Anthropic Falls Below Its Opening Price on IPO Debut, Testing Its Trillion-Dollar Valuation Amid Questions Over Compute Returns and Its Safety Narrative
Anthropic debuted on Nasdaq at an issue price of $83 and opened at $111.50. The stock reached an intraday high of $115.10 before falling as low as $95.05. Citing growth in enterprise customers and approximately $900.9 billion in compute purchasing and leasing commitments, the article examines whether sustained revenue and profits can support its lofty valuation. Claude experienced degraded performance on the day of the listing, while the Claude API and Claude Code operated normally. As of publication, the company had not explained the cause.
Why read this
SelectedAI score 87
Drawing on revenue, customer, and compute commitment data from the prospectus, the article highlights the tension between Anthropic's rapid growth and its long-term capital investment.
Topics
Body · AI translation
Anthropic officially began trading on Nasdaq on November 19, 2026, local time. Its offering price was $83, while its actual opening price was $111.50, about 34.3% above the offering price.
“Trading hadn’t even begun, and people in the gray market had already made 20%,” one netizen remarked. After the market opened, however, the share price initially surged before retreating. It reached an intraday high of $115.10 and at one point fell below its opening price to a low of $95.05.
For a period following the listing, the Nasdaq Composite, Dow Jones Industrial Average, and S&P 500 all declined to varying degrees, while shares of well-known technology companies also fluctuated repeatedly.
Coincidentally, Claude experienced degraded performance on the day of the listing. The status page listed Claude as having “degraded performance,” while the Claude API and Claude Code were operating normally. As of publication, the company had not explained the specific cause.
The IPO came at a rather contradictory moment. AI-related stocks had undergone a correction, and investors were beginning to reassess how much return could actually be generated by data center construction, computing-power procurement, and model training. Yet Anthropic attracted intense attention from the capital markets while simultaneously facing severe scrutiny from all sides.
“Going public will only make us more determined to do the right thing. We will never bow to commercial interests or political pressure!” Amodei’s remarks before ringing the bell were an indirect response to skeptics and a declaration of his resolve to the outside world:
“We will control the pace at which AI systems are trained and deployed, taking more time if necessary until we are confident that they are safe and will remain under human control. We will not yield to commercial pressure, nor will we yield to political pressure. Today puts the company in a stronger financial position, and that will only strengthen our determination to do the right thing. If we can prove that it is possible to build an extraordinarily successful company while making the right choices on safety, then others in the industry will be encouraged to follow our example.”
Some netizens, however, are not optimistic about Anthropic’s future trajectory, and the changes in its share price may also reflect that.
Anthropic’s IPO exceeded SpaceX’s previous base offering size of $75 billion, making it the largest IPO in global capital-market history. According to reports, Anthropic’s record-breaking IPO attracted subscription orders from about 1,200 institutional investors, but share allocations were clearly tilted toward U.S.-based investors, and more than 50% of institutional investors received no allocation.
Anthropic issued 980 million shares in the offering and was expected to raise approximately $81.34 billion. Of those 980 million shares, Anthropic itself sold 956.35 million, generating approximately $79.38 billion; another 23.649 million shares were sold by certain existing shareholders for approximately $1.96 billion, none of which will go to the company. Anthropic also plans to raise $26 billion through a private placement. Combining the public offering and private placement, the total size of the capital transaction is approximately $107.3 billion.
What makes Anthropic distinctive is that it is turning expectations that AI will replace human labor into a rapidly growing business. From Claude Code to enterprise agents, AI is beginning to take over more and more work that previously required professionals. As a result, the addressable market for model companies is expanding beyond software procurement budgets into enterprises’ vast labor-cost budgets.
However, commercial constraints on computing power and an excessive emphasis on the “safety” narrative, among other issues, cast a shadow over the strong share-price performance. Compared with its first-day share price, Anthropic’s performance over the next several quarters may be more important.
1
How Does Anxiety About Unemployment Become a Business?
Dario Amodei’s warnings about AI-driven unemployment have become unavoidable in technology-industry discussions about jobs. Behind these forecasts, however, lies another judgment directly related to Anthropic’s business model: if a growing share of work that once required professionals can be handed over to models, AI companies will no longer be addressing only the software market, but also enterprises’ enormous labor expenditures.
In May 2025, Amodei made a widely discussed prediction in an interview with Axios: over the next one to five years, AI could replace about 50% of entry-level white-collar jobs in the United States and potentially push the unemployment rate to 10%–20%. He specifically cited industries including technology, finance, law, and consulting, arguing that entry-level positions in these fields could be the first to feel the impact.
In June 2026, Amodei further discussed AI’s effects on employment and income distribution in an article titled “Policy on the AI Exponential.” He worried that while AI could drive rapid economic growth, it might also cause persistent labor displacement and income inequality. To address this, he proposed policy measures including wage insurance, employee-retention incentives, and vocational training. He even discussed funding universal income mechanisms by taxing AI companies or raising capital gains taxes.
This anxiety is changing the way people use AI. Employees are proactively learning new tools and trying to integrate models into their daily work, hoping to remain competitive by improving efficiency. Subscribing to models, using agents, and exploring automated workflows are gradually becoming ways to actively adapt and even protect themselves. People fear being replaced by AI, yet they also dare not fall behind it.
This enthusiasm for AI use at the individual level is becoming an important gateway for Anthropic to enter the enterprise market.
Anthropic disclosed in its prospectus that Claude’s adoption within enterprises is driven largely by organic, bottom-up growth: software developers and individual users first use Claude in their own work, then become internal advocates within their organizations, bringing the tool into their teams and departments and promoting broader deployment. Anthropic even regards these individual users as an important customer-acquisition channel for its enterprise business.
This route is particularly noteworthy. Enterprise AI procurement does not necessarily begin with a top-down technology-selection process; it may also start with an employee’s daily work. Once individual users have demonstrated Claude’s value and encouraged their teams to expand its use, a tool originally aimed at individuals has an opportunity to enter an enterprise’s core workflows. Anthropic gains not only individual subscription revenue, but also a growth channel leading to enterprise customers.
Employees use it first, teams then expand its adoption, and enterprises ultimately pay for larger-scale deployment. This bottom-up form of penetration gives Anthropic an opportunity to bypass the lengthy sales and procurement processes of traditional enterprise software and find customers directly through real workplace needs.
The reasons enterprises procure AI are also changing. In the past, software primarily helped employees improve efficiency; now, products such as coding agents and office agents are beginning to independently perform some tasks. Enterprises see the potential to reduce repetitive labor costs, slow new hiring, lower outsourcing expenses, and accomplish more business with fewer people. At the same time, pressure from competitors that have moved first in completing their AI transformations is pushing enterprises to accelerate deployment.
Anthropic sits squarely at the center of this surge in demand. The company’s revenue increased from $41 million in 2023 and $386 million in 2024 to $4.585 billion in 2025. In the first nine months of 2026, it had already reached $33.242 billion, up approximately 1,206% year over year. As of September 30, 2026, more than 11,900 customers had each contributed at least $100,000 in revenue over the preceding 12 months, more than 12 times the number a year earlier.
From individual users to enterprise customers, and from developer tools to deployment in business processes, Anthropic is turning expectations about AI’s impact on the labor market into commercial growth. Employees hope to use AI to remain competitive, enterprises hope to use it to cut costs and increase output, and model companies are beginning to derive ever-growing revenue from this transformation.
The problem is that the more work enterprises are willing to entrust to AI, the more Anthropic must prepare in advance the computing capacity needed to handle it.
2
The continued increase in the workload handled by AI means that demand for model calls and inference will rise in tandem. But computing capacity is not an ordinary resource that can simply be purchased as needed. To keep its models at the forefront and accommodate future demand, Anthropic must secure vast amounts of computing resources in advance; whether it can recoup these investments depends on whether enterprises truly entrust more work to AI.
Amodei is highly optimistic about the future market. He believes today’s AI market is only the beginning. As model capabilities improve, AI applications could expand beyond software development into fields such as scientific research, biotechnology, cybersecurity, and robotics. These new use cases mean AI may have the opportunity to undertake more specialized work, giving Anthropic a glimpse of a market larger than today’s programming and enterprise knowledge-work markets.
But infrastructure investment is already moving ahead even though the market has not yet fully opened up. According to the prospectus, as of September 30, 2026, Anthropic had signed computing-capacity purchase commitments and lease agreements totaling approximately $900.9 billion. Of this, about $451.6 billion is expected to be incurred by the end of 2030, while most of the remainder is expected to be incurred between 2031 and 2045.
This is a resource commitment spanning decades, and it will closely tie Anthropic’s operating performance for years to come to computing capacity supply, utilization efficiency, and customer demand.
Amodei has called computing capacity the most fundamental business dilemma facing Anthropic. Model companies typically need to secure computing resources one to two years in advance, but AI capabilities and market demand change so rapidly that it is difficult to accurately determine how much capacity will ultimately be needed or how much economic value the resources secured in advance will eventually create.
Anthropic’s approach is to arrange long-term computing-capacity purchases based on relatively conservative demand forecasts, then use short-term resources to cover demand that exceeds expectations, seeking to strike a balance between insufficient capacity and overinvestment. As of the end of September 2026, the company held approximately $9.56 billion in cash and cash equivalents and $3.63 billion in short-term investments, as well as credit facilities, providing a substantial financial buffer for continued investment.
A set of noteworthy data has emerged on whether investment in computing capacity can generate sufficient returns.
According to the prospectus, Anthropic’s revenue per dollar of computing expenditure rose from $0.40 in the first quarter of 2025 to $2.08 in the second quarter of 2026, before falling back to $1.72 in the third quarter. The company explained that one important reason was the addition of approximately $5 billion in short-term computing-capacity purchases during the quarter, with these resources priced significantly higher than those acquired under multiyear agreements.
This means that continued growth in model usage and revenue does not necessarily imply that computing costs per unit will keep declining. In particular, when the company must use short-term purchases to meet surging demand, higher resource prices may directly affect operating efficiency.
Profit figures reveal a similar tension. In the first nine months of 2026, Anthropic generated approximately $4.081 billion in adjusted operating profit, with an adjusted operating margin of about 12%, yet it still recorded an operating loss of approximately $1.26 billion. The difference between adjusted profit and the operating loss under accounting standards indicates that the company remains in a phase of rapid expansion accompanied by large-scale investment.
Anthropic has already demonstrated that enterprises are willing to pay enormous sums for model services and has shown early signs of operating efficiency, but it still needs to answer a more difficult question: how can its continually growing AI workloads consistently generate commercial returns sufficient to cover its computing investment?
If growth in demand for enterprise automation slows, long-term computing commitments could become a heavy fixed cost; if demand continues growing rapidly, expensive short-term computing capacity could erode margins. The capital raised through the IPO can strengthen the company’s ability to absorb these pressures, but it cannot eliminate the time lag between computing investment and commercial returns.
Foreign media estimates that Anthropic would need to generate approximately $60.2 billion in annual revenue to achieve a 20% return on its roughly $1 trillion in computing commitments.
Specifically, based on $1 trillion in capital investment, a 40% operating margin and a 17% tax rate, Anthropic would need to generate approximately $60.2 billion in annual revenue to achieve a 20% return on investment. That would exceed Alphabet’s current annual revenue of approximately $47.1 billion and Apple’s approximately $47.8 billion, and would equal about two-thirds of the combined $90.8 billion in revenue generated by all 76 software and IT services companies in the S&P 1500 Index.
Amodei is betting that as model capabilities improve, the computing capacity being built today will ultimately support an AI industry far larger than the current market. But whether fields such as scientific research, biotechnology, cybersecurity and robotics can generate sustained customer demand will take time to establish.
3
As Models Get Cheaper, How Long Can Anthropic Sustain Its Premium?
Even if AI enters an increasing number of enterprise workflows, Anthropic may not be able to capture this additional market by itself.
As model capabilities become increasingly widespread, enterprises have more and more suppliers to choose from, making it increasingly difficult to command a substantial premium solely through superior model performance.
Competition among frontier models is creating a new tension: training more powerful models requires greater investment in computing capacity and R&D, but once competitors catch up with a model’s capabilities, enterprise customers may not be willing to pay higher prices for performance improvements.
Enterprises do not need to buy the most powerful model for every task. For basic information organization, text processing and code generation, lower-cost models can already meet many needs. But in complex programming, financial analysis, professional research and high-risk operations, accuracy, reliability and the ability to complete tasks still determine whether enterprises are willing to pay more.
This means Anthropic must defend markets that demand greater model capability and task reliability and are more willing to pay a premium.
As open-weight models continue to improve, providers such as DeepSeek are also reshaping the balance between model capability and usage costs, challenging both the technical and pricing advantages of frontier closed models. Anthropic also acknowledges that competitors may achieve comparable model capabilities at lower cost, while enterprise customers may procure services from multiple model providers simultaneously and allocate workloads according to task complexity, price and performance. Once the capability gap between models narrows, it will become increasingly difficult to maintain a price premium solely through technological leadership.
Claude Code’s commercial performance has already pointed to another direction for competition. What enterprises are buying from AI is expanding beyond model capability alone to encompass a complete task-execution experience.
A model’s ability to generate code is not the same as its ability to understand an entire codebase, invoke development tools, plan tasks, identify errors and complete repairs. Systems-engineering capabilities such as tool use, context management, codebase comprehension, task planning and error correction also determine whether AI can truly take on work. Products such as OpenAI Codex are also developing similar capabilities, and competition among model companies is gradually extending into product experience, enterprise workflows and tool ecosystems.
For enterprises, model capability is only the starting point. Whether an entire job can be entrusted to AI from beginning to end is what determines whether it is worth shifting the corresponding budget from labor or outsourcing expenses to AI services. An Agent’s ability to complete tasks reliably, integrate into existing workflows and connect to enterprise tools and data systems, as well as whether the cost of human intervention can continue to decline, will all influence corporate purchasing decisions.
Anthropic is expanding its market along this path. It began with chat and API services, then expanded into programming, using Claude Code to turn model capabilities into actual software development work. In the future, it also hopes to extend this capability to scientific research, life sciences, cybersecurity and other professional fields. Each new work scenario the company enters gives it an opportunity to capture additional enterprise budgets.
However, a larger addressable market does not mean that Anthropic will automatically generate corresponding revenue. In the first nine months of 2026, the company’s five largest customers contributed about 25% of its revenue, but some major customers have made no long-term purchasing commitments and can adjust or even stop spending according to their business needs. Enterprises can also allocate their budgets among multiple model providers or switch to lower-cost options after pilot programs end.
This presents Anthropic with a twofold challenge: on the one hand, it must continually improve model capabilities and the product experience so that AI can take on more high-value work; on the other, it must ensure that enterprise customers continue to expand their usage rather than switching to other providers as model prices fall.
Ultimately, a genuine competitive moat must be embedded in workflows. The fact that an enterprise once purchased Claude does not mean it will continue using it indefinitely. Only when Claude can reliably handle work that enterprises are reluctant to hand back to human employees, and when switching costs are sufficiently high, will Anthropic have a chance to turn a one-time purchase into long-term revenue.
4
Chinese Model Companies Will Face a More Stringent Test
Anthropic’s listing also provides an important public-market benchmark for other frontier model companies. When assessing the commercial value of companies such as OpenAI in the future, investors may look beyond model capabilities and user scale to scrutinize enterprise customers’ actual usage, revenue quality, computing costs and whether AI can genuinely take on enterprise work. This influence will also extend to Chinese model providers.
In a report released in September 2026, Rhodium Group compared the valuations and annual recurring revenue (ARR) of Chinese and US model companies. The report estimated that Anthropic’s valuation/ARR multiple was approximately 21 times, compared with about 34 times for OpenAI, 46 times for China’s Zhipu, 50 times for Moonshot AI and 163 times for DeepSeek.
These figures show that the valuations of some Chinese model companies already incorporate high expectations for future growth. Although the companies differ in business structure, revenue scale, stage of commercialization and financing environment, Anthropic’s listing will give the market more opportunities to observe how a frontier model company supports its valuation through enterprise customers, product services and recurring revenue.
If Anthropic maintains a strong performance in the public market, it may further reinforce investor confidence in the commercial value of frontier model companies. For companies that have already gone public, such as Zhipu and MiniMax, as well as other model companies such as DeepSeek and Moonshot AI, Anthropic’s financial disclosures also provide new dimensions for comparison: how quickly enterprise customer numbers are growing, whether customers continue to expand their usage, whether revenue can cover computing investment, and whether model services can generate stable profit margins.
Chinese model providers do not necessarily need to catch up with the leading closed models on every task. Open weights, low-cost inference and local deployment can still offer advantages in a wide range of enterprise scenarios. These approaches may be particularly attractive in practice to enterprises that are budget-conscious, have stringent data-security requirements or want independent control over model deployment.
But low prices and openness also create another commercial problem: widespread use of a model does not mean that its original developer can earn revenue commensurate with the scale of that usage. If enterprises can obtain models at low cost and then build their own Agents and business systems, how can model providers generate recurring revenue from them?
This requires Chinese model companies to think further about how to turn model capabilities into productivity for which enterprises are willing to pay over the long term through API services, enterprise deployment, industry solutions, Agent products and ecosystem partnerships.
Anthropic’s commercial path offers a case worth watching: starting with adoption by individual developers, entering specific workflows through products such as Claude Code, and then expanding commercial revenue through enterprise APIs and Agent services. Its growth shows that model companies have an opportunity to move beyond enterprise software procurement budgets and into labor-cost budgets.
Chinese model companies need to find their own entry points, translate technical capabilities into tangible business outcomes for enterprises, and establish sustainable business models spanning open ecosystems, low-cost inference and enterprise services.
Ultimately, Anthropic’s IPO is testing more than whether Claude can maintain its lead in model capabilities, and more than whether the company can achieve higher revenue and profits.
Over the past few years, one of the AI industry’s grandest commercial visions has been that machines will gradually take over some human jobs, prompting companies to reallocate costs previously used to pay for human labor. That story is now beginning to be priced into capital markets.
Wall Street is putting a price on the future of human labor. What Anthropic must prove next is whether it can turn this vision of the future into productivity that businesses are willing to pay for every day.
5. The “safety” narrative may backfire
However, the tests facing Anthropic may extend beyond its business model and technological competition. For a company that has made AI safety its core mission, another, harder-to-avoid question is whether it can uphold its original safety commitments while continually pursuing more powerful AI.
Amodei’s strategic logic contains a notable contradiction: he believes superintelligence could pose catastrophic risks to humanity, yet he is also convinced that the best way to avert that outcome is for Anthropic to develop superintelligence first. This approach has been described as a “race to the top.” Its basic premise is that the AI race is inevitable and that the eventual winner will gain enormous technological and social influence, while Anthropic is better equipped than its competitors to wield that power responsibly.
But not all employees agree with this logic. Some want the company to voluntarily slow its development of frontier models or even limit the computing resources devoted to training them. Amodei, however, argues that even if Anthropic stopped its research and development, its competitors would not. Rather than withdraw from the race, he believes the company should seek the lead and ensure that superintelligence ultimately rests in the hands of a company more capable of developing it responsibly.
Internal disagreements have also emerged over AI regulation and policy lobbying. Some employees have expressed dissatisfaction with Anthropic’s involvement in promoting draft policies that would restrict state-level AI regulation, arguing that such actions are in tension with the company’s long-emphasized safety mission. When an AI company highlights the risks of the technology while also seeking to influence regulatory rules, the unavoidable question is whether it is promoting more effective safety governance or securing a more favorable competitive environment for itself.
In the run-up to the listing, these doubts spread beyond the company. Hugging Face co-founder Thomas Wolf, xAI co-founder Igor Babuschkin, Harvard University professor Steven Pinker, Turing Award winner Judea Pearl and others signed the MisAnthropic Statement.
The statement said that Anthropic warns that recursive self-improvement (RSI) could lead to a concentration of power and even cause AI to escape human control, while simultaneously racing to develop the technology in an attempt to gain a first-mover advantage. It also accused Anthropic of using its safety image to attract investors, employees and the public while weakening its previous safety commitments and promoting regulatory policies aligned with its own interests. The statement ultimately reached a pointed conclusion: “Anthropic is not a safe AI company.”
This debate has moved beyond AI safety principles themselves to encompass corporate governance, commercial interests and public trust. Amodei firmly believes that he can develop superintelligence responsibly, but management’s own judgment is no substitute for external oversight. Particularly when technological leadership brings greater commercial rewards and social influence, whether safety principles can genuinely constrain the company’s research, development and business decisions will face more rigorous scrutiny.
Anthropic’s IPO therefore faces two tests. Capital markets must determine whether it can turn expectations that AI will replace human labor into sustained revenue and profits; the public must determine whether a company seeking to be the first to develop superintelligence that could transform human society is truly capable of keeping that power under control.
This article is from the WeChat public account “InfoQ” (ID: infoqchina), written by Chu Xingjuan and published by 36Kr with authorization.
AI Trends