https://x.com/i/article/2114877522169212928
Anthropic is going public, so I've summarized 11 interesting details from its IPO prospectus.
Last night, Anthropic finally made its IPO filing public, posting its prospectus on the US Securities and Exchange Commission website.
Many people are saying that to go public, A\ finally got real and gave everyone Claude Opus 5.5, an ultra-good model, to make the IPO go better.
This document was published right before my flight took off last night, so I downloaded it and skimmed the whole thing on the plane, because this may really be the first time we get such a comprehensive look at Anthropic, a world-leading AI company.
This round is planned to raise about $100 billion, with a valuation of roughly $1.5 to $2 trillion. If they really raise that much, Anthropic would be the largest IPO in human history so far. Compare that with the world record set by Musk's SpaceX listing: an initial offering raising $75 billion, after which the underwriters fully exercised their greenshoe option, ultimately raising $85.7 billion.
From any angle, this is a very interesting event.
After reading it, I found it quite fascinating and summarized 11 points I think are pretty fun to share.
1. Ticker symbol 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 angles.
First, when Anthropic introduces its team, there's a line:
Our employees are known internally as "Ants"
Anthropic employees internally call themselves "Ants," a swarm of ants. The prospectus explains that this nickname represents a team culture of mutual collaboration and low ego; the original wording is "ego-light."
The other interesting point is six years ago, on October 26th, the same day.
On the night of October 26, 2020, Ant Group publicly released its IPO prospectus on the Hong Kong Stock Exchange, planning to list simultaneously in Shanghai and Hong Kong, aiming to raise $34.5 billion, with trading set to begin on November 5.
Back then, that was also set to be the largest IPO in human history.
The rest of the story is known to everyone: on November 3, the listing was suspended, and it has been indefinitely delayed ever since.
Six years later, on the same October 26th, another company with the ticker ANT, bearing the name "Ant," is pushing to set the largest IPO record in human history.
As for the listing date, it hasn't been confirmed yet, but the mainstream view is that it will wait until after the US midterm elections on November 3.
ANT, October 26th, November 3rd, there are just too many coincidences in the human world.
2. Explosive growth
In 2023, Anthropic's revenue was $41 million, and in the same year its interest income was $69 million.
It earned more from deposit interest than from selling Claude, by nearly 70%.
That year, it spent $2 million on sales and marketing in total. In 2023, the average price of a 30-second Super Bowl ad was $7 million, so Anthropic's entire year of sales and marketing couldn't even buy 10 seconds of airtime.
On costs, in 2023 it spent $433 million on compute and infrastructure.
Then the picture changed dramatically.
In Q1 2025, revenue was $376 million. Q2 was $787 million, Q3 $1.382 billion, Q4 $2.040 billion. This year, Q1 was $4.731 billion, and Q2 was $11.555 billion.
Over five quarters, it grew 30 times.
In Q2 this year, its average daily revenue was $127 million. The money it earned in all of 2023 has now been earned back in less than 8 hours.
But looking at revenue alone doesn't reveal the company's most important change. I think that change is hidden in the ratio between revenue and compute costs.
I ran the numbers on the same table: how much revenue each $1 spent on compute and infrastructure brings back.
In Q1 2025, it was $0.40. Q2 $0.57, Q3 $0.68, Q4 $0.69. Q1 this year $1.40, Q2 $2.08.
For the first four quarters, it was consistently doing one thing: spending $1 on compute and selling it for less than 70 cents. In Q1 this year, that line crossed 1 for the first time. By Q2, it was spending $1 and earning $2.
Also in Q2, its operating profit turned positive for the first time, at $974 million.
I also looked into why it grew so fast, and they gave roughly three reasons.
1. Improved training efficiency. Claude is increasingly involved in developing the next generation of Claude, which they internally call "closing the loop."
2. Improved inference efficiency. The same model capability requires less and less compute. You can see this from the formidable strength of Claude Opus 5.5.
3. Chips. They train and run models on four types of chips simultaneously: Amazon's Trainium, Google's TPU, and Nvidia's and AMD's GPUs. They built their own scheduling system to squeeze out maximum performance and cost-efficiency.
So this year, Q2 revenue rose 1368% year over year, while compute and infrastructure spending rose only 300%.
At the same time, API users still make up the bulk of revenue.
The customer structure is also changing. In Q1 this year, the top five customers contributed about 36% of revenue. By Q3, that share dropped to about 21%, and no single customer accounted for more than 10%.
And more than half of the Fortune 500 are now customers spending over $100,000 a year with it, yet these companies spend on average less than 0.1% of their total expenditure on Anthropic. Even the top 10 heaviest users account for less than 1%.
From this perspective, AI revenue may have extraordinarily terrifying room for growth in the future.
By the way, the sales and marketing spending that couldn't buy 10 seconds of Super Bowl ad time is $3.168 billion in the first half of this year.
In February, it actually ran an ad during the Super Bowl, with the closing line "Ads are coming to AI. But not to Claude," a direct dig at OpenAI putting ads in ChatGPT.
Altman posted a long tweet in response, saying the ads were pretty funny and he laughed too, but he couldn't understand why Anthropic would do something so "clearly dishonest."
3. A loss of $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, so the loss was roughly five times revenue.
I know many people are wondering: didn't operating profit turn positive in Q2? If it's making money operationally but losing tens of billions net, where did the money go?
The answer is in a line item in the lower half of the income statement called "change in fair value of convertible notes, derivative liabilities, and SAFEs," which was negative $72.4 billion in the first half of this year.
This might be a bit convoluted, so let me explain it in plain terms.
A large part of the money Amazon invested in Anthropic was not given as direct stock purchases but as convertible notes. You can think of a convertible note as an IOU, stating that it can later be converted into stock at an agreed price. As of the end of June this year, the principal on these IOUs was still $6.2 billion.
The issue is that Anthropic's valuation went from $380 billion in February this year to $965 billion in May.
The more the valuation rises, the more valuable these IOUs become, since they can be converted into stock at old prices. Accounting standards require that Anthropic record as a loss on its books however much the IOUs have gained in value.
So in Q2 alone, this item was recorded at $59.6 billion. The prospectus itself explains that this quarter's figure is the largest mainly because the Series H financing pushed up the valuation; another $8.4 billion relates to several convertible note purchase agreements signed this year.
So this kind of loss doesn't involve any actual cash outlay. The prospectus also provides an adjusted measure that excludes these fair value changes, stock-based compensation, and similar items. On that basis, adjusted net income for the first half of this year was a positive $2.671 billion.
4. A novelist's name
In Anthropic's 2025 income statement, there's a rarely seen line item.
It's called "Bartz settlement," for $1.5 billion.
Bartz refers to American thriller novelist Andrea Bartz, author of bestsellers like *We Were Never Here*. In August 2024, she and two other authors led a lawsuit against Anthropic for training Claude on pirated books.
And Anthropic paid $1.5 billion...
The settlement covers over 400,000 books at roughly $3,000 each, and the court formally approved it on July 20 this year.
The most bizarre part of the case is that Judge Alsup had ruled that Anthropic buying physical books itself, scanning them into its database, constituted fair use and was legal. The millions downloaded from pirate sites were not.
Later, the court unsealed over 4,000 pages of case documents, and the *Washington Post* dug up an internal project called Project Panama from them. Starting in early 2024, Anthropic began buying millions of used books in bulk, cutting the bindings with hydraulic cutters, scanning them at high speed, and then sending them off for recycling. The internal documents said, in their words, "our effort to destructively scan all the books in the world," meaning they aimed to destructively scan every book in the world.
Buying books, cutting them up, and scanning them was legal. Directly downloading from pirate sites cost $1.5 billion in penalties.
Insane...
5. 5% to 7% of revenue comes from "policy-violating customers"
Anthropic wrote that part of its revenue has always included, and will continue to include, money from customers who were later found to have violated its usage policy or commercial terms. The reason: detecting, investigating, and banning takes time, so the money gets collected first and the account gets banned later.
It then gave a figure: for the first nine months through September 30 this year, this portion came to roughly 5% to 7% of total revenue.
What this means is, you can understand it as: banned, but no refund...
I did a rough calculation: H1 was 16.286 billion, plus Q3's 16.8 billion from media reports, so about 33 billion dollars for the first nine months. So 5% to 7% works out to 1.6 billion to 2.3 billion dollars.
Then they said that in August and September they stepped up efforts to identify and ban accounts, so many accounts were wiped out at that time (my own old account was among that batch). This will lower the share of their violation-related revenue, and may also drag down future revenue and growth.
6. Making peace with Musk
This one has to be told in chronological order, or you won't appreciate how surreal it is.
In January this year, Anthropic cut off xAI employees' access to Claude through Cursor. An xAI co-founder wrote in an internal email that efficiency would suffer, but that this would force us to build our own coding product.
On February 12, Anthropic announced it had completed a 30 billion dollar funding round, and Musk replied beneath it: "Your AI hates Whites & Asians, especially Chinese, heterosexuals and men. This is misanthropic and evil."
Roughly: you are evil, you stop at nothing...
Three months later, on May 6, Anthropic announced it was renting xAI's entire Colossus 1 data center in Memphis, with over 220,000 Nvidia GPUs.
Musk's tone changed immediately, saying the Anthropic people he met were all extremely capable and very much cared about doing the right thing.
In June, SpaceX went public. According to Reuters' reading of the SpaceX prospectus, Anthropic was then paying about 1.25 billion dollars per month.
But in Anthropic's own prospectus, that figure has changed.
From May to October, Anthropic signed, and then revised several times, contracts with a company called SpaceX.AI LLC. The parentheses note that this company was formerly known as X.AI LLC.
That's the xAI that makes Grok, now a wholly owned subsidiary of SpaceX.
The total contract value is up to 120.5 billion dollars, running through 2029. Working it out, once all the compute is delivered in December, Anthropic will be paying this company roughly 3.1 billion dollars per month.
SpaceX's prospectus says its entire AI business had revenue of 818 million dollars in the first quarter of this year.
Anthropic's compute bill for one month is roughly four times the full quarterly revenue of Musk's AI business.
Whatever the insults, business is business...
7. Signing bills running to 2045
Musk's 120.5 billion dollars isn't even the most striking part. What really opened my eyes to what "heavy asset" means are the others.
Let me list the major ones.
Google: at least 114.5 billion dollars, from April 2026 to July 2033. There's also an agreement for additional TPU compute purchases in 2028 to 2031, which Anthropic itself estimates will cost another 125 to 150 billion dollars. That portion isn't even included in the contractual commitments table.
Amazon: at least 110 billion dollars, from May 2026 to April 2036.
Microsoft: at least 44.9 billion dollars, from November 2026 to May 2033.
SpaceX's AI subsidiary: up to 120.5 billion dollars.
AMD: starting next year, procurement of Helios full racks, expected to exceed 20 billion dollars.
Broadcom: a five-year AI accelerator equipment lease, with total rent of about 161.2 billion dollars, non-cancellable.
The Google, Amazon, and Microsoft contracts all contain the same phrase, "regardless of actual usage," meaning the money must be paid whether or not the capacity is used. Google and Amazon also specify that if usage falls short, the shortfall must be made up.
This kind of clause has a specific name: take-or-pay. It was most commonly found in long-term contracts in energy industries like natural gas and electricity.
Looking at it this way, compute in Anthropic's mind is essentially already on par with energy.
8. Spent 7 dollars and gained control of the company
There's an interesting contrast here.
When SBF invested in Anthropic years ago, the 500 million dollars bought a pile of shares with no voting rights. And it's not just him: big shareholders like Amazon and Google, who invested billions upon billions, received Class N shares upon listing, which likewise carry no voting rights.
So who actually holds the votes at this company?
The answer is hidden in an agreement for a limited liability company attached as an appendix.
Article 10 reads as follows:
Each person on the list contributed 1 dollar to the company. The company used this money to subscribe for 1 share of Class F stock in Anthropic.
There are 7 people on the list: Dario, Daniela, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Chris Olah, who are Anthropic's seven co-founders.
One dollar each, seven dollars in total.
This one share of Class F stock carries no economic rights and entitles its holder to none of the profits. But its voting power is fixed at exactly 50.1% of the company's total votes, covering all matters put to a shareholder vote, including the election of directors outside the trust seats. No matter how new shares are issued in the future, it cannot be diluted.
The agreement specifically includes a sentence saying this 1 dollar is an irrevocable contribution, and the seven people waive any right to get this money back.
Seven dollars, in exchange for more than half the voting power of a company that could be worth over a trillion dollars.
The prospectus's description of these seven people is also telling: "guardians of our mission."
A\ At the foundational level, it's still that A\.
9. For every 50% rise in the share price, 2% goes to charity
Before the IPO, Anthropic had already transferred 840 million Class N shares into a wholly owned subsidiary called the Anthropic Public Benefit Fund, representing 5% of the fully diluted share capital. Buddy Shah, chair of the trust, will also be one of the fund's first managers.
At the price from the May Series H round, this stake is worth roughly 50 billion dollars.
Then, the board approved a mechanism: starting from the IPO offering price, each sustained 50% rise in the share price allows up to another 2% of shares to be put into this fund.
The more the stock rises, the more gets released.
In 2020, several researchers at the Future of Humanity Institute at Oxford wrote a paper called "Windfall Clause," envisioning AI companies committing in advance that, if they earn extraordinary profits, they will hand a large share of them to society.
Six years later, it has appeared in a more concrete form in an IPO prospectus.
A footnote in the founders' letter states that each of the 7 co-founders has pledged to donate 80% of their personal shares, which together amount to 9% of the fully diluted share capital. Then, what employees pledge to donate through the matching donation program is 8%, plus the 5% from the public benefit fund.
The total comes to roughly 22%.
Based on a rough calculation at the Series H price, this 22% stake is worth over 200 billion dollars.
10. Dario's resume doesn't mention Baidu
On page 255 of the prospectus, Dario's resume reads as follows.
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, ultimately rising to VP of Research.
It doesn't mention Baidu, not once...
You see, Dario himself has said that he entered the AI field at the end of 2014 at Baidu, working under Andrew Ng on speech recognition, and it was there that he first noticed that giving models more data reliably produced better results.
This insight later came to be known as Scaling Laws.
Surprising, right? Scaling Laws, which weren't proposed until 2020, had actually first sparked in Dario's mind at Baidu.
Interestingly, though, the prospectus makes no mention whatsoever of Dario's time working at Baidu.
11. A 900% net revenue retention rate
Finally, a few numbers that might keep friends in the SaaS business up at night...
First, the net revenue retention rate is roughly 900%.
What this means is that customers who spent 1 dollar at Anthropic a year ago (excluding subscription users) spent 9 dollars in the same quarter this year.
Anyone in enterprise software knows that 120% is considered excellent for this metric. Snowflake's peak back in the day was 178%, which has been told as a legend for years.
But with AI advancing so rapidly, after all these years in product work, this is genuinely the first time I've seen numbers this outrageous.
Second, revenue per employee. As of September 30 this year, the revenue corresponding to each employee over the past 12 months exceeds 6.6 million dollars, versus 1.4 million a year earlier. Headcount rose from about 940 at the end of 2024 to about 5,300.
Third, what I think best illustrates how competitive this industry is.
The prospectus says that in the third quarter of this year, over 80% of Anthropic's revenue came from model versions released in just the past six months.
Also, OpenRouter's data broke the day before yesterday, showing a dramatic surge in consumption of Claude Opus 5.5.
There's no way around it: the models are just that strong.
Fourth, customers. Customers spending over 100,000 dollars a year numbered roughly 20 at the end of 2023, and by the end of September this year exceeded 11,900. In addition, there are roughly 40 million customers spending less than 100,000 dollars.
Very interesting.
Writing at the end
Since I myself worked in public fund management for several years, I've read quite a few prospectuses and financial reports back then.
But frankly, I've never seen a prospectus as sweeping and astronomically large as Anthropic's.
AI is accelerating everything, both our lives and the capital markets.
What we are witnessing is already a comprehensive restructuring involving capital, energy, software, hardware, and social institutions.
OpenAI is also about to go public in 2027, and by then, it will be another behemoth.
By then, we'll finally have the opportunity to directly compare OpenAI and Anthropic's actual revenue, compute costs, gross margins, cash flow, and the profitability of their model businesses.
Just thinking about it, it's a very interesting thing.
That's all. Since you've read this far, if you think it's good, feel free to like, comment, and repost, which would be a huge help to us~
Thank you for reading my article. See you next time.