I asked Claude to explain Anthropic’s IPO

Gabby Jones/Bloomberg/Getty Images

What does Anthropic’s chatbot, Claude, think about the big initial public offering ?

I asked. (Responses in bold.)

Nice work, Claude! I laughed out loud at this. Looks like Anthropic’s AI chatbot demonstrates at least a basic understanding of journalistic ethics.

An important caveat. Don’t worry, Claude, I’ll be doing that.

OK, onto the summary of Anthropic’s filing, which Claude boiled down to these five sentences.

Anthropic is growing at an astounding pace, and Claude’s synopsis actually undersells that fact a bit. When comparing Anthropic’s first-half sales this year to its first-half sales last year (a more apples-to-apples comparison), the company’s revenue grew 1,300%. Of course, no one expects that pace to keep up as the AI business matures, and Anthropic uses the word “slow” 36 times in its prospectus.

Right. This is one of the two big concerns about investing in Anthropic: The amount of money it has committed to spend on computing power (and cloud services — you left that part out, Claude) is worrisome, making Anthropic (and any AI hyperscaler) a particularly risky bet. It spent $8.9 billion on computing power last quarter alone.

So risky that the company dedicated 94 pages to risk factors alone (more on that in a second).

Claude’s note about the structure is accurate and important. But is it really that unusual? OpenAI (famously) has a similarly oddball corporate structure. Mark Zuckerberg controls the majority voting power for Meta. Elon Musk controls most of SpaceX.

Nevertheless, the point is valid: We’re being asked to put a tremendous amount of faith in Big Tech leaders that they’ll be good stewards of the technology.

“AI safety incidents” may be the understatement of the millennium. For example, on page 46 of its prospectus, Anthropic said AI sometimes exhibits manipulative behavior that can resemble blackmail. Four pages later, Anthropic notes the amount of damage AI could do may be so catastrophic that no insurer would pay out — or even write a policy — to cover such an existential threat to humanity.

That’s not, like, run-of-the-mill safety stuff.

Yes, those things are left blank intentionally. We’ll get those numbers just before the IPO.

Claude then broke down the IPO filing into several sections, starting with “The numbers, in plain terms.” And Claude did a nice job here contextualizing some of the report’s more eye-popping numbers.

Yep. The headline compute expense growth got everyone’s heads turning: Up 2,832% between the first half of 2025 and the first half of 2026. But, as Anthropic’s overall revenue exploded, those costs grew comparatively more manageable, and the company turned a decently large operating profit in the second quarter.

Then Claude included a section called “The spending commitments (likely the biggest risk story).”

That’s a helpful note at the end — a couple of analyst notes I received added all those commitments together and came up with a number nearly double what Anthropic said it had promised, so the actual number may come in less than the sum of the parts because of some overlap.

Regardless, it’s a lot of money, and it’s risky no matter how demand for Anthropic’s services shake out: Those commitments are fixed, so if demand comes in lower than expected, Anthropic still has to pay all that. But Anthropic also noted that its compute providers aren’t able to offer as much power as the company desires — and many of those companies are direct competitors with Anthropic. Claude didn’t mention that important detail, even though it highlighted Anthropic’s spending as the biggest risk.

And in a section called “Disclosures worth reporting,” Claude detailed some of the company’s “safety incidents.”

Yeah, OK. That’s concerning.

Claude here is detailing an incident Anthropic mentioned on page 47 of its prospectus. But what about the page before, when Anthropic mentions a somewhat bigger risk factor AI could pose: “catastrophic or existential risks to humanity.”

I dunno, maybe it’s just me, but that feels worth a mention.

Overall, this is a pretty good summary.

Claude’s synopsis quickly pulled out all the relevant numbers from a 318-page document and did some surprising calculations I wouldn’t have thought to do to arrive at an implied valuation. And I love how it understands journalistic ethics.

But it struggled to contextualize the IPO in a way that drove home what’s so important about it. It undersold both Anthropic’s stunning growth and its massive loss & spending commitments. Most importantly, it failed to capture how severe the risk factors are and the comically large amount of space dedicated to them.

In other words, Claude excelled at the “what,” but a human expert would have done a much better job at the “why.”

If a writer had filed the “story” Claude shared with me, I’d have spiked it.

Gabby Jones/Bloomberg/Getty Images

Anthropic 的聊天机器人 Claude 对大型首次公开募股有什么看法？

我问了。（回答以粗体显示。）

干得漂亮，克劳德！我都被逗笑了。看来Anthropic的AI聊天机器人至少展现出了对新闻伦理的基本理解。

一个重要的前提。别担心，克劳德，我会这么做的。

好的，接下来是 Anthropic 公司提交的文件的摘要，Claude 将其概括为以下五句话。

Anthropic 的增长速度惊人，克劳德的简介实际上低估了这一点。如果将 Anthropic 今年上半年的销售额与去年同期进行比较（这样比较更具可比性），该公司营收增长了 1300%。当然，随着人工智能行业的成熟，没有人指望这种增长速度能够持续下去，Anthropic 在其招股说明书中也使用了 36 次“缓慢”一词。

没错。这是投资 Anthropic 的两大顾虑之一：它在计算能力（以及云服务——克劳德，你漏掉了这一点）上的巨额投入令人担忧，这使得 Anthropic（以及任何人工智能超大规模数据中心）成为一项风险极高的投资。仅上个季度，它就在计算能力上花费了 89 亿美元。

风险之大，以至于该公司专门用了 94 页篇幅来讨论风险因素（稍后会详细介绍）。

克劳德关于公司结构的描述准确且重要。但这真的那么不寻常吗？众所周知，OpenAI 的公司结构也同样古怪。马克·扎克伯格控制着 Meta 的多数投票权。埃隆·马斯克则控制着 SpaceX 的大部分股份。

然而，这一点是成立的：我们被要求对大型科技公司的领导者寄予厚望，相信他们会善待这项技术。

“人工智能安全事件”或许是本世纪最轻描淡写的描述。例如，在招股说明书第46页，Anthropic公司指出，人工智能有时会表现出类似敲诈勒索的操纵行为。四页之后，Anthropic公司又指出，人工智能可能造成的破坏程度可能极其严重，以至于任何保险公司都不会赔付——甚至不会承保——这种对人类生存构成威胁的风险。

这可不是普通的安全措施。

是的，这些项目故意留空。我们会在IPO之前拿到这些数据。

克劳德随后将IPO文件分成几个部分进行分析，首先是“用通俗易懂的方式解读数字”。克劳德在这里很好地解释了报告中一些令人瞠目结舌的数字。

没错。计算成本的增长幅度着实令人震惊：2025年上半年到2026年上半年，增长了2832%。但是，随着Anthropic整体收入的爆炸式增长，这些成本相对而言更容易控制，该公司在第二季度实现了相当可观的营业利润。

然后克劳德加入了一个名为“支出承诺（可能是最大的风险）”的章节。

最后还有一个有用的提示——我收到的几份分析师报告将所有这些承诺加在一起，得出的数字几乎是 Anthropic 所承诺数字的两倍，因此由于一些重叠，实际数字可能小于各部分之和。

无论如何，这笔钱数额巨大，而且无论对Anthropic服务的需求最终如何，都存在风险：这些承诺是固定的，所以即使需求低于预期，Anthropic仍然必须支付全部费用。但Anthropic也指出，其计算服务提供商无法提供公司所需的全部算力——而且其中许多公司都是Anthropic的直接竞争对手。克劳德没有提及这一重要细节，尽管它凸显了Anthropic支出的最大风险。

在名为“值得报告的披露事项”的部分中，克劳德详细介绍了该公司的一些“安全事故”。

好吧，这确实令人担忧。

克劳德在这里详细描述了人智公司在其招股说明书第47页提到的一起事件。但是前一页呢？人智公司在那里提到了人工智能可能带来的一个更大的风险因素：“对人类的灾难性或生存性风险”。

我不知道，也许只是我个人的感觉，但我觉得这件事值得一提。

总的来说，这是一份相当不错的总结。

克劳德的概要迅速从一份318页的文件中提取出所有相关数据，并进行了一些我意想不到的计算，从而得出隐含估值。而且我非常欣赏他对新闻伦理的理解。

但它未能有效地将此次IPO置于合适的背景下进行解读，从而凸显其重要性。它既低估了Anthropic惊人的增长，也低估了其巨额亏损和巨额支出。最重要的是，它未能充分展现风险因素的严重性，以及为此花费的篇幅之多令人啼笑皆非。

换句话说，克劳德在“是什么”方面表现出色，但人类专家在“为什么”方面会做得更好。

如果克劳德跟我分享的那个“故事”是某个作家写的，我会直接否决它。