel]: by the end of this quarter we have something that like tests this right it's called inferencemaxa it's open source all the code is and the results are uh but we run across I think $60 million of GPUs which are donated to us by companies like Nvidia AMD openai Microsoft Amazon on Crusoe Core Weave Together AI uh all these companies are sponsoring GPUs for us to run this we're running VLM and SDG Lang every night on you know nine different kinds of GPUs on a variety of different models and different work uh context lens and all these things right to see the performance and you can see the performance moving every day or pretty often because the software changes all the time and so like the fact that this exists is the cuda boat right it's not that like AMD you can do this on their chips Nvidia can do this on their chips it's oh when the new model comes out how fast does it get to peak performance because you know it's it's a moving target or hey can I implement this KV cache management thing how hard is it how many engineers do I need oh just one great like or 10 great if I need a hundred people to develop it like Google and you know so on and so forth did then that's much harder

[译文] [Dylan Patel]: 到本季度末，我们有一个东西可以测试这个，它叫 InferenceMaxa（注：可能是 Inference Benchmark 相关工具），它是开源的，所有代码和结果都是。我们在大约价值 6000 万美元的 GPU 上运行测试，这些 GPU 是由 Nvidia、AMD、OpenAI、Microsoft、Amazon、Crusoe、CoreWeave、Together AI 等公司捐赠给我们的。所有这些公司都赞助 GPU 让我们运行测试。我们每天晚上在九种不同的 GPU 上运行 vLLM 和 SGLang，测试各种不同的模型、不同的上下文长度等等，以此查看性能。你可以看到性能每天或经常在变化，因为软件一直在变。所以，这类工具存在的事实本身就是“CUDA 护城河”（CUDA Moat）。问题不在于 AMD 或 Nvidia 的芯片能不能做这件事，而在于：当新模型出来时，它多快能达到峰值性能？因为这是一个移动的目标。或者，嘿，我能实现这个 KV Cache 管理功能吗？这有多难？我需要多少工程师？哦，只要一个？太棒了。或者十个？也行。如果我像 Google 那样需要一百个人来开发它，那就难多了。

本章 Dylan 从软件护城河转向硬件竞争格局。他犀利地点评了 AMD 的追赶策略，并深入分析了初创芯片公司（如 Etched、Cerebras）在 Nvidia 阴影下的唯一生存路径——“极端专业化”。他通过对比 Midjourney（图像生成）与 Claude（代码生成）的不同硬件需求，揭示了未来芯片市场可能分裂为“算力导向”与“内存带宽导向”的技术逻辑。

章节 3：芯片竞争格局：AMD、初创公司与专业化路径

📝 本节摘要：

在这一章中，Dylan 首先确认了 AMD 是 Nvidia 唯一的有力竞争者，但指出 AMD 处于“硬件偶尔追平，软件永远落后”的循环中。

>

随后话题转向 AI 芯片初创公司。Dylan 认为，试图在通用芯片领域击败 Nvidia 是徒劳的，因为 Nvidia 拥有无法比拟的供应链和技术封锁。初创公司的唯一机会在于“单点解决方案”（Point Solutions）。他以具体的技术参数解释了这点：图像/视频生成（如 Midjourney）渴望纯算力，而长文本推理（如编码代理）则极度依赖内存带宽。

>

尽管成功率极低（不到 1%），但如果模型架构稳定下来，针对特定工作负载（如推荐系统或特定模态）的专用芯片将有一席之地。

[原文] [Matt Turck]: do you think AMD can uh catch up

[译文] [Matt Turck]: 你认为 AMD 能追上吗？

[原文] [Dylan Patel]: I think AMD will be caught up at times and very behind at other times like currently they're super far behind right because Blackwell is just way better than MI355 um and then you know Rubin comes out and they'll be way way behind but then AMD's new chip comes out and AMD will be caught up or evenlight ly ahead on a hardware perspective software's behind right and you have this like leaprogging and and AMD is a very credible second competitor i don't think they'll go beyond like I think they'll stay in single digits market share single digit percentage market share single digit percentage market share is still pretty good yeah i mean Nvidia's revenue this year is going to be like it's a lot the three gajillion dollars i think it's actually four gajillion

[译文] [Dylan Patel]: 我认为 AMD 有时会追上，有时会落后很多。比如目前他们就超级落后，因为 Blackwell 比 MI355 强太多了。然后你知道 Rubin（Nvidia 下一代芯片）出来时，他们又会落后非常非常多。但随后 AMD 的新芯片出来，他们又会追平，甚至在硬件角度上略微领先。但软件还是落后的，对吧？所以你会看到这种交替领先（leapfrogging）。AMD 是一个非常可信的第二竞争者。我不认为他们能突破……我认为他们的市场份额会保持在个位数百分比。不过个位数百分比的市场份额其实也相当不错了。是的，我是说 Nvidia 今年的收入会是……很多，大概“三亿万万”（three gajillion）美元吧。我想实际上可能是“四亿万万”。

[原文] [Matt Turck]: what about all the startups you mentioned a few so there's a cerebrus on the one end of the spectrum and then newer ones edged and and others if if AMD has a you know uphill battle in front of them like do you think those guys can take significant market share

[译文] [Matt Turck]: 那所有的初创公司呢？你提到了一些，一边是 Cerebras，然后是更新的 Etched 等等。如果连 AMD 面临的都是一场艰难的爬坡战，你认为这些家伙能抢占显著的市场份额吗？

[原文] [Dylan Patel]: you sort of the whole specialization game right you you have to specialize because you're never going to beat Nvidia at their own game right they're going to have the supply chain unlock they're going to get to the newest memory technology or process technology or whatever packaging technology whatever it is sooner than you and they're just going to crush you right if you play their game you have to AMD is trying to play Nvidia's game but AMD is like extremely good at engineering silicon right everyone else has to has to has to try something weird or different right

[译文] [Dylan Patel]: 这就是整个“专业化”的游戏，对吧。你必须专业化，因为你永远无法在 Nvidia 擅长的领域击败他们。他们已经解锁了供应链，他们会比你更早获得最新的内存技术、制程技术或者无论什么封装技术。如果你玩他们的游戏，他们会直接碾压你。AMD 正在试图玩 Nvidia 的游戏，但 AMD 在硅片工程方面极其出色。其他人则必须、必须、必须尝试一些奇怪或不同的东西。

[原文] [Dylan Patel]: and so when you look at Etched or Maddx or Posatron or Cerebrris or Tenstor you go to look at all these companies right there are unique things about what they're doing and it's not clear if AI models will still be within that realm when that comes out right uh does oh now people use like engrams and other sparse attention techniques is that like is does that change like some of the specializations people are doing or hey people are now doing like you know models are now sparse instead of being dense models does that change things there's so many optimizations and changes on the model side and you can't predict what's going to happen with the ML research easily at least you can't

[译文] [Dylan Patel]: 所以当你看到 Etched、MatX、Positron、Cerebras 或 Tenstor 时，你看这些公司，他们正在做的事情都有独特之处。但不清楚的是，当芯片做出来时，AI 模型是否还处于那个范畴内，对吧？比如，噢，现在人们使用 N-grams 和其他稀疏注意力（sparse attention）技术，这是否改变了人们正在做的某些专业化方向？或者嘿，现在的模型是稀疏的（sparse）而不是稠密的（dense），这会改变事情吗？模型端有太多的优化和变化，你很难轻易预测机器学习研究会发生什么，至少你做不到。

[原文] [Dylan Patel]: the thing you're optimizing for today has to be a vision of where AI will be in 2 years and Nvidia's fully accepted they don't know where that's going to be that's why they have a portfolio of chips now not just one GPU line right it's not just Hopper Blackwell Reuben now it's going to be you know it's not Ampure Hopper you know you know it's not that line it's like there's a variety of chips to serve the different markets um and different possible scenarios they think each of them has this vision today but oh it might turn out the general purpose one sucks and and actually AI models have developed in a way where CPX or Grock style chips are the best right well okay now we have a solution for that market

[译文] [Dylan Patel]: 你今天优化的东西必须是基于对 2 年后 AI 发展方向的愿景。而 Nvidia 完全接受了“他们不知道未来会怎样”这一事实，这就是为什么他们现在拥有一个芯片组合（portfolio），而不仅仅是一条 GPU 产品线。不仅仅是 Hopper、Blackwell、Rubin，现在会有各种各样的芯片来服务不同的市场和不同的可能场景。他们认为今天每种芯片都有其愿景，但也可能结果是通用芯片很烂，而 AI 实际上发展成了 CPX 或 Groq 风格的芯片才是最好的，对吧？好吧，那通过这种布局，我们现在针对那个市场也有解决方案了。

[原文] [Dylan Patel]: and so I think that's the challenge with the startups with that said I think they're all taking very interesting bets i think it's I think it's much more exciting than the first wave of AI hardware uh bets graph course rebringing the memory on the chip they sort of just made a bet and they optimized for a certain kind of model all similar kinds of model and it didn't end up working out for a long time right they had to pivot and they had to work on a lot of things and it took a long time I think these companies have like a really clear vision of what they think models will look like right like Etch does Maddx does Posatron does and that's what's really cool about it between the three of them uh these new age so I mean I'm I'm excited for them i'm very very skeptical i don't know what uh what a venture capitalist views as likely chances of succeeding but I think all of them are less than 1% right but you know that's that's that's a but the world where they win is a multi-silicon kind of world where any given customer uses a range of different GPUs it could it could or it could be any given customer has like one workload they care a lot about

[译文] [Dylan Patel]: 所以我认为这就是初创公司面临的挑战。话虽如此，我认为他们都在下非常有趣的赌注。我觉得这比第一波 AI 硬件赌注要令人兴奋得多——比如 Graphcore 那时候把内存放回芯片上，他们某种程度上只是下了一个注，针对某一种特定类型的模型进行了优化，结果很长一段时间都没奏效，对吧？他们不得不转型，不得不做很多事情，花了很多时间。我认为现在这些公司对自己认为的模型未来形态有着非常清晰的愿景，比如 Etched、MatX、Positron 都是如此。这就是他们这三家新时代公司很酷的地方。我是说，我为他们感到兴奋，但我同时也非常非常怀疑。我不知道风险投资人认为的成功概率是多少，但我认为他们所有的成功率都不到 1%。但是，如果他们赢了，那个世界将是一个“多硅片”（multi-silicon）的世界，任何给定的客户都会使用一系列不同的 GPU。或者可能是任何给定的客户只非常在乎某一类特定的工作负载。

[原文] [Dylan Patel]: anthropic clearly does not give a crap about videogen image gen right they just don't care um on the flip side company like midjourney cares a lot about image and videogen right image and videogen is very very like like I mentioned like it's a very like it's not very memory bandwidth heavy it loves loves loves compute right whereas inference of large language models in the style of like you know this these you know say for example coding agents cares a lot about decoding for long streams of time um and that's very memory bandwidth heavy right

[译文] [Dylan Patel]: Anthropic 显然完全不在乎视频生成或图像生成，他们就是不在乎。但在另一面，像 Midjourney 这样的公司非常在乎图像和视频生成。图像和视频生成，就像我之前提到的，它对内存带宽（memory bandwidth）的要求并不高，它非常非常非常喜欢算力（compute）。相反，大语言模型的推理，比如像编码代理（coding agents）这种风格，非常在乎长时间流的解码（decoding），而那是极度依赖内存带宽的。

[原文] [Dylan Patel]: and so there's like that's like a simple example but there's a lot more nuance there in terms of like even like the size of like the matrix multiply you know the tensor cores that you you know the systolic arrays that you use or the ratios of networking and memory memory and like what's that memory hierarchy look like and you know what are you doing for different kinds of attention and like oh like all these sorts of things like there's a lot of specialization here and so some people are betting big on on different types of specialization

[译文] [Dylan Patel]: 这是一个简单的例子，但这其中还有更多的细微差别。比如矩阵乘法的大小、你使用的张量核心（Tensor Cores）、脉动阵列（Systolic Arrays），或者网络与内存的比例、内存层级结构是什么样的，以及你针对不同类型的注意力机制做了什么等等。这里有太多的专业化空间，所以有些人正在押注不同类型的专业化。

[原文] [Dylan Patel]: and I I think like you could clearly see a world where companies do care about different stuff right like like if for example a chip optimized for video and image generation existed today and it was better than Nvidia or Nvidia made it i think Midjourney would absolutely only use that for inference i think for training they'd still use the general purpose thing and as would like Meta and Google would like they should do that right

[译文] [Dylan Patel]: 我认为你可以清楚地看到一个公司关注不同东西的世界。比如，如果今天存在一款专门针对视频和图像生成优化的芯片，而且它比 Nvidia 的好，或者是 Nvidia 制造的，我认为 Midjourney 绝对只会用它来做推理。至于训练，我认为他们还是会用通用芯片。Meta 和 Google 也是如此，他们应该这样做。

[原文] [Dylan Patel]: and hey Meta actually has two lines of AI chips there mtia there's a line that's focused on recommendation systems and then there's a line that's focused on Gen AI the GI one is a new line but that recommendation systems ch line is still continuing right it's not sexy no one cares because there's no and bite dance also has a recommendation system line of chips and it's not really focused on Jedi which is fine because you know this is a $200 billion business or something which is just deciding what ad to serve me right and what order to put my friends stories and you know things like this so so I think like it's perfectly fine for there to be specialized AI chips given the target market is big enough and you have to have vision to know what that target market is unless you're hyperscaler then you can like just like you can just use general purpose until you've like it's clearly there and then you can make your asich right

[译文] [Dylan Patel]: 实际上，Meta 有两条 AI 芯片产品线——MTIA。一条线专注于推荐系统，另一条线专注于生成式 AI（Gen AI）。Gen AI 那条线是新的，但推荐系统那条芯片线仍在继续，对吧？它不性感，没人关心，因为没有热度。字节跳动也有一条推荐系统的芯片线，并不是真的专注于 Gen AI。但这没关系，因为这是一个 2000 亿美元的生意，仅仅是决定给我推送什么广告，或者按什么顺序排列我朋友的故事，诸如此类。所以我认为，只要目标市场足够大，存在专用的 AI 芯片是完全没问题的。你必须有眼光去识别那个目标市场。除非你是超大规模厂商（Hyperscaler），那你就可以先用通用芯片，直到需求非常明确了，然后再制造你自己的 ASIC。

本章进入了宏大的地缘政治话题。Dylan 用非常生动（甚至有些滑稽）的例子描绘了中国举国上下的“半导体热潮”——从以晶圆厂为背景的偶像剧，到专注于生产灯罩或吉他的“专业化城市”。

同时，他严肃地指出了美国制裁下的漏洞（如 ByteDance 在马来西亚租用算力），并强调了 Huawei 作为“世界上最垂直整合公司”的恐怖实力，认为它是西方不可小觑的对手。

章节 4：地缘政治：中国的“芯片热”文化与华为的崛起

📝 本节摘要：

在本章中，对话转向中美芯片战。Dylan 指出，尽管 Nvidia 在中国的销售受限，但中国公司（如字节跳动）正通过在马来西亚等地租用算力来绕过限制。

>

Dylan 观察到一个独特的文化现象：中国已经“全员半导体化”（Semiconductor Pilled）。这种狂热甚至渗透到了娱乐业，出现了以半导体工厂为背景的浪漫偶像剧。他详细描述了中国供应链的极度专业化（如专门生产灯罩的城市），并将这种集群效应与半导体产业类比。

>

最后，他高度评价了华为，称其为“最可怕”的对手，因为华为拥有极度垂直整合的能力，甚至在折叠屏手机等领域击败了三星和苹果。他警告说，这是一场经济战，如果美国输掉 AI 竞赛，中国将凭借其制造和执行能力超越美国。

[原文] [Matt Turck]: fascinating turning to the geopolitical aspect of of uh all of this which is always fun huawei and Nvidia in China last year that was like 10 or 12% of their overall revenue and this year they they were saying that their market share but has basically dropped to not very much is that Huawei chips is that restrictions is that tariffs uh what's happening

[译文] [Matt Turck]: 很有趣，让我们转向这一切的地缘政治方面，这总是很有意思。关于 Huawei 和 Nvidia 在中国的情况，去年中国市场大概占 Nvidia 总收入的 10% 或 12%，而今年他们说市场份额基本上跌得所剩无几了。这是因为 Huawei 的芯片吗？是限制措施？还是关税？到底发生了什么？

[原文] [Dylan Patel]: it's a variety of things actually in in some in some quarters last year uh it was even north of 20 I think but I don't remember exactly but anyways you know if you look at 2022 China was almost the size of the US in terms of buying server hardware right almost not quite but getting there um and it looked like they were going to be the same size as America in like a year or two after that right

[译文] [Dylan Patel]: 实际上是多种原因。去年某些季度，我认为这个比例甚至超过了 20%，但我记不太清了。总之，你看 2022 年，中国在购买服务器硬件的规模上几乎与美国相当，对吧？虽然还没完全赶上，但已经很接近了。而且看起来一两年后他们就要达到和美国一样的规模了。

[原文] [Dylan Patel]: and so there are restrictions Nvidia's been handcapped hand handicapped from selling their best chips to China and so that's obviously impacted the sales a lot because like why would you do that and so when you look at who rents the most GPUs in the world it's three companies right so one of them is obviously OpenAI second one actually they were bigger than OpenAI they are bigger than OpenAI today or no they were bigger than OpenAI than OpenI eclipsed them recently is Bite Dance

[译文] [Dylan Patel]: 确实有限制措施，Nvidia 被束缚住了，无法向中国出售他们最好的芯片，这显然极大地影响了销售，因为（如果不能买最好的）你为什么要买呢？当你看看谁是世界上租用 GPU 最多的公司时，是三家公司。其中一家显然是 OpenAI。第二家，实际上他们曾比 OpenAI 还大，或者说今天依然很大……不，他们曾比 OpenAI 大，后来 OpenAI 超越了他们，这家公司就是字节跳动（ByteDance）。

[原文] [Dylan Patel]: bite Dance runs rents tons of chips from Oracle and Google and and you know many other cloud companies because they couldn't get the chips they need in in China they're mostly just serving Tik Tok right okay well they they're not allowed to buy them and that sucks but you know they're they're allowed to rent them and so okay if I'm not allowed to get the best ones I'm going to rent externally and if Bite Dance is the second biggest renter of GPUs in the world that's substituting demand that would have been built in China in many cases it's instead being built in Malaysia and Oracle has over a gigawatt of capacity in Malaysia that Bite Dance is going to take right so things like this are you know you know hundreds of thousands if not millions of chips tens of billions of dollars of cap capacity that would go to China but it's not that it's going to Malaysia instead as an example

[译文] [Dylan Patel]: 字节跳动从 Oracle、Google 以及许多其他云公司那里租用了大量的芯片，因为他们在中国无法获得所需的芯片。他们主要是在为 TikTok 提供服务，对吧？好吧，他们不被允许购买芯片，这很糟糕，但他们被允许租用芯片。所以，既然我不被允许拥有最好的芯片，我就去外部租用。如果字节跳动是全球第二大 GPU 租户，这就替代了原本会在中国建设的需求，这种需求现在很多都转移到了马来西亚。Oracle 在马来西亚有超过 1 吉瓦（GW）的容量，字节跳动打算吃下这些。所以像这样的事情，涉及到数十万甚至上百万个芯片，数百亿美元的产能，原本会流向中国，但现在流向了马来西亚。

[原文] [Dylan Patel]: another sort of point around this is China's like you know they've had these 5-year plans so and and you know the way these initiatives work from China is there is like some top down ordering but then they just kind of whip the whole like everyone just kind of gets into it and it's really cool like I don't think it's as top down as many people think like I think the entire country is like semiconductor pill right there are dramas where people fall in love in the fab or dramas where people fall in love and they're photovoltaic like solar cell researchers and engineers and it's like it's like this is just the backdrop and it's like actually this is it's like super cool for your like significant other to be that semiconductor engineer or to be that photovoltaic you know uh solar panel researcher as opposed to an influencer as opposed to an influencer right

[译文] [Dylan Patel]: 另一点是，你知道中国有这些“五年计划”。这些倡议在中国的运作方式是，虽然有一些自上而下的指令，但随后他们会动员所有人，大家都投入其中，这真的很酷。我不认为这完全像很多人想的那样只是自上而下的，我觉得整个国家都已经“全员半导体化”（Semiconductor Pilled）了。那里甚至有电视剧，讲人们在晶圆厂（Fab）里谈恋爱，或者讲光伏太阳能研究员和工程师谈恋爱。这就像是背景设定，而且这就好像……你的另一半如果是半导体工程师或者光伏研究员，那是超级酷的事情，而不是什么网红（Influencer）。是的，比起网红要酷多了。

[原文] [Dylan Patel]: like I'm sorry love Island is I I I watched like for 10 minutes cuz I was forced to i was like this is freaking terrible um but you know like um we are so cooked no you know seriously we're cooked we're cooked and I think I think like when you think about like this happens it's like it's diffused into drama even people like like there's multiple dramas like taking place about semiconductor industry and and they're like romance comedy like the entire spectrum right drama like it's like it's like what the heck is going on

[译文] [Dylan Patel]: 抱歉，我看过《爱之岛》（Love Island），因为被强迫看了 10 分钟，我觉得那简直糟糕透顶。但是你知道吗……我们真的完了（We are so cooked）。不，说真的，如果这种文化已经渗透到了电视剧里，甚至有多部关于半导体行业的剧，还是浪漫喜剧，涵盖了各种类型……这简直是，到底发生了什么？

[原文] [Dylan Patel]: there's this one um Tik Tok or not Tik Tok Tik Tok and Instagram like uh person and they're like they they like sing it they're like if you want to if you want to buy things in China make sure you go to the right place and then they just say the most random shit and name the city and then you look into it and you're like wow this city has the entire supply chain for this um and it's like lampshades and then it names the city it's like what the fuck there's a city that specializes in lampshades like it's like and it's like microphone arms like microphones it's like it's like literally there's a city in China that specializes in guitars as well right this one one city that became the guitar capital of the world it's literally everything literally everything there's a city

[译文] [Dylan Patel]: 有一个 TikTok 还是 Instagram 上的博主，他们会唱这种歌，大意是：“如果你想在中国买东西，确保你去对了地方”。然后他们会说一些最随机的东西并报出一个城市名。你去查一下，就会发现：“哇，这个城市真的拥有这东西的完整供应链。”比如灯罩，然后报出一个城市。你会想：“我靠（What the f***），竟然有一个城市专门生产灯罩？”还有麦克风支架、麦克风之类的。真的有一个中国城市专门生产吉他，成了世界吉他之都。简直是所有东西，所有东西都有一个专门的城市。

[原文] [Dylan Patel]: and so like the semiconductor industry I think people don't realize is absurdly specialized i'm not answering your question i'm just going a little bit of a rant because I think people don't understand China semiconductors it's really sick... I think if you were to close your eyes and say or if you were to cut off every country and say there's no more globalism China has the most vertical stack in semiconductors today and they're the best at semiconductors in the world because their fabs could still run somewhat on a lot of things because they have built some of these chemical supply chains

[译文] [Dylan Patel]: 所以就像半导体行业一样，我认为人们没有意识到那是多么荒谬的专业化。我没在回答你的问题，我只是在发泄一下，因为我觉得人们不了解中国的半导体，那真的很疯狂（Sick）……我认为如果你闭上眼睛，或者切断所有国家的联系，说全球化不复存在了，中国拥有当今半导体领域最垂直的堆栈。从某种意义上说，他们是世界上最好的，因为他们的晶圆厂仍然可以在很多方面运转，因为他们已经建立了一些化学供应链。

[原文] [Dylan Patel]: but then when the flip side is like well like you kind of do need specialization that's how that chemical gets the purest best you know most engineered you know or that that slurry of chemicals or that you know that gas or like that tool because every smart person or a lot of them in that country grew up around that culture and like every the supply chain is there... China doesn't have lithography their lithography is like 10 years behind and I think it'll be 5 years behind in a couple years right they're catching up fast

[译文] [Dylan Patel]: 但另一方面，你也确实需要专业化，这样化学品才能达到最纯、最好、最精密的程度。因为那个国家里每一个聪明人，或者很多人，都是在那种文化中长大的，供应链就在那里……中国没有光刻机（Lithography），他们的光刻技术落后大概 10 年，但我认为几年后就会只落后 5 年，他们追赶得很快。

[原文] [Dylan Patel]: an example of that is Huawei right huawei in mobile phones was on par with Apple like entirely yeah and they had become Apple TSMC's biggest customer and they were designing the best thing and they are number one in telecom and their tech is just literally better and so when you think what happens you know is is is China missing anything it's like they're they don't they don't they don't have the best of much you know today in the AI supply chain they have a complete package and a couple years behind and they'll figure out how to make it cheaper slash do more slashcatch up

[译文] [Dylan Patel]: 一个例子就是 Huawei。Huawei 在手机领域曾完全与 Apple 并驾齐驱。他们曾成为 TSMC 仅次于 Apple 的最大客户，他们设计出了最好的东西。他们在电信领域是第一名，他们的技术确实更好。所以当你思考发生了什么，中国缺什么吗？他们今天在 AI 供应链中确实没有最好的东西，但他们有一个完整的方案，虽然落后几年，但他们会搞清楚如何让它更便宜、做得更多、并追赶上来。

[原文] [Dylan Patel]: in the same way I think the thing is Nvidia is like deathly terrified of Huawei because Huawei has caught up to Apple and actually surpassed them as TSMC's biggest customer before they got banned right they did just crush Nokia Sony Sony Ericson etc right like the entire telecom supply chain they just like completely destroyed them and there's so many other areas like they straight up made a folding phone right you know I have a Samsung folding phone they have a folding phone that's better than Samsung's folding phone and it's like bro what like you know you know Huawei's really really cracked and so of course they're terrified of uh and and Huawei is the most vertical company in the world no company is more verticalized than Huawei which then leads to huge innovations

[译文] [Dylan Patel]: 同样地，我认为 Nvidia 对 Huawei 怕得要死（deathly terrified）。因为 Huawei 曾追上 Apple，实际上在被禁之前甚至超过了他们成为 TSMC 的最大客户。他们彻底碾压了 Nokia、Sony Ericsson 等等，整个电信供应链被他们完全摧毁了。在很多其他领域也是如此，比如他们直接做出了折叠屏手机。我有一个三星折叠屏手机，但 Huawei 的折叠屏手机比三星的更好。这简直是：“兄弟，搞什么？”你知道 Huawei 真的强得离谱（Cracked）。所以 Nvidia 当然害怕他们。Huawei 是世界上最垂直整合的公司，没有哪家公司比 Huawei 更垂直化，这带来了巨大的创新。

[原文] [Dylan Patel]: how was Huawei actually building chips well actually they were uh using shell companies to get chips from TSMC and using different methods of like sneaking HBM which is memory from you know Korea through Taiwan to China right like all sorts of crazy stuff we've reported on and and people it's like a whack-a-ole right they shut it down or like tools that get shipped to China and they shouldn't be for you know making leading edge chips but they actually are

[译文] [Dylan Patel]: Huawei 到底是怎么制造芯片的？实际上，他们利用壳公司从 TSMC 获得芯片，并使用各种方法，比如把 HBM（高带宽内存）从韩国经台湾偷偷运到中国。各种疯狂的事情，我们都报道过。但这就像打地鼠游戏（Whack-a-mole），他们封锁了这个，或者那些不该运往中国制造尖端芯片的工具，实际上还是运过去了。

[原文] [Dylan Patel]: at the end of the day this is an economic war right if the US and the West win in AI and control you know more powerful AI systems that have this feedback loop that improved economic growth and weapons systems and whatever else right engineering of grids and cyber attacks and all these sorts of things they have this like advantage over China then China will not rise to be the global hedgeimony but without AI China definitely will rise to be the global hedgeimony they're just going to outrun America

[译文] [Dylan Patel]: 归根结底，这是一场经济战。如果美国和西方在 AI 领域获胜，控制了更强大的 AI 系统——这些系统拥有反馈循环，能促进经济增长、武器系统、电网工程、网络攻击等等——如果他们对中国拥有这种优势，那么中国将不会崛起成为全球霸主。但如果没有 AI，中国绝对会崛起成为全球霸主，他们会直接跑赢美国。

本章聚焦于美国为应对地缘政治挑战而推出的《芯片法案》（CHIPS Act）。Dylan 用数据直观地指出了该法案的局限性：500 亿美元的补贴在半导体这一“人类最复杂供应链”面前显得杯水车薪。

他揭示了一个讽刺的现实：《芯片法案》能够通过，并非因为政客们关心尖端 AI 芯片，而是因为疫情期间汽车芯片（低端微控制器）的短缺导致选民买不到车。尽管如此，他对 TSMC 在亚利桑那的建厂持乐观态度，并用生动的“火箭外科手术”（Rocket Surgery）一词来形容半导体制造的极致难度。

章节 5：美国《芯片法案》与全球供应链的现实

📝 本节摘要：

在本章中，讨论转向了美国的本土化努力。Dylan 指出，美国《芯片法案》的 500 亿美元补贴虽然是好事，但在半导体行业巨大的资本开支（Capex）面前显得微不足道（相比之下，台湾已投入超过 5000 亿美元）。

>

他透露了一个有趣的幕后逻辑：《芯片法案》之所以能通过，主要是因为“准时制库存”（Just-in-time inventory）策略的失败导致了汽车芯片短缺，而非为了尖端 AI 芯片。尽管建设晶圆厂面临巨大的延误和挑战，Dylan 仍对美国实现一定程度的产能回流持乐观态度，即使这无法实现完全的自给自足。他最后抛出了一个“不受欢迎”的观点：全球化其实是件好事。

[原文] [Matt Turck]: where do the uh US onshoring efforts fall in that category what do you make of them from the chipsack to like all the thing that is being built everything looks like it's massively delayed by the way which perhaps is not surprising

[译文] [Matt Turck]: 美国的“回流”（onshoring）努力属于哪一类？你如何看待从《芯片法案》到正在建设的所有设施？顺便说一句，所有项目看起来都严重延误了，这也许并不令人惊讶。

[原文] [Dylan Patel]: I think TSM MC's manufacturing wafers and they're like building real wafers and there's real fabs and like you know there's some other fabs that have been announced and like they're doing well and there's like a bunch of like different kinds of plants like a Korean company making a random gas plant in Texas for you know their chips right like uh for chips and all these like sort of things are happening

[译文] [Dylan Patel]: 我认为 TSMC（台积电）正在制造晶圆，他们在制造真正的晶圆，那里有真正的晶圆厂（Fabs）。还有一些其他的晶圆厂也已经宣布了，进展不错。还有一堆不同类型的工厂，比如一家韩国公司在得克萨斯州建了一个随机的气体工厂，为了他们的芯片，对吧？所有这类事情都在发生。

[原文] [Dylan Patel]: I think the chips act did really well with its $50 billion it's just I don't think people understand the scale of the semiconductor industry is the most complicated supply chain in the world right it's much bigger than you know say manufacturing airplanes it's much bigger than like you know really anything else right

[译文] [Dylan Patel]: 我认为《芯片法案》用那 500 亿美元做得真的很好。只是我觉得人们不理解半导体行业的规模，它是世界上最复杂的供应链，对吧？它比制造飞机大得多，比任何其他东西都要大得多。

[原文] [Dylan Patel]: if you look at the top 10 companies like of the world I think eight of them designed semiconductors right now obviously like Google designed semiconductors but it's like oh wait no but their cost of search would be like 10x higher if they didn't have TPUs and TPUs were super optimized for search right or like you know you you you go down the list right like Meta serves recommendation systems with their chips right like you go down the list it's everyone is making their own chips apple devices would be materially worse if they didn't have their own chips right

[译文] [Dylan Patel]: 如果你看全球前 10 大公司，我认为其中有 8 家都在设计半导体。显然像 Google 设计半导体，你会想“哦等等，不完全是”，但如果他们没有 TPU，他们的搜索成本会高出 10 倍，而 TPU 是针对搜索进行了超级优化的。你顺着名单往下看，Meta 用他们的芯片服务推荐系统；Apple 设备如果没有自研芯片，性能会大打折扣。每个人都在造自己的芯片。

[原文] [Dylan Patel]: um and you just go down the list it's like it's the most complicated supply chain and they they're spending something on the order of like $150 billion roughly in subsidies a year to the chip industry we are doing 50 over like a decade there's a difference in scale here right the collective total amount of like capex that has been spent in Taiwan is like 500 billion plus right across the industry across all the companies that are making semiconductors in Taiwan and Taiwan doesn't have a domestic industry how is $50 billion of subsidies going to change America's needle right

[译文] [Dylan Patel]: 顺着名单看，这是最复杂的供应链。全球每年大概在向芯片行业投入大约 1500 亿美元的补贴，而我们是在十年内投入 500 亿。这里存在规模上的差异，对吧？在台湾，整个行业、所有制造半导体的公司累计投入的资本支出（Capex）总额超过 5000 亿美元。而台湾甚至没有那样庞大的国内终端市场。那么，500 亿美元的补贴怎么能改变美国的现状呢？

[原文] [Dylan Patel]: it does move it a little bit right i I want to be clear like the chips act is awesome i don't understand why like EVs or like solar was given this massive massive like trillion dollar package semiconductors were only given 50 like semiconductors need a lot bigger package to actually incentivize onoring i think what's happened so far has proven that it's working well tsmc is literally making chips for Nvidia and Apple and AMD and others in Arizona today right and I think that's really great

[译文] [Dylan Patel]: 它确实推动了一点点。我想明确一点，《芯片法案》很棒。但我不明白为什么像电动汽车（EV）或太阳能得到了那种巨大的、万亿美元级别的包裹，而半导体只给了 500 亿？半导体需要一个大得多的包裹才能真正激励回流。我认为到目前为止发生的事情证明它运作良好，TSMC 今天确实正在亚利桑那州为 Nvidia、Apple 和 AMD 等公司制造芯片，我认为这真的很棒。

[原文] [Matt Turck]: is is your sense that the broad American government is just uh aware of of all of this that it's uh

[译文] [Matt Turck]: 你感觉美国政府的大多数人是否意识到了这一切？

[原文] [Dylan Patel]: I wouldn't say only passed because the automotive like prices went up because car manufacturers are like the worst because they do just in time inventory right or not worse but like this is just like a thing right just in time inventory systems covid happens sales plummet fabs that were making you know random power IC's or random microcontrollers for engines got repurposed to the boom from COVID which is which was data centers and PCs and smartphones

[译文] [Dylan Patel]: 我不会说它是“仅仅”因为汽车价格上涨才通过的，但汽车制造商简直是最糟糕的，因为他们搞“准时制库存”（Just-in-time inventory）。或者也不是最糟糕，但这就像是个通病。准时制库存系统遇到新冠疫情，销量暴跌。原本制造引擎用的随机电源 IC 或微控制器的晶圆厂，被重新分配去应对新冠带来的繁荣——也就是数据中心、PC 和智能手机。

[原文] [Dylan Patel]: so that stuff was booming and then when people were like "Oh wait actually like you know I have some money i stayed at home i didn't go out i didn't drink i have a lot of I have some cash right let me buy a car." They went out and bought cars and cars started skyrocketing in prices oh let's restart and let's let's Oh yeah can I can you sell me that microcontroller for the engine again it's like "No I I'm making a slightly different microcontroller that works for you know uh let's say a keyboard or a mouse right or whatever." And it's like and and and they actually didn't just leave me flatfooted and they were like a partner through co right you know versus you just left me screw you Ford or whoever Toyota um or automotive OEM

[译文] [Dylan Patel]: 那些东西当时很火爆。然后人们反应过来：“哦等等，其实我有点钱，我呆在家里没出门，没喝酒，我有一些现金，让我买辆车吧。”他们出去买车，车价开始飙升。车企说：“哦，让我们重启吧，你能再卖给我那个引擎用的微控制器吗？”晶圆厂说：“不，我正在做一个稍微不同的微控制器，用于键盘或鼠标或其他什么。”而且，你知道，消费电子厂商在疫情期间是合作伙伴，并没有让晶圆厂空转；相比之下，你（车企，比如 Ford 或 Toyota）当时直接抛弃了我，去你的吧。

[原文] [Dylan Patel]: you know you up that supply chain and so Chips Act did not get passed only got passed because that happened and people are like "Oh my god the semiconductors are why cars can't be made." If that didn't happen we wouldn't even have the chips act it's like it's like silly so like I don't know like I think you know whereas like and and even though that's what was pitched to all the senators like I know people who were running around Capitol Hill just pushing that narrative and story and that's why it finally got passed in reality it was all for advanced leading edge chips right nothing that goes in a car right and so it's like this like funny thing

[译文] [Dylan Patel]: 所以《芯片法案》得以通过，仅仅是因为发生了那件事，人们惊呼：“天哪，因为半导体我们造不出车了。”如果没发生那件事，我们要么根本不会有《芯片法案》。这很愚蠢。尽管这就是当时向所有参议员推销的说辞——我认识那些在国会山跑来跑去推销这个叙事故事的人，这就是法案最终通过的原因——但在现实中，法案的资金全是给先进尖端芯片的，完全不是给汽车芯片的，对吧？所以这真是一件滑稽的事情。

[原文] [Matt Turck]: so in other words do you think my words my words not yours but is it is it hopeless that the US is going to

[译文] [Matt Turck]: 换句话说，你是否认为——这是我的话，不是你的——美国要做成这件事是无望的？

[原文] [Dylan Patel]: I'm very optimistic okay i mean do you think there's a world where the US just decides to invest in semiconductor at the scale that you know I thought we just needed a bigger chips act but look Trump's kind of gotten TSMC to promise to invest a fuckload more and they're moving on it right they're like actually like just building it it's like I'm going to tariff the shit out of you unless you build a fab but it's like we'll build a fab and they're building it right now

[译文] [Dylan Patel]: 我非常乐观。好吧，我是说，你认为有没有可能美国决定以那种规模投资半导体？我原以为我们需要一个更大的《芯片法案》，但你看，特朗普某种程度上让 TSMC 承诺投资更多，而且他们确实在行动，对吧？他们真的在建厂。这就像是：“除非你建个晶圆厂，否则我要用关税搞死你。”然后他们说：“我们会建个厂。”他们现在正在建。

[原文] [Dylan Patel]: the timelines for fabs just takes forever cuz again it's the most complicated thing in the world the cleanest space in the place in the world is not like a hospital or a biotech lab or whatever it's a semiconductor fab and the most expensive tools in the world are not you know any of these medical tools or whatever it's it's semiconductor tools or it's not a rocket it's a semiconductor tool right like everything you know I describe it as um I remember when I was a kid I was like I want to be a rocket scientist and then I was like oh I want to be a surgeon and I'm like wait chips are like rocket surgery but even cooler right

[译文] [Dylan Patel]: 晶圆厂的时间线就是要花很久，因为再说一次，这是世界上最复杂的事情。世界上最干净的地方不是医院或生物技术实验室，而是半导体晶圆厂。世界上最昂贵的工具不是医疗工具，也不是火箭，而是半导体工具。我把它描述为……我记得小时候我想当火箭科学家，然后我想当外科医生，然后我想：“等等，芯片就像是‘火箭外科手术’（Rocket Surgery），甚至更酷。”

[原文] [Dylan Patel]: like I think anyways like sort of like there there are fabs being built in America they won't take America to self-sufficiency i don't think that's a relevant i don't think that's a goal relevant like that's relevant right like globalism is generally just good hot take like in terms of economics we'll turn this into a short a YouTube short globalism globalism is good dude you're gonna get me like cancelled

[译文] [Dylan Patel]: 总之，美国正在建设晶圆厂。它们不会让美国实现自给自足，我不认为那是相关的，我不认为那是一个相关的目标。比如，全球化通常就是好的。这是个“暴论”（Hot Take），从经济学角度来看。我们要把这段剪成 YouTube Short 短视频：“全球化是好的”。老兄，你会让我被取消（Cancelled）的。

本章探讨了 AI 行业面临的外部舆论压力与内部的经济泡沫争议。Dylan 从一个脱口秀现场的亲身经历出发，揭示了公众对 AI 的普遍反感。随后，对话切入华尔街最关心的“资本支出（Capex）泡沫”问题。Dylan 提出了一个核心论点：模型性能是硬件投入的滞后指标，只要模型能力还在通过堆算力提升，当前的巨额投入就是合理的。

章节 6：公众情绪与 AI 资本支出（Capex）泡沫论

📝 本节摘要：

本章首先触及了公众对 AI 的负面情绪。Dylan 分享了自己在脱口秀俱乐部的见闻——演员仅因提及使用 ChatGPT 就遭到嘘声，反映出人们将电价上涨（如新泽西州案例）、垃圾内容泛滥（Slop）以及就业焦虑归咎于 AI,,。

>

随后，Matt 提出了关键的“资本支出（Capex）泡沫”问题：我们是否投资过剩？Dylan 自称是“最大主义者”（Maxi），他认为尽管今年超大规模厂商（Hyperscalers）的资本支出可能高达 5000 亿美元，但这并非泡沫,。他提出了一个关键洞察：模型进步是硬件资本支出的“滞后指标”（Lagging Indicator）。现在的投入（购买 GPU、建数据中心）是为了创造明年更强大的模型。只要 Scaling Laws（缩放定律）依然奏效，模型没有撞墙，这些投资最终都会转化为巨大的经济价值,。

[原文] [Dylan Patel]: whatever happens I don't know like I was at a comedy club on Sunday night and like he's like "Oh I'm I use Chad GPT." And then like there were a couple people who booed and he's like "Yeah I'm one of those guys i know." And like it's like "Wow people hate AI." And that has has not even started right like the actual impact of AI or like New Jersey power prices are up right

[译文] [Dylan Patel]: 无论发生什么，我不知道……比如周日晚上我在一家脱口秀俱乐部，演员说：“噢，我是用 ChatGPT 的。”然后有几个人发出了嘘声。他说：“是啊，我就是那种人，我知道。”这让我觉得：“哇，人们真讨厌 AI。”而真正的 AI 冲击甚至还没有开始，对吧？或者像新泽西州的电价上涨了。

[原文] [Dylan Patel]: uh is it because of a data center new Jersey the governor's election like I think literally fl like there's like an election that changed recently in New Jersey because power prices were up and people blamed a Microsoft Nebius data center in New Jersey for that reason but in reality that data center has nothing to do with power prices going up it's super storm standy like five years ago knocking or whatever how many years ago knocking down the state's electrical infrastructure and then the then improving all these improvements and then those improvements have to be paid by someone and it turns out the consumer has to pay for them with higher power prices right

[译文] [Dylan Patel]: 这是因为数据中心吗？新泽西州的州长选举——我想最近新泽西州的选举局势确实因此改变了，因为电价上涨，人们为此责怪新泽西的一个 Microsoft Nebius 数据中心。但实际上，那个数据中心跟电价上涨毫无关系。真正原因是五年前（或者不管多少年前）的“超级风暴桑迪”（Superstorm Sandy）摧毁了该州的电力基础设施，然后进行了各种修复和升级。这些改进必须有人买单，结果就是消费者必须通过更高的电价来买单，对吧？

[原文] [Dylan Patel]: and so like you know like there's like there's a lot like going on in that regard right um that kind of is uh sad um and and people hate AI and they're blaming AI on it and artists hate AI and like you know you see all this deep fake stuff and like I think I think it'll be the hottest button issue especially as like we're really getting into like I think last year Google spent $3 billion on Whimo and we're waiting for their guide for this year $3 billion on Whimo taxis but their t their Whimos went from like 300k to like 100k or 90k the new Whimo car and they're going to spend more than three because they've just launched in like four cities now right or five cities and and they're testing it a lot and the same a robo taxi like people are going to hate AI for that reason people are going to hate AI because the slop on the internet people are going to hate AI because you know the perceived job replacement people are going to hate AI for all these reasons and so yeah it's going to be a hot button political issue don't you think

[译文] [Dylan Patel]: 所以你知道，这方面发生了很多事情，这挺令人悲哀的。人们讨厌 AI，并把问题归咎于 AI。艺术家讨厌 AI。你看到所有那些深度伪造（Deepfake）的东西。我认为这将成为最热门的争议话题，特别是当我们真的进入……比如去年 Google 在 Waymo 上花了 30 亿美元，我们在等他们今年的指引。Waymo 出租车的成本从 30 万美元降到了 10 万或 9 万美元（新款 Waymo 车）。他们今年的花费会超过 30 亿，因为他们刚刚在 4 个或 5 个城市推出了服务，正在大量测试。同样的，Robotaxi（自动驾驶出租车）——人们会因为这个原因讨厌 AI。人们会因为互联网上的垃圾内容（Slop）讨厌 AI。人们会因为感知到的工作替代而讨厌 AI。人们会因为所有这些原因讨厌 AI。所以是的，这将是一个热门的政治议题，你不觉得吗？

[原文] [Matt Turck]: yeah talking about that so um capex is there a capex bubble are we u investing too much or actually are we investing not enough given what you were saying earlier about uh the the the rate of revenue increase and and therefore implied demand that you expect for this year

[译文] [Matt Turck]: 是的，说到这个，关于资本支出（Capex）——是否存在资本支出泡沫？我们是投资过多了，还是实际上投资不足？考虑到你早些时候提到的收入增长率以及你对今年隐含需求的预期。

[原文] [Dylan Patel]: i'm obviously a maxi i think we're going to need a lot of infra and I think I'm literally paid to like analyze the supply chain and do consulting like that's what my company does so like obviously I'm very biased i think I think we're pretty good at calling when when things go down though right before like a part of the supply chain reb anyways you know again going back to the economics of it it's north of hundred billion dollars of revenue exiting this year for AI from a base of you know sub1 billion gen AI from a base because ads and stuff is like already a multiundred billion dollar AI industry right you know go back to 2023 it was like less than a billion right and 2024 I don't know exactly what number maybe let's call it 10 and 25 was maybe like 30 40 it'll be north of 100 easily

[译文] [Dylan Patel]: 我显然是一个“最大主义者”（Maxi），我认为我们需要大量的基础设施。我也确实是靠分析供应链和做咨询赚钱的，这是我公司的业务，所以显然我有很大的偏见。不过我认为我们在预测行情下跌方面也做得不错，比如在供应链某个部分反弹之前。不管怎样，回到经济学上：今年 AI 的收入将超过 1000 亿美元，而这仅仅是从一个不到 10 亿美元的生成式 AI（Gen AI）基数增长起来的——因为广告之类的已经是数千亿美元的 AI 产业了，对吧？回到 2023 年，它可能还不到 10 亿；2024 年我不知道确切数字，也许叫它 100 亿；2025 年可能是 300-400 亿，之后很容易就会超过 1000 亿。

[原文] [Dylan Patel]: if you're talking about hundred billion of revenue let's say at a 50% gross margin so that's $50 billion of gross profit um and $50 billion of COGS that $50 billion of COGS needs to run on infra which cost roughly if a five if you're talking about fiveyear depreciation call it $250 billion right of infra for hundred billion of revenue mhm okay what is what is the actual spend on AI infra this year it's going to be like it's I mean it depends on what layer if you're talking about energy those are longer lived assets and all these other things right um data centers are longer lived assets the chips are not as much people are putting capex down um and the hyperscalers capex is going to be like $500 billion this year or something like this

[译文] [Dylan Patel]: 如果你说的是 1000 亿美元的收入，假设 50% 的毛利率，那就是 500 亿美元的毛利和 500 亿美元的销货成本（COGS）。这 500 亿美元的 COGS 需要运行在基础设施上，如果按 5 年折旧计算，大概对应 2500 亿美元的基础设施成本，来支撑 1000 亿美元的收入。嗯，好的。那么今年实际的 AI 基础设施支出是多少呢？如果你谈论的是能源，那是长寿命资产；数据中心也是长寿命资产；芯片则没那么长。人们正在投入资本支出，超大规模厂商（Hyperscalers）今年的资本支出大概会是 5000 亿美元左右。

[原文] [Dylan Patel]: and then besides them there's also a lot more hyp uh capex elsewhere um and so you know is it a bubble i mean theoretically like you know it's twice as much as it should be but it's also like well no there's an R&D component to this and the excess spent that wasn't revenue generating last year is what led to models being so good this year um and led to like everyone who can using cloud code and like that changing their life this is like it's not a bubble right i don't think it's a bubble yet

[译文] [Dylan Patel]: 除此之外还有很多其他的资本支出。所以，这是泡沫吗？理论上讲，现在的投入是“应该有”的两倍多。但也并非如此，因为这里面有研发（R&D）的成分。去年那些没有产生收入的“过剩支出”，正是导致今年模型变得如此之好的原因，也导致了所有能用 Claude Code 的人生活发生了改变。这就像……这不是泡沫，对吧？我认为目前还不是泡沫。

[原文] [Dylan Patel]: I think if AI model progress stops and that's the main thing right the moment model progress stops all the spending is for not but so far we've had consistent improvement as you put in more compute you get more performance and better models yeah model performance being the lagging indicator of hardware progress or data center yeah of of capex right yeah ultimately the capex that Microsoft spent in 2024 for OpenAI is what results in in 2025 for OpenAI Cory or whoever is what results in their models being so good this year same with Enthropic and Amazon Google and their models now being so good now is is that capex and actually they still haven't paid for those chips yet because those chips are still have a useful life for another few years right

[译文] [Dylan Patel]: 我认为如果 AI 模型的进步停止了——这才是关键——只要模型进步停止的那一刻，所有的支出就都白费了。但到目前为止，我们看到了持续的改进：你投入更多的算力，就获得更高的性能和更好的模型。

[Matt Turck]: 是的，模型性能是硬件进步或数据中心资本支出的“滞后指标”（lagging indicator），对吧？

[Dylan Patel]: 是的。归根结底，Microsoft 在 2024 年为 OpenAI 投入的资本支出，是导致 OpenAI（或 Cory，或其他谁）在 2025 年模型变得如此之好的原因。Anthropic 和 Amazon、Google 也是如此，他们现在的模型之所以这么好，就是源于之前的资本支出。实际上他们还没完全为那些芯片“买单”（指折旧），因为那些芯片还有几年的使用寿命。

[原文] [Dylan Patel]: I think model progress is very clear um the moment that stops happening right if we hit a wall there's no new research directions um then then it's cooked yeah right and that assumes that better model leads to more demand which is a reasonable assumption yeah for sure

[译文] [Dylan Patel]: 我认为模型的进步是非常清晰的。一旦这种进步停止，如果我们撞墙了，没有新的研究方向了，那就彻底完了（Cooked）。

[Matt Turck]: 是的，对。而且这假设了更好的模型会带来更多的需求，这是一个合理的假设。

[Dylan Patel]: 是的，毫无疑问。

本章 Dylan 以极具破坏力的观点粉碎了关于 AI 基础设施的两个核心谣言：一是“AI 摧毁电网”，二是“AI 耗尽水资源”。

他指出美国电网的停滞是人为的行政惰性，而非物理限制。而在水资源问题上，他抛出了一个荒诞却有力的计算：马斯克那个世界上最大的 AI 数据中心（Colossus），其耗水量仅仅相当于 2.5 家 In-N-Out 汉堡店所售汉堡背后的农业用水量。

章节 7：基础设施瓶颈：电力短缺与用水谣言的真相

📝 本节摘要：

本章聚焦于基础设施的物理限制。关于电力，Dylan 直言美国电网“又慢又蠢”，过去 50 年几乎没有大规模新建电力设施，导致公用事业公司对数据中心的激增措手不及。为了解决设备短缺，人们甚至开始抢购 Mitsubishi 的涡轮机或皮卡车用的 Cummins 引擎。

>

关于水资源，Dylan 驳斥了“AI 正在耗尽水源”的说法，称其为无稽之谈。他通过一个经典的“汉堡包对比”指出：生产燕麦奶或牛肉的农业耗水量远超数据中心。根据他的计算，Elon Musk 的 Colossus 数据中心（目前世界上最大的之一）的耗水量，仅相当于 2.5 家 In-N-Out 汉堡店的隐性水足迹。最后，他讽刺了针对 Meta 数据中心的环保抗议，指出真正的水污染源往往是由于向亚洲出口天然气而进行的水力压裂（Fracking），而非数据中心本身。

[原文] [Matt Turck]: but oh that $50 billion capex was spent in year one what about energy in the in the data center world you had this fun post about the gas replacement for for energy so is uh is AI basically uh uh destroying the grid

[译文] [Matt Turck]: 但是，那 500 亿美元的资本支出是在第一年花掉的。那能源呢？在数据中心领域，你有一篇很有趣的文章是关于用天然气替代能源的。所以，AI 基本上是在摧毁电网吗？

[原文] [Dylan Patel]: it would if the utilities were willing to let it but I think the utilities are so slow and dumb that they don't want to not destroy but like expanding the grid yeah um I think the US could have a way better grid but we just don't want to like no one's made the effort or initiative you know there's not enough power america's not built power for 50 years really right it's like converted from coal to gas and like things like this but like really just have not built wholesale new power on a large scale

[译文] [Dylan Patel]: 如果公用事业公司（Utilities）愿意让它摧毁，那它确实会。但我认为公用事业公司太慢、太蠢了，他们并不想扩建电网。是的，我认为美国本可以拥有好得多的电网，但我们就是不想，没人付出努力或采取主动。你知道电力不足，美国真的有 50 年没建新电厂了，对吧？也就是把煤电转成气电之类的，但真的没有大规模地建立全新的电力供应。

[原文] [Dylan Patel]: and there have been a lot of times where the industry blew up right independent power producers IPs have blown up multiple times in the 2010s when uh Korean and Japanese investors like flooded the market with because they saw such a good return there or before in the early 2000s power was growing a little bit for a little bit and so people overbuilt on power so power industry has been burned a couple times but no one really built power and then you've got data centers now all of a sudden coming online and going from 2% to 10% of the US grid in just a handful of years and so you've got this humongous humongous change in the industry

[译文] [Dylan Patel]: 这个行业曾多次崩溃。独立发电商（IPPs）在 2010 年代多次爆雷，当时韩国和日本投资者因为看到了所谓的高回报而涌入市场；或者在 2000 年代初，电力需求稍微增长了一点，人们就过度建设了。所以电力行业被烧过几次，导致没人真去建电厂了。然后现在数据中心突然上线，短短几年内从占美国电网的 2% 飙升到 10%。所以你面临着行业内这种巨大、巨大的变化。,

[原文] [Dylan Patel]: we don't have the labor right i think ultimately that's the biggest problem is the equipment and the labor and equipment is basically you know again labor and time takes time to build a factory so you can build the things i think the equipment side of things will be solved like more reasonably and one one example was like gas right people initially thought oh you can only use like the two vendors right uh Seammens or G Vernova for gas turbines but they have the they have the best ones the most efficient ones it's like okay well like okay also Mitsubishi exists and they're ramping up production fast oh Ducson and Korea exist and they're ramping up production fast

[译文] [Dylan Patel]: 我们没有足够的劳动力。我认为归根结底最大的问题是设备和劳动力。设备基本上也受限于劳动力和时间，因为建工厂来生产这些设备需要时间。我认为设备方面的问题会得到较合理的解决。一个例子是天然气，人们最初以为只能用两家供应商的燃气轮机，西门子（Siemens）或 GE Vernova，因为它们拥有最好、最高效的产品。但其实三菱（Mitsubishi）也存在，而且他们正在快速提升产能；韩国的斗山（Doosan）也存在，也在快速扩产。

[原文] [Dylan Patel]: oh actually I can just take Cumins engines right like you know if you've ever like ridden a pickup truck or like you know like diesel trucks like everyone loves Cumins right you know you see the Ram on the street and has the Cumins like badge it's like it's like a that's like an aura symbol for a certain kind of redneck from South Georgia which I have a little bit of anyways I I don't have a I don't have a truck i have though um but anyways like the there's like all these engines like people are figuring out how to make the equipment

[译文] [Dylan Patel]: 哦，实际上我也可以直接用康明斯（Cummins）引擎。如果你坐过皮卡或柴油卡车，大家都爱康明斯，对吧？你在街上看到道奇 Ram 皮卡贴着 Cummins 的标，那就像是南佐治亚某种“红脖子”（redneck）的光环象征——我其实也有点这种气质，虽然我没有卡车。总之，有各种各样的引擎，人们正在想办法制造设备。

[原文] [Dylan Patel]: you know solar sucks it's too intermittent wind sucks it's too intermittent nuclear sucks it takes forever to build coal sucks it's way too dirty how do you make power for data centers besides gas and like okay the grid's not willing to put the gas on your site right that's what Elon did now everyone's doing it right

[译文] [Dylan Patel]: 你知道，太阳能很烂，因为太间歇性了；风能很烂，也太间歇性了；核能很烂，因为建设周期太长；煤电很烂，因为太脏了。除了天然气，你怎么为数据中心发电？好吧，如果电网不愿意把天然气送到你的站点，那就像 Elon（马斯克）做的那样（自建发电），现在大家都在这么做。

[原文] [Matt Turck]: this other cool post just uh last week or two weeks ago that was about water consumption uh did you want to talk to that yeah yeah so there's this annoying thing where everyone's like "Oh AI is using all the water oh wow AI and data centers are going to like use up all the water and now we don't have any water." And it's like that's so silly uh water is a distribution problem not a like we don't have enough problem right like you look at California so California has shitloads of water but people decide to make oat milk which consumes like 1,000x the water of like anything else like regular milk even and and cows obviously eat a you know consume a lot of water

[译文] [Matt Turck]: 就在上周或两周前，还有一篇关于水资源消耗的很酷的文章，你想聊聊那个吗？

[Dylan Patel]: 是的，是的。这事儿特烦人，大家都说：“噢，AI 正在用光所有的水！哇，AI 和数据中心要把水用光了，我们没水了。”这太蠢了。水是一个分配问题，不是我们没有足够水的问题。你看加州，加州有成吨的水，但人们决定去生产燕麦奶，那东西消耗的水是其他东西的 1000 倍，甚至比普通牛奶还多。而奶牛显然也消耗大量的水。,

[原文] [Dylan Patel]: but anyways like you know data centers consume very little water actually right so the US grid will get to like 10% of power by like 28 27 is data centers for water consumption it's not even going to crack 1% by the end of the decade and what was the metric um and so so the the comparison we made is because like you know it was a bit of a shit post but it was like serious research yeah basically like we were doing serious research because we keep getting this like question and debunking it and we would do it seriously but then I was like no no no this is like too like complicated like let's make it very simple

[译文] [Dylan Patel]: 总之，数据中心实际消耗的水非常少。到 2027 或 2028 年，数据中心可能占美国电网的 10%，但在水资源消耗方面，到本年代末甚至连 1% 都不到。我们做的对比其实有点像是“钓鱼贴”（shit post），但背后是严肃的研究。因为我们不断收到这类问题并去辟谣，我们本来想严肃地做，但我后来觉得：“不不不，这也太复杂了，让我们把它弄简单点。”

[原文] [Dylan Patel]: so I was like "Guys why don't we just compare it to like hamburgers right cuz cuz you know I've heard that argument from some like vegetarian people before or some Hindus or like I'm Hindu myself although you know and I I I do eat beef sometimes but you know like I I'm Hindu but like you know so so we made this comparison to hamburgers right hamburgers require a shitload of water cuz cows you know when to for them they require a ton of water and when a cow's taking a lot of water it's not the cow itself it's all the feed you're feeding them right because no one grass feeds their cows you know and just lets the rain take care of the grass they like either rain the the grass or most likely they do mass industrial farming of corn soybean alalfa etc which uses shitloads of water right

[译文] [Dylan Patel]: 所以我说：“伙计们，我们要不直接拿它跟汉堡包比吧？”因为我以前听素食主义者或一些印度教徒说过这个论点——我自己也是印度教徒，虽然我有时也吃牛肉。汉堡包需要极其大量的水，因为养牛需要成吨的水。牛消耗的水不仅仅是它喝的，而是你喂给它们的饲料。没人真的完全靠雨水滋润的草来喂牛，他们通常进行大规模工业化种植，种玉米、大豆、苜蓿等，这都要消耗极其大量的水。

[原文] [Dylan Patel]: i think the uh metric was the entirety of Elon Musk's Colossus data center right uses as much water as two and a half in-n-outs um because that's you know you do the calculation on how many how many b what's the average revenue per in-n-out and how many hamburgers does that translate to right if everyone's ordering like a combo right okay let's ignore the drink let's ignore the fries let's just talk about the hamburger let's ignore the bread which does use have grain let's just do the meat and the cheese and all of a sudden all this water is there's so much water right

[译文] [Dylan Patel]: 我们得出的指标是：Elon Musk 的整个 Colossus 数据中心（注：世界上最大的 AI 集群之一）所消耗的水，仅相当于 2.5 家 In-N-Out 汉堡店。计算逻辑是：一家 In-N-Out 的平均收入是多少？这转化为多少个汉堡？如果每个人都点套餐，忽略饮料，忽略薯条，忽略面包（虽然谷物也耗水），只算肉和奶酪，突然间你会发现这里面全是水（隐性水足迹）。,

[原文] [Dylan Patel]: like a single query like all of your AI usage from chat GBT of the average user is like a hamburger right like it's like okay this is nothing right you know because these things are the data centers actually are like they're mostly closed loops and like sure they evaporate some water for like cooling reasons but like by doing evaporative cooling they're using less power right and that's actually better for the environment than uh than not using evaporative cool

[译文] [Dylan Patel]: 比如一次查询，或者一个普通用户在 ChatGPT 上的所有 AI 使用量，大概就相当于一个汉堡。这根本不算什么。因为数据中心实际上大多是闭环循环的。当然，为了冷却它们会蒸发一些水，但通过蒸发冷却，它们消耗的电力更少，这实际上比不使用蒸发冷却对环境更好。

[原文] [Dylan Patel]: there's all all these reasons why this myth or hoax of AI of AI using all the water is just nonsense right like Meta's data center in Louisiana is getting protested because the water it's it's going to be the largest data center in the world it's going to be like four or five gigawatts at least announced so far... but Meta is getting protested because the local population around that area is like "Oh the water's dirty it's because of this meta data center."

[译文] [Dylan Patel]: 有很多理由说明所谓“AI 正在用光水资源”的神话或骗局纯属胡扯。比如 Meta 在路易斯安那州的数据中心遭到抗议，这将是世界上最大的数据中心，至少已宣布的有 4 到 5 吉瓦（GW）。当地居民抗议说：“噢，水脏了，都是因为这个 Meta 数据中心。”,

[原文] [Dylan Patel]: but at the end of the day what actually made the water dirty is that that's an area where you go fracking like fracking is absurdly worse and almost all of that gas is being shipped to an LG terminal and being shipped to Asia like you know you know like Japan or Taiwan or China or Korea and some Europe as well right like like actually all of this water is dirty because of regulation fracking like I support fracking by the way but you know that's that's an insane take too maybe um but like water usage is is is like not a relevant argument

[译文] [Dylan Patel]: 但归根结底，真正让水变脏的是那个地区在进行水力压裂（Fracking）。水力压裂要糟糕得多，而那里几乎所有的天然气都被运往 LNG 终端，然后运往亚洲，比如日本、台湾、中国、韩国，还有一些运往欧洲。实际上水脏是因为缺乏监管的水力压裂——顺便说一句，我是支持水力压裂的——但这（指怪罪数据中心）是个疯狂的观点。水资源的使用根本不是一个相关的论点。

本章深入探讨了 AI 基础设施建设背后的能源投资逻辑与金融结构争议。Dylan 分析了独立发电商（IPPs）的暴利机会，并详细拆解了被市场质疑为“循环融资”（Circular Financing）的 CoreWeave 与 Nvidia 交易模式，将其定义为必要的信用担保而非金融泡沫。

章节 8：能源投资与融资结构：循环交易的争议

📝 本节摘要：

在本章中，Dylan 首先对能源市场进行了预测。他看好独立发电商（IPPs），特别是那些能将发电资产直接与数据中心配对（Behind-the-meter）的厂商。他认为核能建设周期太长（即使在中国也需 5 年），无法满足 AI 爆发式需求，因此天然气仍是主力,。他甚至分享了一个客户案例：通过买下废弃的燃煤电厂并重启，专门为超大规模厂商供电而赚得盆满钵满。

>

随后，话题转向敏感的“循环融资”（Circular Financing）问题（即 Nvidia 投资 CoreWeave/OpenAI，后者再用这笔钱买 Nvidia 芯片）。Dylan 驳斥了“庞氏骗局”的说法，认为这是解决由于信贷紧缩导致新晋数据中心建设者（如 CoreWeave）无法获得银行贷款的必要手段。通过 Microsoft 等巨头的合同兜底（Backstop），或者 Nvidia 的战略投资，这些公司得以完成基础设施建设的冷启动,,。

[原文] [Matt Turck]: are you bullish on the sort of energy companies I'm thinking constellation for nuclear or Vistra I guess is an independent power producer

[译文] [Matt Turck]: 你看好那些能源公司吗？比如做核能的 Constellation，或者 Vistra，我猜它是独立发电商？

[原文] [Dylan Patel]: i think IPS will do well i think IPS can secure contracts at premiums to what they've previously been able to for new power plants that are either uh dedicated or grid connected but come with a pairing of a grid load right for example utilities won't let you just do data centers now but if you come with a a pair right you're like hey I'm going to build this massive data center but we're also going to have this massive uh power generating asset right say you know whatever it is right some IP they're going to partner with and they'll build the load and the uh consumption even if it's connected through the grid for better stability and more reliability um or it's not it's behind the meter i.e not connected to the grid at all um like some part some data centers like partially like Colossus from Elon uh the original one or part of Abene's Texas OpenAI right like Cruso there's a lot of room for power producers to get outsized returns

[译文] [Dylan Patel]: 我认为独立发电商（IPPs）会做得很好。我认为 IPPs 能够以比以前更高的溢价拿到新电厂的合同，无论是专用的还是并网的，前提是它们自带“配对”的电网负载。例如，公用事业公司现在不会让你只建数据中心，但如果你带着“一对”来——比如你说“嘿，我要建这个巨大的数据中心，但我们也会有这个巨大的发电设备”，不管是跟哪个 IPP 合作，他们同时建设负载和消耗端。即使通过电网连接以获得更好的稳定性和可靠性，或者完全不连，也就是“表后”（behind the meter，完全不连接电网），像 Elon 的 Colossus 数据中心的一部分，或者 OpenAI 在得州 Abilene 项目的一部分，或者 Crusoe Energy。电力生产商有很大的空间获得超额回报。

[原文] [Dylan Patel]: i'm not necessarily bullish nuclear um existing nuclear fine yeah it'll it'll it can find a higher buyer higher priced buyer but majority of it will be gas but like you can do like renewables backed by gas and then just turn off the gas and like it's cost more but whatever right or you can do wind backed by gas and why not nuclear takes too long takes too long no one can build nuclear fast even China takes like 5 years to build nuclear right like it's it's complicated it's unsafe right you know I love nuclear I wish it would work it's just not relevant in the time scale that like AI's power is going crazy

[译文] [Dylan Patel]: 我不一定看好核能。现有的核能还好，是的，它能找到出价更高的买家。但大部分将会是天然气。不过你可以做“天然气兜底的可再生能源”，然后把天然气关掉，雖然成本更高，但无所谓对吧？或者用天然气兜底风能。为什么不是核能？因为太慢了，太慢了。没人能快速建设核能，即使是中国建核电站也要花 5 年时间。它很复杂，（在公众眼里）不安全。你知道，我热爱核能，我希望它能行，但在 AI 算力疯狂增长的时间尺度下，它是不相关的。

[原文] [Dylan Patel]: um but yeah there's a lot of interesting stuff like have clients would like had a client buy a coal plant and we were advising them on the transaction based on they just like showed up and they're like "Yeah we want to buy we want to buy power assets we believe in this power story." It's like "Okay great." So yeah so here's all of the like power plants that we know of like you can get some of it from EIA blah blah blah um which are these like and then we like worked through the economics and we looked at the new data centers being built in the region and all this and then they decided to buy a coal plant and they restarted it and they're like making tons of money now because now someone a certain hyperscaler wants to buy the entire pipeline of power and put a load load near it right instead of just being a grid connected asset so it's like a super awesome investment so like you know power is power is going to do great

[译文] [Dylan Patel]: 不过是的，有很多有趣的事情。比如有个客户买了一个燃煤电厂，我们为这笔交易提供咨询。他们当时找过来说：“是的，我们想买，我们想买电力资产，我们相信电力的故事。”我说：“好极了。”然后我们列出了所有我们知道的电厂，你可以从 EIA（能源信息署）等渠道获得信息。然后我们分析了经济性，看了该地区正在建设的新数据中心等等。最后他们决定买下一个燃煤电厂并重启了它。他们现在赚得盆满钵满，因为某个超大规模厂商（Hyperscaler）想买下整个电力管道，并在附近放置负载，而不是仅仅作为一个并网资产。所以这是一个超级棒的投资。正如你所知，电力行业将会表现极好。

[原文] [Matt Turck]: yeah i was going to talk about peace dividends of the whole AI boom uh generally yes right like hyperscalers are paying for uh transmission grid upgrades which people will benefit from right or like you know investors are obviously going to benefit people who work in the industry electricians wages are skyrocketing you know etc right like plumbers wages are skyrocketing so there's like a lot of trades that are doing really well too I think that's definitely also um part of it

[译文] [Matt Turck]: 是的，我本来想谈谈整个 AI 繁荣带来的“和平红利”（peace dividends）。

[Dylan Patel]: 总体上是的。比如超大规模厂商正在为输电网升级买单，大众将从中受益。投资者显然会受益。在这个行业工作的人，电工的工资正在飙升，水管工的工资也在飙升。所以有很多行业也都做得很好，我认为这也是其中的一部分。

[原文] [Matt Turck]: yeah I wanted to come back quickly to uh that um Nvidian core wave deal that you mentioned as we sort of close the discussion on uh on capex and a and a bubble it seems like there is circular deals but also a lot of debt kind of like flushing around so I don't know the specifics of of that deal but like I did hear variations of this where effectively you have a large player guaranteeing the debt being the last recourse uh for a lot of infrastructure build is sort of uh this plus the whole like oracle commitment there there is a fragility into this whole thing that can be a little unnerving what do you make of it

[译文] [Matt Turck]: 是的，我想快速回过头来谈谈你提到的 Nvidia 和 CoreWeave 的交易，作为我们关于资本支出和泡沫讨论的结尾。这看起来像是“循环交易”（circular deals），同时还有很多债务在流动。我不清楚那笔交易的具体细节，但我听过这类变体，即实际上有一个大玩家在担保债务，作为许多基础设施建设的最后追索权（last recourse）。这加上 Oracle 的承诺等等……整件事有一种脆弱性，让人有点不安。你怎么看？

[原文] [Dylan Patel]: i think it's like completely fine and I think like people are like freaking out and making narratives where there really is shouldn't be one it's like well okay Google doesn't have enough data center capacity they need people to build data centers but no one can build a data center because they don't have the capital like don't have you know many cases capital is not the you know they don't have capital right or like no one will give them a loan because they don't trust some random fucking company and it's like but then Google's like well no we've due diligence to them we think they can build it here we'll like even guarantee we'll buy the thing or start using it once they build it you know just having a customer alone spoken for it was enough right

[译文] [Dylan Patel]: 我觉得这完全没问题。我认为人们是在大惊小怪，在不该有故事的地方编造叙事。情况是这样的：好吧，Google 没有足够的数据中心容量，他们需要人来建数据中心，但没人能建，因为他们没有资本。在很多情况下，他们没有资本，或者没人愿意给他们贷款，因为银行不信任某家随机的破公司。但 Google 会说：“不，我们对他们做了尽职调查，我们认为他们能建。看，我们甚至会担保：一旦他们建好，我们就会买下或开始使用。”你知道，仅仅是有一个客户承诺使用就已经足够了。

[原文] [Dylan Patel]: um in the case of Cororewave they were actually able to no backs stop right right they were able to just say "Hey hey look here's our Microsoft contract for this many GPUs i want to put in that data center that data center that data center here's the contract for renting those GPUs i want to hire these people i want to do this." No one will like they don't have any money but then they were able to like have it work out because they were able to get people to lend to them i think like Cororeweave did that and there was no circular financing but that was when there was like the scale of investment was like single digit billions or less than a billion right now the scale of investment is hundreds of billions um and so the question is like oh well if I want data center capacity how do I how do I get data center capacity i just go to everyone who's going to build it looks smart is smart enough to do it but can't afford to do it and tell them I'll I'll take it and in fact I won't just take it i'll go to your debtor and be like I'll guarantee you yeah because you know obviously you're a new company i've vetted you but the debtor hasn't

[译文] [Dylan Patel]: 在 CoreWeave 的案例中，他们实际上甚至不需要兜底（backstop）。他们可以直接说：“嘿，看，这是我们的 Microsoft 合同，包含了这么多 GPU。我想把它们放进那个、那个、还有那个数据中心。这是租用这些 GPU 的合同。我想雇这些人，我想做这个。”虽然他们没钱，也没人愿意借钱，但他们最终搞定了，因为他们能拿着合同让别人借钱给他们。我认为 CoreWeave 以前就是这么做的，那时并没有所谓的“循环融资”，但那时的投资规模是个位数（十亿）或更少。而现在投资规模是数千亿。所以问题变成了：如果我想要数据中心容量，我该怎么做？我就去找每一个看起来聪明、有能力建但没钱建的人，告诉他们：“我会包下的。”事实上，我不只是包下它，我还会去找你的债权人说：“我来担保。”因为显然你是一家新公司，我审查过你，但债权人没有。

[原文] [Dylan Patel]: and so you know like you know you know they don't want me to just be able to walk away because like in the Microsoft Corwave deals Microsoft could have walked away if Corwe fucked it up right there's no I mean yeah there's there's always like uh sort of like cancellation or whatever possibilities and so this is just a further form of guarantee um as far as on like a lot of these back stops as far as on like Oracle getting the money and then OpenAI getting money and Nvidia you know paying and it's a whole circular it's kind of nonsense because it's like Nvidia's getting equity in OpenAI they're basically saying "Hey every gigawatt you buy we'll also buy some equity." Yeah right okay well cool now Nvidia owns an asset which they think is valuable openai right open AAI is turning around and is like trying to rent those uh use the equity they buy what do they what was their use of equity people's cash pay isn't that great right it's mostly just 99 plus% of their spend at the company is probably just compute

[译文] [Dylan Patel]: 债权人不希望我（大客户）能直接一走了之。因为在 Microsoft 和 CoreWeave 的交易中，如果 CoreWeave 搞砸了，Microsoft 本来是可以走人的，虽然总是有取消条款之类的可能性。所以这（现在的模式）只是一种更深层的担保形式。至于那些关于 Oracle 拿钱、OpenAI 拿钱、Nvidia 付钱这种“完全循环”的说法，其实有点胡扯。因为 Nvidia 实际上获得了 OpenAI 的股权。他们基本上是在说：“嘿，你每买一吉瓦（的芯片/算力），我们也买一些股权。”好吧，酷，现在 Nvidia 拥有了一项他们认为有价值的资产——OpenAI。OpenAI 转过身来，试图租用……用他们卖股权换来的钱做什么？他们的人员现金工资也没那么高，对吧？公司 99% 以上的支出可能只是算力。

[原文] [Dylan Patel]: uh so so sort of like it's like okay well then I I raise this money i'm going to do the the whole thing I explained earlier right year one and two I lose money year three four five I hope to make money on it right um and open has been doing that right so I'm going to Okay I'm going to go out there i've raised $50 billion i've raised $10 billion i'm going to raise it i'm going to rent a cluster for five years for $65 billion and I've rented that contract and now I only have enough to pay for the first year to be clear but I think you know you trust me Oracle you think I'm going to grow and you think I'll be able to pay for it oracle's like "Yeah or if you're not I think I'll be able to sell it to someone else." So like okay cool i'm going to spend $50 billion this year yep to build that data center and and and this these this is like for a gigawatt um and so is it like circular that OpenAI is every amount of GPUs they consume and gives an investment that investment is turned around to pay for the first year of the rent to the cluster um or second year then first two years go you know it's sort of like it's fine yeah yeah like it's like it's like it is a little bit funky but like I don't think it's a big deal

[译文] [Dylan Patel]: 所以这就像是：好吧，我筹集了这笔钱，我要做我之前解释过的整个事情。第一年和第二年我亏钱，第三、四、五年我希望能赚钱。OpenAI 一直在这样做。好的，我已经筹集了 500 亿或 100 亿美元。我要去租一个为期五年的集群，价值 650 亿美元。我签了合同，但我现在只有足够的钱付第一年的租金，说清楚点。但我觉得你信任我，Oracle，你认为我会增长，你认为我能付得起。Oracle 会说：“是的，或者如果你付不起，我觉得我能把它卖给别人。”所以，好吧，酷，我今年要花 500 亿美元来建那个数据中心，这是一个吉瓦的规模。那么，OpenAI 消耗的每一个 GPU 都对应一笔投资，而这笔投资又转过头来支付集群的第一年租金，这算循环吗？或者第二年？这有点……这没问题。是的，这看起来是有点怪（funky），但我认为这不是什么大事。

本章标志着访谈进入了对未来的终极展望。Dylan 提出了一个颠覆性的概念：“代币经济学”（Tokenomics）应被重新定义为追踪 AI 模型使用率的指标，而非加密货币术语。

他通过一个震撼的实例（非程序员分析师用 Claude Code 独立完成卫星图像与财报的交叉分析）宣告了“低级脑力劳动的终结”。他大胆预言，初级分析师和 L4 级软件工程师的职位将变得毫无意义，因为 AI 已经能更高效地完成数据清洗、图表制作甚至全栈编程。

章节 9：代码的终结：模型进化与“代币经济学”

📝 本节摘要：

在本章中，Dylan 首先“从加密货币圈夺回”了 Tokenomics（代币经济学） 一词，将其定义为通过 Token 使用量来追踪 AI 采用率的新学科。

>

他分享了一个极具冲击力的案例：他公司的一位非程序员分析师，仅通过向 Claude Code 下达指令，就完成了从卫星图像中计算晶圆厂洁净室面积、抓取财务报表、处理数据并生成图表的复杂任务。这导致了一个残酷的结论：初级分析师（Junior Analysts）和 L4 级工程师不再被需要。

>

此外，Dylan 对模型演进进行了预测。他认为 OpenAI 在强化学习（RL）栈上领先，而 Anthropic 在预训练（Pre-training）上更强。一旦双方互补短板，模型能力将再次飞跃。他认为像 Claude Code 这样的工具标志着工作方式的永久性改变。

[原文] [Matt Turck]: yeah love it contrary intake maybe let's finish with the models and the software side of things we talked extensively about hardware and supply chain and all the things i get a sense that you are super super bullish on uh what's happening next in in AI your roommate Schulto I assume was the roommate that you were talking about earlier on this pod effectively making the point that we're just starting to scratch the surface and there was so much low hanging fruit around you know RL and all the things you were in Silicon Valley circles is that is that your sense as well and what are you tracking on the model side one thing is like you know simple stuff like uh GitHub commits other things are like what's the amount of usage how much are people using like all these sorts of things

[译文] [Matt Turck]: 是的，我很喜欢这个反向观点。也许让我们以模型和软件方面来结束话题。我们广泛讨论了硬件和供应链，我有种感觉，你对 AI 接下来的发展超级超级看好。你的室友 Sholto——我猜就是你之前在播客里提到的那位室友——实际上提出了一个观点，即我们才刚刚触及表面，在强化学习（RL）等方面还有很多唾手可得的果实（low hanging fruit）。你身处硅谷圈子，你也是这么感觉的吗？你在模型方面追踪什么指标？比如 GitHub 提交量这种简单的东西，还是使用量、人们的使用频率等等？

[原文] [Dylan Patel]: i think there's so many different alternative data sources for tracking AI model progress area tokconomics uh token economics tokconomics and so that's like an entire practice for us

[译文] [Dylan Patel]: 我认为有很多不同的替代数据源可以追踪 AI 模型的进展。比如“代币经济学”（Tokenomics），这是我们的一项完整业务。

[原文] [Matt Turck]: are you rebranding the term from crypto

[译文] [Matt Turck]: 你是从加密货币圈重新定义（rebranding）了这个词吗？

[原文] [Dylan Patel]: i yeah I don't believe in crypto people like I've always hated them um so now you're taking the term yeah yeah and Jensen's used it now so I've like I've convinced him to use the word he's used it as sovereigns and so I think I think we've won that's awesome congratulations i've said it to him we've written it in articles it's an entire practice of consulting that I just I started in like 23 2023 uh was token economics and we've been trying to build out these like you know

[译文] [Dylan Patel]: 是的，我不相信加密货币圈的人，我一直很讨厌他们。

[Matt Turck]: 所以你把这个词夺过来了？

[Dylan Patel]: 是的，是的。而且 Jensen（黄仁勋）现在也用这个词了，我已经说服他使用这个词，他把它用在了主权 AI（Sovereigns）上。所以我认为我们赢了。

[Matt Turck]: 太棒了，恭喜。

[Dylan Patel]: 我对他说过，我们在文章里也写过。这是我在 2023 年开始的一项完整的咨询业务，就是“代币经济学”，我们一直在努力构建这些东西。

[原文] [Dylan Patel]: but basically I think the main things are like people who don't code can use cloud code now right i think people don't understand that like even if you don't code you've never had any training in software development you've never take had a job as a software developer you can code let's take an an example of what one of the one of the analysts at my company did right comes from a engineering background but on like semiconductor systems right uh like worked on mechanical systems worked on these sorts of things and they coded this thing which was they wanted to do an analysis of area of clean rooms right clean rooms are the building that you the fab has all the tools in the most complicated kind of building in the world has every all sorts of chemical systems and all this

[译文] [Dylan Patel]: 但基本上，我认为最主要的事情是：不会写代码的人现在可以使用 Claude Code 了。我认为人们不理解这一点：即使你不写代码，你从未受过任何软件开发培训，你从未做过软件开发的工作，你也可以编程。举个例子，我公司的一位分析师做了什么。他有工程背景，但是是半导体系统方面的，比如机械系统之类的。但他“编码”出了这个东西：他想分析洁净室（clean rooms）的面积。洁净室是晶圆厂放置所有工具的建筑，是世界上最复杂的建筑，拥有各种化学系统等等。

[原文] [Dylan Patel]: area of that a company who builds systems builds these systems and revenue of that company right and so it was like okay uh we have this fab data set pointed it at it was like hey here's this fab data set what's the square footage of all of them and we have this like thing that we built which uh just pulls with cloud code separately which for data centers and and and fabs and everything else just calculates the area of something from a from a satellite image right very simple so we have the square footage of all these things points at that here's the company name okay go find the filings so it dig dug through all these filings it it pulled the data right okay great now told it to um compare these two make a chart great

[译文] [Dylan Patel]: 他想分析建造这些系统的公司的洁净室面积以及该公司的收入。所以过程是这样的：好吧，我们有这个晶圆厂数据集，指给它（Claude Code）看，说：“嘿，这是晶圆厂数据集，它们所有的平方英尺面积是多少？”我们有一个用 Claude Code 独立构建的工具，可以从卫星图像计算数据中心、晶圆厂等设施的面积，非常简单。所以我们有了所有这些东西的平方英尺面积。然后指出公司名称：“好的，去找到财务文件（Filings）。”于是它挖掘了所有的文件，提取了数据。好的，太棒了。然后告诉它：“比较这两个数据，做一个图表。”太棒了。

[原文] [Dylan Patel]: oh wait there's this like weird inflection oh that's because they bought a company five years ago can you do a proform of this analysis without those financials of that of that company they acquired okay great and then like we were able to like like figure out an investment case for our clients as well as like you know some other interesting details from someone who's never really coded just using clawed code and it like doing this all and this is like not even their and it wrote the note and they just like they didn't even like work on this full-time for like 3 hours right they just told the model and would go work on other things and told the model and worked on other things they just did this

[译文] [Dylan Patel]: “噢等等，这里有个奇怪的拐点。噢，那是因为他们五年前收购了一家公司。你能做一个剔除那家被收购公司财务数据的预估分析吗？”好的，太棒了。然后我们就能够为我们的客户搞定一个投资案例，以及发现一些其他有趣的细节。而这来自一个从未真正写过代码的人，只是使用 Claude Code。它做了一切，甚至写了报告（Note）。而这个人甚至没有全职做这个，大概只花了 3 个小时？他们只是告诉模型指令，然后去忙别的，回来再告诉模型，再去忙别的。他们就这样搞定了。

[原文] [Dylan Patel]: people don't understand that like the skill sets that like I think like if you go talk to an analyst right a very junior analyst at any right whether it's venture or especially growth venture or public markets or private equity their their job is like finding data cleaning it making charts it's like this is cloud code now you don't need junior analysts just like a lot of companies have stopped hiring L4 engineers because it's useless why would I hire an L4 engineer i just tell Claude to do it

[译文] [Dylan Patel]: 人们不理解技能组合的变化。如果你去和任何一个初级分析师（Junior Analyst）交谈——无论是在风投，特别是成长型风投，还是公开市场或私募股权——他们的工作就是找数据、清洗数据、制作图表。这现在就是 Claude Code 的活儿了。你不需要初级分析师了。就像很多公司已经停止招聘 L4 级工程师一样，因为没用了。我为什么要雇一个 L4 工程师？我只要告诉 Claude 去做就行了。

[原文] [Dylan Patel]: you you sort of like have this has happened and this is a really big like shift I guess like is that like low-level knowledge work just doesn't matter right why would I why would I use Excel when I can just tell Claude to manipulate CSVs why would I use Word when Claude will just generate the markdown and I can copy and paste the markdown directly into our WordPress and then you know and that WordPress is fully formatted now and it's like oh my god like what's the point of Word right

[译文] [Dylan Patel]: 这种事情已经发生了，这是一个非常大的转变。我想说的是，低级脑力劳动已经不再重要了，对吧？当我可以告诉 Claude 去操作 CSV 文件时，我为什么要用 Excel？当 Claude 可以直接生成 Markdown，我可以把它直接复制粘贴到我们的 WordPress 里，而且格式完全正确时，我为什么要用 Word？这就像是，天哪，Word 还有什么意义？

[原文] [Dylan Patel]: um and what's the point of doing all sorts of stuff i think when we look at model progress that's just for Opus 4.5 open's new model I think will be better than Opus 4.5 and it's coming like somewhat soon in Marchish um time frame i maybe February Marchish but yeah um because OpenAI has a better RL stack than Enthropic today it's just their pre-trained models suck compared to Enthropic's pre-training right

[译文] [Dylan Patel]: 做各种琐事还有什么意义？当我们观察模型进展时，这还只是 Opus 4.5（注：可能是指 Claude 3.5 Sonnet 或 Opus 的下一代）。我认为 OpenAI 的新模型会比 Opus 4.5 更好，而且它很快就要来了，大概在三月左右的时间框架，也许二月或三月。因为 OpenAI 目前拥有比 Anthropic 更好的强化学习（RL）栈，只是他们的预训练模型相比 Anthropic 的预训练要差劲（Suck）。

[原文] [Dylan Patel]: and so like if they catch up a lot on pre-training and keep their better RL stack they would actually have a model that's much better right flip side Google has a better pre-trained model than Anthropic or OpenAI but their RL stack sucks so if they catch up on RL like these models are going to get ridiculously and then Anthropic is obviously advancing as well right and so and then and then you look across the ecosystem everyone's advancing really fast progress these moments are happening right you know chat GPT was a moment gibbly was a moment those were more consumer those were less like I mean there chat GBT is everyone using it for work too but like I think cloud code is like a new moment right 4.5 on cloud code is a new moment where the way you work has forever changed

[译文] [Dylan Patel]: 所以，如果他们在预训练上大幅追赶，并保持他们更好的 RL 栈，他们实际上会拥有一个好得多的模型。反过来说，Google 拥有比 Anthropic 或 OpenAI 更好的预训练模型，但他们的 RL 栈很烂。所以如果他们在 RL 上追上来，这些模型会变得强得离谱。显然 Anthropic 也在进步。当你环顾整个生态系统，每个人都在飞速进步。这些“时刻”正在发生。你知道 ChatGPT 是一个时刻，Gibeley（可能指某种图像生成工具或口误）是一个时刻，那些更多是消费者层面的。虽然大家也用 ChatGPT 工作，但我认为 Claude Code 是一个新的时刻，4.5 版的 Claude Code 是一个新的时刻，你工作的方式已经被永远改变了。

[原文] [Dylan Patel]: and so now we're trying to force everyone in my company there's 54 people here i think like half of them have coded the other half we're trying to force them to use like cloud code and it could be like oh well actually you come from a consult a semiconductor consulting background oh you come from like a semiconductor like engineering of like package oh you worked in a fab right like these kind of people they're using cloud code now right and and their productivity is being boosted

[译文] [Dylan Patel]: 所以现在我们正试图强迫我公司的每个人——这里有 54 个人，我想一半人会写代码，另一半我们正试图强迫他们使用 Claude Code。这可能是：“噢，实际上你是半导体咨询背景；噢，你是做封装工程的；噢，你在晶圆厂工作过。”像这类人，他们现在正在使用 Claude Code，他们的生产力得到了极大的提升。

本章作为访谈的尾声，气氛变得轻松而充满“硅谷特色”。Dylan 分享了与 AI 圈知名人物合租的趣闻，通过室友 Sholto “烧钱”开发游戏的疯狂案例，再次印证了 AI 编程能力的质变。同时，他也聊到了另一位室友 Dwarkesh Patel 的硬核工作态度。

章节 10：硅谷轶事：与 AI 研究员的合租生活及结语

📝 本节摘要：

在最后一章中，Dylan 分享了极具画面感的硅谷生活片段。他谈到了自己的室友 Sholto Douglas（Anthropic 研究员），在假期利用 Claude Code 花费了 1 万美元的 API 额度，在一周内“口述”完成了一款以“中美 AI 竞赛”为主题的 RTS（即时战略）游戏，全程没有手写一行代码。这一案例生动地展示了当今 AI 辅助编程的恐怖效率（目前约 5% 的代码已由 AI 生成）。

>

此外，Dylan 还聊到了另一位室友 Dwarkesh Patel（知名播客主持人），形容他是“播客界的播客”，为了准备采访甚至会亲自去训练模型。尽管这群室友是技术狂人，日常对话也充满了普通人的色彩——从吐槽糟糕的约会经历到一起打《帝国时代 2》。访谈在对技术乐观主义的氛围中结束。

[原文] [Dylan Patel]: yeah my uh one of my roommates I was asking him because he's like always been a really low-level good programmer and he started you know I was like he's like he had this um holiday obsession right I mean he was using cloud code for work already right like whatever um but he had this holiday obsession we got into playing Age of Empires 2 myself you know my roommate a handful of people from like Open Eye GDM Anthropic we just would do land parties of AoE 2 over the holidays a bit not not like Christmas but like a little bit before a little bit after you know cuz most of us went home for Christmas

[译文] [Dylan Patel]: 是的，我问过我的一个室友，因为他一直是一个非常优秀的底层程序员。他有过这种“假期痴迷”（holiday obsession）。我是说，他已经在工作中通过 Claude Code 及其它工具使用了，这无所谓。但他有了这种假期痴迷：我们开始玩《帝国时代 2》（Age of Empires 2），我自己、我室友，还有几个来自 OpenAI、Google DeepMind、Anthropic 的人。我们在假期里搞过几次《帝国时代 2》的局域网派对（LAN parties），不是在圣诞节当天，而是前后几天，因为我们大多数人圣诞节都回家了。

[原文] [Dylan Patel]: um but like we'd do these lands my roommate got so obsessed with like the game that during Christmas week cuz he didn't go home he just stayed in San Francisco um he just worked on an RTS game and he built an entire RTS game and I think I kid you not I think he he used like $10,000 of Claude in one week and built an entire RTS from scratch uh about a like but instead of like being a standard RTS where it's like oh Age of Empires for advance through ages or Starcraft it is it is an RTS where it's China versus the US and you're in the AI race and you go from the start of the information age all the way through to you know AGI and like robots and humanoids and and and like all like space fairing civil like it's crazy he built it in a week and he didn't type a single line of code right he can only dictate it to the model

[译文] [Dylan Patel]: 总之我们搞了这些局域网派对，我室友对这个游戏太着迷了，以至于圣诞周他没回家，就待在旧金山。他着手做了一个 RTS（即时战略）游戏，他构建了整个 RTS 游戏。我没开玩笑，我觉得他在一周内用了大概 10,000 美元的 Claude 额度，从零开始构建了一个完整的 RTS 游戏。但这不像标准的 RTS，比如《帝国时代》那样通过时代进化，或者《星际争霸》。这是一款中国对战美国的 RTS，你身处 AI 竞赛中，从信息时代开始，一路发展到 AGI、机器人、人形机器人，以及太空文明。这太疯狂了，他一周就做出来了，而且他没敲一行代码，他完全是口述给模型去做的。

[原文] [Dylan Patel]: and he told me yeah like we have an indicator internally at Enthropic where you see how many people actually write code now there's only a few hold outs left... probably like 5% of code committed today is AI generated if not higher marked as AI generated what's going to happen when normal workers who do spreadsheets and office processing start automating their workflows i think it's a whole new world

[译文] [Dylan Patel]: 他告诉我说，是的，我们在 Anthropic 内部有个指标，看现在有多少人还在真正手写代码，只剩下几个顽固分子了……如今提交的代码中可能大概有 5% 是 AI 生成的，如果不是更高的话（指标记为 AI 生成的）。当处理电子表格和办公室流程的普通工人开始自动化他们的工作流时会发生什么？我认为那将是一个全新的世界,。

[原文] [Dylan Patel]: and speaking of Schultoe we both agreed that he was a a perfect specimen dude I' I've been I'm straight but I've been accused of being uh homosexual which is perfectly fine for for how much I like praise this man because like think about it right he's like 6'4 he's like really good-looking he's like Australian accent sounds amazing like you've heard his I I have like a annoying voice probably his voice sounds amazing he's absurdly good at coding he was an Olympian level fencer like like he picks up any sport he's really good at it right because he's athletic it's like "Holy crap you're a specimen."

[译文] [Dylan Patel]: 说到 Sholto，我们都同意他是个“完美的样本”（perfect specimen）。老兄，我是直男，但我曾被指责是同性恋——这完全没问题——因为我太喜欢赞美这个男人了。你想想看，他身高 6 尺 4（约 1.93 米），长得真帅，那是澳大利亚口音，听起来太棒了——你知道，我的声音可能挺烦人的，但他的声音听起来很棒。他写代码强得离谱，他还是奥运会级别的击剑手。他拿起任何运动项目都能玩得很好，因为他很有运动天赋。这就像：“天哪，你简直是个完美的样本。”

[原文] [Matt Turck]: Yeah yeah this clip and sent him for sure yeah it must be uh you know I guess uh may maybe some people don't follow the playbyplay on on Twitter and like don't haven't haven't heard of like the fact that all of you guys are roommates or you roommate with Scholto and then with Dwarish and Darkish is like the podcasters podcaster so it must be absolutely What's a podcasters podcaster mean

[译文] [Matt Turck]: 是的，这段剪辑肯定要发给他。我想可能有些人在 Twitter 上没跟得那么紧，没听说过你们大家是室友这个事实，或者是你和 Sholto 以及 Dwarkesh 合租。Dwarkesh 就像是“播客界的播客”（podcaster's podcaster），这绝对是……等等，“播客界的播客”是什么意思？

[原文] [Matt Turck]: uh the podcaster that other podcasters uh aspire to to to become or learn from yeah yeah his his when he's preparing you know it's like he's he's he's he's so locked in and he prepares so hard for interviews it's great

[译文] [Matt Turck]: 意思是其他播客渴望成为或从中学习的那种播客。是的，当他准备的时候，你知道，他非常专注（locked in），他为采访做了极其刻苦的准备，这很棒。

[原文] [Dylan Patel]: no he's he's just uh incredible and then and then he might only say like a hundred words on the episode but he's prepared so hard and then like I think people just realized oh wow he's not just like you know it's like oh he just has good guests no no no like he's preparing really hard but you can't tell if you're not like realizing that and then once he started writing more and he started writing more people like oh wow he's actually really really smart it's like yeah cuz he's studying like crazy like it's like "Oh I'm interviewing an AI researcher who worked on this i'm gonna try and train a freaking model." Yeah right it's like that's the level of like commitment he goes to when he records this stuff

[译文] [Dylan Patel]: 不，他简直不可思议。在节目里他可能只说了一百个字，但他准备得非常辛苦。我想人们才刚刚意识到，哇，他不仅仅是请到了好嘉宾，不不不，他准备得非常努力，但如果你没意识到这一点你是看不出来的。一旦他开始写更多东西，人们就会觉得：“哦哇，他实际上真的非常非常聪明。”那是当然，因为他学得像疯了一样。就像：“噢，我要采访一位研究这个的 AI 研究员，那我要自己试着去训练一个该死的模型。”是的，这就是他录制这些内容时的投入程度。

[原文] [Matt Turck]: what do you guys talk about when you bump into each other is that is that AI non-stop or you talk about everything but AI

[译文] [Matt Turck]: 你们碰面时都聊些什么？是不停地聊 AI，还是除了 AI 什么都聊？-

[原文] [Dylan Patel]: with Shoto it's like the Age of Empires game you know because we we got super into it for a bit we talked only about that in his RTS that he made uh with with with Dwarash it's I mean it's all sorts it's like normal roommate stuff it's like "How's your dating life?" "Oh okay you went on a date it wasn't well it didn't go well." "Okay well okay." Yeah you know like oh you know like that's me that's me you know my days don't go well no I'm just kidding um or like it's like oh you want to like have dinner we can invite a few friends like yeah great or like you know it's like all sorts of like normal stuff too um al obviously we also do talk about a lot about tech right like we are like this is our lives um and tech is the most fun thing

[译文] [Dylan Patel]: 和 Sholto 在一起时，主要是《帝国时代》游戏，因为我们有一阵子特别迷那个，我们就只聊那个和他做的 RTS 游戏。和 Dwarkesh 在一起时，就是各种各样的，普通的室友话题。比如：“你的约会生活怎么样？”“噢，好吧，你去约会了，结果不太好？好吧。”你知道，就像……噢，其实那是我，那是我的约会不太顺利，不，我开玩笑的。或者像是：“噢，你想吃晚饭吗？我们可以叫几个朋友来。”“好的，太棒了。”各种各样的普通话题也有。当然，我们也确实聊很多科技，因为这就是我们的生活，而且科技是最有趣的事情。

[原文] [Matt Turck]: awesome well great great San Francisco lore uh Dylan thank you so much uh that was absolutely fabulous really enjoyed it learned a lot so really appreciate uh your coming on the pub thank you so much

[译文] [Matt Turck]: 太棒了，很棒的旧金山传说。Dylan，非常感谢，这绝对是太精彩了，我很享受，也学到了很多，真心感谢你来上这个播客。非常感谢。