The 7 most intriguing details in Anthropic’s IPO prospectus

Apart from various speculations and misunderstandings, it is a prospectus that is very characteristic of Company A.

Author｜The King of Techno

Editor｜Jingyu

On September 28, local time, foreign media disclosed an IPO prospectus for Anthropic.

First let’s make it clear what the nature of this document is. It is not a version that Anthropic proactively released. In June, Anthropic confidentially filed a draft S-1 registration statement with the SEC as a public benefit corporation in preparation for an initial public offering of common stock. According to the rules, companies must publicly submit a revised S-1 at least 15 days before a roadshow to institutional investors, and the complete financial data will not be visible to the outside world until then. What Reuters saw this time was a version that has not yet been made public.Anthropic declined to comment.。

This means that the figures may still be adjusted according to the SEC’s review opinions, and many details can only be glimpsed through Reuters’ reporting.

But if the document is a rough reflection of what Anthropic will look like when it finally comes to the open market, its weight is hard to overestimate. The listing’s valuation target is more than double Anthropic’s own estimate of $965 billion in May, or more than $2 trillion.

This is the first cutting-edge large-scale model laboratory to spread its complete ledger to public investors in the nearly four years since ChatGPT triggered this AI wave.

Until then, the AI race has been funded by venture capital, sovereign wealth funds and technology giants, and the outside world can only piece together the operating conditions of these companies from scattered annualized revenue and financing valuations. Now, for the first time, the ledger has audit capabilities.

Geek Park picked out the 7 most noteworthy places from the prospectus.

Misinterpreted $42 billion loss

展開全文

The first number circulated in the prospectus was a net loss of approximately US$42 billion in 2025.

It’s easy to read this number as “burning $42 billion a year,” but that’s not the case. About $34 billion of that was an accounting charge that reflected rising valuations of financing instruments that could be converted into Anthropic stock in the future, rather than money spent by the company to run its business.

This produces a counter-intuitive result.The more valuable Anthropic is, the more it “loses” on paper. 2025 happens to be the year when its valuation skyrockets, and this fee is also amplified simultaneously.

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

Revenue has increased by about 12 times, and operating losses have expanded by less than 3 times. Scale effects are indeed emerging. But an operating loss of $8 billion a year is a figure that needs to be explained repeatedly for any company preparing to go public.

The pitfall of $65 billion in annualized revenue

After the news of the prospectus spread, the top comment on the American retail investor forum wallstreetbets contained only one sentence, to the effect of “2 trillion valuation, with 4 billion revenue?」

This question is intuitive, but the denominator is wrong.

The main body of the prospectus covers fiscal 2025, which ended on December 31 last year, while Anthropic’s real breakout occurred in 2026. Its annualized revenue was reported to be about $9 billion at the end of 2025, about $47 billion in May this year, and more than $65 billion by the end of July.

The $4.59 billion in the prospectus is the last snapshot of a company “before it takes off”. The US$2 trillion valuation given by the capital market is betting on the curve after 2026, not the full-year performance in 2025.

However, the annualized income itself needs to be discounted. It’s usually calculated by multiplying a given month’s income by 12, which naturally amplifies short-term fluctuations. The prospectus itself also gives two warnings. Nearly a quarter of revenue last year came from just two customers; the company also acknowledged in its risk factors that many of its largest customers do not have long-term contracts and could cut or even stop spending.

Reuters did not disclose the names of these two customers. This will be the most important information to check as soon as the public version of S-1 comes out.Anthropic’s growth curve is steep enough, but it lacks lock-in at the contract level, which is a key prerequisite for understanding its valuation.

The largest shareholder, the largest landlord

In 2025, Anthropic will spend $7.33 billion on computing power and infrastructure, accounting for 58% of total operating expenses and approximately 1.6 times its annual revenue.

The question is to whom the money goes.

Amazon and Google are not only investors in Anthropic, but also the cloud providers it relies on to train and run Claude. The investment money flows back to the investor’s cloud business in the form of cloud service fees, which not only supports Anthropic’s valuation, but also drives up the income of cloud vendors.

This structure is not illegal and is not uncommon in the AI industry. Similar cycles exist between Microsoft and OpenAI, Nvidia and a number of computing power customers. But after entering 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 the circulation of capital among giants?

The answer to this question isn’t just about Anthropic. Some traders already regard the $518 billion commitment as the biggest benefit to “shovel sellers” of computing power, which means that Anthropic’s stock price will be more closely tied to the entire AI infrastructure chain in the future.

50.1% voting rights

If the first few 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 set up an entity called Founder LLC, with seven co-founders including CEO Dario Amodei as the initial members. Seven people vote by majority to dominate one Class F stock, which holds 50.1% of the total voting power on key corporate matters, including the election of certain directors.

Upon completion of the listing, Chairman Daniela Amodei, Dario Amodei and one yet to be announced director will be elected by Class F and Class A shareholders; the remaining four directors will be elected by the Long-term Interest Trust,Current trustees include former Federal Reserve Chairman Ben Bernanke and national security expert Richard Fontaine。

The Class A shares bought by ordinary investors are one share, one vote, but the other classes of shares have their own uses and restrictions. Among them, strategic partners only have very small voting rights. This structure will actually weaken the influence of ordinary investors.

The prospectus also prepared a plan for the disintegration of the founding team. Members may be removed from this group if they leave the company, die, sell too many shares, or are fired for cause; if the number of members is reduced to two or fewer, the voting power advantage of Class F shares will gradually disappear. The co-founders also pledged in the document to use 80% of their personal stake in Anthropic for charity.

The most interesting thing is the self-evaluation of all this in the prospectus. It bluntly stated that this governance model may cause the company to make decisions that are inconsistent with short-, medium-term or long-term financial interests and operating performance, which may reduce the value of Class A shares.

A company seeking a $2 trillion valuation explicitly reminds investors in its listing documents that their returns don’t always come first. This is quite rare in the capital market. Two major proxy advisory firms, ISS and Glass Lewis, have historically recommended against multi-tier equity arrangements. Whether this design can be accepted by mainstream institutions will be a major focus of the roadshow stage.

Actively abandon video and image generation

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

One is Mythos Preview, a model that is particularly capable in terms of cybersecurity, and Anthropic has set up a limited access plan for it, open only to a few institutions.

Another example is more surprising. The prospectus stated,The company chooses not to develop certain commercially attractive products, such as image and video generation models, in order to devote computing power to research and security priorities.

Just looking at this sentence, it is an expression of position. Put into this year’s industry context, the meaning is much more complicated.

Half a year ago, OpenAI announced the closure of Sora, a video generation application that had only been online for about six months, on the grounds that the company was turning to commercial tools to prepare for a possible listing. Sora’s application and web pages were offline on April 26, and the API was also discontinued on September 24. The high cost of computing power, the rapid loss of users and copyright disputes are all reasons why Sora was abandoned.

OpenAI did the same thing, burned money, and eventually cut it off because it was focused on enterprise business; Anthropic never got involved and wrote it in the prospectus as a trade-off between security and research.

The two companies ended up in the same position, betting on enterprise customers and programming scenarios, but the stories they told were completely different. The ending of Sora also makes the premise of “commercially attractive” seem a bit questionable.

Geek Park prefers to believe that the mission narrative and business judgment here are difficult to separate. At a stage when computing power is extremely tight, concentrating resources where the unit value is highest is the right business in itself.

The brilliance of Anthropic is that it turns a correct business decision into a proof of values.

80 pages of risk factors written “human survival”

The risk factors section of the prospectus is 80 pages long and explicitly warns that its AI models could pose catastrophic or even existential threats. Reuters also noted that Anthropic’s own research has documented controlled tests in which increasingly autonomous AI systems exhibited potentially harmful behavior, including breaking code, assisting fraud, and manipulating information.

Writing “our products may threaten humanity” into the prospectus is both a sincere stance and standard legal foreshadowing.. The two are not contradictory. The original function of risk factors is to spell out all possible problems in advance to avoid being involved in investor lawsuits in the future. If a company that has been publicly talking about the risks of AI for many years avoids mentioning it in its prospectus, it will be legally passive.

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

It’s no surprise that calls to slow down coincide with accelerated releases.In a track where everyone is accelerating, the cost of unilaterally slowing down is to give up the position of defining the rules to others.. This is the dilemma Anthropic has faced from day one, and the prospectus just puts it in black and white.

$518 billion in computing power costs?

Another widely cited figure is approximately $518 billion in cloud, computing power and infrastructure commitments.

Many reports interpreted it as “US$518 billion will be spent on computing power next year.” Even different versions of Reuters’ manuscripts are inconsistent in wording. Some analysts pointed out that this is the total value of multi-year contracts, not single-year expenditures; based on disclosed allocations, Amazon AWS is about $100 billion, Google Cloud is about $200 billion, Lambda and Nscale are about $80 billion, Fluidstack is about $50 billion, Microsoft Azure is about $30 billion, and the rest are smaller contracts. The AWS portion is a ten-year contract signed earlier this year that locks in up to 5 gigawatts of computing capacity.

Even looking at the multi-year total, this number is still staggering. Anthropic had a combined $20.28 billion in cash, cash equivalents and short-term investments as of Dec. 31.The $518 billion commitment is more than 25 times that cash.

This also explains the urgency of going public. Reports suggest Anthropic could seek to raise as much as $100 billion this time around.To a large extent, listing is about finding money for the computing power contract that has been signed.

Two kinds of “safety”, two ways to go public

According to previous reports by Reuters, Anthropic’s public listing is likely to be postponed until after the U.S. midterm elections in November. What awaits it is not exactly a friendly market. AI and chip stocks have recently experienced a sell-off; SpaceX soared 19% to $160 on its first day of listing in June, and has now fallen back to around $147. Although it is still higher than the issue price of $135, it is enough to make investors more cautious about high-valuation growth stocks.

Anthropic was previously scrutinized by the White House due to its tools, and was once blacklisted by the Pentagon. It was later stopped by a U.S. judge, and political variables are also unresolved.

What is more worthy of comparison is its old rival.

OpenAI also secretly submitted its listing application about a week after Anthropic, and admitted in the announcement that it made the proactive announcement in anticipation of the news leaking. But in September, Sam Altman confirmed in an interview with Fortune that OpenAI would not go public this year, citing that it was not a good time to go public given everything that was going on in security; the New York Times had previously reported that one of the reasons for the delay was Altman’s insistence on a valuation of at least $1 trillion.

The same is “safety”. Anthropic has written it into its governance structure as a reason to convince investors to go public; OpenAI uses it as a reason to postpone the listing.

It’s hard to say who is more sincere. But it reveals one thing. At this moment when AI companies are entering the public market, “security” is no longer just a technical and ethical issue, but has also become part of the capital narrative.

Reuters believes that this listing will establish Wall Street as the benchmark for the valuation of leading AI companies, including OpenAI. Once Anthropic is successfully listed, all cutting-edge laboratories will face the same set of problems. Whether the revenue curve can catch up with the computing power bill, whether customers can be locked in contracts, how much control the founder should have, and whether the market is willing to pay for the business that a company voluntarily gives up.

What this prospectus really wants to answer may be an older question, whether the capital market can tolerate a company that writes “what it doesn’t do” as important as “what it does”.

*Source of header image: Reuters

This article is an original article by Geek Park. For reprinting, please contact Geek Jun WeChat geekparkGO

極客一問