Anthropic Files IPO Prospectus, Author Summarizes 11 Details

Anthropic is about to IPO, so I've summarized 11 interesting detail figures from the prospectus for everyone. Digital Life Kazik, October 27, 2026, 08:05:45

Last night, Anthropic finally made public their IPO application, posting their prospectus on the U.S. Securities and Exchange Commission's website.

Many people also said that A\, in order to go public, finally acted like a human and pulled out an extremely good model like Claude Opus 5.5 for everyone, just to look better at the time of the IPO.

This document happened to be released right before my flight took off last night, so I scraped it down and roughly scanned the whole thing on the plane, because this may indeed be the first time we have the opportunity to understand Anthropic, this world-class AI company, so comprehensively.

And this time they plan to raise about 100 billion USD, with the valuation likely landing between 1.5 trillion and 2 trillion. If they really can raise that much by then, Anthropic should be the largest IPO in human history so far, compared with the world record achieved when old Ma's SpaceX went public previously: 75 billion USD raised in the initial offering, after which the underwriters fully exercised the over-allotment option, ultimately raising 85.7 billion USD.

No matter how you look at it, this is a very interesting event.

After reading it myself, I found it quite interesting, and I summarized 11 points that I think are pretty fun to share with everyone.

1. Stock ticker ANT

Anthropic is applying to list on Nasdaq this time, with the ticker ANT.

In English, ANT means ant.

This word has two very interesting points.

First, when Anthropic introduces its team, there is a sentence:

Our employees are known internally as "Ants"

Anthropic's employees all call themselves "Ants" internally, a group of ants. The explanation in the prospectus is that this nickname represents a collaborative, not-too-self-centered team culture; the original wording used is ego-light.

And the other interesting point is that six years ago, on October 26, the same day.

On the evening of October 26, 2020, Ant Group publicly released its IPO prospectus on the Hong Kong Stock Exchange, preparing to list simultaneously in Shanghai and Hong Kong, planning to raise 34.5 billion USD, and preparing to begin trading on November 5.

On that day six years ago, this was originally also the largest IPO in human history.

As for what happened afterward, everyone knows: on November 3, the listing was suspended, and since then, there has been no date in sight.

Six years later, on the same October 26, another company with the ticker ANT, under the name "Ant," continues to sprint toward the record for the largest IPO in human history.

As for the listing date, it has not yet been confirmed, but the mainstream view is that everyone will wait until after the U.S. midterm elections on November 3.

ANT, October 26, November 3—there are still too many coincidences in the human world.

2. Exaggerated growth rate

In 2023, Anthropic's revenue was 41 million USD, and in the same year, its interest income was 69 million USD.

The money earned from deposit interest was nearly 70% more than the money earned from selling Claude.

That year, it spent a total of 2 million USD on sales and marketing. The average price of a 30-second Super Bowl ad in 2023 was 7 million USD. Anthropic's sales and marketing expenses for the entire year could not even buy 10 seconds.

On the cost side, in 2023 it spent 433 million USD on compute and infrastructure.

Then, the style changed abruptly.

In the first quarter of 2025, revenue was 376 million USD. In the second quarter, 787 million; in the third quarter, 1.382 billion; in the fourth quarter, 2.040 billion. By this year, the first quarter was 4.731 billion, and the second quarter was 11.555 billion.

In five quarters, it rose 30-fold.

In the second quarter of this year, its average daily revenue was 127 million USD. The money it earned in all of 2023 is now earned back in less than 8 hours.

But looking only at revenue, you actually cannot see the company's most core change. I think this change is hidden in the ratio of revenue to compute costs.

Using the same table, I calculated how much revenue could be earned back for every 1 USD spent on compute and infrastructure.

In the first quarter of 2025, 0.40 USD. Second quarter 0.57, third quarter 0.68, fourth quarter 0.69. This year, first quarter 1.40, second quarter 2.08.

For the entire four quarters before that, it had been doing one thing: spending 1 yuan to buy compute and selling it for less than 70 cents. In the first quarter of this year, this line crossed 1 for the first time. By the second quarter, it became spending 1 yuan to earn 2 yuan.

Also in the second quarter, its operating profit turned positive for the first time, at 974 million USD.

I also dug into why it grew so fast. They themselves gave about 3 reasons.

1. Improved training efficiency. Claude is increasingly participating in the R&D of the next generation of Claude; internally this is called closing the loop.

2. Improved inference efficiency; for the same model capability, less and less compute is needed. This can be seen from the formidable momentum of Claude Opus 5.5.

3. Chips. They themselves simultaneously train and run models on four kinds of chips: Amazon's Trainium, Google's TPU, and Nvidia and AMD GPUs. They built their own scheduling system to maximize performance and cost-effectiveness.

So this also led to this: in the second quarter of this year, revenue rose 1368% year over year, but spending on compute and infrastructure rose only 300%.

At the same time, in the overall revenue mix, API users still account for the largest share.

The customer structure is also changing. In the first quarter of this year, the top five customers contributed about 36% of revenue; by the third quarter, that proportion had fallen to about 21%, and no single customer accounted for more than 10%.

Moreover, more than half of the Fortune 500 are already customers that spend more than $100,000 a year with it, but on average these companies spend less than 0.1% of their total expenditure on Anthropic. Even the top 10 that use it most spend less than 1%.

By the way, its sales and marketing expenses—which back then could not even afford a 10-second Super Bowl ad—were $3.168 billion in the first half of this year.

In February this year, it actually did go run a Super Bowl ad. The line at the end, "Ads are coming to AI. But not to Claude.", was a direct jab at OpenAI's plan to put ads in ChatGPT.

Altman posted a long tweet in response, saying these ads were pretty funny and he laughed too, but he could not understand why Anthropic would make something so "clearly dishonest."

3. Lost $80.7 billion

Anthropic's net loss in the first half of this year was $80.7 billion.

In the same half year, revenue was $16.286 billion, and the loss was roughly five times revenue.

I know many people are curious: wasn't it said that operating profit in the second quarter was positive? So with operating profit on one side and a net loss of tens of billions on the other, where did the money go?

The answer is a line item in the lower half of the income statement, called fair value changes in convertible notes, derivative liabilities and SAFEs. In the first half of this year, this line was negative $72.4 billion.

This may be a bit convoluted, so let me explain it in plain language.

A large part of the money Amazon invested in Anthropic was not used to directly buy shares, but was provided in the form of convertible notes. You can think of a convertible note as an IOU that says it can later be converted into shares at an agreed price. As of the end of June this year, the principal on these notes was still $6.2 billion.

The problem is that Anthropic's valuation was $380 billion in February this year and became $965 billion in May.

The faster the valuation rises, the more valuable this IOU that can be converted into shares at the old price becomes. Accounting rules require that Anthropic record as a loss on its own books however much more the note is worth.

So in the second quarter of this year alone, this item recorded $59.6 billion. The prospectus's own explanation is that the figure was largest this quarter mainly because the Series H financing pushed up the valuation, and another $8.4 billion was related to several note purchase agreements newly signed this year.

So this kind of loss actually does not require spending a single cent of cash. The prospectus also gives an adjusted measure: excluding items such as these fair value changes and equity incentives, adjusted net profit for the first half of this year was positive $2.671 billion.

4. A novelist's name

In Anthropic's 2025 income statement, there is a rarely seen line item.

It is called "Bartz settlement", in the amount of $1.5 billion.

Bartz is the American suspense novelist Andrea Bartz, who wrote bestsellers such as "We Were Never Here". In August 2024, she and two other authors led a lawsuit against Anthropic for using pirated books to train Claude.

And then Anthropic paid $1.5 billion...

The settlement covered more than 400,000 books, about $3,000 per book, and the court formally approved it on July 20 this year.

The most surreal part of this case is that Judge Alsup at the time ruled that Anthropic buying paper books with its own money and scanning them into a database was fair use and legal. The millions of books downloaded from pirate sites were not.

Later the court unsealed more than 4,000 pages of case files, and The Washington Post dug out an internal project called Project Panama. Starting in early 2024, Anthropic bought millions of secondhand books in batches, used hydraulic book cutters to cut the books open, scanned them at high speed, and then sent the books for recycling. The exact wording in an internal document was, "our effort to destructively scan all the books in the world."

Anthropic wrote that its revenue has always included, and will continue to include, some money from customers later found to have violated usage policies or commercial terms, because discovering, investigating and banning accounts takes time: the money is collected first, and the account is banned later.

Then it gave a number: in the first nine months ended September 30 this year, this portion of revenue was about 5% to 7% of total revenue.

What does this mean? You can understand it as: ban the account but do not refund the money...

I did a rough calculation: 16.286 billion in the first half, plus the 16.8 billion in the third quarter reported by the media, gives about 33 billion USD for the first nine months, so 5% to 7% is 1.6 to 2.3 billion USD.

Then it said that in August and September they stepped up detection and banning, so a lot of accounts were killed off at that time (my old account also fell in that batch), which will lower the share of their revenue from violations and may also drag down future revenue and growth.

6. Making peace with Musk

This one has to be told in chronological order, otherwise you won't appreciate how surreal it is.

In January this year, Anthropic cut off xAI employees' access to Claude through Cursor. A co-founder of xAI wrote in an internal email that efficiency would suffer, but this also forces us to build our own coding product.

On February 12, Anthropic announced the completion of a 30 billion USD funding round, and Musk replied underneath, "Your AI hates Whites & Asians, especially Chinese, heterosexuals and men. This is misanthropic and evil."

Roughly meaning you are evil and beyond redemption...

Three months later, on May 6, Anthropic announced it was renting all of xAI's Colossus 1 data center in Memphis, more than 220,000 Nvidia GPUs.

Musk's tone immediately changed too, saying the people he met at Anthropic were all very capable and cared deeply about doing the right thing.

In June, SpaceX went public. According to Reuters' reading of SpaceX's prospectus, at that time Anthropic had to pay about 1.25 billion USD per month.

By the time of Anthropic's own prospectus, this amount had changed again.

From May to October, Anthropic signed and then revised several contracts with a company called SpaceX.AI LLC, with a parenthetical note that this company's former name was X.AI LLC.

That is the xAI that makes Grok, now a wholly owned subsidiary of SpaceX.

The total contract value is up to 120.5 billion USD, with a term through 2029, so by that calculation, once all the compute is delivered in December, Anthropic will have to pay it about 3.1 billion USD per month.

SpaceX's prospectus states that the revenue of its entire AI business in the first quarter of this year was 818 million USD.

Anthropic's monthly compute fee is roughly four times the quarterly revenue of Musk's entire AI business.

All I can say is, curse all you want, but business is business...

7. A bill signed all the way to 2045

Musk's 120.5 billion USD is not even the big one. What comes next is what really opened my eyes and made me reunderstand what "heavy assets" means...

Let me list a few of the big ones.

Google: at least 114.5 billion USD, from April 2026 to July 2033. In addition, there is an agreement for additional TPU compute purchases from 2028 to 2031; Anthropic itself estimates it will spend another 125 to 150 billion USD, and this part is not even included in the contractual commitment table.

Amazon: at least 110 billion USD, from May 2026 to April 2036.

Microsoft: at least 44.9 billion USD, from November 2026 to May 2033.

SpaceX's AI subsidiary: up to 120.5 billion USD.

AMD: starting next year, purchasing Helios full racks, expected to exceed 20 billion USD.

Broadcom: a five-year AI accelerator equipment lease, with total rent of about 161.2 billion USD, irrevocable.

The Google, Amazon and Microsoft contracts all contain the same phrase, regardless of actual usage: whether you use it or not, you have to pay. Google and Amazon also wrote that if usage falls short, the difference must be made up.

This kind of clause has a special name: take-or-pay. Previously it was most commonly seen in long-term contracts in energy industries such as natural gas and electricity.

Judging by this approach, in Anthropic's mind, compute is basically already equivalent to energy.

8. Spent 7 dollars and took control of the company

There is a very interesting contrast.

When SBF invested in Anthropic back then, what he got for 500 million USD was a pile of non-voting shares. In fact, it was not just him: major shareholders like Amazon and Google, who put in billions or tens of billions of USD, received Class N shares after the IPO, which likewise have no voting rights.

Every person on the list contributes 1 dollar to the company. The company uses this money to subscribe to 1 Class F share of Anthropic.

There are seven people on the list: Dario, Daniela, and Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish and Chris Olah, the seven co-founders of Anthropic.

One dollar each, seven dollars in total.

This one Class F share has no economic rights whatsoever and cannot receive a single cent of profit, but its voting power is fixed at 50.1% of all the company's voting power, covering all matters submitted to shareholders for a vote, including the election of directors other than trust seats, and no matter how many new shares the company issues in the future, it cannot be diluted.

The agreement also specifically includes a sentence: this 1 dollar is a non-revocable capital contribution, and the seven people have waived any right to get this money back.

Spend 7 dollars to get more than half of the voting power of a company that could be worth hundreds of billions or even a trillion dollars.

The prospectus's characterization of these seven people is also very interesting: "guardians of our mission."

A\, at its underlying philosophy, is still that same A\.

9. Every 50% rise in the stock price, 2% is set aside for philanthropy

Before the IPO, Anthropic had already allocated 840 million Class N shares to a wholly owned subsidiary called the Anthropic Public Benefit Fund, accounting for 5% of the fully diluted share capital. The trust chairman, Buddy Shah, will also be one of the fund's first managers.

At the price of the May H round, this block of shares is worth about 50 billion dollars.

Then, the board approved a mechanism: starting from the IPO issue price, every time the stock price rises 50% on a sustained basis, up to another 2% of shares can be placed into this fund.

The more the stock price rises, the more is put out.

In 2020, several researchers at the Oxford Future of Humanity Institute wrote a paper called the "Windfall Clause," which envisioned AI companies committing in advance that once they earned extraordinary profits, they would distribute a large portion of them to society.

Six years later, it appeared in a more concrete form in an IPO prospectus.

The footnote to the founders' letter states that each of the seven co-founders has pledged to donate 80% of their personal shares, totaling 9% of the fully diluted share capital. Then employees pledged 8% through a matching gift program, plus 5% from the public benefit fund.

Overall, about 22%.

Roughly calculated at the H round price, this 22% stake is worth more than 200 billion dollars.

10. Dario's resume does not mention Baidu

On page 255 of the prospectus, Dario's resume is written like this.

PhD in biophysics from Princeton, postdoc at Stanford School of Medicine, October 2015 to June 2016 at Google Brain, July 2016 to December 2020 at OpenAI, where he ultimately became Vice President of Research.

Not a single mention of Baidu...

You should know that, as Dario himself said, he entered the AI field at the end of 2014 at Baidu, working with Andrew Ng on speech recognition, and it was there that he first noticed that the more data given to a model, the more steadily better it performed.

This thing later also came to be called Scaling Laws.

Didn't expect that, right? The Scaling Laws, only proposed in 2020, were actually a flash of insight Dario had back at Baidu.

But interestingly, in the prospectus, Dario completely omitted his experience working at Baidu.

11. 900% net revenue retention

Finally, a few numbers that friends in SaaS may lose sleep over...

First, net revenue retention, about 900%.

That means a customer who spent 1 dollar at Anthropic a year ago (excluding subscription users) spent a total of 9 dollars in the same quarter this year.

Anyone in enterprise software knows that 120% on this metric is considered excellent; Snowflake at its peak was 178%, and that has been told as a legend for years.

Second, revenue per employee. As of September 30 this year, Anthropic's revenue per employee over the past 12 months exceeded 6.6 million dollars, up from 1.4 million a year earlier. Headcount rose from about 940 at the end of 2024 to about 5,300.

Third, I think this best illustrates how brutal this industry is.

According to the prospectus, in the third quarter of this year, more than 80% of Anthropic's revenue came from model versions released only in the past six months.

OpenRouter also released data the day before yesterday showing that consumption of Claude Opus 5.5 has surged.

No way around it, the model is just too strong.

Fourth, customers. Customers spending more than 100,000 dollars a year numbered about 20 at the end of 2023 and more than 11,900 by the end of September this year. In addition, there are about 40 million customers spending less than 100,000 dollars.

Very interesting.

In closing

Because I myself actually worked in public funds for several years, and back then I also read some prospectuses and financial reports.

But frankly speaking, a prospectus as sweeping and astronomical as Anthropic's is truly something I have never seen.

What we are witnessing is already a comprehensive restructuring involving capital, energy, software, hardware and social institutions.

Soon OpenAI will also go public in 2027, and by then, it will be another giant.

By that time, we will finally have the chance to directly compare OpenAI and Anthropic's real revenue, compute costs, gross margins, cash flow and the profitability of their model businesses.

Just thinking about it is a very interesting thing.

>/ Author: Kazike