HomeArticle

Should we give Anthropic 1.4 trillion US dollars?

36氪的朋友们2026-11-17 09:21
It's time to do the multiple-choice questions.

The artificial intelligence industry is ushering in a landmark moment.

In November 2026, Anthropic, which was founded just over five years ago, launched its initial public offering (IPO), attempting to enter the capital market with a valuation of over US$1 trillion.

According to the information provided in the prospectus, the total basic circulating share capital of Anthropic after the completion of the IPO is approximately 130.7 billion shares. Based on the offering price range of US$70 to US$78, the corresponding base market value is approximately US$9.16 trillion to US$10.2 trillion. If approximately 49.2 billion potential shares such as employee stock options are included, the corresponding fully diluted valuation is approximately US$12.6 trillion to US$14 trillion.

A company whose core product is a large language model has obtained such a huge market valuation in such a short period of time, which is in itself a rare phenomenon in the history of the technology industry.

However, Anthropic's listing not only represents the capital market's recognition of the prospects of artificial intelligence, but also means that a problem that has long been covered up by technological breakthroughs and financing booms must finally undergo public scrutiny: Artificial intelligence has created huge market demand, but can it really generate profits that match its valuation?

We hope to discuss several topics through this article:

1. With US$70 billion in annualized revenue, adjusted operating profit of US$4.1 billion, and a net loss of US$104 billion all appearing in the prospectus, what is the basis for judging the business model?

2. Every inference of the model consumes computing power, and the most active users do not necessarily contribute the most revenue.

3. Model companies organize capital in the manner of energy, telecommunications and infrastructure, so the valuation logic shifts from "how powerful the model is" to capital cost, investment return cycle and return on invested capital.

4. Companies that control user relationships, workflows and ecosystems may find it easier to turn intelligence into revenue than model companies.

5. It is now widely agreed to price intelligence based on "getting work done", but there is no consensus on who should bear the economic losses caused by task failures and errors.

6. The value created by artificial intelligence does not necessarily equal the revenue and profit of model developers.

7. After all costs including chips, electricity, R&D and financing are settled, how much is left for the people who "dig for gold" is the real thing that Anthropic's listing needs to test.

01

700 Billion vs 13 Billion: Revenue Miracle and Profit Doubts

In the previous artificial intelligence gold rush, the most certain winners have always seemed to be enterprises that sell "shovels". NVIDIA provides GPUs, TSMC undertakes advanced chip manufacturing, cloud computing companies build data centers, and power and energy enterprises provide basic conditions for the increasingly large computing facilities.

No matter which large model company wins in the end, these infrastructure suppliers can benefit from the expansion of the entire industry.

In contrast, although enterprises directly developing large models such as OpenAI and Anthropic have achieved astonishing user growth and revenue growth, they have always faced a more complex economic problem: can the cost of developing intelligence, the price of selling intelligence, and the capital investment required to maintain technological leadership form a sustainable business cycle?

Anthropic's prospectus provides an important basis for understanding this issue.

According to the listing documents updated in November 2026, the company's revenue in 2025 was approximately US$4.585 billion, and its revenue in the first nine months of 2026 has reached US$33.242 billion, with approximately US$17 billion in revenue realized in the third quarter alone. This means that its revenue in just one quarter has already exceeded the sum of the entire first half of the year. If the single-quarter revenue of the third quarter is simply annualized, Anthropic's revenue run-rate is close to US$680 billion to US$700 billion.

This growth rate far exceeds the development pace of the traditional software industry, indicating that generative artificial intelligence is no longer an experimental technology that mainly relies on capital subsidies to maintain user scale, but is forming a commercial market with actual payment capacity.

It is particularly noteworthy that Anthropic's revenue mainly comes from enterprise customers and services billed based on intelligent usage, rather than simply relying on fixed monthly fees from individual consumers. This shows that the focus of large model commercialization is shifting from selling the right to use chat tools to selling intelligent capabilities that can participate in production activities.

Large models have not only dug for gold, but the mining volume is calculated in tons. However, the explosive growth of revenue has not automatically eliminated the profitability problem.

Anthropic's computing power and infrastructure expenditures in the first nine months of 2026 reached approximately US$18.8 billion, accounting for approximately 57% of its total revenue of US$33.242 billion.

There is still a loss at the operating level, with a GAAP (Generally Accepted Accounting Principles in the United States) operating loss of approximately US$1.26 billion to US$1.3 billion in the first nine months. The net loss in the first nine months was as high as US$104 billion, but a large part of it came from the fair value revaluation of convertible financing instruments and related liabilities, not the cash loss from daily operations. At the same time, the company disclosed that the adjusted operating profit has turned positive to approximately US$4.1 billion (the adjusted operating profit margin is approximately 12%), and the efficiency of revenue relative to computing power costs has also improved significantly.

Therefore, it is obviously too simplistic to assert that Anthropic's business model is not feasible just based on the huge net loss; but it is also insufficient to conclude that large models have solved the profitability problem just because the adjusted operating profit has turned positive.

The key lies in whether enterprises can generate sufficient free cash flow (FCF) in the long term after undertaking continuous R&D investment, computing power procurement, infrastructure construction and financing costs.

More and more companies are beginning to emphasize the quantifiable productivity improvements brought by artificial intelligence, including accelerating workflows, automating routine tasks, expanding service capabilities, and improving cost control. Obviously, the demand for artificial intelligence is strong, but Wall Street is losing patience with the costs required to meet this demand.

The most vivid example is Meta. The company's third-quarter financial report shows that although artificial intelligence drove Meta's advertising revenue growth by 26%, the company's quarterly free cash flow turned negative, which is the first net cash outflow in more than a decade. At the same time, Meta also raised its capital expenditure forecast to support its aggressive artificial intelligence infrastructure expansion plan.

The entire Wall Street's anxiety about the return on investment (ROI) of AI infrastructure has been instantly amplified.

In the past two years, the market would give blind premiums as long as it heard "AI capability improvement and increased user stickiness"; at the current stage, when capital expenditures are approaching hundreds of billions of dollars or even trillions of dollars across the industry, Wall Street is beginning to calculate like actuaries: When exactly can these computing power investments be converted into real, net increased free cash flow?

If the investment return cycle is stretched too long, or the incremental revenue brought by AI cannot hedge the annual depreciation and electricity bills of hundreds of billions of dollars, the market will impose a valuation penalty.

According to the forecast of Gartner, a market research institution, if enterprises continue to maintain the current expenditure rate, by the end of 2030, more than US$3.5 trillion in artificial intelligence infrastructure investment may face the risk of asset impairment. It is expected that artificial intelligence service providers will invest nearly US$1 trillion in artificial intelligence infrastructure in 2026, and another US$6 trillion between 2027 and 2030.

Gartner pointed out that during this period, the artificial intelligence revenue brought by enterprise customers and consumers may not be enough to support these investments to achieve profitable returns.

02

Diseconomies of Scale

The huge mismatch between supply, demand and finance reveals the fundamental difference between large models and the traditional software industry.

In the past few decades, the reason why Software as a Service (SaaS) has become a business model favored by the capital market is that after the software development is completed, the marginal cost of providing services to more users is relatively low. As the scale of customers expands, enterprises can usually continuously increase profit margins through scale effects. But large models do not fully follow this logic.

Every model inference, code generation, long text analysis, and intelligent agent execution of tasks requires consumption of actual computing resources. Especially when artificial intelligence develops from answering questions to independently completing complex work, a task may need to repeatedly call the model, retrieve external information, execute programs and verify results, and the cost will increase accordingly. For enterprises that provide fixed monthly subscription, the most active users are not necessarily the users with the highest profit contribution, but may be the users who consume the most computing power.

This is why major model manufacturers have to "screen users" by limiting the context length, setting the upper limit of call times or adopting tiered billing.

In the scenario where heavy users may change from "assets" to "burdens", to judge whether a large model enterprise's business model is mature, we can't only look at the number of paying users, subscription revenue growth or user activity, but also observe the unit economic benefit of different user groups: how much revenue each user contributes, how much inference resources are consumed, how the profit margin changes after the use intensity increases, and whether the enterprise can improve this relationship through more reasonable pricing and technical optimization.

Large models are not destined to fail to form economies of scale. Model compression, inference optimization, chip upgrading, caching technology, and task allocation between different models can continuously reduce the unit computing cost. The improvement of operating efficiency disclosed by Anthropic also shows that technological progress can change the cost structure.

However, artificial intelligence also faces a special demand expansion mechanism: when intelligent services become cheaper and more powerful, users may not just complete original tasks at lower costs, but begin to require machines to undertake more work that could not be completed in the past.

The Jevons Paradox in economics provides an enlightening reference here: the improvement of resource utilization efficiency does not necessarily reduce the total resource consumption, but may stimulate demand growth. For large model enterprises, the decline in unit inference cost does not necessarily mean the decline in overall computing power expenditure. Cheaper intelligence may bring more extensive applications, and may also bring larger computing demand.

The disclosure of future computing power procurement commitments in Anthropic's prospectus highlights this contradiction. The total scale of long-term computing power and infrastructure arrangements that the company has signed and disclosed is close to US$9 trillion, a large part of which is expected to occur before 2030.

These figures do not mean that Anthropic has paid the corresponding amount, nor can they be simply regarded as current debt, but they reflect an important change: cutting-edge artificial intelligence enterprises are transforming from traditional asset-light software companies to heavy-asset technology enterprises that highly depend on chip manufacturing, grid capacity, data centers and long-term capital financing. Although large model companies are still betting on algorithm innovation and research talents, they are increasingly restricted by the supply of physical infrastructure.

If an enterprise must continuously invest hundreds of billions of dollars to maintain technological leadership, even if its revenue growth is amazing, it cannot guarantee that its return on capital is sufficient to support the current valuation.

Artificial intelligence is presenting a special industrial form: it generates revenue at the speed of the software industry, but may need to organize capital in the manner of energy, telecommunications and industrial infrastructure industries. In the past, investors only needed to pay attention to which company had the most advanced models, the most users and the fastest revenue growth; but in the future, capital cost, computing power procurement price, infrastructure utilization rate and investment recovery cycle may also determine the long-term competitiveness of enterprises.

03

Several Possibilities of Model Commoditization

Compared with Anthropic, OpenAI's business layout is more extensive, trying to become a general intelligent assistant for individual users, and also trying to become a basic platform for enterprises and developers to use artificial intelligence. But OpenAI also faces similar contradictions.

Although OpenAI's annualized revenue (ARR) and single-quarter revenue have reached the level of tens of billions of dollars, multiple public financial disclosures and industry forecasts point out that restricted by the sky-high training cost of next-generation cutting-edge models, huge inference expenditures, and hundred-billion-level computing power lease agreements, its cumulative operating losses in recent years and the expected cash flow gap in the next few years are still as high as tens of billions or even hundreds of billions of dollars.

Therefore, the key is not whether it can attract users, and the biggest challenge is how to convert extensive usage demand into revenue sufficient to cover continuous capital investment.

Meta shows another path.

Different from OpenAI and Anthropic, Meta does not mainly rely on selling large models to obtain revenue, but embeds artificial intelligence into its existing advertising, recommendation and social platform systems. The improvement of model capabilities can improve the efficiency of content distribution, ad matching and marketing automation, thereby directly enhancing the profitability of its mature business.

For Meta, artificial intelligence is first and foremost a technology that improves the efficiency of existing commercial systems, not necessarily a new product that requires independent charging. This model has obvious advantages: it can use ready-made customer relationships, user scale and advertising transaction systems to realize monetization, without having to fully rely on consumers to form new payment habits for intelligent services.

However, Meta also faces the problem of expanding data center investment, and huge capital expenditures are putting pressure on free cash flow. This shows that having a mature business model cannot eliminate the capital-intensive feature of artificial intelligence, but only provides a more stable revenue base for undertaking these costs.

Meta's case also reveals a deeper contradiction in the commercialization of large models: The economic value created by the model is not necessarily directly obtained by the model provider.

When artificial intelligence is more and more widely embedded in various products and services, enterprises that master user relationships, application scenarios and commercial channels may find it easier to realize value monetization than enterprises that only provide model capabilities. This also makes whether basic models will gradually become commoditized a key issue for understanding the profitability prospects of large model enterprises.

If the models developed by different enterprises gradually approach in performance, reliability and user experience, customers will pay more and more attention to price and service cost. Even if model providers continue to make technological progress, they may find it difficult to maintain high profits for a long time due to competition.

The diffusion of technical capabilities is beneficial to consumers, but not necessarily to the original technology developers. In history, many important technologies eventually became widely used infrastructure, but the popularization of the technology itself does not guarantee that its inventors obtain the maximum economic benefits.

However, model commoditization does not mean that all artificial intelligence services will become commoditized.

What enterprises really buy is often not a model that can answer questions, but an intelligent system that can understand internal business, call proprietary data, execute complex processes and meet security and compliance requirements. When artificial intelligence deeply enters the financial, R&D, legal and customer service processes of enterprises, changing suppliers is no longer just replacing a chat interface, but may involve data migration, system integration, risk control and organizational training.

Therefore, the more solid commercial moat in the future may not lie in the model parameters themselves, but in the combination of the model and the enterprise's workflow, as well as the resulting developer network, data environment and service ecosystem.

In this sense, Anthropic and OpenAI do not necessarily stay in the position of basic model suppliers forever. They are also developing complete services for consumers and enterprises, trying to build differentiated advantages through product experience, workflows and customer relationships.

Whether it is Anthropic lowering the threshold for enterprises to deploy agents through Claude Managed Agents, penetrating into the software development process with the help of Claude Code, or OpenAI promoting artificial intelligence to integrate into the daily operations of enterprises, both companies are exploring the same direction: from providing model capabilities and dialogue tools to providing intelligent services that can embed business processes and continuously execute tasks.

At the same time, the development of intelligent agents may further change the pricing logic of artificial intelligence. The traditional subscription model sells the right to use software for a certain period of time, billing by token sells the consumption of computing resources, while intelligent agents may actually sell the ability to complete a certain job.

If an AI system can independently complete legal document review, software testing or enterprise financial analysis, what customers really care about may not be how many tokens it consumes, but whether the task is completed, whether the result is reliable, and how much cost is saved as a result.

Charging by tasks, workload or even actual results may become an important part of the future business model. But this model will also bring new liability issues: when artificial intelligence fails to complete the task, or its error causes economic losses, who should bear the costs and risks? The closer intelligent services are to real labor, the more difficult it is for their business models to completely follow the rules of the traditional software industry.

04

Who Can Keep the Gold They Dug?

From a more macro perspective, the artificial intelligence industry is undergoing a restructuring of the value distribution structure.