Anthropic’s Sprint to Listing: 75 Billion Cloud Commitment – A Moat or a Hanging Sword?

Shiwen: The news that Anthropic submitted its S-1 last week spread. Its revenue in the first half of the year jumped from $1.1 billion a year ago to $16.3 billion, while its net loss reached $80.7 billion. Even more striking is its cloud commitment for the next 7 to 10 years, which could amount to up to $750 billion. In this episode, we invite Yongliang to talk about what these numbers mean for an AI company on the eve of listing.

Yongliang: Shiwen, good. Let me introduce myself first. I have worked in technical management for 17 years, managing a team of up to 70 people. Now, I mainly provide technical and infrastructure consulting for various companies. Looking at the numbers in Anthropic’s draft prospectus, my first reaction is not about the valuation but that they include the computing power needed in the next few years in the financial reports all at once.

Shiwen: That’s an interesting perspective. Most people look at revenue and valuation first.

Yongliang: Revenue is certainly important, but cloud commitments better reflect a AI company’s real situation. Every dollar it earns now may be tied to several dollars it will have to pay to cloud providers and chip landlords in the future. This ratio is rare in traditional industries.

Shiwen: Let’s start with revenue. $16.3 billion in the first half, compared to only $1.1 billion last year—this growth rate is rare in tech history. What do you think?

Yongliang: The growth is indeed rapid, but we need to break it down. Of the $16.3 billion, a large portion comes from key clients. The prospectus reveals that about 20% of revenue comes from two major clients. If either of these clients reneges on the contract or starts developing its own technology, revenue stability will be immediately affected.

Shiwen: Is this concentration of customers very dangerous for a listed company?

Yongliang: Very dangerous. What listed companies fear most is not slow growth but unpredictable growth. Two major clients can control one-fifth of the revenue, indicating that Anthropic’s bargaining power isn’t as strong as外界 thinks. It may have to make concessions on contract terms, pricing, and service scope to retain these orders.

Shiwen: The net loss of $80.7 billion is also impressive, but you said nearly $72 billion were non-cash expenses?

Yongliang: Yes, mainly due to changes in the fair value of convertible bonds and other financing instruments. This is complex from an accounting perspective, but it reflects early financing costs. The more valuable the company becomes, the greater the book loss. Looking only at operational costs, $9 billion was spent on computing and infrastructure in the first half, exceeding the total for 2025’s entire year of $7.3 billion.

Shiwen: A large part of the loss is book loss, but cash consumption is also real.

Yongliang: That can be understood this way. More importantly, Anthropic told investors that its revenue in the third quarter reached $16.8 billion, a 45% increase from the second quarter. This news pushed valuation expectations higher again. The market now focuses not on whether it makes money but on whether it can continue to grow rapidly.

Shiwen: How long can this growth last?

Yongliang: That’s the problem. Rapid growth requires continuous investment, and its approach isn’t linear—it locks in computing power for about the next decade through long-term contracts. This strategy is rare even in the internet era.

Shiwen: This prospectus was secretly submitted on June 1st and only confirmed via the official blog on October 22nd. Why was it kept secret for so long?

Yongliang: Secret submissions are common practice, giving companies and the SEC enough time to modify documents and avoid repeated work after publication. For Anthropic, it also takes time to reassure investors. After all, a net loss of 80.7 billion dollars and a cloud commitment of 750 billion dollars—any one of these figures is impressive.

Shiwen: There’s news that Anthropic achieved adjusted profitability in the third quarter. Is this reliable?

Yongliang: This information comes from sources cited by Reuters, not directly from the prospectus. Even if true, there are many accounting adjustments between adjusted profitability and actual profit. Many companies adjust items like stock incentives, depreciation, and financing costs to present a better figure. Investors focus on free cash flow and operating profits, not adjusted numbers.

Shiwen: Let’s discuss these cloud commitments in detail. The undisclosed, irrevocable portion is about 450 billion dollars, with Google Cloud accounting for 111 billion, AWS 110 billion, Microsoft 31 billion, and NVIDIA chip rentals up to 86 billion. How did these figures come about?

Yongliang: These are pre-orders for computing power. AI companies need large amounts of GPUs and cloud resources for model training, but supply is tight, so leading companies sign long-term agreements to lock in capacity. Anthropic has secured capacity from cloud providers and chip rentals to ensure it has cards available in the coming years.

Shiwen: What’s different about these commitments compared to regular cloud purchase contracts?

Yongliang: Regular contracts are usually paid based on actual usage. Long-term commitment contracts often include minimum consumption, prepayment, or irrevocability terms. The prospectus mentions that most of these commitments require payment regardless of whether they’re used. Even if training demands decline next year, Anthropic still has to continue paying.

Shiwen: What does 750 billion dollars mean?

Yongliang: If spread over 10 years, it’s an average of 75 billion dollars per year. Yet Anthropic’s revenue in the first half of the year was only 16.3 billion dollars, and its full-year revenue will likely be less than 40 billion dollars. Its cloud commitment spending could be close to or even exceed twice its annual revenue. This kind of leverage is hard to imagine in traditional industries.

Shiwen: But why sign so many?

Yongliang: One reason is that without signing, they can’t secure capacity. The world’s top GPU capacity is limited, and cloud providers also use it themselves. Another reason is that long-term contracts offer better prices. But for a company just going public, this means betting on future cash flows and strategic flexibility.

Shiwen: Could these commitments become a drag on future profitability?

Yongliang: Yes. If model training efficiency improves or demand drops in the future, these commitments will become sunk costs. More problematic is that AI technology evolves quickly. Computing power bought today might not be needed tomorrow, but the contracts remain.

Shiwen: Such large investments naturally lead to discussions about the AI infrastructure bubble. What do you think?

Yongliang: Whether it’s a bubble or not depends on whether there’s enough demand to support it in the end. Currently, large model companies are competing with each other in capital expenditure, and costs for data centers, electricity, and chips are rising. If AI application revenue growth slows over the next few years, these locked-in capacities will become a burden.

Shiwen: But if not locked in now, they might not be able to buy them in the future.

Yongliang: This is the mentality of a gambling table. Everyone is betting that they will be one of the few remaining. OpenAI recently raised a private equity round worth up to $30 billion, with a valuation of $1.25 trillion. Anthropic’s valuation is discussed between $1.5 trillion and $2 trillion, and Reuters says it’s more likely to be around $1.5 trillion.

Shiwen: What are the differences in the two approaches?

Yongliang: OpenAI chooses to continue with private equity fundraising without rushing to go public, reducing financial disclosure pressure. Anthropic, on the other hand, decides to go public, which means it must face more transparent financial scrutiny. The fact that the cloud commitments are fully disclosed in the prospectus is a signal: it wants the market to see its cards, and it links future growth to these commitments.

Shiwen: What practical impact does going public or not go public have on Anthropic’s financing ability?

Yongliang: After going public, Anthropic can raise funds through bonds and additional shares from the public market, providing more diverse sources of funding. However, the cost is that it must disclose financial reports quarterly, with the market keeping an eye on its profit path. If growth falls short of expectations, stock prices will fluctuate significantly.

Shiwen: Does Anthropic’s decision to go public also relate to the payment schedule outlined in its cloud commitments?

Yongliang: It’s closely related. Long-term commitment contracts usually require upfront or phased payments, and after going public, the funds raised can directly cover these expenses. Otherwise, relying solely on private equity fundraising would mean higher costs and more stringent terms.

Shiwen: Besides financial numbers, what other notable risks are mentioned in the prospectus?

Yongliang: One is the risk of customer concentration, with about 20% of revenue coming from two major clients. The other is so-called catastrophic AI incidents, which may exceed insurance coverage. This sounds distant, but it shows that Anthropic itself acknowledges that the systemic risks associated with large models have not been fully priced.

Shiwen: What exactly are the insurance issues for catastrophic AI incidents?

Yongliang: Once a large model experiences serious security incidents, such as abuse or uncontrollable output, the losses could far exceed what ordinary insurance covers. Insurance companies currently lack mature pricing models for such risks, so Anthropic may face a situation where there are no policies available. This risk is difficult to resolve through market mechanisms in the short term.

Shiwen: Cloud providers are both suppliers and potential competitors. How can this contradiction be resolved?

Yongliang: It cannot be resolved in the short term. Hyperscalers also develop models themselves; Google has Gemini, and AWS and Microsoft are also launching their own solutions. They rent computing power to Anthropic while also competing for its customers. In the future, if computing power becomes scarce, priority ordering will be very practical.

Shiwen: The suppliers that Anthropic currently relies on may become its biggest competitors in the future.

Yongliang: Yes. The prospectus itself mentions that hyperscalers increasingly view Anthropic as a competitor rather than a customer, which could lead to difficulties in obtaining computing power. This is not unfounded worry but a structural change that is already happening.

Shiwen: If the AI infrastructure bubble really breaks, who will be affected first?

Yongliang: The first to be affected are those companies that have made huge long-term commitments but fail to keep up with revenue growth. Next are chip leasing companies and data center operators, whose customers are starting to default or demand renegotiation. Finally, there are cloud providers, which, despite their size, face pressure from overcapacity that lowers prices.

Shiwen: Will Anthropic be the first to be affected?

Yongliang: Not necessarily. Its revenue growth is still high, and it has major customers and capital support. But if its revenue growth slows in the next two years while cloud commitments continue to grow rigidly, the pressure will be very great.

Shiwen: What lessons can readers who don’t invest directly take from this?

Yongliang: One perspective is that competition in the AI industry has shifted from model capabilities to capital and supply chains. Whoever can lock in more computing power and train models at lower costs will have a chance at the next round. Another perspective is that these numbers remind us that the commercialization of large models is not yet stable; behind high revenue growth is equally high capital expenditure.

Shiwen: So going public isn’t the end, but the start of a new round of financing and competition.

Yongliang: Exactly. For Anthropic, going public brings funds, but it also brings quarterly disclosure pressure. When its cloud commitments, customer structure, and reasons for losses are exposed, the market will expect it to provide a profitable path with stricter standards.

Shiwen: Let’s summarize.

Yongliang: Anthropic’s prospectus shows the true nature of a top AI company: impressive revenue growth, but equally impressive losses and long-term commitments. $750 billion in cloud commitments is both a moat and a sword hanging over its head. In the next few years, its task is not to prove it’s still growing, but to show that this growth can cover such massive fixed costs. Only then will this prospectus be truly recognized by the market.

Shiwen: Good, that’s all for this episode. If you have anything else to say about Anthropic going public, feel free to leave a comment.

【Column】