On September 28 local time, foreign media disclosed a draft IPO prospectus from Anthropic.

First, the nature of this document needs to be made clear. It is not a version that Anthropic voluntarily made public. In June of this year, Anthropic, as a public benefit corporation, confidentially submitted a draft S-1 registration statement to the U.S. Securities and Exchange Commission in preparation for an initial public offering of common stock. Under the rules, the company must publicly file an amended S-1 at least 15 days before its roadshow to institutional investors, and only then will the full financial data become visible to the outside world. What Reuters saw this time was a version that has not yet been made public, and Anthropic declined to comment.

This means the figures may still be adjusted in response to SEC review comments, and many details can only be glimpsed through Reuters' account.

But if this document broadly reflects what Anthropic will ultimately hand to the public market, its significance is hard to overstate. The valuation target for this listing is more than double the $965 billion Anthropic itself estimated in May, meaning more than $2 trillion.

This is the first time in the nearly four years since ChatGPT ignited this round of the AI wave that a frontier large-model lab has laid its complete books open to public investors.

Until now, the funding for this AI race has been borne by venture capital, sovereign wealth funds and technology giants, and the outside world could only piece together these companies' operating conditions from scattered annualized revenue and financing valuations. Now, for the first time, the books have an audited basis.

From the prospectus, we picked out seven of the most noteworthy points.

The Misread $42 Billion Loss

The first number to spread from the prospectus was the net loss of about $42 billion in 2025.

This figure is easily read as "burning $42 billion in a year," but that is not actually the case. About $34 billion of it is an accounting charge, reflecting the rising valuation of financing instruments that may convert into Anthropic stock in the future, rather than money spent on the company's operating business.

This produces a counterintuitive result. The more valuable Anthropic becomes, the more it "loses" on paper. 2025 happened to be a year in which its valuation soared, and this charge was magnified accordingly.

What truly reflects the operating situation is another set of numbers. Revenue grew from $386 million to $4.59 billion, an increase of 1,088%; operating losses widened from $2.98 billion to $8.06 billion.

Revenue rose about 12-fold, while operating losses expanded less than 3-fold, and economies of scale are indeed emerging. But an annual operating loss of $8 billion is a figure that any company preparing to go public would need to explain repeatedly.

The Pitfall of $65 Billion in Annualized Revenue

After news of the prospectus spread, the top-voted comment on the U.S. retail forum wallstreetbets was just one line, roughly meaning "a $2 trillion valuation paired with $4 billion in revenue?"

This challenge is intuitive, but it uses the wrong denominator.

The main body of the prospectus covers fiscal 2025, ending December 31 last year, while Anthropic's real explosion occurred in 2026. According to reports, its annualized revenue was about $9 billion at the end of 2025, about $47 billion in May this year, and by the end of July had exceeded $65 billion.

The $4.59 billion in the prospectus is the last snapshot of a company "before takeoff." The $2 trillion valuation given by the capital market is betting on the curve after 2026, not on the full-year results of 2025.

However, annualized revenue itself needs to be discounted. It is usually estimated by multiplying one month's revenue by 12, which naturally amplifies short-term fluctuations. The prospectus itself offers two cautionary signals. Nearly a quarter of last year's revenue came from two customers; the company also acknowledged in its risk factors that many of its largest customers do not have long-term contracts and may cut or even stop spending.

Reuters did not disclose the names of these two customers, and that will be the information most worth checking first once the public version of the S-1 is released. Anthropic's growth curve is steep enough, but it lacks contractual lock-in, which is a key premise for understanding its valuation.

The Largest Shareholder, the Largest Landlord

In 2025, Anthropic's spending on computing power and infrastructure reached $7.33 billion, accounting for 58% of total operating expenses and about 1.6 times its full-year revenue.

The question is where this money went.

Amazon and Google are not only Anthropic's investors, but also the cloud service providers on which it relies to train and run Claude. Investment funds flow back to the investors' cloud businesses in the form of cloud service fees, both supporting Anthropic's valuation and boosting the cloud vendors' revenue.

This structure is not illegal, nor is it rare in the AI industry. Similar circularity exists between Microsoft and OpenAI, and between Nvidia and a host of computing-power customers. But once it enters the public market, the questions investors need to answer become very specific. How much of Anthropic's growth is real end-demand, and how much is capital circulating among giants?

The answer to this question concerns more than just Anthropic. Some traders have already regarded the $518 billion commitment as the biggest positive for the computing-power "shovel sellers," meaning Anthropic's stock price will become more tightly bound to the entire AI infrastructure chain in the future.

50.1% Voting Power

If the previous details are about money, then the governance structure is about who has the final say. This is the most "Anthropic" part of the entire prospectus.

According to the prospectus, Anthropic will establish an entity called Founder LLC, whose initial members are seven co-founders including CEO Dario Amodei. The seven control one Class F share by majority vote, and this one share has 50.1% of total voting power on key corporate matters, including the election of some directors.

After the listing is completed, Chairman Daniela Amodei, Dario Amodei and a yet-to-be-announced director will be elected by Class F and Class A shareholders; the other four directors will be selected by a long-term benefit trust, whose current trustees include former Federal Reserve Chairman Bernanke and national security expert Richard Fontaine.

The Class A shares bought by ordinary investors carry one vote per share, but the other share classes each have their own uses and restrictions, among which strategic partners hold only a very small amount of voting power. This structure effectively weakens the influence of ordinary investors.

The prospectus also prepares for the disintegration of the founding team. Members leaving the company, dying, selling too many shares or being dismissed for cause may all be removed from this group; if the members fall to two or fewer, the Class F shares' voting advantage will gradually disappear. The co-founders also pledged in the document to devote 80% of their personally held Anthropic equity to charity.

The most intriguing part is the prospectus's self-assessment of all this. It states plainly that this governance model may lead the company to make decisions that conflict with short-term, medium-term or long-term financial interests and operating performance, which may in turn reduce the value of Class A shares.

A company seeking a $2 trillion valuation explicitly warns investors in its listing document that their returns are not always the first priority. That is quite rare in the capital markets. The two major proxy advisory firms ISS and Glass Lewis have historically recommended against multi-class share arrangements, and whether this design can be accepted by mainstream institutions will be a major point of interest during the roadshow stage.

Voluntarily Giving Up Video and Image Generation

To prove that this governance structure is not just talk, the prospectus gives several examples.

One is Mythos Preview, a model with particularly strong cybersecurity capabilities, for which Anthropic set up a limited-access program open only to a small number of institutions.

The other example is more unexpected. The prospectus says the company chose not to develop certain commercially attractive products, such as image and video generation models, in order to direct computing power toward research and safety priorities.

Taken alone, this sentence is an expression of stance. Placed in this year's industry context, its meaning becomes much more complicated.

Half a year ago, OpenAI announced the shutdown of Sora, a video generation app that had been online for only about six months, citing the company's shift toward commercial tools as it prepared for a possible listing. Sora's app and web version went offline on April 26, and its API stopped service on September 24. High computing costs, rapidly churning users and copyright disputes were all reasons Sora was abandoned.

The same thing: OpenAI did it, burned money, and ultimately cut it, citing a focus on enterprise business; Anthropic never entered it, and framed it in the prospectus as a trade-off made for safety and research.

The two companies ultimately arrived at the same position, both betting their chips on enterprise customers and coding scenarios, yet telling very different stories. Sora's outcome also makes the premise of "commercially attractive" seem somewhat questionable.

We are more inclined to believe that the mission narrative here and the commercial judgment are hard to separate. At a stage when computing power is extremely tight, concentrating resources where unit value is highest is itself the right business.

Anthropic's cleverness lies in writing a correct business decision as proof of values.

80 Pages of Risk Factors Mentioning "Human Survival"

The risk factors section of the prospectus runs 80 pages and explicitly warns that its AI models may pose catastrophic or even existential threats. Reuters also mentioned that Anthropic's own research has documented controlled tests in which increasingly autonomous AI systems displayed potentially harmful behavior, including sabotaging code, assisting fraud and manipulating information.

Putting "our products may threaten humanity" into a prospectus is both a sincere stance and standard legal groundwork. The two are not contradictory. The original function of risk factors is to make clear in advance all the places where things could go wrong, so as to avoid investor lawsuits later. A company that publicly discusses AI risk year-round would instead be legally exposed if it avoided the topic in its prospectus.

The real tension lies elsewhere. Dario Amodei once called on the global AI community to slow the pace of releasing new capabilities, but Anthropic still launched the new model Opus 5.5 last week in response to the momentum after OpenAI released GPT-6 Astra.

Calling for slowing down and accelerating releases at the same time is not surprising. On a track where everyone is speeding up, the cost of unilateral deceleration is ceding the position of defining the rules to others. This is the dilemma Anthropic has faced since its first day of existence, and the prospectus merely puts it in black and white.

A $518 Billion Computing-Power Cost?

Another widely cited figure is the commitment of about $518 billion for cloud, computing power and infrastructure.

Many reports interpreted it as "spending $518 billion on computing power next year," and even different versions of Reuters' articles were inconsistent in wording. Some analysts pointed out that this is the total value of multi-year contracts, not single-year spending; according to disclosed allocations, Amazon AWS is about $100 billion, Google Cloud about $200 billion, Lambda and Nscale about $80 billion, Fluidstack about $50 billion, Microsoft Azure about $30 billion, and the rest are smaller contracts. The AWS portion is a ten-year contract signed earlier this year that can lock in up to 5 gigawatts of computing capacity.

Even measured as a multi-year total, this number is still astonishing. As of December 31 last year, Anthropic's cash, cash equivalents and short-term investments totaled $20.28 billion. The $518 billion commitment is more than 25 times that cash.

This also explains the urgency of going public. Reports say Anthropic may seek to raise as much as $100 billion this time. To a large extent, the listing is about finding money for computing-power contracts already signed.

Two Kinds of "Safety," Two Paths to Listing

According to earlier Reuters reports, Anthropic's public listing will likely be postponed until after the U.S. midterm elections in November. What awaits it is not a particularly friendly market. AI and chip stocks recently went through a round of selling; SpaceX rose 19% to $160 on its first day of trading in June and has now fallen back to around $147, which, although still above its $135 issue price, is enough to make investors more cautious about high-valuation growth stocks.

Anthropic had also previously come under White House scrutiny over its tools and was once blacklisted by the Pentagon, a move later halted by a U.S. judge, leaving political variables still unresolved.

More worth comparing is its old rival.

OpenAI confidentially filed for a listing about a week after Anthropic and admitted in its announcement that it made the move public because it expected the news to leak. But by September, Sam Altman confirmed in an interview with Fortune that OpenAI would not go public this year, saying that given everything happening around safety, now is not a good time to list; The New York Times also reported earlier that one reason for the delay was Altman's insistence on a valuation of at least $1 trillion.

Both invoke "safety," but Anthropic writes it into its governance structure as a reason to persuade investors to go public; OpenAI uses it as a reason to postpone listing.

It is hard to say who is more sincere. But it reveals one thing: at the moment AI companies move toward the public market, "safety" is no longer just a technical and ethical issue, but has also become part of the capital narrative.

Reuters believes this listing will establish Wall Street's benchmark for valuing leading AI companies, including OpenAI. Once Anthropic is successfully listed, every frontier lab afterward will have to face the same set of questions. Can the revenue curve catch up with the computing-power bill, can customers be locked in by contracts, how much control should founders have, and will the market be willing to pay for a business a company voluntarily gives up.

What this prospectus truly has to answer may be a more ancient question: can the capital market tolerate a company that writes "what it will not do" as importantly as "what it will do."