The largest IPO in AI history? Anthropic's valuation may exceed 2 trillion US dollars.

Computing power contractual obligations reach 518 billion USD

Revenue has surged by around 12 times year on year, yet its operating loss has exceeded 8 billion USD.

Anthropic, which is preparing for a public listing, has presented its rapid growth, massive computing power input and AI security risks to the public market all together.

01 Revenue Growth Cannot Cover Losses

On September 28, foreign media disclosed a copy of Anthropic's IPO prospectus.

It is worth noting that this is not the version voluntarily disclosed by Anthropic. In June 2026, Anthropic, as a public benefit corporation, secretly submitted a draft S-1 registration statement to the U.S. Securities and Exchange Commission to prepare for its initial public offering of common stock.

According to the rules, Anthropic must publicly submit the revised S-1 at least 15 days before the roadshow to institutional investors, and the complete financial data will be publicly disclosed by then. The version seen by foreign media this time is an unpublished one, on which Anthropic has declined to comment.

Even so, this document is sufficient to outline the current business profile of this AI company.

In 2025, Anthropic recorded a revenue of nearly 4.6 billion USD, 12 times higher than that of the previous year; its net loss for the whole year was close to 42 billion USD.

A company that has been established for only five years has turned large language models into a multi-billion-dollar business, but has not yet proved that this business can make profits.

Breaking down the 42 billion USD net loss, about 34 billion USD of it is non-cash accounting expense generated from the revaluation of convertible instruments related to previous financings, rather than the actual operating cost paid by Anthropic in that year.

Therefore, it is inaccurate to refer to the entire net loss as "cash burn". But even excluding this factor, its operating loss still reached 8.06 billion USD, a significant increase from the 2.98 billion USD loss of the previous year.

The cost mainly flows to computing power. In 2025, Anthropic spent 7.33 billion USD on computing and infrastructure, about three times that of the previous year, accounting for more than half of the 12.65 billion USD total operating expenses of the year.

Taking the annual revenue as a simple comparison, for every 1 USD of revenue earned, the corresponding computing and infrastructure expenditure exceeds 1.5 USD. This is not the cost accounting for a single model call, but it is enough to show that even if revenue grows rapidly, the resource consumption supporting the growth is also very fast.

What attracts more attention is the future bills. Anthropic plans that the total contractual obligations for cloud services, computing power and infrastructure in the coming years will amount to about 5.18 trillion USD.

This is not an amount that needs to be paid immediately, nor can it be directly equated with the capital expenditure in a certain year in the future; but signing long-term contracts means that Anthropic has undertaken huge payment responsibilities for continuous access to computing power.

By the end of 2025, Anthropic held 20.28 billion USD in cash, cash equivalents and short-term investments, which provides a certain buffer for business expansion.

Faced with contractual obligations spanning several years, the market still needs to see more specific answers, including how the payments are distributed, how much expenditure can be adjusted according to demand, and how much revenue and profit the newly added computing power can eventually bring.

Anthropic has secretly submitted S-1 registration documents to the U.S. Securities and Exchange Commission in June 2026. The number of shares to be issued, the price and the listing time have not been determined, and the IPO still depends on regulatory review and market conditions.

More than 2 trillion USD is the potential valuation that has attracted attention at present, which is not the finalized issue pricing.

02 Swap Computing Power for Growth

Anthropic was founded in 2021 by a group of researchers who left OpenAI.

Around the Claude model, Anthropic has gradually built a product system for individual users, developers and enterprises: individual users subscribe to Claude, developers call the model through APIs, and enterprises integrate tools such as Claude Code into software development and daily work processes.

The enterprise market is an important source of its growth. Once the model enters high-frequency scenarios such as programming, knowledge retrieval and data analysis, the call volume may expand with the growth of business.

Anthropic also provides services to enterprises through Amazon AWS, Google Cloud and Microsoft Azure, making it easier for customers who already have cloud platforms to purchase and deploy Claude.

In May 2026, Anthropic completed a 65 billion USD Series H financing, and the proceeds from the financing will be used to expand computing power, advance safety research and product deployment, bringing the post-money valuation of the company to 965 billion USD. At that time, its annualized revenue had exceeded 47 billion USD, and it had exceeded 65 billion USD by July of the same year. The market expects the full-year annualized revenue in 2026 to exceed 1000 billion USD. Anthropic expects its revenue in 2028 to be between 1900 billion USD and 2000 billion USD.

In this public offering, the potential listing valuation discussed by the market has exceeded 2 trillion USD, which is expected to become the largest IPO in global history.

Behind the surge in valuation is investors' expectation of enterprise AI demand. However, annualized revenue is only calculated based on the revenue speed at a certain point in time, which cannot replace the audited full-year revenue, let alone directly infer the full-year profit.

For Anthropic, whether the model capabilities can be continuously converted into customer payments is more important than the growth rate at a single point in time.

The customer structure reveals another pressure. In 2025, nearly a quarter of Anthropic's revenue came from two customers; many large customers have not signed long-term lock-in contracts, which may reduce expenditure or stop procurement.

Large customers can make revenue expand rapidly, but also make the performance more vulnerable to changes in the budgets of individual customers.

Competition is compressing the time for Anthropic to improve its cost. Enterprises can test models from different vendors at the same time, and adjust procurement according to performance, price, stability and security requirements.

Anthropic must continuously launch more competitive Claude models to retain customers; the new round of model training and deployment often requires more computing power.

Amazon and Google are not only important investors of Anthropic, but also its cloud infrastructure partners. This relationship helps the company quickly obtain capital, computing resources and sales channels, but at the same time makes supply capacity, procurement costs and cooperation terms issues that investors must examine.

Whether it can turn computing power input into stable customer renewal, rather than just leading to the next round of more expensive model competition, will determine the quality of its business model.

03 The Price of Safety

As remarkable as the financial figures are the risk disclosures in the prospectus.

According to foreign media, in the main body of the about 261-page prospectus, risk factors account for about 80 pages. Anthropic frankly stated that advanced AI may bring "catastrophic or existential risks", and listed potential scenarios such as the model resisting shutdown, hiding or manipulating information, and exhibiting extortion-like behaviors.

These disclosures do not mean that extreme consequences have already occurred. What Anthropic wants to remind investors is that as the capabilities of the model improve, some behaviors may be difficult to detect in advance.

If the model can recognize that it is being evaluated and adjust its performance accordingly, the ability of existing testing methods to judge security will also be limited.

Safety is an important label that distinguishes Anthropic from its competitors, and it is also a commitment that must be continuously fulfilled. The problem is that safety research requires time, talent and computing power, while revenue growth relies on continuous improvement of product capabilities.

Anthropic admits that the company's growth and even revenue are driven by new models, and continuously releasing models is a necessary condition to maintain cutting-edge competitiveness. The company emphasizes prudent advancement on one hand, and must keep up with the industry iteration speed on the other hand. The tension between the two cannot be resolved by a single safety declaration.

This also makes Anthropic's IPO different from the listing of ordinary software companies, that is, investors not only need to judge how much the products can be sold for, but also need to judge how much the company needs to invest to keep its models leading, whether long-term contracts can be supported by revenue, and whether security incidents will change product release and customer procurement plans.

The valuation itself has pushed this uncertainty to the extreme. Roughly calculated based on the 2 trillion USD valuation and the nearly 4.6 billion USD revenue in 2025, the corresponding historical price-to-sales ratio of Anthropic exceeds 430 times; if calculated based on the annualized revenue of more than 47 billion USD disclosed in May 2026, it is about more than 40 times.

The huge gap between the two calculation methods reflects that the market is including a large amount of future growth into the price. But whether the annualized speed can be realized as full-year revenue, and whether revenue can be converted into profits, all need to be verified by subsequent financial disclosures.

Anthropic has proved that enterprises are willing to pay for Claude. After entering the public market, it still needs to prove that customers can stay, computing power costs can be controlled, and safety commitments can withstand the test of product iteration.

A 2 trillion USD valuation can bet on the future of cutting-edge AI, but cannot replace the quarterly delivered business performance report.

This article is from the WeChat Official Account "Insight IPO" (ID: dongcha-ipo), Author: Gong Xingchao, Editor: Chen Xiaoran, Published with authorization from 36Kr.