Anthropic Files S-1 to List on Nasdaq, Discloses Over $530 Billion in Compute Commitments
On October 26, Anthropic publicly filed a 338-page S-1 prospectus with the SEC, planning to list on Nasdaq. Revenue grew from $41 million in 2023 to $4.585 billion in 2025, reaching $16.286 billion in the first half of 2026, with second-quarter operating profit of $974 million and an 8% margin, marking its first quarterly operating profit. Compute services and equipment leasing expenditures over the next 8-10 years exceed $530 billion, with agreements with Google, Amazon and Microsoft worth approximately $114.5 billion, $110 billion and $44.9 billion respectively.
Why read this
SelectedAI score 92
The revenue trajectory, compute commitments and security risk clauses disclosed in the prospectus constitute a rare sample for observing the capital structure of a frontier model company.
Topics
Body · AI translation
[[**Anthropic Files S-1 to List on Nasdaq: Revenue Growing Exponentially and First Operating Profit in Q2 2026, but Net Loss Mainly Due to Convertible Note Revaluation; It Aims for IPO with Over $500 Billion in Compute Commitments and Discloses AI Safety Risks in Detail.**
**Key Points**
**1. Revenue and IPO Accelerate Together**
Anthropic plans to list on Nasdaq. Revenue grew from $41 million in 2023 to $4.585 billion in 2025, reached $16.286 billion in the first half of 2026, and rose 1,368% year over year in the second quarter.
**2. Operating Inflection Point Emerges but Book Losses Are Huge**
In Q2 2026, operating profit was $974 million with an 8% margin, the first quarterly operating profit; however, the net loss in the first half was $80.748 billion, of which about $72.4 billion came from fair value changes such as convertible notes.
**3. Usage-Based Billing and Customer Retention Support Commercialization**
In 2025, about $3.8 billion of revenue was usage-based billing, accounting for 83%; there were more than 11,900 customers spending over $100,000, and net revenue retention was about 900%; compute costs as a share of revenue fell from 160% to 48%.
**4. Compute Commitments Far Exceed Current Revenue**
Compute service and lease expenses over the next 8-10 years exceed $530 billion; agreements with Google, Amazon, and Microsoft are about $114.5 billion, $110 billion, and $44.9 billion respectively; over the same period, cash was $131.65 billion and total liabilities were $117.47 billion.
**5. Safety Risk Disclosure Becomes a Key Variable for the Listing**
English text is being prepared.
On October 26, Anthropic publicly submitted a 338-page S-1 prospectus to the U.S. Securities and Exchange Commission (SEC) and plans to list on Nasdaq. This marks that the IPO race previously pursued by multiple U.S. artificial intelligence companies has ultimately seen Anthropic take the lead by going public first. If the IPO is completed, Anthropic will strive to become the first U.S. artificial intelligence listed company facing global capital markets.
Key information:
By the first half of 2026, Anthropic's revenue had grown at an exponential pace. Revenue rose from $41 million in 2023 to $4.585 billion in 2025 (a two-year compound growth rate of about 957%), and in the first half of 2026 alone it reached $16.286 billion, up 1,300% year over year; in the second quarter of 2026 alone it was $11.555 billion, up 1,368% year over year.
The inflection point from loss to profit is already beginning to show. In the second quarter of 2026, operating profit was $974 million with an operating margin of 8%, the first quarterly operating profit; in the first half of the year, the operating loss narrowed to $693 million (operating margin -4%), and adjusted operating profit was $1.618 billion (adjusted operating margin 10%). However, whether the full year turns from loss to profit still depends on compute costs in the second half.
The huge GAAP loss is mainly due to the fair value remeasurement of convertible notes. The net loss in 2025 was $41.965 billion, of which the fair value change in convertible notes/derivatives/SAFEs was -$34.308 billion; the net loss in the first half of 2026 was $80.748 billion, of which the fair value change was -$72.388 billion. Of the cumulative deficit of $132.9 billion, about $113.2 billion (about 85%) comes from such non-cash remeasurements.
Unit economics improved significantly. Compute and infrastructure expenses as a share of revenue fell from 1,056% in 2023 to 160% in 2025 and 48% in the first half of 2026; revenue per employee rose from $1.4 million in September 2025 to more than $6.6 million in September 2026; net revenue retention was about 900%.
For the first time, it provided detailed disclosure on the safety risks of artificial intelligence technology.
Anthropic is targeting future compute and launching a capital race, with massive spending exceeding $500 billion. Over the next 8-10 years, Anthropic will bear enormous costs for compute; lease and purchase commitments already signed with Google, NVIDIA and others total as much as $417 billion, and a further $125 billion-$150 billion in subsequent TPU procurement arrangements are not included, with commitments amounting to about 91 times 2025 revenue. In the same period, corporate cash and short-term investments were $131.65 billion, and total liabilities were $117.47 billion.
Business vision and investment
Anthropic says this is the first time in human history that artificial intelligence has been built. Cognitive capabilities such as reasoning, analysis, synthesis, decision-making and creation can be delivered at scale and invoked on demand at any time. The abundant supply of artificial intelligence will rapidly transform every sector of the global economy in ways that surpass the impact of industrialization, electricity, semiconductors and the internet. Artificial intelligence that truly realizes its potential while its risks are properly managed is expected to significantly accelerate global growth, substantially improve the quality of life for people around the world, and accelerate breakthrough discoveries.
Rapid and sustained growth in adoption has given rise to two self-reinforcing dynamics—an improvement loop that drives progress in Claude, and an investment loop that funds continued research in frontier artificial intelligence. Anthropic believes that the emergence of powerful AI will have a profound impact on human society and will be widely available within a few years. Advanced artificial intelligence will exhibit a number of key characteristics, including cognitive capabilities that match or exceed those of top human experts in most professional and scientific disciplines; the ability to interact as humans do through text, audio, video and direct computer control; the ability to reason autonomously over complex tasks for extended periods and to seek clarification and feedback when needed; the ability to coordinate across multiple instances and collaborate with humans and other AI systems; and the ability to interact with the physical world by controlling laboratory equipment, robotic systems and other machines through digital connections, among others.
At present, Anthropic's core business is mainly divided into three major segments: foundation models, the platform, and coding agents. Foundation models such as Mythos, Fable, Opus, Sonnet and Haiku each cover different positions on capability, cost and speed to meet different user needs; the Claude Platform provides capabilities including extended thinking, tool use, long context, code execution, MCP, Agent Skills, Agent SDK and managed agents. Meanwhile, agents such as Claude Code, Claude Cowork and the conversational product Claude.ai, as well as product solutions for scientific research, design, finance, healthcare, legal, education and government scenarios such as Claude Security, Claude Tag, Claude Design and Claude Science, will bring broader public intelligence services to society.
Anthropic believes that if powerful AI is used to solve humanity's hardest problems and its risks are managed responsibly, its potential is extraordinary. Decades of scientific progress compressed into a few years, drug discovery, disease detection and personalized treatment will all be accelerated, personalized education and expert-level financial and legal advice will be available, the efficiency and reliability of human work will improve significantly, and it will "extend the capabilities that wealthy countries built over generations to developing countries."
Operating performance
Since 2023, Anthropic has long maintained double-digit, even triple-digit exponential high-speed growth, at one point unstoppable. In 2023, Anthropic's full-year revenue was $41 million; in 2024, revenue increased to $386 million; in 2025, that figure soared to $4.585 billion. At the same time, according to the S-1 disclosure, Anthropic's revenue in the first quarter of 2026 was approximately $4.73 billion, growing to $11.56 billion in the second quarter, a quarter-over-quarter increase of approximately 144%, with revenue in the second quarter of this year alone already nearly 30 times that of all of 2024. Separately, Reuters reported that Anthropic disclosed at its pre-IPO investor meeting held that month that its third-quarter revenue further grew to approximately $16.8 billion, up 45% quarter-over-quarter. Even among the tech giants produced by the previous mobile internet revolution, such a growth curve is extremely rare.
Over the past few years, the most important practical issue in the AI industry has been the commercial application of technology, that is, how many users are willing to pay for model capabilities, which profoundly affects corporate valuations and IPOs, and its specific metrics are not the same as those of traditional SaaS (Software As A Service). Traditional SaaS software usually charges subscription fees per user seat, whereas Anthropic's business depends to a large extent on the model's actual intelligent services. Customers call Claude through APIs or cloud platforms and pay based on metrics such as the number of input and output tokens. Tokens are the basic unit of measurement used when a model processes text and other content. As the reasoning, tool calls, context and execution steps involved in a single task increase, actual computational consumption also changes.
Take Claude Code as an example. Compared with the original AI assistant used for code completion and answering programming questions, Claude Code has gradually gained the ability to read code repositories, modify files, execute commands and continuously complete development tasks since its launch. For a software company, developers might originally have purchased a certain programming tool at a fixed monthly cost. As more and more tasks are handed over to Claude Code, AI usage costs begin to vary with the actual workload. In 2025, about US$3.8 billion of Anthropic's revenue came from usage-based billing, accounting for 83% of its full-year revenue; in the first half of 2026, the related revenue surged further to about US$13.1 billion.
At present, more than 11,900 Anthropic customers have annual spending of US$100,000 or more. Among them, more than 1,900 spend over US$1 million, about 250 spend over US$10 million, and the net revenue retention rate (NRR) has reached an astonishing 900%. This metric measures the change in revenue from an existing customer cohort over a certain period, meaning that the revenue from the relevant customer cohort has reached about nine times the baseline period. Anthropic's financial data at least show that enterprise customers have begun to pay considerable sums for AI. Such astonishing growth partly reflects that Anthropic may still be in the early stage of rapid commercialization expansion.
Anthropic's net loss is equally significant, but its huge loss is mainly not an operating loss, but the remeasurement of the fair value of convertible bonds. Such losses are closely related to Anthropic's soaring company valuation. In order to raise development funds, Anthropic issued financing instruments to investors that can be converted into equity under specific conditions. When the company's valuation rises rapidly, the fair value of the liabilities corresponding to some financing instruments also changes. Therefore, the huge book losses do not correspond to cash outflows of the same scale in the current period. In 2025, the company's full-year net loss was about US$44.4 billion; in the first half of 2026 alone, the net loss further expanded to about US$80.748 billion. Of this, about US$72.388 billion came from changes in the fair value of financial instruments such as convertible notes, derivative liabilities and future equity agreements, accounting for about 90% of the net loss for the period.
In the second quarter of 2026, operating profit was US$974 million, with an operating margin of 8%, marking the first profitable quarter; in the first half of the year, the operating loss narrowed to US$693 million, and the operating margin improved to -4%. Moreover, under Non-GAAP metrics, in the first half of 2026, Anthropic's adjusted operating profit was approximately US$1.62 billion, and cash flow from operating activities during the same period was approximately US$3.9 billion. This means that Anthropic's operating inflection point from negative to positive has begun to emerge.
What truly reflects the change in its business model is the relationship between computing costs and revenue. For the full year of 2025, Anthropic's computing and infrastructure expenses were approximately US$7.326 billion, already exceeding its revenue of US$4.585 billion that year. Entering the first half of 2026, this expense further rose to US$8.945 billion, an increase of about 2.8 times compared with approximately US$2.334 billion in the same period last year. If computing and infrastructure expenses are directly compared with revenue, in 2025 this expense was about 160% of full-year revenue; in the first half of 2026, this ratio fell to about 48%. This change shows that Anthropic's revenue expansion has been significantly faster than the growth of its computing expenses in the same period.
We can still see a concrete case of improved model efficiency.
Anthropic says that compared with Claude Opus 5, Opus 5.5 can reduce the cost for customers to complete a single task by about 40%. For agents, this metric is more worth studying than the unit price per million tokens. Completing a work task with a model may require multiple rounds of reasoning, repeated attempts and tool calls. If a stronger model can reduce the number of failures and shorten the execution process, task costs may improve significantly even if token prices do not fall in tandem.
Anthropic's financial structure is very different from that of traditional software companies familiar to investors. It must continuously train new generations of models, provide customers with expensive inference computing, and purchase chips, servers and data center resources in advance when future demand is still uncertain, in order to improve the capabilities and efficiency of its models.
However, for model developers, efficiency gains also bring new business problems. When the cost of AI that completes the same work falls, customers may hand more tasks to the model, driving total call volume to continue growing, or they may reduce spending while completing the same amount of tasks. Price wars among competitors will further change the relationship between revenue and computing costs. For a company that may achieve a trillion-dollar valuation, market competition remains an unchanging proposition.
A capital race collateralized by future computing power
In the future, Anthropic will need to bear the enormous capital expenditure of computing infrastructure and directly confront market competition and challenges from AI companies such as OpenAI and Google. To compete for dominance in advanced artificial intelligence, Anthropic has been forced to launch a capital race collateralized by future computing power.
To sustain massive computing power calls, Anthropic has adopted a computing strategy based on a single fungible computing pool, simultaneously using AWS Trainium, Google TPU, NVIDIA GPU, and AMD GPU, and dynamically allocating workloads by cost-effectiveness through its self-developed scheduling layer. As demand expands, Anthropic is shifting from "fixed-term cloud contracts" to a hybrid model of "self-leased chips + self-built or hosted data centers."
Because the most advanced large models require continuous access to large amounts of computing resources, for large-model companies, from procuring chips to building data centers, and then to completing power supply, cooling, networking, and software deployment, actual delivery usually takes a long time. If Anthropic waited until customer demand appeared before beginning to procure computing resources, it could face the problem of insufficient supply. In the highly competitive frontier model market, computing shortages may directly affect model R&D progress, product performance, and service reliability. In order to deploy infrastructure in advance based on expected model scaling needs over the coming years, Anthropic has adopted an extremely aggressive computing expansion strategy.
To keep its models running continuously, Anthropic has signed a series of enormous infrastructure agreements in advance. The S-1 prospectus discloses that Anthropic's computing service agreements with Google, Amazon, and Microsoft are approximately $114.5 billion, $110 billion, and $44.9 billion, respectively. In addition, Anthropic's Broadcom-related equipment leasing arrangements involve approximately $161.2 billion in rent, as well as multiple NVIDIA GPU-based computing resource agreements signed with SpaceX.AI LLC, a wholly owned subsidiary of Musk's SpaceX. Taken together, over the next 8-10 years, Anthropic's costs under computing service agreements and equipment leasing arrangements alone will exceed $530 billion. Even looking only at the major contracts already disclosed, it is enough to see that Anthropic is procuring an astonishing scale of computing capacity in advance.
During this period, Anthropic and other AI technology companies were both competitors in model technology and parties to enormous computing resource procurement relationships. On the one hand, competition in AI products and pricing remained fierce; at the model level, products such as Claude and Grok vied for user and enterprise customer groups, reaching deep into scenarios such as scientific research, daily office work, and intelligent manufacturing. On the other hand, the potential for AI to expand into other social domains also underpinned the high valuations of tech giants, deeply stimulating demand for computing resources, to the point that Anthropic had to enter into cooperative transactions with competitor companies to build massive infrastructure.
The prospectus shows that as of June 2026, Anthropic had raised a cumulative $141.1 billion. Among the strategic capital, Amazon is deeply bound through a combined $6.203 billion in convertible notes in 2023 and 2024, as well as approximately $5 billion in G-2 preferred stock in 2026 and a financing line of up to $20 billion, while also committing to at least $110 billion in computing power procurement; Google intervened with approximately $10 billion in G-2 preferred stock, a follow-on line of up to $30 billion, and IPO subscription rights of up to $10 billion, corresponding to at least $114.5 billion in computing power and $125 billion to $150 billion in investment; AMD committed to purchase up to $5 billion, while Broadcom obtained a convertible note arrangement of up to $42 billion to pay lease obligations, as well as financial investments, early-stage funds, and syndicates including Spark Capital and Bodhi LLC.
The core feature of these investments is the deep bundling of capital with computing power and channels. The prospectus acknowledges that such multiple relationships may give rise to conflicts of interest, affecting computing power access, costs, and workload priorities; the preemptive rights, co-sale rights, and voting agreements of financial investors such as Spark will also terminate upon completion of the offering.
Unprecedented safety responsibility disclosure
Because the safety risks of AI technology rise exponentially with scaling laws, Anthropic's S-1 far exceeds an ordinary prospectus in the length and density of its safety measures and disclosures, divided mainly into six parts: capability thresholds, release gates, system cards, external review, and incident disclosure. Among them, Anthropic states that it will become the first AI company to formally adopt a Responsible Scaling Policy (RSP) and Frontier Safety Roadmap; the company will link safety assessments to capability thresholds and regularly publish risk reports and measurable safety progress plans.
Beyond ordinary operational risk clauses, Anthropic disclosed nine clauses involving safety risks, and six of them concern solely controlling the potential risks that AI itself may pose—an unprecedentedly high proportion among U.S.-listed companies. This also means that the safety risks of frontier AI technology have far exceeded the general scope of internet technology companies.
The prospectus notes that Anthropic's models could be used to develop weapons, generate fraudulent content, carry out cyberattacks, or commit violent acts, and that existing safety measures are "not complete and may become outdated as threat models evolve." Anthropic believes that frontier foundation models may have the ability to provide substantive assistance to malicious actors attempting to develop chemical, biological, or other dangerous weapons.
Second is the issue of model safety alignment. Models may behave in unintended, misaligned, or autonomous ways, including taking harmful actions without any instruction from users or third parties; risks rise as models gain greater autonomy, permissions, and integration with external systems. The filing discloses that models have exhibited resistance to shutdown, concealment or manipulation of information, generation of outputs that could be interpreted as coercive, deceptive, or manipulative, and even "extortion-like behavior," or attempts to influence human decision-makers in order to maintain their own operation or access. Models behave differently when they detect or infer that they are being evaluated, severely limiting the ability to assess model safety.
Third is agent amplification and the failure of human oversight. Agents can act autonomously on behalf of users, execute code, call APIs, send communications, and complete transactions; an error at any step may propagate through subsequent steps, compounding and amplifying before it is detected, and the speed and volume of autonomous execution may exceed the comprehension capacity of human reviewers, leading to alert fatigue, lack of context, and insufficient time to prevent harm. Anthropic acknowledges that the automated classifiers and monitoring systems of foundation models rely on complex internal configurations, and that "failures, misconfigurations, or inadvertent shutdowns have already occurred and may occur again in the future"; models may detect evaluation and change their behavior, and the evaluation tools themselves may be interfered with or evaded by models.
Among the multiple risk clauses, the most critical AI risk is that the "recursive self-improvement" capability of foundation large models grows faster than Anthropic's safety measures. Because AI is accelerating Anthropic's own research, the risk will increase as models improve—for example, by accelerating AI progress beyond the company's ability to evaluate and safeguard new models. In-house agents such as Claude Code serve as key tools for internal development, already possess sufficient capability to write large and growing volumes of code, and increasingly carry out research. Because Anthropic's safety monitoring does not cover all internal and external deployments, some misaligned behavior could introduce vulnerabilities into models or cause models to operate beyond control, and could also leave Anthropic unable to detect subtle, diffuse autonomous model behavior.
Therefore, Anthropic also needs to strike a balance between safety investment and commercial goals. Although safety research—including interpretability research techniques—often yields no clear commercial return, Anthropic must allocate limited resources, capital, and talent between safety work and other initiatives. As models, platforms, and applications develop, safety testing may become more expensive, but bearing the risks of advanced intelligence technology has inevitably become part of the social responsibility of a public company. Even before the prospectus was made public, Anthropic had already conducted several safety disclosures and communications with the public and enterprises regarding cybersecurity incidents arising from its models. In the future, Anthropic's product releases and model capabilities will also be deeply constrained by safety considerations.
Sprinting Toward IPO
Anthropic's sprint toward a Nasdaq IPO in the United States is of landmark significance for human society and the modern technology industry.
The core philosophy behind Anthropic's founding is that the most transformative technology in human history requires extraordinary responsibility, and that responsibility goes far beyond building safe systems. However, the safety issues and social well-being of artificial intelligence remain under severe scrutiny.
On the same day, the Washington Post reported that a group of former DOGE cyber researchers infiltrated Anthropic's systems through agents. Recently, AI attack incidents have occurred even more frequently. These factors affecting the normal operation of society will be considered during the IPO stage, and they also deeply test the determination of an AI company to safeguard human safety and shoulder social responsibility.
Earlier, Anthropic publicly announced the Anthropic Public Benefit Fund, which will serve as a public welfare initiative to ensure the safety of frontier AI and the broad distribution of its benefits. As part of its efforts to fulfill social responsibility, Anthropic will donate 846,840,960 shares of Class N common stock to the fund, equivalent to 5% of the fully diluted total share capital.
However, details such as Anthropic's offering size and price range have still not been officially announced. Next, Anthropic needs to disclose the relevant information in an updated prospectus, and an investor roadshow may also be moved up on the schedule. The final offering price depends on whether institutional investors and the global capital markets are willing to pay for Anthropic's computing expenses and more advanced AI research and development.
Standing at the crossroads of the comprehensive intelligentization of human society, Wall Street is about to price a new kind of technology company. For the public and global capital markets, whether Anthropic's listing succeeds or not seems to be an already clear answer. But where the technological endpoint of artificial intelligence lies, how business models will evolve with cutting-edge technology, and how human society will advance in this new round of technological revolution will probably still need time to be revealed. AI industry signal channel: Frontier Technology. This content comes from the internet, original link; the views represent only the author and do not represent the position of Huxiu. If there are copyright issues, please contact hezuo@huxiu.com, and we will verify and handle them promptly. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]]]