Anthropic's Official Prospectus Unveiled: Can AI Large Language Models Achieve Sustainable Profitability?

At 9 p.m. Beijing time last night (9 a.m. Eastern Time on October 26), Anthropic officially filed its prospectus with the U.S. Securities and Exchange Commission. The company plans to list on the NASDAQ under the stock symbol ANT. For the first time, one of the world's two leading AI model companies has publicly revealed its full financial and business performance.

The prospectus nearly revived enthusiasm for AI stocks in the U.S. market on the day it was released. The NASDAQ Composite Index rose 0.82%, while the PHLX Semiconductor Index climbed 2.82%. Just last Thursday, the semiconductor index had fallen 6.72% in a single day. This autumn, investors have grown uneasy about the AI market, and one company after another has put its planned stock listing on hold. Anthropic is still moving forward, while OpenAI plans to go public in 2027.

The prospectus matters because it provides official answers for the first time to several questions that have sparked intense debate about AI over the past three years. Can large AI models really make money? Will computing costs consume all their revenue? Are businesses actually using these technologies? The listing will also give Wall Street a benchmark for valuing leading AI companies. It will bring stock market investors into a field that has mainly been funded by venture capital firms, government investment funds and major technology companies.

At least in the first half of this year, Anthropic's results were strong. Its total revenue for 2023 was only $41 million, but in the first half of this year it reached $162.86 billion, 14 times the amount for the same period last year. Its operating profit under generally accepted accounting principles (GAAP) was $9.74 billion in the second quarter. The reported net loss of $807.48 billion in the first half of the year mainly came from accounting changes in the estimated value of financing instruments. This did not involve any cash leaving the company.

Many people may see this as a beautiful financial report from a fast-growing company. But I think that view is only half right. By the end of June, Anthropic had signed agreements to buy computing power and leases that had not yet started, worth about $5.87 trillion in total. It will pay these amounts gradually over nearly 20 years. The total is more than four times its cash and short-term investments. After June, the company signed more agreements and increased the size of existing ones to secure more computing power. These changes added at least another $2.45 trillion to its future payment commitments.

This figure will probably keep growing. Demand for AI has repeatedly been higher than expected, even by Anthropic itself. A September column in the Financial Times said that a year and a half ago, Anthropic expected its full-year revenue in 2027 to be only $120 billion. Its revenue in the first half of this year has already passed that amount. If demand keeps rising this quickly, the company will need to buy even more computing power.

The prospectus shows that Anthropic's revenue can now cover the computing power it is already using. But it does not yet answer where the money for the next rounds of computing power will come from.

So far, cloud service providers and chipmakers have paid most of the costs of buying chips and building data centers upfront. Anthropic then pays them according to its contracts. After the company goes public, stock market investors will help fund part of these costs. How much investors are willing to pay, and what price they will accept, will ultimately depend on whether Anthropic can keep selling its most powerful model for more than competing models cost.

01 Revenue has temporarily grown faster than computing costs

Anthropic's business has three layers. At the bottom are its AI models. In the middle is a platform that developers use to build applications. At the top are its own products, including the Claude.ai chat service and the Claude Code programming tool. Its models fall into four levels: Mythos, Opus, Sonnet and Haiku. Each has different abilities and prices, so customers can choose the one that suits their tasks.

In the first half of 2026, 80% of Anthropic's revenue came from charges based on usage. The company measures usage in tokens, and the price depends on which model the customer chooses. Anthropic calls this approach "pay for intelligence." The remaining 20% came from subscriptions. Looking at how customers buy its services, 48% of revenue came through the cloud marketplaces run by Amazon, Google and Microsoft. These cloud providers collected the money on Anthropic's behalf.

Businesses are Anthropic's main source of revenue. By September 30, 2026, more than 11,900 customers had paid over $100,000 in the previous 12 months. A year earlier, there were more than 950 such customers. By the end of June, the number had already exceeded 6,000, and these customers accounted for about 74% of revenue in the first half of the year. The prospectus did not name any customers that accounted for more than 10% of total revenue, but the company's dependence on a small number of large customers is declining. In 2025, two customers each accounted for 12% of revenue, together contributing about $1.1 billion. In the first half of 2026, only one customer accounted for more than 10%, making up 11% of total revenue.

Anthropic's revenue was $386 million in 2024 and rose to $45.85 billion in 2025, an increase of 1,088% from the previous year. In the first half of 2026, revenue reached $162.86 billion, 14 times the amount in the same period last year. Looking at individual quarters, quarterly revenue grew about 30-fold over five quarters.

According to the prospectus, revenue in the third quarter of this year was 86 times the $197 million recorded in the fourth quarter of 2024. This suggests third-quarter revenue of about $16.9 billion. The quarter-on-quarter growth rate fell from 144% in the second quarter to about 46%. But the actual increase in revenue has not collapsed. In my view, the third quarter added about $5.4 billion more revenue than the second quarter, while the second quarter added about $6.8 billion more than the first quarter. The size of the starting revenue base has played a part in the slower growth rate.

The prospectus did not explain why quarter-on-quarter revenue growth slowed in the third quarter. However, I think several factors may have played a role. Customers often use less computing power during summer holidays. From August to September, Anthropic also stepped up its efforts to remove customers who broke its rules. These customers accounted for about 5% to 7% of revenue in the first nine months. In June, the U.S. Department of Commerce introduced worldwide export restrictions on Fable 5 and Mythos 5. Anthropic disabled both models for all customers until it updated its security measures and the restrictions were lifted.

This rate of revenue growth is almost without precedent in human history. People often compare today's large AI models with the railways and the internet. But when we compare how quickly major companies in these fields grew their revenue, neither of those technological revolutions matched the growth seen in the large-model industry.

The prospectus clearly says that when a new model is released, business customers tend to switch to the latest and most powerful version. After Opus 4.5 launched in November 2025, quarterly revenue growth accelerated from 48% in the previous quarter to 132%. In the third quarter of 2026, more than 80% of revenue came from models released within the previous six months.

Looking at the quality of its revenue, Anthropic's business of serving customers directly is highly profitable on its own. The prospectus does not report gross profit margin directly. However, calculations based on the additional tables in the notes to its financial statements show that revenue in the first half of the year was $162.86 billion, while the cost of providing those services was $23.80 billion. This means that for every $1 of revenue, the computing power needed to serve customers directly cost only about 15 cents, giving the business a gross margin of about 85%. This cost figure does not include fees paid to cloud providers for sales channels. Much of the company's spending goes toward developing next-generation models, and the computing power used to train models is counted as research and development (R&D) spending. R&D expenses in the first half of the year were $97.41 billion, or about 60% of revenue.

In the second quarter of 2026, Anthropic reported a GAAP operating profit of $9.74 billion, with an operating margin of 8%. After excluding share-based compensation and other items, the operating margin was 19%. The company had still reported an operating loss of $16.67 billion in the first quarter. This level of profitability is unusual for the company. Technology companies usually spend the most and suffer their biggest losses when they are growing fastest. Anthropic is doing the opposite. Its total operating loss for the first half of the year narrowed to $693 million.

A closer look at the company's three financial statements shows that the timing of payments and revenue is an important reason for its profits. Anthropic gets most of its computing power through long-term contracts lasting several years. These are usually arranged 12 to 18 months in advance. Once a supplier delivers a batch of computing power and it is ready to use, Anthropic must pay according to the contract, regardless of how much it actually uses. But the revenue it earns depends on how much computing power its customers use during the current period.

As a result, the amount Anthropic must pay for computing power in any given quarter was largely decided more than a year earlier. In the first half of this year, customer usage grew much faster than the computing power the company had booked in advance. This meant each $1 spent on computing power generated several times more revenue than before. In the first quarter of 2025, Anthropic spent $2.51 on computing power and infrastructure for every $1 of revenue. By the second quarter of 2026, that amount had fallen to just $0.48.

But this only applies to the computing power the company currently has. The prospectus warns that when the next large batch of new computing power becomes available, the costs will be recorded before the revenue it helps generate. This could sharply reduce profit margins. Since the company's current profits are still relatively small, it could return to making losses.

The net loss of $807.48 billion that attracted attention last night mainly came from an accounting adjustment to the value of financing instruments. Companies that raise money through tools such as convertible notes often record this kind of item before going public. When a company's valuation rises, the shares that investors could receive by converting these notes become more valuable. The company must then increase the recorded value of the related liability, and the increase is counted as a loss.

As for cash flow, Anthropic had $1,316.51 billion in cash and short-term investments at the end of June this year. Its financing activities brought in a net $1,107.74 billion in the first half of the year, most of it from issuing preferred shares. During the same period, cash flow from operating activities was positive at $3.929 billion. This was the first time this figure had been positive in any reporting period covered by the prospectus. It had been negative $2.637 billion for the whole of 2025. This suggests that the company's operations can now generate cash, with the amount generated in half a year reaching this level.

That may sound like a huge amount of money, but Anthropic has many more bills to pay. By the end of June, its signed commitments to buy computing power and leases that had not yet started totaled about $5.87 trillion. The company will pay these amounts gradually over nearly 20 years. The current version of the prospectus leaves blank the amount due before 2030. The total commitments are more than four times its cash and short-term investments.

Both its profits and its operating cash flow show that Anthropic's revenue can now cover the computing power it is using. However, the money it earns from its operations is nowhere near enough to cover its next roughly $5.87 trillion in computing power commitments. Where the money will come from, and how quickly it will arrive, is the most practical problem Anthropic must solve after going public.

02 Can Anthropic keep charging a premium for its most powerful model?

In January 2020, Jared Kaplan, Sam McCandlish, Tom Brown, Dario Amodei and others published a paper called "Scaling Laws for Neural Language Models" at OpenAI. This paper is one of the key theoretical foundations of today's large AI models. The same group of people later left OpenAI in 2021 because of disagreements over ideas and founded Anthropic.

At the beginning of its prospectus, Anthropic presents a bold vision. For the first time in human history, intelligence has become something people can build. Skills such as reasoning, analysis, decision-making and creativity can be produced on a large scale and accessed whenever needed. Anthropic believes the effect on the global economy will be greater than that of industrialization, electricity, semiconductors and the internet.

There are two ways to understand this statement. One is that Anthropic's leaders are painting an impressive but empty picture of the future. The other is that they truly believe what they are saying. These people helped develop the scaling laws at OpenAI. After leaving to found Anthropic, they made "advance safety and capabilities together" a core principle of the company. Now, they are betting on that principle by signing computing power contracts worth far more than the cash they hold.

The main reason behind Anthropic's growth miracle is the revenue it earns from its most advanced models. Business customers are willing to pay high prices for the latest and most powerful models. The company's huge, long-term computing power contracts are based on the same expectation. The bet is not simply that "demand for AI will grow," something few people doubt. The real bet is that the most advanced models will continue to command the highest prices.