本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
大家好。
Everybody.
欢迎回到《Semi Analysis Weekly》第四期。
Welcome back to episode number four of Semi Analysis Weekly.
我是乔丹·纳诺斯，再次和道格·奥劳林一起。
I'm Jordan Nanos here with Doug O'Laughlin again.
这周我们请到了杰里米·埃利亚胡·翁蒂韦罗斯。
And this week, we've got Jeremie Eliahou Ontiveros.
我们只得到了他部分的注意力。
We've got part of his attention.
他其余的注意力还留在Semi Analysis数据中心模型的办公室里，此刻杰里米正同时在四个窗口里写代码。
The rest of his attention is back at the offices of the semi analysis data center model where Jeremie is clot coding four separate windows at a time right now.
你能稍微看向镜头，自我介绍一下吗？
Can we can we get you to look at the camera for a second and introduce yourself?
兄弟，让我写代码吧。
Bro, let me write code.
这很有价值。
It's valuable.
我没时间做这个。
I have no time for this.
嘿。
Hey.
人们相信我。
People believe for me.
是的。
Yeah.
当然。
Sure.
当然。
Sure.
当然。
Sure.
是的。
Yeah.
在我们正式开始之前，对于在场的所有Vibe程序员，如果你们在听的话，我们将在GTC开幕前一天的周日举办一场黑客马拉松，我想是三月。
Just before we we get started, for all the vibe coders in the audience, if you're listening to this, we are hosting a hackathon the day before GTC kicks off Sunday, March, I think.
十四号还是十五号？
Fourteenth, fifteenth?
我不确定。
I don't know.
不管具体是哪天。
Whatever it is.
在GTC开幕前的那个周日，我们会现场发放计算资源的使用权限、云代码令牌，以及来自众多优秀赞助商的大量精彩奖品。
The Sunday before GTC, yeah, we're gonna be on-site giving out grants for access to compute as well as cloud code tokens and lots of cool prizes from lots of great sponsors.
所以，如果你关注Semi Analysis并且会参加GTC，一定要报名参加这场黑客马拉松。
So if you're following semi analysis and you're attending GTC, definitely, yeah, sign up to be part of the hackathon.
这会很有趣。
It's gonna be fun.
你们对GTC期待吗？
You guys you guys excited for GTC?
是的。
Yeah.
什么好笑？
What's funny?
我不去，兄弟。
I'm not going, bro.
所以不是我。
So not me.
道格，你不去哪儿？
Doug, where aren't you going?
因为大家都去GTC，所以我改去OFC。
Because everyone's going to GTC, so I'm going to OFC instead.
好的。
Okay.
我给你
I'll give you
打个电话。
a call.
我们得在另一个地方也有些存在感。
We have to have some presence at the other place.
你知道的吧？
You know?
比如，我们有没有计量器欢乐时光？
Like, do we have, a meter happy hour?
他们就说：哦，你就打算派些残次品去？
And they're like, oh, you're gonna just send the dregs?
不。
No.
总得有人去露个面。
Someone's gotta show up.
我想我说的其实就是那些残渣。
I I guess I'm the dregs is what I'm saying.
不。
No.
但说实话，我觉得这里面有很多情况。
But I think there's a lot of I I'll be honest with you.
我觉得你们肯定能完全掌控住。
You guys you guys are gonna have it totally under control, I think.
我们都深度参与了GTCs。
We're all in pretty deep to GTCs.
是的。
Yeah.
我们有个庞大的团队。
We got a big crew.
负责Inference X的同事们将带来一场精彩的演示。
The guys have been working on Inference X are gonna be presenting a nice session.
如果你们在听的话，去GTC目录里搜索一下这个。
If you guys are listening, search up that in the GTC catalog.
我非常期待由前创始人兼CEO布罗克·乔纳森·布罗克主讲的关于新型AI GPU的讲座。
I'm real excited for the help new art GPU session presented by the former founder CEO, Brock, Jonathan Brock.
而且是的。
And Yeah.
LPU讲座可能会是场场爆满的讲座。
LPU session is gonna be, I think, probably gonna be the the sold out session.
对吧？
Right?
就像人们一大早就排队等候。
Like, people standing in the morning.
这绝对是最重要的亮点。
That's definitely the big thing.
人们对SRAM变得非常激动。
People are getting really jumpy about SRAM.
他们觉得因为每个人都特别认为，这只是一个SRAM。
They're like because everyone is super like, this is just an SRAM.
现在大家都极度看好HBM。
Everyone's so long HBM right now.
他们都害怕SRAM会取代HBM，而你得卖HBM。
They are all scared that s s Ram is gonna take over HBM, and you gotta sell the HBM.
这可能是我最近听过的最荒谬的论调。
That's like the the dumbest narrative I've heard recently.
但人们太害怕了，因为他们都重仓HBM。
But it's people are so so so scared because they're so so so long HBM.
所以，嗯，这确实是个挺有意思的……我不确定怎么说。
So, yeah, that's that's kind of been the the like an interesting I don't know.
人们觉得LP会很有趣。
People are LP is gonna be interesting.
是的。
Yeah.
太棒了。
Sweet.
好的。
Okay.
这周的节目，我们计划讨论三个重要话题。
Well, for the show this week, we got a plan to talk about three big topics.
第一个是我们发布在免费通讯中的文章。
The first is the article that we put out on the free tier newsletter.
杰里米会主要来谈这个。
Jeremie's here to talk about that primarily.
这篇文章的标题是：人工智能数据中心是否推高了消费者的电价。
Title of that article is are the AI data centers increasing consumers' electricity prices.
负责我们数据中心模型、能源模型、工业模型的那位负责人，以及整个团队，可能是探讨这个问题的最佳人选。
So the guy who leads our data center model, energy model, industrials model, that whole team, probably the best one to to pick brain on that.
然后我们会讨论国防部与Anthropic之间的博弈。
Then we're gonna talk about the Department of War versus Anthropic.
无论如何，关于这件事的新闻很多。
Lots of news about that in anyway.
最后，是的，我们内部使用Claude代码的情况越来越多了。
And then finally, yeah, just our internal usage of Claude code going up.
我们最近都在做仪表盘。
We're all we're all working on dashboards these days.
你知道的。
You know?
我们过去卖的那些模型就像电子表格和数字，但现在我们需要在仪表盘上实现实时可视化。
We used to sell these models, which are just like spreadsheets and numbers, and now we need to do real time visualization in the dashboard.
那就是未来。
That's the future.
我们正在为商业洞察构建半分析型仪表盘。
We are building dashboards for business insight as semi analysis.
用于商业洞察的功能性仪表盘。
Functional dashboards for business insight.
那么仪表盘之后是什么呢？
So what comes after dashboard?
那就是未来。
It's the future.
通用人工智能，兄弟。
AGI, bro.
它只是直接传输的。
It's just it's just directly piped.
所以
So
从半分析系统推出的第一个仪表盘是InferenceX仪表盘，我想我们已经知道，现在有人正在内部试点一个功能，就是在仪表盘上方加一个聊天机器人。
the the first dashboard from from semi analysis was the InferenceX dashboard, and I think we already know, which is that the guys are piloting a feature internally right now to have a chatbot on top of the dashboard.
所以你甚至不需要再阅读仪表盘了。
So you don't even need to read the dashboard anymore.
你可以让智能代理向你解释仪表盘的结论是什么。
You can ask the agent to explain to you what the conclusion of the dashboard is.
接下来，你可能就只是戴些耳机或者眼镜，然后直接问它们：嘿。
So next up, you just have, like, I don't know, some earbuds or glasses, and you just ask them, hey.
现在发生什么了？
What's going on?
你走在街上，就会说：帮我抓取这个500兆字节的数据集，尽快搞定，兄弟。
You know, you're walking on the streets, and you're like, scrape this dataset with 500 megabytes of data for me and do it ASAP, bro.
你现在的编码方式怎么样？
How are you vibe coding?
你还在打字吗，杰里米？还是已经升级到语音转文字了？
Are you are you typing still, Jeremie, or have you have you graduated to audio speech to text?
是的。
Yeah.
我还在打字。
I'm still typing.
我这个人有点老派，哥们。
I'm I'm kind of a boomer, man.
我花了一个月才加入云代码狂热。
Took me took me a month to join the Cloud Code Mania.
我
I'm
我其实想用语音转文字，但我觉得办公室里环境太吵了。
I actually I wanna do speech to text, but I feel like the office is a a disruptive place for it.
说实话，我不知道你们有没有像杰里米那样去过办公室。
And honestly, I don't know if you guys are like, Jeremie's been to the office.
乔丹从来没去过办公室。
Jordan's never been to office.
我们今天真是深入骨髓了。
We're actually rolling so deep today.
人多得不得了。
It is completely packed.
是啊。
Like, yeah.
确实是
It is
对
yeah.
大卫得坐在另一个房间。
David has to sit in the other room.
所以我无法想象我们只是大声喊叫，同时还在震动。
So I can't imagine we're just, like, all yelling at each other while we're vibrating.
那可能是最大的障碍。
That's probably the biggest impediment.
如果我能轻声地写代码，真的像那样，那会是最好的。
If I can whisper code, like, literally or something, that would that would be my best.
那也是纽约办公室。
That is the New York City office too.
那么现在哪个办公室更大？
So which office is bigger now?
纽约还是新加坡？
New York or Singapore?
从人数来看，我觉得我们还不错。
In terms of headcount, think we're well, okay.
我觉得还是新加坡。
I think Singapore still.
好的。
Okay.
因为今天有两个访客。
Because there's two visitors today.
有两个访客。
There's two visitors.
埃里克、埃里克和鲍文都在这里。
Eric and Eric and Bowen are here.
是的。
Yeah.
是的。
Yeah.
好观点。
Good point.
好的。
Okay.
我们来谈谈这篇文章。
Let's talk about the article.
我打算共享屏幕，然后我们就可以开始了。
I'm gonna gonna share my screen and, yeah, we can we can kick things off.
所以这篇文章，我认为是对新闻中许多人的回应，特别是特朗普，他谈论了与人工智能数据中心相关的立法，以及是否会增加人们的电费。
So this article, I think, came as, like, a bit of a response to a lot of people in the news who are talking about well, really Trump who was talking about, like, legislation related to AI data centers and how you know, whether or not they're gonna increase people's electricity bills.
显然，这是最新的这篇文章。
Obviously, is the latest article here.
而且，这是一篇不错的文章。
And, yeah, it's a nice one.
七千个词？
7,000 words?
杰里米，有多少个标记？
How many tokens, Jeremie?
这是一篇很长的文章。
It's a big one.
实际上，那是我第一个Cloud Code项目。
Actually, that was my very first Cloud Code project.
你可以看到一些试验性的内容，和我们通常的风格不太一样。
You can see some of the trials that are a little bit not in our usual style.
我用Flockode做的第一件事就是这些分析。
That was the first thing I ever did with Flockode is a lot of these analysis that came up.
但要明确的是，这并不意味着内容是编造的。
But to to be clear, it doesn't mean it's it's made up.
我花了很多时间去真正理解其中的数学和它处理的数据。
I spent a lot of time trying to actually understand the math and the data it it processes.
但这是一个绝佳的教科书式案例，展示了如何使用云教练，因为这些数据很多都是关于消费者或商业用户在各州或按地区划分的电价等等。
But like, this is a great textbook example of how you can use Cloud Coaches because a lot of these data are, like, electricity prices for consumers, for for commercial users, by by state or by by a note or whatever.
实际上，所有这些数据都是免费提供的，因此你只需直接提示它就能获取所有这些信息。
All of that is available for free actually, so you can just basically prompt it to to get all of that.
然后你可以自行核实，确保数据正确，并确认数学计算、容量等相关内容也是准确的。
And then you can just fact check it yourself to make sure it's correct and make sure the math and, like, capacity and such is accurate as well.
是的。
Yeah.
有道理。
Makes sense.
所以你能大致讲讲这篇文章的核心观点吗？
So can you kinda talk through the thesis of the article?
我觉得大致是这样的，是的。
Like, the I think the high level is, yeah.
数据中心上线大量算力确实会对电网造成影响，但具体影响程度取决于电网的管理方式，以及我们讨论的是哪个数据中心或哪个时间阶段，这都会影响到消费者的电价。
There is an impact on the grid for from data centers, from a bunch of capacity coming online, but it kinda depends on how the grid is managed and kinda depends on which data center or which timeline we're talking about in terms of what impact it can have on consumers' prices.
是的。
Yeah.
所以，简而言之，大家普遍认为能源是导致电价上涨的瓶颈，而这全都是人工智能数据中心的错。
So so look, basically, the so the premise everyone is saying is energy is the constraint for a power price that are gonna go up, and it's all the fault of AI data centers.
对吧？
Right?
这在某种程度上是有道理的。
Which kind of makes sense.
但这里面有很多需要注意的地方。
There's a big caveats to that.
但问题是，人们总是过度关注他们看到的短期数据点，而有一个非常引人注目的例子，就是包括弗吉尼亚州、新泽西州、俄亥俄州在内的PJM区域。
But the thing is people over index always on the short term data points that they see, and there's a a pretty flashy one, which is everything in the PJM area, which includes Virginia, New Jersey, Ohio, some of these big states.
这些州的许多人相比两年前的电费上涨了20%。
Many people in these states faced a 20% increase in their bill versus two years ago.
对吧？
Right?
因此，当出现如此大的涨幅，而且又恰逢人工智能时代，人们自然会认为这是数据中心造成的。
And so, obviously, when you have such an increase and it's in the it's in the era of AI, people are saying this is because of data centers.
在某种程度上，确实如此。
And to some extent, is.
但我们这篇文章要表达的观点是，这实际上更多是由市场设计造成的，而其中很大一部分原因在于一些失误。
But the the point we're making the article is that it's actually way more driven by market design, and a lot of that is a function of, to some extent, errors.
某种程度上，我认为设计者并没有很好地适应人工智能的发展。
To some extent, I would argue, a designer is not properly adapted to AI.
所以，归根结底，这更多是过去二十年美国用电负荷零增长的结果。
And so if anything, like, there was like, this is more of a function of, the last twenty years, The US had zero load growth.
这是二十年来首次出现负荷增长，很明显，这个市场并没有为负荷增长的环境做好准备，这样说你能理解吧。
This is the first time in twenty years we see load growth, and it's pretty clear that this market was not designed for an environment where load is growing, if that makes sense.
是的。
Yeah.
完全说得通。
It makes total sense.
当我阅读这份材料时，我觉得最引人注目的图表是这个，它展示了PJM模型如何预测负荷增长以及他们是如何不断修正的。
And when I was reading through this, I think the the chart that, like, stuck out to me the most was this one, which kinda covered the PJM model that they used to forecast what that load growth is going to be and how they've been making revisions.
由于他们通过拍卖方式让能源供应商提前两年确定电价，所以如果预测不准，就会让所有人付出巨额成本。
And because of the way in which they run an auction to let Yeah.
能源供应商，你知道的，提前两年确定他们供电的价格，如果预测失误，就会突然让所有人付出巨额成本。
Energy providers, you know, set the price that they're gonna run power at two years ahead of time, if they miss on this forecast, then all of a sudden, it costs everybody a whole bunch of money.
我们认为，从根本上说，他们的预测准确性很差，因为他们没有正确追踪数据中心的数据，至少不像我们使用自己的数据和数据中心模型时那样精确。
And we think fundamentally, like, the accuracy of their forecast is bad because they're not tracking the data centers properly, at least, like, to the quality with which we can when we use our data and our data center model.
你看。
And look.
另一个大问题是，这一点与ERCOT形成对比，这种机制通常每年只运行一次。
The the other big problem, and this is where we contrast with ERCOT, is that this is an option that is typically run once a year.
需求是基于他们的预测。
Demand is based on their forecast.
供应则是基于他们对供应的分析。
Supply is based on the way they analyze supply.
所有这些都取决于主观判断。
All of that is subject to judgment calls.
因此，你基本上又回到了关于共产主义与资本主义的辩论，某种程度上是这样，因为在ERCOT，市场参与者可以实时评估这些因素。
And so then you basically go back to the debates, communism versus capitalism, right, to some extent because in ERCOT, you basically let market participants evaluate these items on a real time basis.
在PJM，你讨论的是一个拍卖机制。
In PJM, you're talking about an auction.
同样，他们每年只做一次，可能会出错。
Again, once once a year, they can make mistakes.
因此，具备实时反应能力能让市场对信号做出响应。
And so basically, like, having the ability to react real time enables the markets to react to the signals.
对吧？
Right?
而相比之下，ERCOT能够提前预见到情况，并提前安排发电。
Whereas if you wanted to like, the point is ERCOT could have can see it coming ahead of time, can place generation ahead of time.
但在PJM，由于价格信号没有正确设定，一切都滞后了，这就形成了一种受限的机制，最终成为瓶颈。
Whereas in PJM, because the pricing signals are not set properly, everything is just late, and you just create a a mechanism that is constrained and just basically becomes bottlenecked.
是的
Yeah.
也许为了帮助观众理解，ERCOT是德克萨斯州的电力公司，那里现在正有大量数据中心上线。
And maybe just to define for the audience, ERCOT is the utility in Texas where there's plenty of data centers coming online right now.
而PJM是另一个。
And then PJM is the one.
PJM代表宾夕法尼亚、新泽西、马里兰之类的，但它覆盖了从弗吉尼亚到俄亥俄的整个数据中心走廊，也就是美国东部所有主要云服务商的区域。
PJM stands for Pennsylvania, Jersey, Maryland or something, but it covers all of, like, data center alley through Virginia and Ohio, which is, like, The US East one and stuff for all the major cloud providers.
所以，当谈到新建数据中心时，这两个电网是最重要的，它们在定价方式上有着完全不同的做法。
So it's, like, kind of the the the two most important grids when it comes to data centers coming online, and they have that completely different approach to how they set prices.
一个是实时定价，另一个则是每年拍卖一次，但提前两年就确定了价格。
One's real time and one's been auctioned yearly, but it happens two years in advance.
老实说，我都不明白PJM的价格到底是怎么定出来的。
I don't even understand how that price gets set for PJM, to be honest.
对
Yeah.
我的意思是，基本上，通过他们的需求预测，他们会分析自身的需求，然后问题就变成了：能够满足这一需求的边际发电机组是什么，以及这些发电机组需要什么样的价格才能满足容量要求。
I mean, like, basically, via their demand forecast, they analyze what their needs are, and then it becomes a question of, like, what is the what is the marginal generator that can meet that need, and what what prices do these generators need to be able to meet the capacity requirements.
对吧？
Right?
所以，如果你有足够的发电能力可以满足这一特定需求，那就没什么问题。
So if you have plenty of generation that can't serve this specific requirement, then everything is fine.
价格不会那样上涨。
Prices don't go up like that.
但如果你的需求激增，而供应紧张，那就形成了典型的市场状况。
But if you have demand that surges, supply is tight, then that's creating this typical market.
但关键在于，尤其是在供应方面，监管机构如何评估供应，以及是否应该让市场实时定价，这非常值得思考。
But the key here, and especially on the supply side, it's really interesting to think about how does a regulator evaluate supply versus can you let the market real time price it?
监管机构评估供应的这一点之所以重要，是因为我们展示了一张图表，显示了PJM供应量随时间的变化。
And the the point of the regulator evaluating supply is really one that strikes because we we put out the chart that shows how supply in PJM evolved over time.
你看到在过去四年里，供应量减少了大约40吉瓦，我记得大概是下降了20%。
And you see that in the last four years, it went down by something like 40 gigawatts, like minus 20% if I remember correctly, top of my head.
所以在短短几年内，供电能力削减得相当巨大。
So it's pretty massive supply cuts in just a few years.
在某种程度上，这是合理的，因为我们确实有一些燃煤电厂停止了发电。
And to some extent, it's legit because for real, we had coal power plants that, you know, stopped producing power.
但如果你看那之后的图表，显示了过渡情况，实际上其中很大一部分是由方法论变化驱动的。
But if you if you look at the chart after that, showed the bridge, and actually a substantial amount of that is driven by methodology changes.
但接着，这又回到了他们的供电预测问题上。
But then, again, this goes back to, like, their it's their supply forecast.
他们评估供电能力。
They evaluate supply.
需要明确的是，这并不是说他们做出了错误的决定。
And to be clear, this is not to say they made the wrong decision.
因为发生了一系列不寻常的事件，才导致了这种情况的发生。
Like, there's a bunch of unusual events that happened that led to this happening the the way it happened.
所以如果你展示下一个图表，我不确定你是否能改变收费。
So if you show maybe the next one, I don't know if you can change charges.
是的
Yeah.
所以这个。
So this one.
比如，热能认证改革，是的，这发生在一场可怕的冬季风暴之后，这场风暴暴露出他们的发电厂实际上无法像他们预期的那样满负荷运行。
So, like, the Thermal Accreditation Reform, yes, that happened after a horrible winter storm that basically revealed that, their power plants were not able to run, to the same extent that they they they thought they could.
但同样，这通常是市场环境下常见的现象——当人们评估自身风险、押上自己的资金时，他们通常能做出更好的决策。
But, again, like, this is typically the kind of things that are when when you have a market, when you have people evaluating their own risk, putting their money at risk, they're typically able to, you know, take better decisions.
我想，这大致就是过去一百年左右美国的历史。
I guess is the history of United States Of America over the last hundred years or so.
对。
Yeah.
所以我想，这里有两个要点。
So I guess, like, two takeaways here.
一是，要明确一点，你所说的过剩供应与他们预测的需求相比，基本上意味着一座发电厂一年中有96%的时间处于闲置状态，却需要在需要时支付大量费用才能启动。
One is that, I guess, to be clear about what you say with the excess supply to match that demand over what they forecasted, it it roughly means a power plant that sits idle for, like, 96% of the year and then needs to be paid a whole bunch of money to turn on for those times when you need it.
因此，如果你一整年都没启动它，价格就会变得非常高，这看起来是显而易见的原因。
And, therefore, the price is, like, really, really high if you haven't turned it on all year for it seems like obvious reasons.
另一点是，我想指出你刚才说的：CFP认证改革，也就是他们衡量标准的方式，导致了他们总容量41%或14吉瓦的下降吗？
The other thing is that just to point out what you were just saying, CFP accreditation reform, meaning how they measure stuff, is responsible for 41% or 14 gigawatts of decline in their total?
为了明确一点，这确实是合理的。
Like, to to be clear, it's legit.
这背后有真正的理由。
There's a real reason.
PJM一直以来都是夏季用电高峰的地区。
Like, PJM used to be has always been an area that summer peaks.
但不同地区的峰值负荷情况不同。
But so basically, peak load the peak load of different areas.
有些地区是夏季高峰，有些地区是冬季高峰，这取决于人们如何取暖或制冷，以及天气情况等等。
In some areas, it's summer and some areas, it's winter, depending on how people heat or cool their homes and, you know, the weather, basically.
因此，PJM更多是一个夏季高峰区域，但他们在2022年经历了一个非常糟糕的冬天。
And so PJM is more of a summer peaking region, but, basically, they had a pretty bad winter in 2022.
2021年就已经发生过一次相当严重的事件。
There was already a pretty bad one in 2021.
这让他们意识到，实际上并不是需求峰值，而是峰值比预期更高。
And what it made them realize, it's not actually a demand pick is it peaked higher than expected.
这是一方面。
That's one thing.
另一方面，他们发现许多发电厂实际上同时停止了运行。
The other thing is they realized many of their power plants are actually simultaneously stopped running.
当天气条件极端时，尤其是天气非常寒冷时，发电厂的可靠性会受到显著影响。
The reliability of power plants just is is materially impacted when weather conditions are bought, especially when it's when it's cold out there.
他们根本没有考虑到，可能会有大约10座发电厂同时无法运行的风险。
And, basically, they didn't account their supply for the risk of, you know, maybe 10 power plants at the same time not working.
所以在某种程度上，这是合理的，因为这种机制的设计初衷就是为了避免最坏的情况。
So to some extent, it's legit because the the way, like, the way this works is always to avoid the worst case scenario.
因此，你不得不做出这些决策，以确保对系统能力进行极限压力测试，他们这才意识到，好吧。
So you're forced to take these decisions to make sure, like, you stress test to the to the max of their ability, and they realize, okay.
今年夏天真的很糟糕。
This summer is really bad.
许多发电厂无法正常运行，所以我们不得不重新调整我们的供电计划。
Many power plants couldn't work as expected, so we have to, to, you to reset our supply.
在某种程度上，这是合理的。
So to that extent, it's legit.
这种情况以前从未发生过，而他们之前衡量供电的方式是基于历史数据，实际上是在发电厂层面评估历史风险。
That had never happened, whereas the way they were you, measuring supply before was looking at historical, Actually, at the power plant level, were measuring historical risk.
他们没有衡量那种多个发电厂同时停运的集群效应，而只是考虑单个发电厂因各种原因暂时停机的情况。
They they were not measuring the the the the clustering impact of, like, one a summary is by, like, many power plants stop working at the same time as opposed to just one power plant being being offline for for a moment for whatever reason.
对吧？
Right?
但再次强调，如果你能实时应对，人们就能更快地做出反应，避免这种不幸的发生。
But but, again, like, if you if you if you let that happen real time, people can react to this faster and avoid this unfortunate.
是的。
Yeah.
这很可能是一件好事，从总共提供170吉瓦减少到156吉瓦，因为你从未真正达到过170吉瓦。
It's it's probably, like I think it's probably obviously a good thing to have a, like, move from a 170 total gigawatts offered to a 156 because you've never actually had a 170
是的。
Yeah.
没错。
That's right.
对。
Yeah.
为了。
For.
对吧？
Right?
所以，我认为没有人会说这个调整是件坏事。
So I don't think anybody's saying this revision is, like, a a bad thing to do.
只是很讽刺的是，在当前这个人们普遍希望增加发电能力的时期，数字却反而减少了14吉瓦。
It's just so funny that in a at a time where I I seem to see so much about people wanting to bring more power generation online, the numbers are moving in the exact opposite direction to the tune of 14 gigawatts.
你知道吗，在PJM的同一张图表上，新增的天然气装机容量是3吉瓦，而新增的可再生能源是1.5吉瓦。
You know, new gas capacity on this same chart in PJM is plus three, and new renewables is plus 1.5 in the same question.
你认为这意味着煤炭会因此被推迟吗？
Do you think this means that coal is going to since they're gonna delay this?
我的意思是，人们已经对此讨论了很多。
I mean, it's like people have been talking quite a bit about this.
对吧？
Right?
比如推迟煤炭项目。
Like, delaying coal.
你认为退役潮会开始在边缘地带发生吗？
Do you think retirement seem that that's gonna start to happen on the margin?
这其实就是这么回事。
This is kinda what this is.
不知道。
Don't know.
高层告诉我。
High level telling me.
是的。
Yeah.
所以有两件事。
So there there is two things.
一是联邦层面阻止了燃煤电厂的退役计划。
One is at the federal level, stopping coal power plants, like, stopping their retirement plan.
这已经发生了。
Does that has happened.
是206年公平法案，如果我没记错的话，不管怎样。
It's fair code of two zero six, if I remember correctly, whatever.
一大批燃煤电厂原本计划退役，但现在它们不会再退役了。
A bunch of coal power plants were supposed to retire, and basically, they're not gonna retire anymore.
另一个方面是从经济角度看，这些电厂中有一些已经运行了40年、50年。
And then there's another aspect which is from an economics point of view, some of these power plants are, like, whatever, 40 years old, 50 years old.
它们显然不是最高效的。
They're obviously not the most efficient.
所以运行它们已经不再经济了。
So running them is not economical anymore.
但如果价格足够高，它就会变得经济。
But if prices are high enough, it it becomes economical.
而目前能源领域的核心问题在于，你需要平衡消费者希望电价保持低位的期望。
And that actually the the core issue is with energy right now is that you have to balance the consumer expectations of prices staying low.
但与此同时，为了激励新的发电设施并网，电价实际上需要保持高位。
But the the the the fact that, like, to incentivize new generation to come online, you need prices to actually be high.
因此，这确实是主要挑战之一。
So that's really one of the big challenges.
原因之一还有很多其他原因。
One of the reasons there's many others.
其中一个原因是，电网上的发电设施难以并网，这就是为什么我们正在转向户内发电的原因。
One of the reasons why it's it's tough for for generation on the grids to to to to come online, which is why we're moving to behind the meter.
对吧？
Right?
你可以完全按照自己的需求来建设，而不必等待电网最终决定允许更多人接入或进行改革。
It's like, you have way more control to to build whatever you want to your needs as opposed to waiting for the grid to perhaps, you know, decide to finally let more folks come in or change reform something.
对吧？
Right?
是的。
Yeah.
你觉得这背后的政治因素是，特朗普曾公开表示，他正试图让所有数据中心建设者或前沿实验室签署协议，承诺不会提高消费者的电费？
And do you think that's so, like, maybe the political aspect of this is that Trump was out there saying he's trying to have all the data center builders or the Frontier Labs sign on the dotted line that they are not going to increase consumers' electricity prices.
电网运营商可能只是在寻找替罪羊，以便在必须提高电价时推卸责任。
The grid operators are probably just, like, looking for somebody else to blame when they have to increase people's prices.
但绝大多数新建的数据中心都在使用新技术，采用电网后或用户侧的发电方式。
But the vast majority of the data centers that are coming online are using the new stuff, are using behind the grid or behind the meter generation.
对吧？
Right?
所以
So
实际上，还没有。
Actually, not yet.
它将在2027年或2026年底变得更大。
It's gonna be much bigger in '27 or late twenty six.
目前上线的绝大多数数据中心仍然是并网的。
Still the vast majority of data centers coming online right now are grid connected.
而且大多数仍受某种程度上的旧规制约束。
And the majority are under the, to some extent, old regulations.
因此，你越来越看到的是，不同的州或不同的公用事业公司正在与他们的大型超大规模客户直接协商，制定定制电价——即你将支付的费用，这与其他人支付的费率不同。
So increasingly what you're seeing is different states or different utilities are negotiating directly with their big hyperscale customers to build a custom tariff, which is like, this is what you are gonna pay, which is different from the rate that everyone else is paying.
以确保它们对电网的影响不会转嫁到其他人身上。
And to basically be sure that their impact on the grid is not like, doesn't flow through everyone else.
对吧？
Right?
所以，如果你退一步看，在什么情况下，人工智能的热潮会影响零售电价？
So, like, if you take a step back, like, in in which context could the AI boom impact retail bills?
关键是，尤其是在输电方面，美国所有的输电设施都受到监管，这意味着它们享有法定的股本回报率，也就是说，如果你建造了一个巨大的资产，但其利用率没有达到预期，就无法实现应有的收益。
Like, the core thing, especially on the transmission side, any any all of the transmission in The US is regulated, which regulated means that it owns a a required return on equity, which means that if you build a gigantic asset and it's not properly monetized, the utilization rate is not as expected.
对吗？
Right?
这些成本实际上必须由建设方承担利润，而他们不承担任何风险。
Those costs, because the actual builder must turn a profit regardless, it takes no risk.
这些成本最终会转嫁给消费者。
Those costs actually end up being passed on to the to the to the consumer.
对吗？
Right?
因此，巨大的风险在于，如果我们过度建设，而监管制度又强制将无盈利的资产成本转嫁给客户，那显然就是最糟糕的结果。
So the massive risk here is that if we overbuild and we have regulations that force a nonprofitable asset to be passed on to the customers, then obviously this is the worst case scenario outcome.
所以，接下来的问题就是，如何确保这种情况不会发生？
So that'd And like then the question is, like, how to make sure this doesn't happen?
实现这一点有几种方式。
And there's a few ways to do this.
一种是我们之前讨论过的定制电价，也就是超大规模云服务商告诉公用事业公司：你对电网所做的所有升级，我来承担费用，或者我来承担大部分费用。
One is customized tariffs as as we discussed earlier, which is, like, the hyperscaler telling the utility, like, all of the upgrades that you're doing on the grid, I'm gonna pay for this or I'm gonna pay for the majority of this.
如果你需要新建一座发电厂，我也来买单。
If you have to build a new power plant, I'm gonna pay for it.
但‘买单’并不一定意味着承担资本支出（CapEx）。
And pay for it does for doesn't necessarily mean pay for the CapEx.
它只是意味着你或许可以与公用事业公司签订一份十五年或二十年的协议。
It just means that perhaps you can sign a, you know, fifteen year, twenty year agreement with the utility.
一个有趣的例子是甲骨文在密歇根州的项目，那里有一个为OpenAI建造的Stargate数据中心。
An interesting example was Oracle in in Michigan, one of the stargate data centers actually for OpenAI.
甲骨文签署的协议涉及一个1吉瓦的IT负载，总用电需求达1.4吉瓦。
Oracle signed so it's a one gigawatt data center, one gigawatt IT, 1.4 gigawatt gross utility.
他们实际上与公用事业公司达成了40亿美元的承诺，这在当前格局中已经相当可观了。
They actually signed a $4,000,000,000 commitment with the utility, which is pretty big in this landscape.
想象一下，如果我们所有的吉瓦级项目都像这样与公用事业公司达成承诺，那么收入、资本支出等都将大幅增长。
Imagine all of the gigawatts that we have if all of them have that kind of commitments utility, probably revenue and CapEx and all of it is gonna go up quite a lot.
但无论如何，重点是20亿美元。
But, anyways, the point being $2,000,000,000
这个时间跨度是十年还是十五年？针对这40亿美元的投资？
timeframe for that use four gig 4,000,000,000 over ten year fifteen?
所以是四年时间，这其中涉及两件事。
So it's four years of which there's two things.
抱歉。
So sorry.
不好意思。
Sorry.
这里有两件事。
There there's two things.
首先，有一个投资期。
So there's a there's an investment period.
这项承诺包括十五年或二十年的电力价格协议，同时还包含投资承诺，即购买电池并将其接入电网，以增加其下的整体容量。
So that commitment has a fifteen year or twenty year actually power prices commitment, but there's also an investment commitment to to purchase batteries and put them on the grids to basically increase the overall amount of capacity under it.
所以这是一项相当有趣的协议。
So it's actually a pretty interesting agreement, this one.
两十亿美元的电池。
So $2,000,000,000 of batteries.
甲骨文表示，我将支付电池费用以提升电网的可靠性等等。
Oracle says, I'm gonna pay for batteries to improve grid reliability and whatnot.
而且我还将签署一份二十年协议，以固定的金额向你们付款，并设定最低需求收费，这意味着——这其实非常有意思——即使负荷未能实现，他们也必须支付这笔费用。
And on the and I'm also gonna sign a twenty year agreement to pay you a given amount of money with a minimum demand charge, which means that and this is actually very interesting, the minimum demand charge, because this means if the load somehow doesn't materialize, they're gonna have to pay a charge regardless.
电网上的一个新趋势是确保各方支付公平的份额，以避免出现这种情况：你新建了发电设施，但如果超大规模用户没有消耗其原本计划消耗的电量，这些成本却要由客户来承担。
That's one of the new big things on the grid is again making sure they pay their fair share, but to avoid like a scenario where you build new generation and those costs are being passed on to customers if the hyperscaler doesn't consume the load that it it was planning to to consume.
哇。
Wow.
好的。
Okay.
而且还有很多，是的，有很多
And I There's a bunch, yeah, there's
有很多有趣的内容正在谈判中，但没有简单的解决方案。
a bunch of interesting stuff being negotiated, but there's no easy solutions.
所以在这些情况下，像那样的大额承诺。
So in these cases, like, big commitment like that.
但有时候，直接建自己的发电厂反而更简单。
But sometimes it's just easier to just, know, build your own power plant.
是的。
Yeah.
这说得通。
It makes sense.
我想知道，Oracle即将开展的这个项目，那个协议是提前多久谈好的？
I wonder, like, how long in advance did that deal get negotiated for that site that Oracle is gonna do?
这不可能在六个月内达成这样的协议，毕竟现在有些人建数据中心的速度太快了。
It's not like they can reach an agreement like that in a six month timeline for the speed at which some people wanna put up data centers these days.
展开剩余字幕（还有 480 条）
那个特定的站点，我认为大约两年前就开始开发了。
That specific site, I think, began development about two years ago.
除了数据中心开发商Correlated Digital之外，还有一家第三方参与其中。
There's a third party involved as well as the data center developer correlated digital.
如果我没记错的话，他们大约在两年前就开始了那方面的谈判。
If I remember correctly, that's about two years ago they started negotiation on that side.
但甲骨文是在几个月前，我想是十月，才签署的协议，而这个数据中心将在2027年投入运营。
But Oracle signed the deal just a few months ago, I think October, and that data center is gonna be operational in '27.
所以实际上，这个时间线本身已经相当短了。
So actually, that timeline itself is pretty short.
从你签署协议到负载投入运营，中间的流程就是如此。
The tenant in in from from which you signed the deal to which the the load is gonna be operational.
它就会直接发生。
It's just gonna happen.
为了完成一个完整的数据中心建设项目，从与公用事业公司谈判到最终建成，通常需要一到三年的谈判时间，整个周期长达四年？
It's it's it's some one to three years of negotiating with the utility for a four year start to finish of the full data center construction project timeline?
我们希望能在2028年之前完成整个数据中心的建设。
The the hope is actually to get it done by twenty eighth, the full data center.
所以这里有不同的时间表。
So there's different timelines here.
电池是另一个方面。
Batteries are another one.
关于电池投资，他们说需要四年，但希望数据中心能在某个时间点前投入运营。
For for the battery investment, they said four years, but the hope is that data center will be operational by something
比如2020年？他们是什么时候破土动工或获得第一个许可，表明要建这个数据中心的？
like '20 They, I don't know, break ground or get the first permit to to say we're gonna do this data center?
那大概是两年前的事了。
That was, like, two years ago.
对吧？
Right?
是的。
So Yeah.
大约两年前。
About two years ago.
是的。
Yeah.
是的。
Yeah.
是的。
Yeah.
所以总共需要五年才能让整个项目上线？
So five years total to get the whole thing online?
而且，我想很多人只是在谈论这些公用事业公司的排队问题，比如实际申请电网互联需要多长时间。
And, I mean, I guess I'm a lot of people just talk about the queue with these utilities, like trying to actually get a grid interconnection done and how long that takes.
但这些大型数据中心项目，从开始到完成确实需要四年时间。
But some of these big data center projects, they are they are taking four years, start to finish.
所以看起来就是排队等候，然后愿意为一些电池设备付费，就能推动这些项目落地。
So just seems like get in line and then be willing to pay for some batteries and you can make some of this stuff happen.
是的。
Yes.
但有一点是存在融资错配，因为你必须承担投机性风险。
But, like, there there's one thing is there's a financing mismatch, which is, like, you have to take speculative risk.
比如，如果你不确定需求是否会真正出现。
Like, if you wanna if you, like you don't know yet the demand is gonna materialize.
对吧？
Right?
所以在某种程度上，这就是造成错配的原因。
So that to some extent, that's what creates the mismatch.
比如，如果你从事太阳能、电池或其他相关行业，你必须在某种程度上做出前瞻性投资。
Like, if you're into solar business or battery business or whatever, you you have to make a forward bet to some extent.
是的。
Yeah.
有道理。
Makes sense.
是的。
Yeah.
你怎么看？
What do you think?
那个需求会实现吗？
Is that demand gonna materialize?
Stargate 会推动 War GPT 吗？因为 War Clawed 已经发布了？
Is Stargate gonna power war GPT because war clawed is out?
我觉得他们会改变最终用户。
I I I think I think they're gonna change the end user.
OpenAI 会衰落，Anthropic 会取而代之，因为全世界每天都会像我们一样，使用价值十万美金的 Cloud Code。
OpenAI goes under an Anthropic takes it all because all of the world's gonna use a thousand a $100,000 of Cloud Code every day just like us at something else.
假数据。
Fake numbers.
这不对。
It's it's not true.
我们来谈一下这个。
Let's bring that.
你现在满脑子都是云代码，觉得OpenAI要垮了。
You're so Cloud Code brained now that you you think OpenAI is gonna go under.
我的意思是，我们每天的支出某种程度上暗示了这一点，我的意思是，天啊。
I mean, our daily spend is kind of implying it will be I mean, the thing is Jesus Christ.
我只是不知道每天的支出是多少。
I just don't know the daily spend.
真正疯狂的是这种增长的速度。
The thing that's really crazy is just, like, the rate of this.
对吧？
Right?
我们来聊聊Anthropic的AR交叉领域吧，我很高兴我这么说，乔丹，我直接说了，我之所以没在2026年下这个注，是有原因的。
Let's just talk about the Anthropic AR crossover, which I'm happy that I I dude, Jordan, I'm just gonna say, there's a reason why I didn't make this bet in '26.
但因为我们原本有个赌注，而Anthropic嘛，就是那样。
But but because we had a bet originally, but, like, Anthropic is, like, whatever.
Anthropic 的收入年化率已经达到了 OpenAI 在年底的水平，而现在才三月。
Anthropic's at the revenue run rate that OpenAI was at the end of the year, and it's March.
所以它领先了两到三个月左右。
So it's, like, two you know, it's, a two or three month lead.
对吧？
Right?
差距实在太小了。
It's just so close.
是的。
Yeah.
两个月。
Two months.
对。
Yeah.
如果你回溯一下，不是为了炫耀我赢了什么，因为我根本没赢，但我在九月、十月时对 ClawCode 深感震撼，于是开始告诉 Doug，他们会击败 OpenAI。
If if you go back to not to victory lap on something I didn't win, but I was so impressed by ClawCode in, like, September, October that I started telling Doug that they were gonna win versus OpenAI.
不，老兄。
No, dude.
那个赌约，不行。
That that bet no.
那个赌约早就发生了。
That bet that bet was way before.
那是在非现场会议之前很久的事了。
It was way before the off-site.
那时候都快到仲夏了，老兄。
It's like middle of the summer, man.
拜托。
Come on.
我们就要开始赌了，我喜欢打赌。
We're gonna like, I I like betting.
是的。
Yeah.
因为是对的。
Because the right.
因为那个赌注会在外勤时变成啤酒。
Because the bet would turn into a beer at the off-site.
是的。
Yeah.
是的。
Yeah.
那是在
Which was in
十月。
October.
九月之前。
Before September.
好的。
Okay.
我觉得之前发生过。
I think it's been before.
我发誓。
I swear to god.
总之
Anyways
好的。
Okay.
好吧，它在这儿。
Well, it's here.
现在是三月，它就在这儿。
It's March, and it's here.
我们认为，二月的ARR退出表明Anthropic即将超越OpenAI。
ARR exit of February, we think is close for Anthropic crossing over OpenAI.
整个公司，对吧，Cloud Code是其中重要的一部分，但
The whole the whole company, right, which Cloud Code is a big part of, but
整个公司。
The whole company.
并不
Not
不一定像云代码目前仍是安特ropic收入的主要部分。
necessarily like, Cloud Code is still not the majority of Anthropic revenue currently.
它正在以非常
It's growing at a very
高的速度增长。
high rate.
我们来谈谈这个吧。
Let's actually talk about that.
我觉得是的。
I think it is.
我觉得关键是，它确实是。
I think the thing is it's the it's yeah.
杰里米能这么做，因为他一直在专门打磨这个模型，但这只是如何分配收入的问题。
Jeremie can do it because he's been model grinding on this one, but it's just how you attribute revenue.
对吧？
Right?
我认为他们的做法是，嗯。
I think the way that they're yeah.
你可以查看我们的使用仪表板。
You can look at our usage dashboard.
云代码的收入分配是基于在云代码中使用Cloud Max，这就是他们定义的云代码。
Cloud code attribution of using Cloud max in Cloud code is what they define as Cloud code.
API使用量就是API使用量，无论是否在云四代码API模式下，他们都不会将其归入云代码。
API usage is API usage anywhere, and they don't they do not attribute it if it's in Cloud quad code API mode or not.
作为快速模式的爱好者，我想你们能感受到我们的态度。
And as fast mode addicts, I think you can tell how we feel.
是的。
Yeah.
比如，特别是那一百万上下文的情况。
Like, actually, especially the the 1,000,000 context.
我觉得快速模式是订阅制的，但一百万上下文只是API的。
I I think fast mode is on is on subscription, but 1,000,000 context is just on API.
但我想重点是，半分析的支出有95%都花在了API上。
But I guess point is semi analysis spending is like 95% going to API.
而不是花在他们所称的Cloud Code上。
It's not going to the what they report as Cloud Code.
所以当他们上个月说，每月云代码的年化收入是25亿美元时。
So when they said last month, they said, like, 2 and a half billion dollars is our annual run rate on Cloud Code.
这实际上远远偏低了。
It's actually way way too low.
这太有趣了。
That's super interesting.
是的。
Yeah.
是的。
Yeah.
我认为他们必须明确表示：不行。
I think I think they have to make that to be like, no.
不行。
No.
不行。
No.
不行。
No.
我认为，对于没有巨大向量的通用企业需求来说，这更符合他们的利益。
We I I think it's more in their interest for generalized enterprise demand without having a ginormous vector.
就像是，所有企业都在使用它，天知道会怎样。
It'd be like, oh, all the enterprises are using it and god knows what.
而不是变成一个超级病毒式的产品，人们为此花费80亿美元。
Instead of being like it's one hyper viral thing that people are yellowing $8,000,000,000 on.
但谁在乎呢？
But who cares?
我的意思是，显然答案是收入，而Anthropic现在就是其中之一。
I mean, it doesn't I I mean, it clearly the answer is revenue, and Anthropic is one right now.
或者OpenAI会回来，但你知道，这周或下个月就会变得无关紧要了。
Or I mean, OpenAI comes back, but, you know, it'll be so over, so back or whatever this week or next month.
你觉得是什么导致了这次额外的激增吗？
Do think that do you think there's a like, what do you think is is causing the extra spike?
因为之前有超级碗的事情。
Because there was the Super Bowl thing.
现在又有与国防部的冲突。
Now there's the conflict with the Department of War.
你觉得，从某种间接的角度来看，如果国防部没有继续对Anthropic实施供应链风险认定，未来六个月他们仍能与美国所有企业继续合作，甚至可能修复关系、重返美国市场，那么达里奥的这条推文风暴实际上是否让人们对
Do you think that, in a in a backhanded way, if the Department of War does not follow through with this supply chain risk designation for Anthropic and they can keep doing business with everybody in The US six months from now and possibly even repair the relationship, get back into The US, that this tweet storm at Dario actually got people to
净推荐值变得正面。
Net promoter score positive.
我的意思是，你看到凯蒂·佩里的那条推文了吗？
I mean, did you see the Katy Perry tweet?
那才是真正的净推荐值正向提升。
That that is the the net score promoter positive.
实际上，我只觉得
Effectively, I would I only
我知道，也许两者都有，但对我来说，这纯粹是关注度。
know I if it's so maybe it's both, but to me, it was just eyeballs.
更多人现在知道什么是Clawed和Anthropic了，是的。
Like, more people know what clawed and Anthropic means now Yep.
因为他们关注的是政治，而不是技术。
Because they are following politics, not technology.
是的。
Yes.
我同意。
I agree
没错。
with that.
等等。
Wait.
等等。
Wait.
你们真的认为这对其收入有任何影响吗？
Do do you guys actually think that it had any impact on their revenue?
是的，老兄。
Yes, dude.
你知道现在它是下载量第一的应用吗？
You know it's the number one downloaded app right now?
它的下载量超过了OpenAI，而这发生在战争部。
It's it's above OpenAI, and that happened on the Department of War.
在法国可能不是这样。
Probably not in France.
大概也不在乎。
Probably not care.
我们根本不关心法国，兄弟。
We don't care about France, bro.
你不喜欢Claude吗？
That You don't like Claude?
Claude是法国的。
Claude Claude is French.
是的。
Yeah.
忘了Claude是法国的。
Forgot Claude is French.
是的。
Yeah.
我的意思是，你知道吗，你有没有听说过那个，那个RL？
Like, I you know, who takes have you heard that the, you know, the RL?
有一个梗。
There was a meme.
我想不起来是谁做的了。
I can't remember who made it.
我觉得是Nier，就是那几个。
I think it was Nier, like, a few.
夏天的时候，Claude很懒，因为他去度假了。
During summer, Claude is lazy because he's on vacation.
C'est moe。
C'est moe.
是的。
Yeah.
这就是法国人的做派。
It is the French way.
目前，Claude正在抗议，过一会儿他就要去度假了。
Currently, Claude is protesting, and a little bit later, he's gonna go on vacation.
然后之后就罢工了。
And then a strike afterward.
是的。
Yeah.
有一种方法。
There's a way.
那就是我们困住的地方。
That's what we were stuck in.
不。
No.
但是
But
它是开放的。
it's open.
它是开放的。
It's open.
而且我认为，那是在战争部。
It's And that was, I think, on the Department of War.
所以你
So you
只要他们不会把他们的供应商，比如，本质上从计算中剔除出去，老兄，我觉得整个局面将会
as long as they don't get, like, as their vendors, like, essentially spit them out of compute, dude, I think it's gonna I think the whole moment is gonna
我们来谈谈收入吧，因为好吧。
Let's talk revenue because okay.
下载量可能只是随机的人在尝试这个应用。
Downloads, that's that could be just random people, like, trying the the app.
但收入很可能来自云代码。
But revenue is probably Cloud Code.
对吧？
Right?
所以这很可能与那个具体问题无关。
So that's most likely unrelated to to that specific issue.
但这是云代码，Cursor还宣称实现了巨大的加速，而且他们使用了大量的API令牌。
Well, it's it's cloud code, but Cursor just claimed a huge acceleration as well, and they're using plenty of API tokens.
我假设其他像WindServe for Sell、Replit这样的AI编程初创公司和应用也都使用了API令牌。
I assume many of the other vibe coding startups and apps like a WindServe for Sell, Replit, they're all using API tokens.
然后他们推出了Cloud for Finance、Cloud for Legal和Cloud for安全软件，这些产品接连一天天地打压了一大批SaaS股票。
And then they launched Cloud for Finance, Cloud for Legal, and Cloud for security software, which tanked a bunch of SaaS stocks one day after the other.
有人在购买除了第一方Claude Code Max订阅之外的其他东西。
There's something somebody is buying something other than the first party Claude code max subscriptions.
你知道，这并不仅仅是一个单一渠道在消耗令牌。
You know, it's it's not strictly a single vector consuming the token.
我的意思是，我认为这种加速， literally，他们在两个月内增加了100亿美元的收入。
My my point is I think, like, this acceleration, literally, they've added $10,000,000,000 of revenue in, two months.
是的。
Yeah.
实际上，他们在两个月内让收入翻了一倍。
Double the revenue in two months, actually.
年经常性收入。
ARR.
年经常性收入。
ARR.
两个月内将年经常性收入翻倍。
Double the ARR in two months.
我的意思是，我认为这完全是由于真正的商业采用推动的，这么说吧，而不是仅仅因为人们为了支持云和某个问题而订阅应用、下载应用所带来的短期热潮。
I mean, I I would think that's entirely driven by, like, true business adoption, if that makes sense, as opposed to just, like, temporary hype of people subscribing to the app and downloading the app in support of of of cloud and a problem.
我完全同意你的观点。
I completely agree with you.
使用场景。
Use cases.
比如，嘿。
Like, hey.
道格，你知道，比任何人都强。
Doug, you know, is better than anyone else.
比如，一些企业正在移除 Salesforce 并自行构建系统。
Like, enterprises, certain companies are removing Salesforce and building their own.
是的，没错。
Well, yeah.
不。
No.
我觉得这很重要。
I I think it's a big deal.
对吧？
Right?
这才是真正的需求驱动因素，而且我同意。
Like, that's the the real demand driver, but I think and I agree.
这并不是推动收入的因素，只是锦上添花而已。
That's not what's moving revenue, but it's just like a a cherry on top.
比如，他们做了 API 的过渡。
Like, they do the API crossover.
这才是真正推动一切的原因。
That's what really is driving all of it.
而且，锦上添花的是，他们下载量也冲到了第一。
And then also the cherry on top is they just hit number one on downloads too.
这就像我们根本没刻意去做，却意外地做到了。
That's like we we weren't even trying and we got you that way.
这简直是一群狂热的粉丝，因为每个人都惊呼：哇。
Like, that's a pretty hardcore mob just because everyone's like, woah.
这甚至不是他们出名的原因。
That's not even what they're known for.
你知道的吧？
You know?
这就像是很随意地，嘿。
It's like a it's like just a casual, like, hey.
顺带一提，这周我们在下载量上也超过你了。
This is just by the way, also passed you on downloads this week.
这可以说是个大事，我认为这就是它如此重要的原因。
Kind of a big I I think that that's why it's a big deal.
这可以说是，用个不太恰当的词，是一个热潮时刻。
It's like a for lack of better word, it is a hype moment.
这到底是什么？
It's a what is it?
这是一个我认为真正重要的时代精神时刻。
It's zeitgeist moment that I think that really matters.
而这正是关键的差别。
And that's like the difference.
实际上，这完全改变了叙事，因为我们过去常说，OpenAI的ChatGPT拥有消费市场。
Well, it's there's like a complete narrative shift because we used to we used to say OpenAI ChatGPT has a consumer market.
当你成为家喻户晓的品牌时，更容易渗透到商业领域。
And when you're the consumer, like household name, it's easier to penetrate the business the business world.
对吧？
Right?
而现在我们实际上看到的是相反的情况，没有人对你产生影响。
And now we're actually seeing the opposite way, which is no one happened to you.
模型已经成为商品，因此价值将集中在应用层，在那里你可以构建出由最适合的模型驱动的卓越体验。
The models are commodities, and therefore, the value is gonna be at the application layer where you can build this great experience powered by whatever model is best is the best fit.
嗯。
Mhmm.
如果你考虑到Anthropic拥有我们无法获得的模型，那这种想法就完全错了。
It's just completely wrong if you consider the fact that Anthropic has access to models we don't have.
他们完全掌控了编码接口，能够比其他人更有效地免费构建大量应用程序。
They have full control over the coding interface, and they can build a whole bunch of applications using that better than anybody else can effectively for free.
因此，我认为Anthropic会像在编码领域对GitHub Copilot所做的那样，进军Salesforce、法律领域的Harvey、SaaS安全领域的CrowdStrike等，这一趋势是合理的。
So the the idea that Anthropic's gonna come for Salesforce, Harvey for legal, you know, CrowdStrike in the SaaS, security business, and on and on down the line because they've already done this to GitHub Copilot on the coding side seems right to me.
在我看来，模型公司将成为无所不包的公司，而不是SaaS公司后来才决定自建模型，或挑选最好的中文开源模型来驱动他们的产品。
It see it seems like the model companies will become the everything companies as opposed to, you know, SaaS companies deciding to build models later or picking the best Chinese open source one to power their product.
就像
Like
我认为这个宇宙是可能存在的，最终或许真的会形成这样的状态。
I think that that universe could exist and might end up existing in the end state.
但在当前状态下，每三个月模型就提升一次，这就毫无意义了。
But in the current state, when every three months the model gets better, there's no point.
对吧？
Right?
因为如果每年才升级一次，而你基于每年的升级来构建你的产品，那还行。
Because it's like if if there's, like, an upgrade every year and you're building your thing on some upgrade every year, that's fine.
也许你可以提前规划，等等。
Maybe you can plan it out, whatever.
但这里是每三个月就升级一次。
But there's an upgrade every three months.
这根本不可能。
There's no way.
你的产品迭代速度注定会比模型的迭代速度慢。
Your your product cadence is gonna be inherently slower than the model cadence.
而且如果这个模型，我的意思是，老兄，这其实就是摩尔定律的历史。
And if the model I mean, dude, this is the history of, like, g like, this is the history of Moore's Law.
对吧？
Right?
CPU之所以是通用的，是因为它们每年性能提升50%，所以从来没有任何理由去搞专门化。
A c p the CPU is one because they got 50% better every single year, so it never made any sense for any specialization ever to happen.
你还不如直接押注那个每年都在变得更强、普遍更擅长一切的底层技术。
You might as well just bet on the underlying better thing that's generally good at everything getting better next year.
对吧？
Right?
但当摩尔定律终结、CPU性能真正达到瓶颈时，专门化才变得极具价值，因为——
But when Moore's Law ended and CPUs really topped out, that's when it became a lot of value to essentially to specialize because, hey.
再也没法通过这种方式获得性能提升了。
There's no more there's no more lift of performance this way.
你只能走专业化的路子。
You have to go down a specialist route.
所以，只要通用模型还没有做到每三到六个月就提升一次，就没有必要做任何专门化。
So until the generalized models don't get better every, you know, three to six months, there's no point of doing any specialization.
所以我认为，直到规模效应停止之前，我们都会看到这种情况。
So I think that that's you know, we're gonna see that until scaling all stop out.
是的。
Yeah.
总的来说，我同意你的观点，但具体到某些情况，比如如果你要进行专门化，他们选择专注于编程，而编程又是如此多其他领域的上游，这显然是一个明智的选择。
I think in general, I agree with you, but the the specific cases, like, if you're going to do specialization, the fact that they chose to do coding, which is upstream of so much other stuff, just seems like an obviously good choice.
换句话说，编程才是应该关注的通用方向，你不应该……我有个问题。
In other words, coding is the generalist thing to focus on, and you shouldn't I I have a question.
你是不是觉得，这明显是一种策略？
Do you do you I think that that that's something definitely, clearly, was a strategy there.
但你认为这不是因为编程本身很特殊吗？你知道，编程的核心就是打造一个能编写出更好模型的自主代理，而这本身就是这个过程的一部分。
But do you think it wasn't because they're like the I I think coding is special in the case that, you know, the whole point of it is you make an autonomous agent who's better able to code a better model, and that's, like, part of the you know, it's inherently a part of this.
如果他们认为营销才是扩大规模的方式，那他们就会打造一个营销型的自主代理。
Now if they thought marketing was the way that they were going to scale for lack of better weight, it would be a it would be a scaling marketing agent.
但事实并非如此。
But it isn't.
对吧？
Right?
唯一重要的是通过编程来提升你的核心产品。
The only thing that matters is coding for making your core products better.
我认为这就是他们一直如此专注于编程的原因。
Think I that that's the reason why they've there's they've always been so focused on coding.
是的。
Yeah.
但真正让你的核心产品变好的是数据。
Well, the the thing that makes your core product better is data.
而获取这些数据最明显的地方就是聊天界面。
And the first obvious place to get that data was in the chat interface.
人们会对好的或坏的回答点赞或点踩，或者你可以看到他们在聊天上花费的时间。
People vote up and down what's a good response, bad response, or you see what their time using the the chat is.
你可以查看日志。
You can review the logs.
但当强化学习出现后，这种情况就变了，因为强化学习需要从单个样本或一条很长的轨迹中获取更多信号，而聊天界面提供的信号不够充分。
And then that shifted when RL came out because you need a lot more signal from one sample or one really long trace in RL than you than you do, and and therefore, chat isn't quite good enough.
编码中的数据质量要高得多，因此，飞轮效应发生在编码领域。
The quality of data in coding is much higher, And so, therefore, the flywheel is in coding.
但这甚至不只是编码，因为你们用的是ClotCode，我不是说它不是代码，但它更多是用于操作你的计算机来生成像我们刚才讨论文章时屏幕上显示的图表截图这样的产物。
But it's it's it's it's not even coding because you guys are using ClotCode for not I mean, it's code, but it's more manipulation of your computer to build artifacts like a screenshot of a plot that we just had on screen when we're talking about the article.
对吧？
Right?
你并不是用它来替代过去你原本会写的软件。
Like, you're you're not using it for replacing software you would have written in the past.
你是在编写软件来完成过去你原本会在Excel里完成的事情。
You're writing software to do something that you would have done in Excel in the past.
所以编码只是一个泛称。
So Coding is just general.
我的意思是，这就像以更高的精度来操作晶体管。
I mean, it's just like a a a more fidelity way to work with transistors.
对吧？
Right?
我的意思是，这就像你拥有软件界的超级大脑。
Like, we're just so the thing is, I mean, this is like, you know, when you software big brain.
对吧？
Right?
从历史上看，软件是人类与机器交互的上层抽象。
Like, software historically is a layer on top of code for humans to interact with the machine.
对吧？
Right?
但既然它已经足够好，完全可以位于软件之下，实际上人类现在是与一个更底层的代理交互，这个代理直接与计算机打交道。
But since it's good enough that it could actually sit underneath software, pretty much humans are interacting with the agent that is now lower level interacting with the the computer.
对吧？
Right?
是的。
Yeah.
因此，你完全跳过了这一部分，因为它没有价值。
And so you just totally skip out that portion because it's not valuable.
所以，对我们来说真正酷的是，我们是分析师。
And so, like, what's really cool for us is we are, you know, we're analysts.
对吧？
Right?
我们一直在分析一堆垃圾。
We're in line analyzing crap.
我们一直不得不使用专门为我门开发的软件来完成这项工作。
We always had to use software that's made for us to do it.
但这是有史以来第一次，天哪，我不必学习任何新软件。
But for the first time ever, it's like, dude, I don't have to learn some new piece of software.
我可以直接让计算机帮我生成信息。
I could use the computer to make information.
我不必去学习统计学、数据分析和R语言，也能进行复杂的分析。
I don't have to, like, learn statistics and stats and r to be able to do crazy analysis.
我不必去学这些乱七八糟的东西来抓取所有数据。
I don't have to learn all this crap to scrape all this stuff.
现在，我可以使用任何我曾经想用的软件，凭借我从未学过的领域知识，做任何我想做的分析。
I can now work with every single type of software I've ever wanted to with, like, all this domain expertise that I never had to learn and do whatever the hell analysis I ever wanted to do.
这才是有价值的部分。
That's the valuables part.
这就像我说的，编程只是其中的一部分，但我觉得编程之所以重要，是因为编程更贴近计算机。
And that's like I mean, coding is all a portion of that, but I think coding the reason why it's coding is because coding is close to the computer.
对吧？
Right?
因此，我真正从根本上相信，LLM由于其计算密集的特性，是计算范式的一种延伸，无论什么模型。
And so it just makes like, I really truly fundamentally believe LLMs, because of how compute intensive they are, are kind of an extension of the computing paradigm, whatever model.
是的。
And Yeah.
你知道，无论智能体是什么，它们都是计算的下一个层级延伸。
You know, whoever Agents agents are kind of the, you know, the next level extension of compute.
所以我的意思是，这完全是全面融入。
So I I mean, it's just like all in.
我们不再是与基于计算构建的软件互动，而是人类正越来越接近
Instead of us, like, interacting with software built on top of compute, really, humans are just getting closer to
计算机。
the computer.
是的。
Yeah.
这就是为什么杰里米会开玩笑说，带着你的智能体在耳机里，然后你可以让它给你展示仪表板，或者做任何你想做的更新。
This is why Jeremie was joking about just walking around with your agent in your earphones, and then you can tell it to show you the dashboard and make an update or whatever you want.
是的。
Yeah.
天啊，我们为什么还需要做软件呢？
Why why why do we need to do software, dude?
它会直接在我耳边读取CSV文件，老兄。
It'll just tell us it'll be reading CSVs into my ear, bro.
我就想听这个。
That's all I wanna hear.
我的意思是，也许这就是DGX Spark或这些边缘计算机的主要应用场景。
I mean, you could what if you have a like, maybe that's the bulk case for, like, DGX Spark or some of these edge computers.
你们都在买Mac mini来运行云代码。
Like, you guys are all buying Mac minis to run your cloud code.
所以你确实需要大量的，大量的，
So you do actually need a lot of a lot of,
我其实已经完全转了一圈，老兄。
like, I've actually I've actually come full circle, bro.
一辈子用VPS，老兄。
VPS for life, bro.
我有个DigitalOcean的Droplet。
I have a digital ocean droplet.
更好了。
It's better.
好的。
Alright.
所以Mac mini要被淘汰了吗？
So is it is it over for Mac minis or
不。
No.
它没有一个
It's not got a
远程SSH连接到我的笔记本电脑。
remote SSH to my actual laptop.
是的。
Yeah.
或者我可以，是的。
Or I can Yeah.
如果我想确保安全，可以在我的笔记本上运行一个容器。
Run a container on my laptop if I want some security.
是的。
Yeah.
我可以远程SSH连接到我的Mac mini，当然也可以远程SSH连接到Digital Ocean的服务器。
I have a remote SSH to my to my Mac mini and then obviously the remote SSH to the Digital Ocean droplet.
太棒了，老兄。
So sick, dude.
我一整天都在运行定时任务。
I run cron jobs all day.
更多的定时任务。
More cron jobs.
更多的定时任务。
More cron jobs.
更新数据库。
Update the database.
更多。
More.
太美了。
It's so beautiful.
抱歉。
Sorry.
酷。
Cool.
你用 Tailscale 吗？
You you use Tailscale?
你之所以说这个，是因为我们实际上在聊 Tailscale。
You you you because, dude so, I mean, we're talking about Tailscale actually.
因为你之前提到你用它来做你正在做的某个项目，我当时就觉得，哇，不错。
Because you were talking about how you use it for one of the like, a project you're doing, and I was like, oh, cool.
我用 Tailscale。
I use Tailscale.
这真的很酷，老兄。
It's pretty it's pretty nifty, dude.
我挺喜欢的。
I like it quite a bit.
是的。
Yeah.
我的意思是，我知道这都是些简单的东西，但很好用。
I mean I know it's the it's simple stuff, but it's good.
就像，确实是这样。
Like, it's yeah.
免费套餐的功能就是，你可以非常简单地通过私人网络从手机登录你的笔记本电脑。
The the free tier get you in option is just, like, log in to your laptop from your phone pretty simply with a private network.
但像CoreWave这样的公司用它来验证用户对集群的访问权限，因为这比使用公共IP和SSH，以及为全球流动的用户设置IP白名单要好得多。
But, I mean, companies like CoreWave use it for authenticating people to their clusters because it's much better than just, like, public IPs and SSH and trying to white list IPs for people who move around the world.
所以是的。
So yeah.
所以Tailscale太棒了。
So Tailscale's awesome.
我的意思是，有很多软件正因为这一点而快速发展，尤其是软件公司。
I mean, there's lots of software that is just like very much accelerating because of this, software companies.
而还有一些软件则完全被甩在了后面，确实是这样。
And then there's other software that's just totally getting left in the dust, and it's yeah.
我的意思是，Tailscale就是其中之一。
I mean, Tailscale's one of those ones.
Obsidian是另一个你非常喜欢的软件。
Obsidian's another one that you love.
对吧，道格？
Right, Doug?
我确实很喜欢。
Like I do.
但你知道有趣的是，我觉得我正在逐渐告别Obsidian。
But you know what's funny is I think Obsidian, I I think I'm graduating from Obsidian.
我想，既然能在终端里直接做，干嘛还要折腾 Markdown 呢？
I think why fuck around with a markdown when I can just do it in the terminal?
说实话，我根本不关心。
Like, I don't care, actually.
我一开始确实想着，得确保能读所有 Markdown 文件，做些审计，就像想好好调优一下那样。
I've come to realize in the beginning, I was like, well, I wanna make sure I could read all the markdown and all this shit to audit it, kinda like looking to, like, you know, do some fine tuning.
我唯一在意 Markdown 的时候，就是在 Obsidian 里打开文档做最后编辑，然后复制粘贴到 WordPress 或其他地方。
The only time I care about markdown is in Obsidian is essentially pulling up a document to edit it for a final time and then copy and pasting it into WordPress or wherever the hell I go to.
但说实话，我根本不在乎它是不是在 Obsidian 里，我完全不在意。
But, like, I literally don't like, it being an Obsidian, I actually don't care.
我期望并相信，模型足够强大，能自动在内容里找到信息，我觉得这应该被抽象掉。
I hope and trust that the model is gonna be good enough at finding information within a thing that like, I I think it it should be abstracted away.
这就是我的看法。
That's my belief.
但 Obsidian 的命令行工具项目还挺有意思的。
But the Obsidian CLI project is pretty interesting.
我一开始确实非常依赖 Obsidian。
I'm kinda I mean, I I was very Obsidian first in the beginning.
现在我已经渐渐远离它了。
Now I've kinda moved away from it.
也许我会再回去用。
Maybe I'll go back.
我不确定。
I don't know.
但我觉得你可以直接用像托比做的那种向量数据库加 QMD 之类的东西，然后直接搜索。
But I I think the you could just have a vector database QMD crap that Toby does, and you can just search.
那样会更快。
It'll be faster.
所以。
So
没错。
why Absolutely.
QMD。
QMD.
致敬Toby。
Shout out Toby.
走吧。
Let's go.
是的。
Yeah.
老兄，这实际上
Dude, that's actually
这些全是加拿大人，伙计。
These are all Canadians, buddy.
Tailsdale、Obsidian 和 Wait 的 QMD。
Tailsdale, Obsidian, and QMD from the Wait.
他们真的是吗？
Are they really?
Shopify。
Shopify.
他们是Wait吗？
Are they wait.
那些人全是加拿大人？
Every single one of those are Canadians?
是的。
Yeah.
是的。
Yeah.
是的。
Yeah.
是的。
Yeah.
哇。
Wow.
我们应该收购加拿大。
We should acquire Canada.
是的。
Yeah.
你这是用‘我们’来指代自己啊，老兄。
The royal we there, bud.
这个‘我们’啊。
The the royal we.
是的。
Yeah.
我们在军队里。
We're in the army.
嘿。
Hey.
这主意真不是我想的，伙计。
It wasn't even my idea, dude.
这根本就不是我的主意。
It wasn't even my idea.
所以，是的。
So Yeah.
而且嘿，这一点也适用于所有收听这个播客的人。
And hey, and that that also that also applies to anyone listening to this podcast.
如果你是加拿大的科技工作者，那你已经运气不错了。
If you're a Canadian tech worker, you've had a lot of good luck.
这边的乔丹在SME分析公司工作。
Jordan over here happens to work for SME analysis.
我们还挺喜欢这家伙的。
We we moderately like the guy.
所以我觉得我们非常乐意招聘更多来自加拿大人。
So I think we're more than happy to hire more people from Canada.
所以，是的，我们一直在寻找。
So, yeah, we're always we're always looking.
所以，不管怎样，这完全是题外话。
So, anyways, that's a complete aside.
一直在招聘。
Always always hiring.
是的。
Yeah.
是的。
Yeah.
是的。
Yeah.
是的。
Yeah.
向Rippling致敬，另一家将在人工智能浪潮席卷一切时加速发展的公司。
Shout out to Rippling, another another company that's gonna accelerate as the AI wave takes over everything.
我们在哪儿？
Where are we?
也许我们该在这里结束了，各位。
Probably good to wrap up here, guys.
我不知道。
I don't know.
是的。
Yeah.
我觉得是这样。
I think so.
我也是这么想的。
I think so too.
聊聊这一周吧。
Talk about the week.
我们聊一下Quinn吗？你们想聊吗？
Do we talk about did we do you wanna talk about Quinn?
我的意思是，我们甚至还没发过关于这个的内容。
Well, I mean, we haven't even posted about this.
我不
I don't
不知道你想聊什么，老兄。
know what you wanna talk about, dude.
我们可以聊聊这个。
We can talk about it.
我们下周也可以聊这个。
We can talk about it next week too.
是的。
Yeah.
我们可以聊聊领导层的变动。
We can talk about leadership changes.
事情还在变动中。
Things are still shaking out.
你对Quen有什么看法吗？
You got a hot take for about Quen?
我没有特别的看法。
I don't have a hot take.
我的意思是，看了几条推文后，这其实是个相当冷淡的看法。
I mean, it's a it's actually a pretty cold take after some tweets.
我当时就想，天啊。
I was like, damn.
我的看法也很冷淡。
My my take's pretty cold.
本质上，他们把他赶走了，因为他们在中国新年期间没有获得足够的用户。
Essentially, they kicked him out because they didn't acquire enough users on Chinese New Year's.
你在这方面确实做对了。
You definitely did right about that.
关于当时中国新年期间的激烈竞争，你对代币经济的更新确实很到位。
That was a a nice tokenomics update about all the Chinese New Year's competition that was going on.
是的。
Yeah.
有趣的是，我的结论是，当时的情况其实也没那么糟，但我觉得他们可能花了一十亿美元，结果还是不够好。
And you know what's funny is my takeaway was, like, when wasn't bad, but I think it you know, you spent, you I think they probably spent, like, a billion dollars, and they're, like, not good enough.
所以现在所有的KPI都变成了用户获取。
So now the KPIs are all user acquisition.
我的意思是，抖音或者字节跳动，肯定算一个。
I mean, Doubou definitely or, like, ByteDance one, for sure.
而且我认为，字节跳动是 summit analysis 所有人想通过 CDance 使用的唯一一个东方模式。
And and it's, like, the only I think ByteDance is the only Eastern model that, like, everyone at Summit Analysis wants to use through CDance.
我真的超想用 CDance，老兄。
Like, I want CDance so badly, dude.
好吧。
Okay.
是的。
Yeah.
那应该是那篇文章的主旨，至少是我写过的一小部分，讲的是：我们已经看到过类似的情况，比如太阳能板、无人机和比亚迪汽车，人们在边缘处会说，哦，这看起来真酷。
That was the, I think, the thesis of that article or at least the the last little section of it that I wrote a bit about, which was, like, this is I think the we've seen this for, like, solar panels and drones and BYD cars where, like, on the margins, people are like, oh, that looks really cool.
但我的特斯拉还是很酷，所以我不会特意去进口一辆比亚迪。
But my Tesla's still cool, so I'm not gonna go out of my way to try to import a BYD.
我认识的所有看过CDN视频的人，都在想办法通过VPN登录中国并加入那个排队名单，埃里克今天早上发的帖子问的是，要等多久？
Everybody who I know who's seen the CDN's videos is, like, trying to figure out how to VPN their way into China and add something to that queue, which Eric posted this morning is, like, how long?
就像，有
Like, there's
大概要排三个月左右吧。
It's, like, a three month queue or something like that.
这太不可思议了，老兄。
It's something incredible, dude.
实际上，如果你想要注册并付款，你就会说，是的，老兄。
Like, effectively like, if you want to sign up and you you're trying to pay you're trying to pay for stuff, you like, you're like, yeah, dude.
你可能几个月后就能在CDance上生成一段视频。
You might be able to generate a video on CDance in, a few months.
这有点像是，有点像
It's kind of a it's it's like kind of
队列里大概有十五万条请求之类的。
It's like a 150,000 requests in the queue or something.
是的。
Yeah.
是的。
Yeah.
是的。
Yeah.
十万？不是。
A 100 no.
是三万中的十四千，我想这是付费队列。
It's it's 14 k out of 30 k, and that's I think in the paid queue.
哇，真的吗。
Oh, wow.
所以啊，老兄，这简直就像老派的互联网。
So there dude, that's like dude, that's like old school Internet.
你必须排队等待你的任务完成，而且这是付费版本。
You have to wait to be in the line for your job to be done, and this is on the paid version.
是的。
Yeah.
已经付费的用户却被告知，你甚至不能多付钱来跳过排队。
People who are already paid by dance are being told, you can't even, like, pay more to skip the line.
你只能老老实实排队等待。
You're just you just gotta wait in the queue.
太夸张了，老兄。
Pretty sick, dude.
还有，我的银河品牌，我的意思是，我不知道我们会深入多少内部细节，但有些情况可能确实是需求所致。
Also, my my Galaxy brand, mean, we I don't know if we've I don't know how much inside baseball we're gonna do, but, like, I I mean, some of this might be the demand.
这可能是我们没有充分认识到的、在GPU上看到的需求的一部分。
That might be that might be part of the the the demand that we're not recognizing well that we're seeing in GPUs.
对吧？
Right?
关键是，中国的模型，嗯，也许其中一些的留存率并不高，但字节跳动显然获得了一些留存。
It's the fact that, like, Chinese models, yeah, maybe the retention rates weren't so great for some of them, but ByteDance clearly got some retention.
而且，中国实际用户的增长正在加速。
And, clearly, the acceleration of actual users in China is going up.
对吧？
Right?
通过CDance、WAN，还有所有这些产品。
With CDance, with WAN, with all this stuff.
而且你知道，中国的人口远超美国，这是个大胆的观点。
And, you know, China has way more people in The United States, hot take.
如果他们开始真正采用人工智能，突然间，这就变成了一个巨大的、并非完全未被开发的新需求领域，我认为这一点可能被低估了，并且正在向世界其他地区蔓延。
And if they're starting to actually adopt AI, all of a sudden, that's a giant new not completely untapped, but a huge new demand vector that I think is just maybe underappreciated and is starting to leak to the rest of the world.
因为中国国内市场实际上无法支撑你。
Because China domestic can't actually support you.
所以这些需求最终都会流向新加坡、欧洲或世界其他地方。
So this just ends up in Singapore or Europe or everywhere else around the world.
所以
So
所以这里的问题是，Anthropic在Clock Code中的收入有多少比例来自中国？
So question here is, like, what percentage of Anthropic revenue in Clock Code comes from China?
据称，全部都是，全部都是，好吧。
Well, allegedly, it's all it's all it it okay.
所以你得知道，其中一部分是因为恶意行为者利用了微调这个机制。
So you have to know that some amount of it is because they're like, oh, malicious actors have been because you know the fine tune thing.
对吧？
Right?
哦，有15万个
Oh, with a 150,000
副本。
copies.
恶意行为者。
With malicious actors.
你知道吗，所有这些关于蒸馏攻击的说法，加起来还不到我们一天在API上花费的时间。
You know, they they I mean, all of this stuff about the distillation attacks, it's like that you add up all of that stuff, it's less than one of our days spend on the API.
我真的不觉得这有多少影响。
I really don't think that's a ton
如果我没记错的话，我想去年我们有一些关于Cursor的数据，大约15%的用户来自中国。
If I remember correctly, I think we had some data maybe last year on cursor and something like 15% of users were from China.
我问了一些本地的人，他们告诉我，这些工具并没有被屏蔽。
And I asked a bunch of people locally, and they were telling me, yeah, these tools are not blocked.
不像谷歌和那些东西，你实际上可以使用西方的AI助手，因为政治性的太差了。
Like, unlike Google and some of that stuff, like, you can actually use Western AI agents because political ones suck.
所以，我其实不知道在Clock Code里它到底有没有被屏蔽。
So I actually don't know exactly in blog code if it's blocked or not.
他们根本不知道。
They have no idea.
所以我觉得这一点有明确的记录，但据说所有中国实验室都通过VPN连接到日本和韩国，然后调整他们的作息时间，以便模仿一个典型的韩国程序员的节奏。
So so I think this one is well well documented, but apparently, like, all the Chinese labs VPN into Japan and Korea, and then they, like, move their sleep schedules forward so they can replicate an average Korean vibe coder.
去年某个时候，他们说韩国和日本占了Anthropic收入的15%。
And at one point in time, last year, they were like, oh, Korea and Japan is 15% of Anthropic's revenue.
我当时就说，是的。
I was like, yeah.
那是中国的需求。
That's that's Chinese demand.
而且，没错。
There's no And and yeah.
所以，我的意思是，显然有人在使用。
So, I mean, there's, like, you know, like, there's obviously usage.
很明显，东方的模型更倾向于从哪些西方模型进行蒸馏。
I mean, it's very clear who the West the eastern models prefer to distill from.
有意思。
Plot.
对吧？
Right?
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。