Anthropic's Prospectus Discloses $65 Billion ARR and $80 Billion Net Loss

Anthropic may not be the largest IPO in history, but it is certainly one of the most anticipated. On the evening of October 26 Beijing time, people waiting for Anthropic to publish its S-1 prospectus crashed the website of the U.S. Securities and Exchange Commission (SEC).

The company is only five years old. In 2021, seven former OpenAI members left to start the company; five years later, Anthropic has overtaken its former parent in both valuation and revenue. In May this year, it surpassed OpenAI with a post-money valuation of $965 billion, becoming the world's most valuable AI startup. At the end of July, its ARR reached $65 billion, compared with about $40 billion for OpenAI in the same period.

The prospectus shows that in the first half of 2026, Anthropic's revenue grew about 13-fold year over year to $16.286 billion, of which API usage-based billing accounted for 80% and subscription revenue for 20%. The book net loss in the same period exceeded $80 billion, but this was mainly affected by changes in the fair value of financing instruments; on an adjusted basis the company was already profitable, with an operating margin of 10% and a gross margin of 85%.

Shortly before this prospectus was made public, Reuters reported that, because the company's quarter-over-quarter revenue growth had slowed, some investors attending Anthropic's investor day discussed lowering their IPO valuation expectation from the previous $2 trillion to around $1.5 trillion. That figure is only an expectation in investor discussions; the final offering valuation has not yet been determined. If issued at a $1.5 trillion valuation, it would be lower than SpaceX's $1.77 trillion valuation at its IPO in June this year.

Investors' concerns are not unfounded. Even Anthropic, which is running at the front, is not yet enough to prove that large models are a good business.

High computing costs, a lead that could shrink at any time, and long-term returns that have yet to be verified could all change the profit curve of this business. The answer given in the prospectus is also far more complex than an adjusted-profit report card.

01. The most profitable large-model company has an operating margin of only 10%

Several brokerage analysts who have long followed U.S.-listed tech companies told "Dingjiao One" that they are cautious about Anthropic's financial data. This prospectus showing adjusted profitability not only failed to strengthen market confidence, but instead exposed problems in the business model of frontier model companies.

The first problem is reflected in the revenue structure and growth model.

The prospectus shows that from 2023 to 2025, Anthropic's revenue grew from $41 million to $4.585 billion, a two-year compound growth rate of about 957%. Revenue in 2024 and 2025 was about 9.4 times and 11.9 times that of the previous year, respectively, equivalent to revenue roughly multiplying tenfold each year on average.

In the first half of 2026, Anthropic's revenue further rose to $16.286 billion, an increase of about 1,300% from $1.163 billion in the first half of 2025. At least judging from the data in the prospectus through the end of Q2 2026, the revenue growth slowdown investors care about has not yet occurred.

The change in ARR (annualized revenue) behind revenue better reflects the frightening speed of Anthropic's growth. At the beginning of 2025, Anthropic's ARR was only $1 billion, and by the end of the year it was about $9 billion. In April this year, Anthropic officially disclosed that the company's ARR had exceeded $30 billion, surpassing the $25 billion OpenAI had disclosed two months earlier.

Enterprise customers are the main factor driving the company's growth.

The prospectus shows that as of September 30, 2026, Anthropic's customer base included more than 11,900 customers that contributed $100,000 or more in revenue over the past 12 months (TTM); more than 1,900 customers that contributed $1 million or more; and about 250 customers that contributed $10 million or more.

By revenue type, Anthropic's revenue is mainly divided into API usage-based billing and subscriptions. Usage-based revenue rose from $288 million in 2024 to $3.8 billion in 2025, with its share rising from 75% to 83%, and further reached $13.1 billion in the first half of 2026, with its share slipping slightly to 80%. Subscription revenue accounted for 25%, 17% and 20% in the same periods. Usage-based billing occupies the main position, and revenue is closely tied to how frequently and at what scale customers continue to call the models.

The prospectus also discloses a detail that is easy to overlook. Model versions released in the six months ended September 30, 2026, contributed more than 80% of revenue in the three months ended that date.

A U.S. dollar fund investor told "Dingjiao One" that this is fairly common in the large-model field, and revenue growth relies heavily on new models. The company must constantly update versions and keep its lead in order to sustain revenue growth. Once it goes a long time without an update, or a new model falls slightly behind, revenue could take a big hit.

Another hidden risk in the prospectus is reflected in profit margins.

From 2023 to 2025, Anthropic's net losses were $1.531 billion, $8.309 billion and $41.965 billion, respectively. In the first half of 2026, the net loss widened to $80.748 billion, about 6.2 times the $12.969 billion in the same period a year earlier.

This huge loss needs to be broken down. In the first half of 2026, fair value change losses on convertible notes, derivative liabilities and SAFEs reached $72.388 billion. Rising company valuation pushes up the fair value of related liabilities, creating non-cash losses; the principle is similar to the common fair value losses on preferred shares at Hong Kong IPO companies, but the instrument structures and accounting treatment differ. The related remeasurement will stop only after the instruments are converted, repaid or settled; this is not a loss generated at the operating level.

At the operating level, break-even has already been achieved. In the first half of 2026, Anthropic's GAAP operating loss was $693 million, adjusted operating profit was $1.618 billion, the adjusted operating margin was 10%, and adjusted net profit was $2.671 billion.

This is basically consistent with figures previously reported by foreign media. In August this year, foreign media reported that Anthropic achieved positive adjusted operating profit for the first time in the second quarter. At its investor day in October, the company also disclosed third-quarter revenue of $16.8 billion, up about 45% quarter on quarter, achieving adjusted profitability and positive cash flow for the second consecutive quarter.

As can be seen, Anthropic's profit margin is not ample; profitability depends mainly on rapid revenue scale expansion, and also on compute efficiency and expense control. Scale growth in turn bets on frequent model updates and sustained leadership. Once the pace of model capability improvement slows, revenue, profit and subsequent R&D investment may produce chain reactions.

Of course, large-model companies do not have only the business model of selling APIs; embedding advertising is a high-margin business model, and OpenAI is doing it. But Anthropic promised in its prospectus to keep Claude free of advertising at all times, explaining that advertising could bring considerable revenue, but the company wants Claude's only incentive to be providing helpful answers, avoiding pressure to pursue user engagement driven by revenue targets.

This choice preserves the product positioning, and also means that Anthropic at this stage still needs to rely on model invocation and subscription revenue to prove its profitability.

02. Carrying a compute bill of more than $800 billion, locking in the next decade

AI large models at this stage are still not an easy business, and Anthropic is well aware of this. That is why this prospectus is unusual: it opens with a large amount of space devoted to AI safety and the company's unique equity setup to avoid the risk of AI losing control, attempting to explain how it maintains long-term goals beyond commercial returns.

Because Anthropic was registered as a public benefit corporation (PBC) when founded, it is legally required to consider shareholder interests, the groups affected by the company's actions, and the public benefit goal of responsibly developing and maintaining advanced AI at the same time. This arrangement writes the safety goal into the company's charter, but it cannot automatically eliminate commercial pressure, nor can it guarantee that the company will always prioritize safety.

In terms of governance structure, the Long-Term Benefit Trust (LTBT), through Class T common stock, generally has the right to elect and remove Class T directors occupying a majority of the authorized board seats; Class T has no economic rights. The Founder LLC established by the seven co-founders will receive one share of Class F common stock before the listing, corresponding to 50.1% of the voting power on general shareholder matters, and likewise has no economic rights.

The public benefit goal and founder control are both written into the post-listing structure, and public shareholders need to accept this special allocation of power. This structure answers who the company wants to guard its long-term mission, but whether the mission can be delivered depends on whether the company has enough money to keep investing; model R&D, talent retention and compute procurement all require long-term funding.

The size of the fundraising in this listing and the specific amounts for each investment have not yet been disclosed. The prospectus lists the general uses as working capital, operating expenses and capital expenditures; capital expenditures cover compute and infrastructure investment, and the funds will also be used to recruit and retain highly skilled talent. The company may also acquire or invest in related businesses, products, services and technologies, but currently has no material acquisition or investment agreements and commitments.

Image source / pexels

On talent, as of September 30, 2026, Anthropic had about 5,300 employees, with revenue per employee over the past 12 months exceeding $6.6 million, clearly higher than the average of the largest U.S. listed companies.

On compute, the prospectus discloses several long-term agreements. The company's five-year equipment lease rent related to Broadcom is about $161.2 billion, excluding interest; rental spending to obtain NVIDIA GPU compute from SpaceX AI is up to $120.5 billion; service commitments paid to Nscale are up to about $44.6 billion.

In addition to the above contracts, Anthropic has also committed to paying Amazon at least $110 billion between 2026 and 2036, and about $114.5 billion to Google between 2026 and 2033, plus an estimated $125 billion to $150 billion in additional TPU compute spending. Long-term arrangements related to Microsoft and Lambda also reach $44.9 billion and about $41.3 billion, respectively. The different agreements are inconsistent on minimum payments, maximum spending, delivery conditions and cancellation clauses; the compute-related bills disclosed in the prospectus total about $850 billion.

These contracts changed the company's cost flexibility. According to the prospectus, most compute will be purchased 12 to 18 months in advance, costs will be recognized once the compute becomes available, and near-term spending is relatively fixed. If demand forecasts are too low, the company may miss out on revenue and R&D progress; if forecasts are too high, it may leave capacity idle, and the hardware and facilities locked in today may lose their cost advantage several years from now.

Anthropic's relationships with cloud providers have also become more complicated as a result. Amazon, Google and Microsoft are simultaneously shareholders and compute suppliers, and may also be customers, distribution channels and backers of competitors. They provide Anthropic with capital and infrastructure, while Anthropic locks in future spending through long-term purchase commitments. An IPO can replenish capital, but it cannot detach this business from this input network made up of talent, chips, cloud providers and long-term contracts.

03. Pricing large models now has a benchmark

Anthropic's prospectus is not only a performance showcase for a frontier model company, but also provides the large-model industry with a new valuation anchor.

Looking back at Anthropic's valuation changes, several milestones in 2026 are particularly critical.

In February, Anthropic completed a $30 billion Series G funding round at a post-money valuation of $380 billion. The company disclosed at the time that its ARR was about $14 billion, with Claude Code contributing $2.5 billion; in April, on the private secondary market platform Forge Global, Anthropic's implied valuation briefly touched $1 trillion, while OpenAI was about $880 billion over the same period. In the same month, Anthropic disclosed that its ARR had exceeded $30 billion.

Image source / Unsplash

Then, one month later, Anthropic completed a $65 billion Series H funding round at a post-money valuation of $965 billion, the most recent publicly disclosed valuation before the company's IPO. According to Bloomberg, Anthropic disclosed to investors that by the end of July its ARR had risen to $65 billion.

Since the final IPO market capitalization cannot yet be determined, using the post-money valuation of the most recent funding round and the corresponding annualized revenue as a reference, Anthropic's valuation-to-ARR ratio is about 20 times.

By comparison, among the two model companies already listed in the Chinese market, Zhipu disclosed an ARR of about $1.8 billion in September, with year-end guidance of $3 billion; MiniMax's ARR exceeded $800 million in August, with a year-end target of $1 billion. Moonshot AI's ARR was reported to have exceeded $1 billion in August, with a year-end target of $2 billion.

Calculated based on the closing price on October 26 and year-end ARR expectations, the corresponding P/ARR for Zhipu and MiniMax is about 10 times and 7.8 times, respectively, far below Anthropic's 20 times.

However, it should also be noted that the revenue types and statistical periods of ARR vary among companies. Whether API, subscriptions, on-premise deployment, customized services and cloud revenue sharing are included differs from company to company. Large-model-related companies generally have two types of ARR calibers: pure API revenue and MaaS revenue. Cloud providers commonly use MaaS revenue; the difference is that MaaS includes API revenue as well as revenue from things such as model fine-tuning, on-premise deployment and customization.

Zhipu's early disclosed ARR was pure API revenue, while the $1.8 billion disclosed in September was on a full-business basis; Alibaba Cloud commonly uses MaaS revenue.

Another difference is the statistical period. Previously, both Zhipu and MiniMax specifically noted in communications with investors that the former calculates on a monthly basis and the latter on a weekly basis. Anthropic calculates on a monthly basis, counting API plus subscription revenue.

So where does the multiple gap mainly lie? Growth rate is not the key. Domestic companies' ARR is also growing rapidly; the gap is more reflected in scale and monetization efficiency per unit of compute. Anthropic can already make money from compute, while domestic companies are still working to make revenue cover compute.

Looking further, compute cost is the main expense for a model company. In accounting treatment, the compute cost of early-stage model R&D is usually recorded as an expense affecting profit, while inference compute is recorded as a cost affecting gross margin. Anthropic's gross margin in the first half of 2026 was as high as 85%, a unique advantage of high pricing and closed-source models.

Domestic large-model companies mainly adopt open-source models and follow a low-price route. With R&D expenses not differing much, the difference in gross margin directly widens the profit levels of the two.

But this yardstick itself is also shaking. Anthropic's expected IPO valuation has dropped from about $2 trillion to around $1.5 trillion; Zhipu's market capitalization briefly reached HK$1.33 trillion intraday on June 22 and is now about HK$240 billion; MiniMax's stock price has fallen more than 80% cumulatively from its March high. From the primary market to the secondary market, valuations of large-model companies are correcting.

Therefore, this is also a "Red Queen's race"-style competition: "You have to run as fast as you can just to stay in place." Only by continuously updating the most frontier models can this fragile business model be maintained, and what this tests behind the scenes is R&D capability, sustained cash-generation capability and capital reserves.

This also explains why a company whose core narrative centers on safety and the public interest would actively seek to go public. Anthropic's offering results will become an important pricing case for frontier model companies and will also provide a reference for OpenAI, whose IPO timing points to 2027. Entering the IPO process also means that for Anthropic, the focus is no longer limited to revenue growth brought by model leadership; it must also prove whether it can leave sufficient shareholder returns after long-term compute investment.

This article comes from the WeChat public account "Dingjiao" (ID: dingjiaoone), author: Wang Hanxing, editor: Wei Jia, published with authorization from 36Kr.