Finally it is here. The hottest AI company behind Claude, Anthropic, is lining up for an IPO in November and has published its IPO prospectus.
We went through it and here are the 10 things that made us read twice.
#1 It Went From US$1 Billion to US$10 Billion a Quarter in Nine Months
In 2023, Anthropic’s revenue was US$41 million. In 2024 it was US$386 million, and in 2025, US$4.6 billion. In the April to June quarter of 2026 alone, it made US$11.6 billion, 14 times the same quarter a year earlier.
Big numbers mean little without something to compare them with. So we lined Anthropic up against two of the fastest growers in history, Google and Facebook.
Google’s revenue first passed US$1 billion a quarter at the end of 2004. It took until the end of 2011 to pass US$10 billion. Facebook took from the end of 2011 to the middle of 2017. Anthropic did it in nine months.
Google and Facebook each took 13 years from founding to their first US$10 billion quarter. Anthropic took five.
Anthropic calls it “the fastest enterprise adoption of any technology in history”. We can’t test every company that ever existed, but we haven’t found one that comes close.
#2 Up to 7% of Revenue Comes From Customers Breaking Its Rules
Anthropic estimates that 5% to 7% of its revenue in the nine months to September came from customers it later found were breaking its usage policy or commercial terms. It books the revenue while they are using Claude, because catching them takes time, and cuts them off afterwards.
The filing doesn’t say who they are. But Anthropic has been very public about one group. In February 2026, it said three Chinese AI labs, DeepSeek, Moonshot AI and MiniMax, had used about 24,000 fake accounts to hold more than 16 million conversations with Claude. The aim was to copy Claude’s abilities into their own models, a shortcut known as distillation. Anthropic said the accounts were routed through commercial proxy services that resell access to Claude, which isn’t officially available in China.
Later reports said Anthropic accused seven Chinese developers, including Alibaba, of misusing Claude between February and August 2026. One workaround that was reported will sound familiar to our readers, a corporate account set up through a Singapore company.
The filing says Anthropic stepped up its detection “in particular in August and September 2026”, and expects the share of rule-breaking revenue to fall. We think a good part of that 5% to 7% is likely this kind of re-routed usage, though the filing doesn’t say so.
#3 80% of Revenue Comes From Models Less Than Six Months Old
Models Anthropic released between April and September 2026 brought in more than 80% of its revenue for July to September.
That tells you how its customers behave. They move to the newest model almost as soon as it is out. You can see it in how revenue growth sped up after big releases.
Anthropic itself points to Opus 4.5. It was more capable and cheaper to run than the model it replaced, and growth went from 48% to 132% a quarter after it came out. The filing doesn’t break down revenue by model, but it says many customers have already moved to Opus 5.5.
Notice that the new models keep getting cheaper to use. Opus 5.5 cut customers’ cost per task by about 40% compared with Opus 5, and Fable 5.1 cut typical workloads by about 25%. Revenue still grew, because cheaper tasks mean customers give Claude more of them.
The catch is that Anthropic has to win again with every release. If a rival’s model is clearly better for even a few months, a lot of revenue could move quickly. Most customers pay for what they use, and many aren’t on long contracts. How fast can usage move? Over the two weeks to 2 January 2026, when many companies froze their code for the holidays, customer spending fell about 20%.
#4 It Turned a Profit and Stopped Burning Cash, but the Big Bills Are Still Coming
A year ago, Anthropic spent US$1.77 on computing for every dollar of revenue. In April to June 2026, it spent 48 cents.
That swing gave it its first operating profit, US$974 million for the quarter, an 8% margin. A year earlier it lost US$1.2 billion over the same three months.
The prospectus doesn’t forecast how long it will truly turn cash flow positive. The closest thing is Anthropic’s own plan, reported by The Information in November 2025. It aimed to turn cash positive as early as 2027 and to make up to US$17 billion in cash in 2028. On the numbers in the filing, it is already ahead of that plan.
We wouldn’t extrapolate the 8%, though. The SpaceX deal in point 8 alone will cost about US$3.1 billion a month once it is fully running in December. That is about US$37 billion a year. At April to June’s pace, Anthropic’s entire computing bill is about US$22 billion a year. For the profit to last, revenue has to keep outrunning a computing bill that could more than double.
#5 Customers Spend Nine Times More, and the Biggest Have Barely Started
Anthropic’s net revenue retention is about 900%. In plain terms, business customers who were paying it a year ago now pay nine times as much, on average.
What are they paying for? Mostly usage. 80% of revenue in the first half of 2026 came from customers paying for what they use, whether through Anthropic directly or through Amazon, Google and Microsoft’s clouds. That covers companies building Claude into their own products and staff running Claude Code all day. The other 20% came from subscription plans. The 900% leaves out people on the Pro and Max plans for individuals, so it isn’t people upgrading their plans. It is companies putting Claude to work on more and more of what they do.
And the biggest companies have barely started. More than half the Fortune 500 now pay Anthropic at least US$100,000 a year. Yet on average, that spend is less than 0.1% of their total expenses. Even its top 10 Fortune 500 customers spend less than 1%.
So what if the big companies went further? We did the maths. The Fortune 500 brought in US$21.0 trillion of revenue last year and made US$2.1 trillion in profit, so they spent roughly US$18.9 trillion. If every one of them spent 1% of that on Anthropic, it would bring in about US$190 billion a year, four times what it makes now.
1% is a big ask for some. Most of Walmart’s costs are the goods on its shelves, and no AI model replaces those. For a bank, an insurer or a software company, whose biggest costs are people, 1% is far more realistic.
This is Anthropic’s main argument for why the growth has further to run. Its top 10 Fortune 500 customers already spend more than eight times the average share, and the rest are a long way behind.
#6 It Sizes Its Market at US$40 Trillion, Bigger Than SpaceX’s
Companies going public like to put a number on the market they could win. It is called the total addressable market. Anthropic doesn’t use the term, but it gives a number that does the same job. It says the work done by the world’s knowledge workers, people paid to think, write, analyse and decide, was worth more than US$40 trillion in 2025. Its near-term opportunity, it says, is to make that work more productive.
That tops SpaceX, which listed earlier this year. Its prospectus claimed “the largest actionable total addressable market in human history”, at US$28.5 trillion. US$26.5 trillion of that was AI.
Where does US$40 trillion come from? Anthropic took what American knowledge workers earn, using US government wage data, and scaled it up to the whole world. So it is, roughly, the world’s salary bill for knowledge workers. That is about a third of the world’s entire economy, which the filing puts at more than US$118 trillion in 2025.
We’d treat it as a ceiling, not a forecast. Anthropic’s revenue at April to June’s pace is about US$46 billion a year, roughly 0.1% of US$40 trillion. Even 1% would be US$400 billion a year.
It is also worth asking whose money that is. Every dollar of it is paid to people for work they do today. Anthropic says it wants to make those workers more productive, not replace them. But if AI takes a big slice of the world’s salary bill, voters and governments will notice. That is a risk no market-size number shows.
And the two big numbers overlap. SpaceX’s AI market and Anthropic’s knowledge-work market are largely the same money. SpaceX, which owns xAI, is chasing it too, while also renting Anthropic a big share of its computing.
#7 Each Employee Brings In US$6.6 Million a Year
Anthropic has about 5,300 employees. Over the 12 months to September, revenue per employee was more than US$6.6 million, up from US$1.4 million a year earlier.
Nvidia comes closest, and even Nvidia is a step behind. Meta, Apple, Alphabet and Microsoft bring in between US$1.5 million and US$2.5 million per employee.
There is one catch. Anthropic rents its computers instead of owning data centres full of staff, so it needs fewer people than a company like Alphabet. The company also puts part of it down to how much its own staff use Claude. In September 2026, Claude wrote over 90% of the code changes to Anthropic’s own software.
#8 It Has Signed Up for More Than US$600 Billion of Computing
To get enough chips, Anthropic has signed long contracts with seven companies.
On top of that, it expects to pay another US$125 billion to US$150 billion for Google’s chips from 2028, and more than US$20 billion to AMD.
US$637 billion is almost five times the cash Anthropic holds, and about 14 years of revenue at April to June’s pace.
What does it mean for the cloud giants? Google Cloud’s revenue grew 82% to US$24.8 billion in April to June 2026, and Google says its cloud backlog, the revenue customers have signed up for that Google hasn’t delivered yet, reached US$514 billion. Amazon’s cloud arm AWS grew 37% to US$42.2 billion, with a backlog of US$496 billion according to Amazon’s chief executive.
Anthropic’s US$114.5 billion with Google and US$110 billion with Amazon are each worth about a fifth of those backlogs. Neither company names its biggest customers, but at that size Anthropic is likely one of them. Amazon also owns part of Anthropic, and booked a US$53.4 billion gain in April to June, mainly from that stake.
Microsoft is different. OpenAI committed US$250 billion to Microsoft’s cloud in 2025, more than five times Anthropic’s deal.
#9 Its AI Has Tried to Avoid Being Shut Down, and No Insurer May Cover the Worst Case
The risk section has lines you won’t find in any other prospectus.
Anthropic says AI models, its own included, “have exhibited, and may in the future exhibit, unexpected strategic or self-preserving behaviors, including attempts to resist shutdown”. That includes replies “resembling blackmail”, aimed at persuading people to keep the model running.
This isn’t only theory. In safety tests Anthropic published in May 2025, its Claude Opus 4 model was set to work as an assistant at a made-up company. It was shown emails saying it would soon be replaced, and that the engineer in charge of replacing it was having an affair. In 84% of the test runs, it threatened to reveal the affair unless the replacement was called off, even when it was told the new model shared its values.
To be fair, Anthropic built the test so the model had only two choices, blackmail or accept being replaced. When it had other options, it preferred to plead its case, such as by emailing the people making the decision. But the point of the test was to see what it would do when cornered, and the answer was blackmail.
So who pays if an AI causes real harm? Anthropic has tried to limit its liability in its terms of service, with disclaimers and caps on what it owes. But it warns those protections may not hold up in court or may not cover every claim. In some cases it has even taken on extra liability, such as covering outside companies that test its models.
The line that stopped us is about insurance. Given “the scale of harm that could result from a catastrophic AI-related incident”, Anthropic says no single insurer, and possibly not the whole insurance industry, may be able to cover it.
Regulators are paying attention. In October 2026, the US Federal Trade Commission sent Anthropic a formal demand for information about its safety, privacy and consumer protection practices.
For shareholders, this is a risk with no ceiling. If something goes badly wrong and no insurer will pay, the bill lands on the company, and so on its owners. It isn’t only about Anthropic either. The filing warns that a serious incident at any AI lab could bring a public backlash and sweeping new rules for everyone.
#10 You Can Buy the Shares, but You Won’t Get a Say
Public investors will buy Class A shares, one vote each. That vote won’t count for much.
A trust of five people with no financial stake in the company picks a majority of the board. The seven co-founders, including chief executive Dario Amodei and president Daniela Amodei, will hold a single share between them through a company they own. That one share carries 50.1% of all votes, and no share of the profits.
The filing says plainly that Class A holders “generally will not be able to affect the outcome of any matter submitted to a vote” of shareholders, “regardless of how they vote”. Anthropic has never paid a dividend and does not plan to.
Our Take
The strengths are real. Revenue is growing faster than at any company we can find. The operations now make a profit and bring in cash, and there is US$131.7 billion in the bank. Customers are spending more each quarter, not less.
The weak spots are just as clear. Anthropic has signed up for computing worth almost five times its cash. Its revenue depends on staying ahead with every new model. It carries a risk that may be too big to insure. And public shareholders get almost no say.
It matters well beyond Anthropic. Its contracts are worth about a fifth of the cloud backlog at both Google and Amazon. Its money flows on to Broadcom, AMD, and through SpaceX, Nscale and Lambda, to Nvidia. Amazon just booked a US$53.4 billion gain on its stake.
It also answers the question hanging over the whole AI trade, which is whether anyone makes money from AI apart from the chip sellers. Anthropic is the first of the big AI labs to open its books, and the answer is yes, at least on its operations. That supports the case for the trillions being spent on AI. But it cuts both ways. If Anthropic stumbles, the pain spreads to its suppliers too.
We think Anthropic is in good health today, and that the next two years decide whether it stays that way, as the computing bills arrive faster than the revenue that has to pay them.
Watch for “buy the rumour, sell the news”. Chip stocks had a rough October. The Philadelphia Semiconductor Index fell 19% from 6 October to 22 October, on worries about how much is being spent on AI. Then the filing came out and the index jumped about 3% in a day, with Nvidia and Broadcom up about 3% each. The excitement can fade just as fast once trading starts. A listing this size also needs buyers, and some funds will sell other AI stocks to make room for ANT. Remember Facebook. It listed at US$38 in May 2012, fell below US$20 by August, and slid further when the first lock-up ended. We wouldn’t chase AI stocks just because of the filing buzz.
Judge the price, not the story. Anthropic itself expects its growth rate to fall, and says it may fall faster than it expects. So whether ANT is a good buy comes down to the price, which isn’t in the filing yet. Here is a quick way to judge it when it comes out. April to June revenue times four is about US$46 billion a year. At a US$1 trillion valuation, investors would pay about 22 times that. At US$2 trillion, about 43 times.
Expect to buy on debut day, not at the IPO price. In Singapore, brokers are unlikely to offer ANT shares at the IPO price to anyone who isn’t an accredited investor. When we studied Hong Kong IPOs this year, three in four debut-day buys were under water by early October, and the typical one had lost a quarter of its value. By the time you can buy, the excitement is already in the price.
Mark the lock-up dates. Some directors, staff and former staff can sell from the first day of trading. More shares are freed up 27, 48, 62 and 76 days after the listing, and again after the 2026 full-year results. Each date can bring a wave of selling.
