Author: Hu Yong, Professor at the School of Journalism and Communication, Peking University
Editor: Su Yang, Tencent Technology
The artificial intelligence industry is迎来一个具有标志性意义的时刻.
In November 2026, Anthropic, which was established just over five years ago, launched its initial public offering (IPO), attempting to enter the capital market with a valuation exceeding one trillion dollars.
According to information provided in the prospectus, the total basic circulating shares of Anthropic after the IPO is approximately 13.07 billion shares. Calculating at an issue price range of $70 to $78, this corresponds to a basic market value of about $916 billion to $1.02 trillion. If we combine approximately 4.92 billion potential shares from employee options, it corresponds toa fully diluted valuation of approximately $1.26 trillion to $1.4 trillion.
A company with large language models as its core product achieving such a massive market valuation in such a short time is itself a rare phenomenon in the history of the technology industry.
However, Anthropic's listing not only signifies the capital market's recognition of the prospects of artificial intelligence but also means that a long-standing issue obscured by technological breakthroughs and financing booms must finally face public scrutiny:Artificial intelligence has created huge market demand, but can it truly generate profits that match its valuation?
We hope to explore several topics in this article:
1. $70 billion in annualized revenue, adjusted operating profit of $4.1 billion, and a net loss of $104 billion all appear in the prospectus; what is the basis for judging the business model?
2. Each inference of the model consumes computing power, and the most active users may not necessarily contribute the highest revenue.
3. Model companies organize capital like energy, telecommunications, and infrastructure, thus shifting valuation logic from “how strong the model is” to capital cost, investment return cycle, and capital return rate.
4. Companies that master user relationships, workflows, and ecosystems may find it easier to convert intelligence into revenue than model companies.
5. There is now a consensus to price intelligence based on “getting the job done,” but there is no consensus on who bears the economic losses caused by task failures or errors.
6. The value created by artificial intelligence does not necessarily equal the income and profit of model developers.
7. After settling costs like chips, electricity, research and development, and financing, how much can those who “mine gold” actually retain? This is the real issue that Anthropic's listing must verify.
01 $70 billion vs. $1.3 billion: Miracle of Revenue and Question of Profit
During the previous artificial intelligence gold rush, it seemed that the most certain winners were the companies selling “shovels.” NVIDIA provides GPUs, TSMC handles advanced chip manufacturing, cloud computing companies build data centers, and electricity and energy companies provide the foundational conditions for increasingly large computing facilities.
No matter which big model company ultimately prevails, these infrastructure suppliers can benefit from the expansion of the entire industry.
In contrast, companies like OpenAI and Anthropic, which directly develop large models, have seen astonishing user and revenue growth, yet consistently face a more complex economic problem: can the cost of developing intelligence, the pricing for selling intelligence, and the capital required to maintain technological leadership form a sustainable business cycle?
The prospectus from Anthropic provides important insights into this issue.
According to the updated listing documents from November 2026, the company expects approximately $4.585 billion in revenue for 2025, with revenue already reaching $33.242 billion in the first nine months of 2026, and achieving approximately $17 billion in revenue in the third quarter alone. This means itsrevenue in just one quarter has surpassed the total for the entire first half of the year. If we simply annualize the third quarter's revenue, Anthropic's run-rate revenue is approaching $68 billion to $70 billion.
This growth rate far exceeds the development pace of the traditional software industry, indicating thatgenerative artificial intelligence is no longer merely an experimental technology reliant on capital subsidies to maintain user scale but isforming a commercially viable market with actual payment capacity.
Notably, Anthropic's revenue primarily comes from enterprise clients and services charged based on intelligent usage, rather than merely relying on fixed monthly fees from individual consumers. This shows that the focus of large model commercialization is shifting from selling chat tool licenses to selling intelligent capabilities that can participate in production activities.
Large models not only struck gold, buttheir extraction amounts are measured in tons. However, the explosive growth in revenue has not automatically resolved the profitability problem.
Anthropic's computing and infrastructure expenses in the first nine months of 2026 reached approximately $18.8 billion, accounting for about 57% of its total revenue of $33.242 billion.
On the operational front, losses still arose,with GAAP (Generally Accepted Accounting Principles) operating losses in the first nine months reaching approximately $1.26 to $1.3 billion. The net loss in the first nine months was as high as $104 billion, but a significant portion came fromthe fair value re-evaluation of convertible financing instruments and related liabilities, rather than cash bleeding from daily operations. Meanwhile, the company revealed thatadjusted operating profit has turned positive at approximately $4.1 billion (with an adjusted operating profit margin of around 12%), and the efficiency of revenue in relation to computing costs has shown significant improvement.
Therefore, it is overly simplistic to assert that Anthropic's business model cannot exist based only on its enormous net losses; similarly, it is also insufficient reasoning to claim that the large model has solved profitability issues simply because the adjusted operating profit has turned positive.
The key question is whether the company can generate sufficient free cash flow (FCF) in the long term after sustaining continuous investments in R&D, computing procurement, infrastructure construction, and financing costs.
More and more companies begin emphasizing measurable productivity improvements brought about by artificial intelligence, including accelerating workflows, automating routine tasks, expanding service capacities, and enhancing cost control. Clearly, demand for artificial intelligence is strong, yet Wall Street is losing patience with the costs required to meet that demand.
The most striking example is Meta. The company's third-quarter financial report revealed that although artificial intelligence drove a 26% increase in Meta's advertising revenue, its quarterly free cash flow turned negative, marking the first cash outflow in over a decade. At the same time, Meta raised its capital expenditure expectations to support its aggressive artificial intelligence infrastructure expansion plans.
The entireWall Street's anxiety over the ROI (return on investment) of AI infrastructure investments has intensified immediately.
In the past two years, as soon as the market heard “AI capabilities improved, user stickiness increased,” it would blindly provide a premium; but at this current stage, as capital expenditures approach hundreds of billions across the industry, even nearing a trillion, Wall Street has begun calculating like actuaries:when exactly can thesecomputing investments translate into real net increases in free cash flow?
If the investment return cycle is drawn out too long, or if the revenue increases from AI cannot offset annual depreciation and electricity bills totaling hundreds of billions, the market will impose valuation penalties.
According to market research firm Gartner’s predictions, if companies maintain the current spending pace, by the end of 2030, more than $3.5 trillion in AI infrastructure investments may face asset impairment risks. It is estimated that AI service providers will invest nearly $1 trillion in AI infrastructure by 2026, and $6 trillion more between 2027 and 2030.
Gartner pointed out that during this period,the AI revenue generated by enterprise clients and consumers may be insufficient to support these investments to yield profitable returns.
02 Diseconomies of Scale
The huge mismatch of supply and demand and finances reveals the fundamental differences between large models and traditional software industry.
For the past few decades, the key reason that software as a service (SaaS) has become a favored business model in the capital market is that after software development is completed, themarginal cost of providing services to more users is relatively low. As the scale of customers expands, companies can usually progressively improve profit margins through economies of scale. However, large models do not completely follow this logic.
Each model inference, code generation, long text analysis, and intelligent agent task execution consumes actual computational resources. Especially as artificial intelligence evolves from answering questions to autonomously completing complex tasks, a single task might require repeatedly calling the model, retrieving external information, executing programs, and verifying results, leading to increased costs. For companies offeringfixed monthly subscriptions, the most active users may not necessarily be the highest profit contributors; they could actually be the ones consuming the most computing power.
This is why major model vendors have to “filter users” through context length limitations, call frequency caps, or tiered billing.
In situations where heavy users can turn from “assets” into “burdens,” determining whether a large model company’s business model is mature cannot rely solely on the number of paying users, subscription revenue growth, or user engagement; it must also observe the unit economic benefits of different user groups: how much revenue each user contributes, how much inference resources they consume, how profit margins change with increased usage intensity, and whether the enterprise can improve this relationship through more reasonable pricing and technological optimization.
Large models are not inevitably unable to achieve economies of scale. Model compression, inference optimization, chip upgrades, caching technologies, and task allocation among different models can continually lower unit computational costs. Anthropic's disclosed improvements in operational efficiency also indicate that technological advancements can alter cost structures.
However, artificial intelligence also faces a peculiar mechanism of demand expansion: as intelligent services become cheaper and more powerful, users may not only complete existing tasks at a lower cost but also begin to demand machines to undertake more tasks that were previously unfeasible.
The Jevons Paradox in economics provides an enlightening reference here: improved resource utilization efficiency does not necessarily reduce the total consumption of resources; rather it may stimulate demand growth. For large model companies, decreasing unit inference costs does not necessarily mean overall computing expenses will decrease. More affordable intelligence may lead to broader applications and potentially larger computing demands.
The disclosure in Anthropic's prospectus regarding future computing procurement commitments highlights this contradiction. The long-term computing and infrastructure arrangements the company has signed and disclosed total nearly $900 billion, a considerable part of which is expected to occur before 2030.
These figures do not mean that Anthropic has already paid the corresponding amounts, nor should they be simply viewed as current debts; instead, they reflect a significant shift: leading artificial intelligence companies are transitioning from traditional light-asset software firms to heavy-asset technology enterprises that highly depend on chip manufacturing, grid capacity, data centers, and long-term capital financing. Although large model companies continue betting on algorithmic innovation and research talent, they are increasingly constrained by the supply of physical infrastructure.
If a company must continually invest hundreds of billions to maintain technological leadership, then even with astonishing revenue growth, it cannot ensure that its capital return rate is adequate to support its existing valuation.
Artificial intelligence is presenting a unique 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. Investors have traditionally only needed to focus on which company has the most advanced models, the most users, and the fastest revenue growth; but in the future, capital costs, the prices of computing procurement, infrastructure utilization rates, and investment payback periods may equally determine a company’s long-term competitiveness.
03 Possible Paths for Model Commoditization
Compared to Anthropic, OpenAI has a broader commercial layout, attempting to serve as both a general intelligent assistant for individual users and a foundational platform for businesses and developers using artificial intelligence. However, OpenAI also faces similar contradictions.
Despite OpenAI's annual recurring revenue (ARR) and single-quarter revenue reaching the level of tens of billions of dollars, multiple public financial disclosures and industry projections point out that constrained by the astronomical training costs of next-generation cutting-edge models, heavy inference expenditures, and leasing agreements involving hundreds of billions, its accumulated operational losses in recent years and anticipated cash flow gaps for the coming years still soar into the hundreds of billions or even thousands of billions of dollars.
Thus, attracting users is not the key; the biggest challenge lies in how to convert extensive usage demands into sufficient revenue to cover ongoing capital inputs.
Meta illustrates another path.
Unlike OpenAI and Anthropic, Meta primarily does not rely on selling large models for revenue but integrates artificial intelligence into existing advertising, recommendation, and social platform systems. Enhancements in model capabilities can improve content distribution, ad matching, and marketing automation efficiencies, directly boosting the profitability of its established businesses.
For Meta, artificial intelligence primarily serves as a technology to enhance the efficiency of existing business systems, rather than necessarily a new product requiring separate charges. This model has obvious advantages: it leverages existing customer relationships, user scale, and advertising transaction systems to monetize without relying entirely on consumers to establish new payment habits for intelligent services.
However, Meta also faces the issue of continuously expanding data center investments, with massive capital expenditures putting pressure on free cash flow. This indicates thathaving a mature business model does not eliminate the capital-intensive characteristics of artificial intelligence; rather, it only provides a more stable revenue base to shoulder these costs.
Meta's case also reveals a deeper contradiction in the commercialization of large models: the economic value created by models may not necessarily be directly obtained by model providers.
As artificial intelligence is increasingly integrated into various products and services, companies that manage user relationships, application scenarios, and business channels may find it easier to realize value than those that provide only model capabilities. This makes the question of whether basic models will gradually commoditize a key issue in understanding the profitability outlook for large model companies.
If models developed by different enterprises become increasingly comparable in terms of performance, reliability, and user experience, clients will start focusing more on prices and service costs. Even if model providers continue achieving technological advancements, they may find it challenging to maintain high profits in the long run due to competition.
The diffusion of technological capability benefits consumers, but it does not necessarily benefit the original technology developers. Historically, many significant technologies have eventually become widely used infrastructures, but the popularity of technology does not guarantee that its inventors will gain the highest economic returns.
Nonetheless, the commoditization of models does not imply that all artificial intelligence services will become commoditized.
What businesses truly purchase is often not a model capable of answering questions, but an intelligent system that can understand internal processes, access proprietary data, execute complex workflows, and meet security and compliance requirements. As artificial intelligence deeply enters the financial, research and development, legal, and customer service processes of enterprises, changing suppliers will no longer merely involve switching a chat interface but may involve data migration, system integration, risk management, and organizational training.
Therefore, stronger business moats in the future may not lie in the model parameters themselves but in the integration of models with enterprise workflows, as well as the resulting developer networks, data environments, and service ecosystems.
From this perspective, neither Anthropic nor OpenAI necessarily have to remain in the position of basic model suppliers forever. They are also developing complete services aimed at consumers and businesses, attempting to establish differentiated advantages through product experiences, workflows, and customer relationships.
Whether through Claude Managed Agents lowering the barriers for enterprises deploying intelligent agents, leveraging Claude Code to deepen software development processes, or OpenAI promoting the integration of artificial intelligence into daily operations, both companies are exploring the same direction: transitioning from providing model capabilities and conversational tools to offering intelligent services that can be embedded into business processes and continually executed tasks.
Meanwhile, the development of intelligent agents may further change the pricing logic of artificial intelligence. Traditional subscription models sell usage rights for a certain period, while billing by tokens is based on computing resource consumption, whereas intelligent agents may fundamentally sell the ability to complete specific tasks.
If an AI system can independently review legal documents, perform software testing, or conduct financial analysis for enterprises, what clients care about may not necessarily be how many tokens it consumed but whether the tasks were completed, whether the results are reliable, and how much cost was saved by doing so.
Charging based on tasks, workload, or even actual outcomes may become an essential part of future business models. However, this model will also introduce new questions of responsibility: when artificial intelligence fails to complete tasks, or when errors cause economic losses, who should bear the costs and risks? The closer intelligent services come to genuine labor, the more difficult it becomes to apply the rules of traditional software industries completely.
04 Who Can Keep the Gold They Unearthed?
From a macro perspective, the artificial intelligence industry is undergoing a restructuring of its value distribution structure.
The early and most evident beneficiaries are chip, manufacturing, and infrastructure suppliers, as the expansion of the entire industry translates into demand for computing resources. However,as model capabilities diffuse, profits may gradually shift towards companies that can control enterprise workflows, user entry points, and application ecosystems.
If computing power remains scarce, infrastructure suppliers will still possess considerable bargaining power; if model supply becomes abundant, application enterprises may gain more room for value distribution; and if a few platforms can integrate models, data, payments, and workflows, they may form a new platform economy.
Ultimately, who reaps the greatest profits depends on which link in the industrial chain can create and maintain sustained scarcity.
Historically, Microsoft’s long-term competitive advantage did not just come from licensing revenues of the Windows operating system but also from the developer ecosystems, software compatibility, and corporate customer networks formed around the operating system; Google did not build its business empire by charging users for searches but transformed the user entry created by searches into advertising revenue.
Today's Anthropic and OpenAI face a crucial choice: does the base model merely require constant massive capital investments to maintain its leading technological product, or can it become the foundational platform that connects developers, enterprise clients, and application ecosystems?
If they can establish a stable developer network on top of the models, deeply integrate enterprise workflows, and create a service system that is difficult to replace, they may gain competitive advantages and pricing capabilities that transcend the performance of the models themselves. However, whether this platform transformation can produce profits sufficing to cover ongoing computing investments remains to be verified by long-term operating performance.
According to the recent regulatory filings submitted by Anthropic, this is the scale their forthcoming IPO could potentially reach under the best circumstances: a $1.4 trillion valuation (on a diluted basis) and $90 billion in fundraising.
This ambitious IPO is, in fact, an appraisal by investors of the artificial intelligence industry. It provides a chance forwider investors —they can choose to support the company continuing to burn cash or opt to avoid this company, which is expected to become a significant barometer of the artificial intelligence industry, and reject giving a historical valuation of $1.4 trillion.
Regardless, the IPO file from Anthropic reveals an undeniable issue: to maintain revenue growth, the company must endure continually rising computing costs. This clearly makes it more challenging to persuade investors. If a highly sought-after artificial intelligence star company consistently performs poorly after going public, can the market still bear it? Time will tell.
One question is easily overlooked: the economic value created by artificial intelligence does not necessarily translate into income and profits for model developers.
When companies utilize artificial intelligence to enhance productivity, the cost savings may remain with users; the conveniences consumers gain through free or low-cost intelligent services may reflect as consumer surplus rather than revenue for model providers. Even if artificial intelligence significantly boosts the productivity of society as a whole, model developers may not obtain commensurate economic returns for the value they create.
There is no inherent correspondence between the social value of technology and the financial gains of developers.
AI indeed struck gold. But after settling costs like chips, electricity, data centers, R&D investments, and financing, how much can those who unearthed the gold actually retain? This is the core question that the business model of large models has yet to answer.
Anthropic's listing does not mark the end of this question but rather brings it into a more rigorous phase. Previously, the market could envision the future based on technological breakthroughs, user growth, and financing valuations; post-listing, companies must continuously validate their commercial value through revenue, profits, cash flow, and capital returns.
Whether the models are smarter is certainly important, but more significantly, whether companies can economically sustainably produce, distribute, and sell intelligence.
免责声明:本文章仅代表作者个人观点,不代表本平台的立场和观点。本文章仅供信息分享,不构成对任何人的任何投资建议。用户与作者之间的任何争议,与本平台无关。如网页中刊载的文章或图片涉及侵权,请提供相关的权利证明和身份证明发送邮件到[email protected],本平台相关工作人员将会进行核查。