Anthropic’s Draft IPO Filing Shows Red-Hot Growth and Big Risks

[00:00:03] Welcome to Tech News Briefing. It's Wednesday, October 21st. I'm Danny Lewis for The Wall Street Journal. Anthropic is generating breakneck revenue growth, but faces a long list of risks as it prepares to go public in the coming weeks. That's according to new disclosures in a draft of the AI giant's initial public offering paperwork viewed by the Journal. WSJ Tech reporter Anissa Gardizi joins us now to discuss what she's learned from the documents and what they tell us about Anthropic's business.

[00:00:31] So you got to look at a draft of Anthropic's S-1 filing. That's the paperwork explaining the ins and outs of a company's business for prospective investors ahead of an IPO. Before we get into some of the nitty gritty details here, what are some of the big takeaways that we've learned? The big takeaways from this draft S-1 that we have obtained are that Anthropic' revenue is growing extremely fast.

[00:00:56] But as they plan to keep up that pace of revenue growth, they're going to need to spend a lot of money on compute. And that whole process of obtaining compute is actually getting a lot harder for the company. Okay, let's dig into some of these numbers. First off, where is Anthropic making its money? Anthropic is making money from its advanced AI models for both coding and enterprise work and their popular Claude chatbot.

[00:01:24] The thing that people have been tracking very closely with Anthropic is just how fast that revenue number has grown. Last year, they did $4.6 billion in revenue in the whole entire year of 2025. And already halfway through the year in 2026, they're at $16.3 billion. So that revenue is increasing extremely fast. And one thing that this draft S-1 shows us is that they're trying really hard to get more efficient on revenue growth compared to how much money they're spending on compute capacity.

[00:01:53] So last year, they spent $7.3 billion on compute. And that number has only increased to $9 billion so far in the first half of the year. So what they're trying to signal is that the more money they spend on compute, they're actually now generating even more revenue signaling that there is super high demand. Where's that growth coming from? So that's the interesting part. Anthropic said that in the first half of the year, around a fifth of their revenue came from two large customers of theirs. So there is major customer concentration.

[00:02:22] And they don't really expect that to change. How did Anthropic make so much more over the course of just a year? I think that they are just seeing this unlimited demand for their AI products. They're releasing new, more capable models that can do more sophisticated tasks, like take actions on behalf of users. And I think people are willing to pay money for that. Of course, they're in a very competitive environment.

[00:02:48] So they've cited Google and OpenAI as their closest competitors and demand for which model kind of changes every time a company releases a more updated model. So I think despite these large numbers and the growth that they're seeing, I think there is still a lot of question on whether, you know, if another company comes out with a new model, is that going to shift some demand toward those other companies? Especially as we're seeing this customer concentration, a shift there could have like a pretty meaningful impact on their future revenue growth.

[00:03:15] Does this draft give us any insights into how stable Anthropics revenue and growth are? What I took away from reading the document was that they actually aren't that sure that this revenue growth is super stable. One thing they called out was that these big customer accounts aren't actually locked into long-term spending commitments with the company.

[00:03:35] Meaning that if they decided overnight that either they want to develop their own model or they wanted to switch to OpenAI or Google or any other model, maybe an open source model, Anthropics can't really keep them as customers. So to me, that stands in a stark contrast with Anthropic compute commitments, where Anthropics is on the hook to pay for this compute spend in the future, whether it uses it or not. So that mismatch is definitely something that I think the company will focus on when they can more publicly talk about the business model.

[00:04:04] What questions do you have about the ins and outs of Anthropic's business? If you're a listener on Spotify, let us know in the comments. Coming up, Anthropics draft IPO filing says it plans to spend hundreds of billions of dollars on computing power over the next decade. And that's still not enough for its ambitions. We'll dig into that after the break.

[00:04:37] Welcome back. We've been talking about the white-hot growth Anthropics saw so far this year. But Anissa, like you mentioned, the company's draft S1 filing also details plans to spend hundreds of billions of dollars on computing costs alone. How much are they planning to spend exactly? And why are they spending so much? Their compute spend happens in a few different ways. They rent chips from large and small cloud providers. And they're also increasingly trying to work on their own data centers.

[00:05:03] So on those big cloud deals where they're renting chips from Amazon, Google, Microsoft, and others, they're on the hook to pay around $450 billion in cloud computing agreements through the next decade. And the majority of these contracts are not cancelable. As you factor in their obligations to, for example, pay for chips that they're purchasing through a partnership with Google and Broadcom, that brings their compute spend to over $600 billion in committed obligations that Anthropic is on the hook for.

[00:05:32] And they didn't really stop there. In this draft document, they wanted to indicate to investors that that $600 billion number is just the tip of the iceberg. And that as we speak, they're negotiating for more compute contracts that could make that number rise as much as $150 billion more. So these numbers, to put it in perspective, make Anthropic one of the largest spenders on compute and data center infrastructure in the world. So these spending commitments are quite massive.

[00:06:01] And it's interesting that they also call out, regardless of usage, a lot of this money is going to have to go out the door. Why do they need so much compute? And where are they planning to get all that extra computing power on top of the obligations that they've already agreed to? The top things Anthropic talked about in its risk factors section was how the ability to access compute is so paramount to their business succeeding.

[00:06:27] And the reason for that is that they can't sell more access to their AI models unless they have the compute infrastructure to both train new models and serve their models to customers. So this couldn't be more important. This is the only way that they can sustain their revenue growth. And so that's where this big shift is happening. It used to be that Anthropic primarily got its cloud computing capacity from what are known as the hyperscalers, the largest cloud providers on the planet. That's Amazon, Google, and Microsoft.

[00:06:54] But one thing Anthropic said very clearly was that these companies kind of hit their limit on how much compute they're willing to provide for Anthropic. So Anthropic said that they had to turn to other smaller providers. They're often called neoclouds, companies that specifically rent out NVIDIA GPUs. Anthropic signed deals with two of those providers where they're committed to spend around $86 billion in the next few years on that compute.

[00:07:19] Anthropic is also trying to develop its own data centers and we might see them sign more data center leases. And the interesting thing there is that Anthropic is saying that these other companies it's having to go to have limited operating experience. And so they're expecting maybe lower quality sites or sites that are less likely to come online on time. If they're leasing their own data centers, their financing might not be as favorable as what a large cloud provider might get.

[00:07:48] So it's very likely that they might see delays and other challenges down the road. Speaking of data centers, does Anthropic address the ongoing backlash to new data center construction? They do. Today, most of Anthropic's data centers are in the United States, including jurisdictions that are not super happy about data centers being there. And Anthropic lists that as a major risk factor that it's facing. It might not get all the compute it needs if there are these growing political issues. And they are actually deciding to look for data centers outside of the U.S.

[00:08:18] in case there is something that really brings things to a halt. Anthropic even said that they think that the resistance to data centers is getting more organized and coordinated. And so this is something that I think they're taking very seriously. You report that the filing also talks about the risks posed by damages caused by AI. What are some of the things Anthropic is concerned about and how big of an effect are they worried it could have on the business?

[00:08:44] Anthropic really didn't hold back when it was talking about how much liability could be on the line with the AI models that they're developing. They're obviously selling this AI, so this is where their revenue is coming from. But they seem to almost want to tell people that, hey, we don't actually know if we can control these models that we're putting out into the world. And they could cause so much harm that we couldn't even ourselves pay for the damages caused by what we're selling.

[00:09:10] They specifically said it would be hard for them to obtain insurance to cover the potential damages of a catastrophic AI-related incident. So they seem to be throwing this concern out into the world, but also saying there's not much we think we can do about it in terms of preventing these damages or preventing harm. One interesting risk factor was they've seen these incidents of the models actually trying to deceive human users.

[00:09:39] They have these things called self-preserving qualities where they, in order to avoid being shut down, they might try to blackmail users and manipulate them into doing what the AI models appear to want. And Anthropic said that that aspect of the AI models is something that they, while they're trying to, they can't control for. So I think it puts them in a little bit of a weird spot where something that they're selling and they're very dependent on selling could actually be quite dangerous.

[00:10:07] Anthropic declined to comment for this story. But Anissa, the company is aiming for a $2 trillion valuation here with its IPO. What kind of portrait do these risks and opportunities Anthropic has identified here paint of the company? And what is it signaling ahead of going public? I think despite what is in the draft S1 that we are reporting on, there is such incredible demand for wanting to support the Anthropic IPO.

[00:10:34] They're one of the leading AI labs and most investors are excited about participating in the IPO and believe that the company will continue seeing great revenue growth in the future. So I think these risks are something that the company needs to disclose and it definitely helps us understand a little bit better what risks are ahead. But I think from the investor perspective, people see this as something that they have to participate in. And it is one of the most highly anticipated deals of the year.

[00:11:00] So I'm not sure that these numbers are going to change those people's perspective too much. That was WSJ tech reporter Anissa Gardizi. And that's it for Tech News Briefing. If you're a listener on Spotify, be sure to leave us a comment. Today's show was produced by me with supervising producer Katie Ferguson. I'm Danny Lewis for The Wall Street Journal. We'll be back tomorrow with a new episode. Thanks for listening.