On the eve of Anthropic's IPO, how do large model businesses work out their financials?

In the past few years, large model companies such as OpenAI and Anthropic have continuously refreshed their financing scale and valuation, but the outside world has always found it difficult to clearly figure out how much revenue and profit these tens of billions of dollars of investment can bring in return.

Now, Anthropic has taken the lead in making its accounts public.

Recently, Anthropic publicly submitted its S-1 prospectus to the U.S. Securities and Exchange Commission (SEC), with plans to list on the NASDAQ.

According to the prospectus, in the first half of 2026, Anthropic's revenue reached 162.86 billion U.S. dollars, more than three times the full-year revenue of last year. In the second quarter, the company recorded an operating profit of 974 million U.S. dollars, marking its first operating profitability in the disclosed quarterly data.

More importantly, the revenue growth rate of Anthropic has far outpaced the growth rate of computing power costs. Compared with the first half of 2025, the proportion of Anthropic's computing power and infrastructure expenses in revenue dropped from 201% to 55% in the first half of this year. The large model business, which was previously considered to require long-term capital burn, has begun to show the possibility of profitability.

But this set of accounts also reveals the other side of the picture.

Anthropic's current revenue is highly dependent on enterprises' model invocation volume, while model performance and prices are still changing rapidly. To maintain competitiveness, the company needs to continuously launch new models and invest more in R&D and computing power. According to foreign media reports, Anthropic has signed multi-year agreements related to computing power and infrastructure worth more than 600 billion U.S. dollars.

At the same time, nearly half of Anthropic's revenue comes from cloud platform sales channels. Tech giants such as Amazon and Google not only provide it with computing power, but also help it sell models, while competing with Claude for enterprise customers.

From rapid revenue growth to positive quarterly operating profit, Anthropic has gone through the initial verification stage of large model commercialization. However, the larger the revenue scale, the higher the investment in computing power and R&D. How much cost do model companies actually need to pay to maintain competitiveness, and how much profit can they retain from the growing revenue?

01 With half-year revenue exceeding 160 billion U.S. dollars, how does Anthropic make money?

With a user scale far smaller than that of OpenAI, Anthropic has achieved a revenue lead over its counterpart.

In the second quarter of 2026, Anthropic's revenue reached 115.55 billion U.S. dollars, up 144% quarter on quarter. According to The Wall Street Journal, OpenAI's revenue in the same period was 6.7 billion U.S. dollars, up 18% quarter on quarter. Although the two companies adopt different revenue recognition methods for some cloud platform sales, the gap in growth rate is still obvious.

Where does Anthropic's growth exactly come from?

The prospectus shows that in the first half of 2025, Anthropic only had more than 600 customers contributing over 100,000 U.S. dollars in annual revenue, and the number exceeded 6,000 in the same period of 2026. The half-year revenue brought by these customers increased from 870 million U.S. dollars to 120 billion U.S. dollars, contributing about 74% of the company's revenue in the first half of the year.

At the same time, individual customers are spending more and more money.

In the first half of 2026, about 80% of Anthropic's revenue came from usage-based billing, while subscription revenue only accounted for 20%. The more tasks enterprises entrust to Claude for processing, the higher the Token consumption, and the higher the revenue it usually brings.

Claude Code further amplifies this feature. In the past, programmers mainly used AI to ask questions and generate codes, but now they can let Claude Code continuously perform tasks such as writing, testing and debugging, and model invocation has begun to run through the entire development process.

The ever-increasing AI usage of enterprises is also reflected in customer bills. As of the end of September, Anthropic had more than 1,900 customers contributing over 1 million U.S. dollars annually, of which about 250 contributed over 10 million U.S. dollars. The net revenue retention rate of customers in the same period was about 900%.

Anthropic is also trying to get customers to use more of its own products.

Among customers contributing over 100,000 U.S. dollars annually, the proportion of those who use both the Claude model platform and applications such as Claude Code and Cowork rose from about 30% at the end of 2025 to over 50% in the third quarter of 2026.

From invoking models, to using programming tools, to integrating AI into more businesses, every new usage scenario added by enterprises may bring new revenue.

However, although usage-based billing can drive rapid revenue growth, it also makes model manufacturers more vulnerable to changes in customer usage volume.

The prospectus discloses that in the third quarter of 2026, more than 80% of Anthropic's revenue came from new model versions released within the previous six months.

This indicates that the latest models have generated the vast majority of revenue. For Anthropic, new models need to not only attract new customers, but also make old customers willing to keep paying. Once competitors launch products with better performance and lower prices, enterprises may reallocate their invocation volume.

What's more, Anthropic's revenue from large customers is not fully locked in. In the first nine months of 2026, its top five customers contributed about 25% of its revenue, while many large customers are not bound by long-term contracts and can reduce or even stop their spending.

OpenAI is also accelerating its competition for the enterprise market. This April, OpenAI disclosed that the revenue proportion of its enterprise business has exceeded 40%, and programming Agents such as Codex are competing with Claude Code for developers and enterprise customers. The two sides started from different markets, but now they are increasingly competing for the same pool of enterprise budgets.

Enterprises' willingness to allocate more budget to AI has become an important foundation for Anthropic's rapid revenue growth. But to continuously obtain this revenue, Anthropic has to continuously launch more competitive models.

02 Achieving operating profitability for the first time, can Anthropic keep making profits?

While its revenue is growing rapidly, Anthropic is getting closer to full profitability.

The prospectus shows that in the first half of 2026, Anthropic's net loss was as high as 807.48 billion U.S. dollars, but its operating loss in the same period was only 693 million U.S. dollars. In the second quarter, the company even recorded an operating profit of 974 million U.S. dollars, with an operating margin of about 8%.

The huge gap between net loss and operating profit mainly comes from accounting treatment. In the first half of the year, Anthropic recorded a book loss of about 724 billion U.S. dollars due to the fair value change of financing instruments such as convertible bonds. This part of the loss is mainly related to the rise of the company's valuation, and does not involve cash outflow of the same scale.

Compared with net loss, operating profit can better reflect the actual operating status of Anthropic.

In the past few years, large model companies have left the public with the impression that their revenue grows fast, but their capital burn speed is even faster. Now, at least judging from the single-quarter performance, Anthropic has started to make money.

The key to profit improvement is that computing power costs have begun to be diluted by revenue.

In the first half of 2025, Anthropic's computing power and infrastructure expenses reached 2.334 billion U.S. dollars, equivalent to 201% of the revenue in the same period. In other words, for every 1 U.S. dollar earned, about 2 U.S. dollars of computing power expenses need to be paid.

A year later, this proportion dropped to 55%. In the first half of 2026, computing power expenses increased by 283% to 8.945 billion U.S. dollars, while the revenue in the same period increased by about 1300%.

Computing power expenditure is still increasing, but revenue is growing faster. This also makes the large model business begin to show economies of scale. Computing power resources usually need to be procured in advance. When enterprises' invocation volume increases rapidly and the existing computing power is more fully utilized, revenue may outpace costs.

However, it is still difficult to judge whether the profitability in the second quarter can be sustained.

First of all, Anthropic must continuously increase its investment in computing power.

As mentioned earlier, more than 80% of Anthropic's revenue in the third quarter came from new models released in the past six months. To maintain such a product update speed, the company needs to continuously train more powerful models. As enterprises entrust more work to AI, the demand for inference is also growing.

The sharp drop in the proportion of computing power expenses in revenue in the first half of this year does not mean that the same trend can be maintained in the future. After the new batch of computing power is put into use, relevant expenses may continue to rise, and the profit margin will also be affected.

Secondly, model prices are also under continuous pressure.

OpenAI, Google and open-source model manufacturers are all competing for enterprise customers. As the performance gap between models narrows and inference costs drop, enterprises have more choices and greater bargaining power.

Price cuts can stimulate invocation volume, but if the growth of invocation volume cannot catch up with the price decline, revenue will be affected. This point is particularly important for Anthropic, which is highly dependent on usage-based billing.

Greater pressure comes from the long-term computing power contracts that have already been signed.

According to foreign media reports, Anthropic has signed multi-year agreements related to computing power and infrastructure worth more than 600 billion U.S. dollars. Among them, the contracts signed with Amazon and Google involve consumption commitments of at least 110 billion U.S. dollars and 114.5 billion U.S. dollars respectively.

These expenses are distributed over many years in the future and cannot be directly included in the current period cost. But some contracts set minimum consumption requirements, and even if the computing power is not fully utilized, Anthropic may still need to pay according to the agreement.

For Anthropic, locking in computing power in advance can guarantee the needs of model R&D and commercialization, but it also increases the uncertainty of future operations. Once the revenue growth rate slows down, a large amount of underutilized computing power may drag down profits.

Therefore, Anthropic's achievement of operating profitability in the second quarter is a positive signal, but there is still a long way to go before it can prove that large models can achieve sustained profitability.

03 The business of model companies still cannot bypass tech giants

Anthropic's revenue has reached the scale of tens of billions of U.S. dollars, but nearly half of its revenue still needs to be completed through the sales channels of other tech companies.

The prospectus shows that in 2023, Anthropic's revenue from sales through cloud platforms only accounted for 11%, rose to 47% in 2025, and further reached 48% in the first half of 2026.

This revenue mainly comes from cloud platforms such as AWS, Google Cloud and Microsoft Azure.

For Anthropic, selling Claude through cloud vendors can save a lot of customer acquisition costs for enterprises. Large enterprises usually have already purchased cloud services such as AWS and Google Cloud, and can directly access Claude through their existing accounts and procurement systems.

But the relationship between Anthropic and cloud giants is far more than just sales cooperation.

Amazon and Google are not only computing power suppliers of Anthropic, but also important sales channels. Anthropic purchases computing power from them, and enterprises buy Claude through these cloud platforms, so the two companies can obtain both computing power service revenue and model sales sharing at the same time.

At the same time, Amazon and Google are also developing their own AI models to compete with Claude for enterprise customers.

Although the two sides adopt a joint sales model, where Anthropic signs contracts directly with customers and retains model pricing rights, and cloud vendors collect channel fees. But the real question is, to what extent can Anthropic operate independently of these partners?

Amazon, Google and Microsoft all have their own models and AI products, and also control a large number of enterprise customers. When the performance and prices of Claude and other models are close, which product the cloud platform chooses to focus on promoting may affect Anthropic's sales.

The same goes for computing power supply. Anthropic reminds in its prospectus that these partners may prioritize meeting the needs of their own businesses or other customers, affecting the cost and supply of computing power that Claude can obtain.

What's more, the competition in the enterprise AI market has gradually extended from the models themselves to applications and services.

Google has Workspace, and Microsoft has Microsoft 365 and Copilot. These products have been integrated into the daily office workflows of enterprises and have a mature customer base.

For enterprises, choosing AI services cannot only focus on model performance, but also consider compatibility with existing systems, data security and usage costs.

Anthropic is also looking for opportunities at the application layer. Products such as Claude Code and Cowork allow Anthropic to directly serve programmers and office staff, rather than only providing underlying models to other software companies.

If you only sell APIs, customers can switch between different models according to performance and price. But if Claude Code has been integrated into the enterprise's development workflow, involving code management, tool configuration and team collaboration, switching to other products will incur additional costs.

Anthropic is also complementing its enterprise service capabilities. In addition to cooperating with consulting firms such as Accenture and Deloitte, it has also established Ode with Anthropic in partnership with private equity institutions to help medium-sized enterprises develop and deploy Claude applications.

However, expansion to the application layer also makes Anthropic start competing for business with its own customers.

Some software companies originally purchased Claude APIs to develop AI products such as programming assistants and office tools. As Anthropic launches Claude Code and Cowork, the products of the two sides may overlap.

Anthropic also explicitly reminds in its prospectus that as the company expands its product scope, some SaaS customers may feel that their own business is threatened, reduce procurement or even switch to other model suppliers.

This puts Anthropic in a complex relationship. Anthropic needs cloud vendors to provide computing power and sales channels, and also needs software developers to continuously invoke Claude. But as its own products continue to expand, the competition between it and some partners and customers will become more and more direct.

OpenAI, Google and Microsoft are also expanding to multiple links of models, applications and enterprise services.

Anthropic has proved that independent model companies can generate huge commercial revenue. But in a market where tech giants control computing power, channels and application entrances at the same time, relying only on leading model capabilities is probably not enough to build long-term competitive advantages.

This article is from the WeChat official account "Market Value Ranking" (ID: shizhibang2021), written by the Market Value Ranking Team, and published with authorization from 36Kr.