What Anthropic’s S-1 Says About $400 Billion in Compute Commitments

Let's look more closely at the economic side of where all this is going. Something important came out today: Anthropic's S-1 filing. As I read through it and thought about what it means, I noticed a few interesting things. Some relate to Jim's point about how we should think about these problems. Others may reveal specific challenges for leading AI companies that could cause problems for them later. First, Anthropic's revenue for the first half of the year was $16.3 billion, a huge increase from the year before. Its revenue for all of 2025 was $4.6 billion. In the second quarter, the company reported a positive operating income, meaning its income from normal business activities was greater than its operating expenses. But that number needs context: the company has more than $400 billion in future commitments to pay for computing power, and those commitments apparently cannot easily be canceled.

The Financial Times reports that Anthropic is expected to seek a valuation of $1.5 trillion to $1.75 trillion. But expectations are easy to make before a company actually enters the market. The company's governance structure also caught my attention. The founders' Class F shares control 50.1% of the votes. A long-term benefit trust chooses most of the board members, and future board members and trustees are chosen by the people already in those roles. This means ordinary shareholders have little real power to change how the company is governed from the outside. Looking at all this, I see two things. On the one hand, revenue is growing very quickly, and Anthropic makes some of the world's most capable AI models. On the other hand, the company has committed to spending an enormous amount of money and will need to find a way to pay for it. Jim, you've been studying these issues for a while. What did you think of the S-1? Have you had a chance to read it?

Yes, I had some time to look through it today. One interesting thing was that Anthropic compared itself to pharmaceutical startups. These companies make large investments at the beginning and hope to develop a drug that will bring in money for many years. That seems to be the kind of long-term income Anthropic hopes to earn from its products, too. Another thing that stood out to me as someone who works in finance was the risk section. It discusses the possibility that the company's work could pose an existential risk to humanity, meaning it could threaten humanity's continued existence. I've asked how the head of a state pension fund, a city government, or another organization responsible for managing other people's money should read that warning. How could they say, 'This company might end humanity, but look at its financial forecasts for 2030. We really need to invest in it'? There's another concern. Imagine buying the stock and then seeing it fall by 50%, as has happened after some major technology IPOs, including Facebook's IPO, or in the case of SpaceX. You could face a lawsuit. Lawyers might point to the risk warning and ask why you thought you were qualified to manage other people's money when you invested in a company that disclosed such a serious risk. They might ask, 'What part of the warning about ending humanity did you not understand?' Because of this, public pension funds and other organizations managing public money might face pressure not to invest. We'll see whether my concerns turn out to be justified. There may also be enough private investment money to make up for that and allow Anthropic to complete the deal at the price it wants.

They'll raise the money because there's so much excitement around the company. But the question is whether they'll raise it at a valuation of $1.75 trillion or $2 trillion. I'd lean toward the lower end. I have a few observations. First, the S-1 doesn't show gross margin. A few weeks ago, leaked financial information suggested that the company's gross margin was 85%. Gross margin is calculated by comparing revenue with the direct costs of earning that revenue. But Anthropic has grouped its costs together, so we can't separate the cost of training models from the cost of running them, also called inference, or the cost of research and development. This makes the accounts hard to understand and is unlikely to inspire confidence. The bigger issue is that open-weight AI models, whose model weights are available for people to use, are taking up a growing share of computing activity. I don't see that trend stopping. DeepSeek, a company from China, released 4.1 Flash. Its open-weight models can be run in the United States, so users don't have to send their traffic to China. The model costs about 90% less and performs almost as well. And by 'almost as well,' I mean it may be good enough for the job. You can simply run the model three times to get a better result.

These are some of Anthropic's biggest challenges. Another problem is the software environment people use as their main workspace. In financial services, the main workspace used to be the Bloomberg Terminal. For coding, it was Claude or Codex. But now, the main workspace is becoming a harness, a system that lets multiple AI agents work with one another. Claude doesn't have a harness of its own that works independently of its models. The harnesses that use Claude favor Claude rather than working equally well with different models. As a result, Anthropic is becoming less well positioned among developers over time. That concerns me.

I want to add that the competition isn't just coming from DeepSeek. The AI model market is becoming much more divided. If you look at the different models and how people actually use them, I think there are three main groups. First, some customers use the most advanced models directly from companies like Anthropic and OpenAI. To Anthropic's credit, many businesses use Claude for their internal development work or access it through its API, a tool that lets other software connect to the model. These are important customers. The second group wants the best model for the best price and uses services such as OpenRouter to access different models. Ram, you know that DeepSeek Flash is a good model. I think its newest version, 1017, performs about as well as Grok High but costs much less to run. There are also models from Mimo, Qwen, and GLM that offer roughly similar performance at very low prices. On top of that, Kimi has just released K 3.1, which competes with Muse Wonder and Claude models while costing much less to run as an open-weight model. This makes it harder for Anthropic to maintain its competitive advantage.

The third group consists of people who want to run everything in their own Amazon Bedrock environment or on systems they host themselves, and who may train their own models. Coinbase, for example, was doing this with a version of Qwen for fraud prevention, running the work inside its own systems. So, as the market becomes more divided, one thing I'm watching closely at Anthropic is whether its revenue keeps rising at an accelerating rate or whether its growth starts to slow because customers are moving to these other options. That will tell us a lot about whether Anthropic can fund its future spending commitments.

Its revenue growth is still accelerating. It could reach $100 billion in revenue by the end of the year. The company has also lowered its prices per token, the basic units of text that AI models process. As you said, OpenRouter can send requests to whichever provider offers the best combination of low cost and high performance. Anthropic saw a spike in demand. So the sequence was that negative news first came out through leaked information, then the S-1 was released, and now there's positive news about token prices. But the bigger issue is competition on OpenRouter. It gives different coding models a prominent place, and the competition is fierce. Ken Griffin of Citadel made a comparison to the airline industry that I think is right. I've made the same comparison myself. Airlines earn money on each additional sale, can have low costs for each extra customer, and often rely on debt to finance their businesses. If revenue falls while fixed costs remain, the business can quickly get into trouble. Those costs will remain for Anthropic because its long-term commitments are difficult to escape. Anthropic and OpenAI have little choice but to keep making the best models in the world if they want to charge higher prices. That's difficult to sustain. We're also seeing more specialized models designed for specific tasks. For example, a video-generation model released today costs one-quarter as much as Higgsfield's video-generation service. Specialized competitors could take business away from larger companies. OpenRouter will direct token usage and spending toward the providers that offer the lowest costs.

If you enjoyed this segment, please like and subscribe, and join us every Monday at 4:30 p.m. Eastern time. I'm Austin Campbell, host of Bips and Bips, with my friends Ram Ahluwalia and Chris Perkins and our excellent guests. Each week, we discuss the economy, cryptocurrency, and how these worlds come together, covering the issues that move markets and shape the financial world.