Anthropic要求华尔街相信AGI

今天在Decoder上，我与芯片分析师Joey Brookhart和AI研究公司SemiAnalysis一起，帮助我挖掘Anthropic IPO存档(一个被称为S-1的文件)数字内的故事。

Anthropic要求华尔街相信AGI

SemiAnalysis的Joey Brookhart帮助破解Anthropic S-1

Anthropic刚刚上市了它的S-1 在Nasdaq号上公开，它可以在下个月开始交易，估值在1.5万亿到2万亿之间。 这将立即成为地球上最有价值的公司之一。

但价格标签其实不是全部故事。 本周早些时候出版的S-1文件在数百页上充满了细节，最终描绘了一个前沿AI实验室经济学的画面。 正如你可能想象的那样，这幅画既非常复杂，又建立在一系列非常，非常雄心勃勃的关于未来和AI会怎样对待它的假设之上。

因此我想让乔伊上节目，帮助我们破解这份文件告诉我们的内容，Joey和他在SemiAnalysis的同事实际上在7月从外部建立了一个Anthropic财务模型，因此他一直和任何人一样热切地等待着这个文件，我兴奋地听到他对我们学到的

SemiAnalysis ' Joey Brookhart在Anthropic S-1上，开始了

这次采访经过了轻度编辑，时间长而清晰。

Joey Brookhart 你是SemiAnalysis的分析师你也是SemiAnalysis的Tokenomics模型的领头人，欢迎来到解码器。

是的，非常感谢，尼莱。

我很高兴能和你谈谈我们只是讨论大Anthropic S-1被提交后的几天。 这也是我们第一次从金融角度审视一个领先的前沿模式。 你有何看法？你读到什么了？

这是一个令人赞叹的S-1。 从投资日算起已经三个星期了。 一些人对9月[每年的经常性收入]数字有些失望，但我们在这里看到的是一家公司，它有效地规模达到了惊人的水平，价值为1万亿美元，价值为700亿美元，时间很短。

以毛利润、计算利润和自由现金流通，我认为这份档案没有多少内容。 这是你看到很多半导体复合体以及新云和计算库存的原因在过去两天里再次凝聚起来

这很有趣，我不认为我能挑出S-1本身。 我想谈谈一些细节因为里面有很多值得思考的数字我其实只是想从大局开始你经营一个叫Tokenomics的东西。 我要向人们解释Tokenomics是什么。

我看着Anthropic和它可以去的地方。 我看着达里奥·阿莫代说，他可以看到自己1万亿美元的收入。 钱是从哪里来的？因为很多这些企业都是为了吃其他的生意，觉得别人必须挣点钱来支付Anthropic来创造这些收入。 我还没有看到这段话非常清晰。

Verge的订户们，头在这里。 不是订户吗？你可以在这里注册。

这很有趣。 如今，即使我们在两大实验室之间的规模是1200亿美元的ARR，它仍然是一个极其狭窄的市场。 客户的前1%是每个雇员花费5万到10万美元。 你更广大的财富500或全球2000

许多支出，特别是Anthropic的支出，都是特定的编码。 这几乎是一个垂直的公司。 我们展望未来到2027年你想想这个更广泛的收养和白领工作这个典型的企业Q4预算季节，我们实际上还没有经过一个这样的实验室，我们想到了网络等其他案例，可能类似于编码，这些都是与领先企业签订的八位数加合约，你得到其他垂直像生物。

为了达到1万亿美元，我们确实需要转向一些白领工作，从人工智能增强到一些地区的劳动力储蓄或替代。 以Microsoft的买断提议为例

我有一个最大的问题，不仅仅是关于S -1，而是整个行业。 我理解，如果有一个或两个主要的实验室。 有两家大公司拥有一个技术，可以自动化一大堆东西。 我们将看到自动化在各部门的运行情况，但不只是两个主要实验室。 有相当的几个。 有开放重量模型。

我从许多CEO那里听说，他们非常紧张地将所有数据交给OpenAI或Anthropic，现在有律师事务所在预告AI上运作这也许是历史上最有趣的转折

你怎么看这个游戏？因为这个S-1的压力很多是数字本身，我想谈谈这些，但后来又出现了外部压力，即“你必须增长，增长并不那么短 ”。 那么，你是如何看到这一切发挥出来？

这非常有趣，因为当我们今天看编码的时候，比如，我想我们在Q3中看到这个，OpenAI 有点赶上了，有些地方的编码成本非常低，我可以在两者之间移动。 我们从企业那里听到的，或许也是从OpenAI方面听到的——显然我认为他们在一定程度上搞错了Fable的发布，缺乏零数据保留就是为什么我们看到一些减速。 也许可以推动更多的企业领养。

对企业来说，我们肯定看到企业，更甚者，许多AI本地人，进行风俗训练和开放重量模型培训

显然，在Google和[Meta超级智能实验室]的DeepMind是资金非常充足的实验室，它们具有相当相似的训练计算能力，它们带来了一吨的计算能力。 更不用提SpaceX了，明年Eron 称的[大约]10千兆瓦净值

在上半年，Anthropic带着它逃走。 占该行业新ARR净额的90%。 我们在Q3中看到，OpenAI很可能实际上是工业净新ARR的大部分。 开源从可能10亿美元的象征性推论成本，ARR，在Q1增长到这里的100亿美元加在Q3.

(笑声) 是的，这个行业的竞争力大增了，但是也有一些新的使用案例，其中一些垂直化的使用案例真的是我们将要看到增长的前进。 因为这显然不能让我们从这里得到三倍的编码。

这很有趣。 这一概念被称为Tokenomics. 你在SemiAnalysis 的另一位联合组长旁听练习解释一下它是什么，因为对我来说，它是一个明智的思考方式，“我们要投入一些钱，也许我们会得到一些钱。 ”

我认为它是[AI]的单位经济学，甚至仅仅是广义的想法——从宏观到微观经济学，经营AI业务，甚至看AI需求的一切。 因此，我们覆盖了超标器，现在一些新云以及一个单独的产品。

但也只是人工智能实验室 Anthropic在两年前的推论中，现在他们有85%的正比值，人们不明白他们如何分配计算。 我们如何看待推论商业模式？

这种总比值对话的一部分是每兆瓦的推断收入。 为什么所有进行推断的公司都看到大量收益推论和计算效率？因此，我们试图模拟这些企业，显然当Anthropic是私人的，我们处理的信息我们可以从媒体和类似的东西。

但现在我们有了具体的审计数字，很容易解释这个单位的经济学。 还有企业的不同之处 OpenAI是一种与Anthropic大不相同的业务，仅仅在消费与订阅之间，然后是历史权重与消费者与企业之间。 [我们]试图从需求层面了解这些企业的经济。

几年前我和GoogleCEOSundar Pichai谈起推论的费用，I 称，“你怎样赚钱？” 他只是非常明确。 He 称，“如果有一件事Google的优点，它会使这些东西降低成本曲线。 ”

这是一个非常钝的答案。 我理解Google如何做到这一点。 他们自己制造芯片。 Anthropic运行在其他人的云端。 S-1的核心部分是他们承诺花费多少在其他人的云：SpaceXAI,Google,Amazon. 其效率从何而来？

也有真正的象征性吞吐量。 他们拥有非常多样化的芯片车队。 在Trainiums,TPUs，和GPU之间，在所有这些方面，我们看到一吨象征性的效率。 例如，在Trainium芯片上的吞吐量一年比一年多5到10倍

处理硬件层面的效率和软件，以及模型如何与硬件互动，有很多象征性的吞吐量收益，如果你每秒做X个符号，现在你做了5到10倍，仅仅一年之后。 他们会告诉你的是

Max Kan，他和我在SemiAnalysis公司共同领导Tokenomics模型，我甚至发表了一些研究。 现在，如果Anthropic的每兆瓦收入为6500万美元，这在这个时候是去年的中年。 我们的称在一块[Nvidia]上的 Opus 5.5 布莱克威尔，每兆瓦的收入可以达到1亿到1.2亿美元。

现在他们平均每兆瓦支付1 000多万美元，你能够做的是再投资培训。

人们并不意识到这一点，但只有大约30%的计算结果实际上是推算收入。 其余用于实验和研究、培训模式和培训后模式。 因此，在构建下一个模型方面，有很多计算，更多的是研发方面，以及你在这个行业中实际获得回报和辨识性的地方。

事实上，我期望在S-1中看到的是动态的发挥更强烈的—— Anthropic制造一大笔钱，卖掉他们现在拥有的模型，并且将来还有更多的客户来卖他们现在拥有的模型。 这将会减少一大堆的差额。 因为Anthropic认为他们正在建立一个数字神

这是一个非常特别的Anthropic故事。 公司正在追杀AGI. 在那里工作的人会很高兴地告诉你，在S-1中，Claude可能是有意识的，这一切可能会下地狱。 Anthropic在自身概念上是一个非常独特的公司。 但你并没有真正看到

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Today on Decoder, I’m joined by Joey Brookhart, an analyst at the chip and AI research firm SemiAnalysis, to help me dig into the story inside the numbers of the Anthropic IPO filing — a document known as an S-1.

Anthropic is asking Wall Street to believe in AGI

SemiAnalysis’ Joey Brookhart helps break down the Anthropic S-1 and why he thinks the company could be worth $10 trillion.

Anthropic just filed its S-1 to go public on the Nasdaq, and it could start trading as soon as next month at a valuation of anywhere between $1.5 trillion and $2 trillion. That would immediately make it one of the most valuable companies on the planet.

But the price tag really isn’t the whole story here. The S-1 document published earlier this week is hundreds upon hundreds of pages filled with details that finally paint a picture of the economics of a frontier AI lab. As you might imagine, that picture is both really complicated and built on a series of very, very ambitious assumptions about the future and what AI will do to it.

So I wanted to have Joey on the show to help break down what this document tells us, both about Anthropic’s finances and the broader AI economy. Joey and his colleagues at SemiAnalysis actually built a model of Anthropic’s finances from the outside back in July, so he’s been waiting for this filing about as eagerly as anyone, and I was excited to hear his take on what we’ve learned and, importantly, what we still don’t know.

Okay: SemiAnalysis’ Joey Brookhart on the Anthropic S-1 and the costs and risks of building frontier AI. Here we go.

This interview has been lightly edited for length and clarity.

Joey Brookhart, you are an analyst at SemiAnalysis. You’re also co-lead on SemiAnalysis’ Tokenomics model, which is going to come up a lot in this episode. Welcome to Decoder.

Yeah, thanks so much, Nilay.

I am very excited to talk to you. We’re talking just a couple days after the big Anthropic S-1 was filed. It’s our first look inside of the finances of one of the leading frontier models. What’s your take? What’s your read into it?

It was an awesome S-1. It’s been a rough three weeks since the investor day. Some people were somewhat disappointed in the September [annualized recurring revenue] numbers, but what we saw here was a company that has efficiently scaled to an amazing level, with a $1 trillion-plus valuation, $70 billion-plus of ARR in a very short amount of time.

In gross margins, compute margins, and free cash flow, everything is what you’d want to see. So I don’t think there was much to nitpick in this filing. It’s the reason you’ve seen a lot of the semiconductor complex, and the neoclouds and compute stocks, really rally again in the last two days.

It’s funny, I don’t think I can nitpick the S-1 itself. I want to get into some of the details because there are a lot of numbers in there that are worth thinking about. I actually just want to start at the really big picture. You run a thing called Tokenomics. I’m going to explain what Tokenomics is to people.

I look at Anthropic and where it could go. I look at Dario Amodei saying that he can see his way to $1 trillion in revenue. Where does that money come from? Because so much of these businesses are about eating every other business and it feels like someone else has to make some money to pay Anthropic to generate these revenues. I haven’t actually seen that laid out very clearly.

Verge subscribers, don’t forget you get exclusive access to ad-free Decoder wherever you get your podcasts. Head here. Not a subscriber? You can sign up here.

It’s so interesting. Today, because even the scale we’re at is $120 billion of ARR between the two leading labs, it’s still an extremely narrow market. The top 1 percent of customers are the ones that are spending $50,000 to $100,000 per employee. Your broader Fortune 500 or Global 2000, which typically do a lot of IT spend, isn’t spending much.

A lot of the spending, especially at Anthropic, is coding-specific. It’s almost a single vertical company. We look in the future to 2027, you think to this broader adoption and white collar work. This typical enterprise Q4 budgeting season, we actually haven’t gone through with one of these labs yet in any meaningful way because their ARR did really start to accelerate until the beginning of the year. We think of other cases like cyber, which can be similar to coding, where these are eight-figure-plus contracts with leading enterprises that can scale very quickly due to all the cybersecurity incidents we’re seeing. You get to other verticals like bio.

To get to $1 trillion, we do have to move to some of that white collar work, where it goes from AI augmentation to labor saving or replacement in some areas. You look at things like Microsoft’s buyout offers to employees as an example of this, but there’s no wealth of other verticals that they can get into and really start to move the needle.

This is the single biggest question I have, not just about the S-1, but the entire industry. I would understand that if there are one or two leading labs. There are two big companies that have a technology that can automate a whole bunch of things. We will see how well that automation goes sector by sector, but there are not just two leading labs. There are quite a few of them. There are open weight models.

I’ve heard from a lot of CEOs who are very nervous about giving all their data to an OpenAI or Anthropic and they would like to run some things on-prem. There are law firms running on-prem AI now, which is the funniest turn in maybe all of history.

How do you see that playing out? Because so much of the pressure on this S-1 was the numbers themselves, which I want to talk about, but then there’s the external pressure of, “You’ve got to grow and growth is not as short.” So how do you see all of that playing out?

It’s very interesting because when we look at coding today, for example — and I think we saw this in Q3 with OpenAI kind of catching up — there are areas of coding where there are very low switching costs, and I can just move between that. What we hear from enterprises and maybe, too, on the OpenAI side — obviously I think they somewhat botched the Fable release, and the lack of zero data retention is why we saw some of the slowdown. That’s something they can fix with Fable 5.5 and maybe drive more enterprise adoption there.

That’s a huge thing for enterprises when we talk with them. We definitely see enterprises now, and more so, too, a lot of the AI natives, doing custom training and post-training on open weight models, and then running either their inference off of that or even their products for internal or external use cases.

Obviously, DeepMind at Google and [Meta Superintelligence Labs] are very well-funded labs that have pretty similar training compute capacity, and they’re bringing on a ton of it. Not even to mention SpaceX as well, with what Elon said on [about] 10 net new gigawatts next year and half of that is the long-term internal use case.

In the first half of the year, Anthropic ran away with it. They were 90 percent of the industry’s net new ARR. We saw in Q3 that OpenAI was probably actually the majority of industry net new ARR. Open-source has grown from probably $1 billion of token inference costs, ARR, in Q1 to $10 billion-plus here in Q3.

So yes, the industry’s gotten much more competitive, but there are all these new use cases, too, and some of these verticalized use cases are really where we’re going to see the growth going forward. Because it’s obvious coding isn’t going to power us to even triple from here.

It’s interesting. The concept is called Tokenomics. You run that practice alongside another co-lead at SemiAnalysis. Explain what it is because to me it’s such a smart way of thinking about, “We’re going to put some money in and maybe we’ll get some money out.”

I think of it as the unit economics [of AI], or even just the broader idea — everything from a macro- to microeconomics of running an AI business or even looking at AI demand. So we cover the hyperscalers, now some of the neoclouds as well in a separate product.

But then just the AI labs, too. For investors, or just people that are interested in the business, Anthropic had negative 94 percent gross margins on inference two years ago. Now they have a positive 85 percent margin, and people don’t understand how they allocate compute. How do we think about the inference business model?

Part of that gross margin conversation is the revenue per megawatt in inference. Why are all companies that are running inference seeing massive gains in inference and compute efficiency? So we try to model out these businesses and obviously when Anthropic was private, we’re dealing with information we can get from the press and things like that.

But now that we have concrete audited numbers, it’s easy to explain the unit economics of this. And also how the businesses differ. OpenAI is a much different business than Anthropic just in both their consumption versus subscription and then historical weight versus consumer versus enterprise. [We are] trying to understand across the demand level the economics of these businesses.

Several years ago when I talked to Google CEO Sundar Pichai about the cost of inference, I said, “How are you ever going to make money?” He was just very clear. He said, “If there’s one thing Google’s good at, it’s bringing this stuff down the cost curve.”

It was a very blunt answer. I understand how Google can do it. They make their own chips. Anthropic is running on everyone else’s clouds. The core part of the S-1 is how much they’ve committed to spend on other people’s clouds: SpaceXAI, Google, Amazon. Where does their efficiency come from?

There’s really token throughput. They have a very diverse chip fleet. Between Trainiums, TPUs, and GPUs, across all of those, we see a ton of token efficiencies. Token throughput, for example, on Trainium chips is up 5x to 10x year-over-year right now.

Dealing with both the hardware level efficiencies and the software — and how the model interacts with the hardware — there are a lot of token throughput gains where if you’re doing X tokens per second, now you’re doing five or 10 times that just a year later. What they’ll tell you is they see continued compute efficiencies as well going forward.

Max Kan, who co-leads Tokenomics model with me at SemiAnalysis, and I have even published some research on this. Right now, if Anthropic’s at $65 million of revenue per inference megawatt, this was mid-teens last year at this time. What we’ve said on maybe Opus 5.5 on a [Nvidia] Blackwell, you can get to $100 to $120 million of revenue per megawatt.

The big thing here is they’re paying on average now just a little over $10 million per megawatt in [cost of goods sold], or just in compute costs. What you can do with those margins is then reinvest in training.

People don’t realize this, but only about 30 percent of the compute is actually in inference-serving revenue. The rest is for experimentation and research, training models, and post-training models. So there’s a lot of compute that is building the next model and that’s more of the R&D side and where you actually get your returns and defensibility in this business.

Actually, what I was expecting to see in the S-1 was that dynamic play out much more strongly — that Anthropic makes a bunch of money selling the models they have now and there are more customers in the future for the models they have now. That’s going to throw off a whole bunch of margin. And then, because it is Anthropic and they think they’re building a digital god, they’re going to train the next models as much as they can.

That is a very particular Anthropic story. The company is chasing AGI. The people who work there will happily tell you that in the S-1 that Claude might be conscious and this could all go to hell. Anthropic is a very unique company in its conception of itself. But you don’t really see that

Auto-fetched and translated by AI Atlas. All rights belong to the original author.