Anthropic, the company that develops the Claude artificial intelligence models, filed for an initial public offering with the US Securities and Exchange Commission (SEC) on 26 October. The filing is called an S-1. In it, a company discloses audited financial statements, risks and its governance before it sells shares to the public. The shares are to be listed on Nasdaq under the ticker ANT. The filing does not yet include the share price or the number of shares to be sold.
The filing runs to more than 400 pages. This overview gathers the most important parts in numbers and charts: where Anthropic stands today and what the filing says about where artificial intelligence is heading. Data for 2023 to 2025 is audited; this year’s is not, which is normal for an S-1. Quarters are marked with Q: Q1 is January to March, Q2 April to June, and so on.
A number in square brackets, such as [1], shows a chart when you rest the mouse on it; a click or tap opens a larger one. All charts are also collected at the end of the article.
In brief
| What | Number |
|---|---|
| Revenue in Q2 2026 | $11.6bn, 14.7 times more than in Q2 2025 |
| Operating result in Q2 2026 | positive for the first time, 8% margin |
| Operating cash flow in Q1 and Q2 2026 | $3.9bn |
| Net loss in Q1 and Q2 2026 | $80.7bn, almost all on paper |
| Customers paying at least $100,000 a year | more than 11,900 (a year earlier more than 950) |
| Cash and short-term investments, 30 Jun 2026 | $131.7bn |
| Non-cancelable compute commitments, 30 Jun 2026 | about $265.6bn |
| Employees | about 5,300 |
| Co-founders’ share of shareholder votes | 50.1% |
Revenue is growing faster than the cost of compute
In 2023 Anthropic had $41 million of revenue, in 2024 $386 million. In 2025 it had $4.6 billion. This year growth accelerated further [1]. The filing compares it with the previous quarter. Revenue growth on the previous quarter was:
- Q4 2025: 48%,
- Q1 2026: 132%,
- Q2 2026: 144%.
The company attributes the acceleration to new models. The growth resembles a multi-stage rocket: each new model is a new stage that ignites before the previous one burns out. Models released in the six months to the end of September 2026 brought in more than 80% of revenue in Q3 2026.
Compute means the chips, servers and power a model needs to be trained and to answer customers. Anthropic secures it through multi-year contracts, 12 to 18 months before it receives it. The cost therefore rises when the capacity arrives, not when the customers do. In Q1 2025 compute cost was 251% of revenue, that is two and a half times revenue. In Q2 2026 it was only 48% of revenue.
Why did this turn around so quickly? The cost of compute for each quarter is mostly fixed in advance, by contract. Revenue, however, grew much faster: in Q2 2026 it was 14.7 times higher than in Q2 2025, while the cost of compute was only 4 times higher. As a result, the operating result swung from a $1.2 billion loss in Q2 2025 to a $0.97 billion profit in Q2 2026. When costs grow more slowly than revenue, profit improves faster than revenue. This is called operating leverage.
| Quarter | Operating margin | Adjusted margin |
|---|---|---|
| Q1 2025 | −249% | −221% |
| Q2 2025 | −152% | −129% |
| Q3 2025 | −224% | −93% |
| Q4 2025 | −139% | −84% |
| Q1 2026 | −35% | −13% |
| Q2 2026 | 8% | 19% |
The margin shows how much of every $100 of revenue is left as operating profit. A negative margin means a loss: −249% means the loss was two and a half times revenue. The adjusted margin is the company’s own measure. It leaves out employee pay in shares, donations in shares and a one-off copyright settlement (the Bartz case, $1.5 billion in Q3 2025). Because it leaves out these costs, it is higher than the ordinary margin in every quarter: the loss is smaller or the profit larger [2]. In Q2 2026 the ordinary margin was 8% and the adjusted margin 19%.
Pay in shares is a real cost even though no cash leaves the company. New shares reduce the stake of existing owners (dilution). In Q2 2026 it amounted to $1.26 billion [3].
The company warns that the margin will fluctuate in future. The reason is how it buys compute. It starts booking the cost of new compute when it receives it, often in large blocks at once. Revenue from that compute grows gradually, as customers start to use it. In a quarter when a large new block arrives, costs therefore jump before revenue and the margin falls. Once customers fill the new capacity, the margin rises again. It is similar to a new factory: the machines cost money from day one, but output is small at first.
Who pays, and how
The money comes mainly from businesses. At the end of Q3 2026 there were more than 12 times as many customers paying at least $100,000 a year as a year earlier [4]. In Q1 and Q2 2026 they brought in 74% of revenue. They include all ten of the largest US companies on the Fortune list. Existing customers keep spending more: net revenue retention (NRR), which shows how much the same group of customers pays today compared with a year ago, is about 900% according to the company.
- Pay for usage. A model cuts text into tokens, small units it reads and writes, and the customer pays for the tokens used. In Q1 and Q2 2026, 80% of revenue came from usage and 20% from subscriptions.
- Through the clouds. 48% of sales go through the cloud marketplaces of Amazon, Google and Microsoft.
- Large customers. More than 1,900 customers pay at least $1 million a year, and about 250 at least $10 million. The five largest bring in about 25% of revenue, none more than 10%.
- Policy violators. The company estimates that 5% to 7% of revenue from Q1 to Q3 2026 came from customers who violate its usage policy. This revenue will decline.
- Holidays. In the two weeks to 2 January 2026, usage fell by about 20% compared with the two weeks before.
Why an $80 billion loss if the business makes money
Most of the loss comes from convertible notes [5]: a loan that converts into shares at the IPO. Anthropic issued $6.2 billion of them to Amazon. They are carried in the books at the value of the shares they will convert into, so their value rose to $108.6 billion by 30 June 2026. The company booked the increase as an expense, but no cash left the company. It is like promising a friend part of your house: when the house rises in value, your obligation in the books grows, even though you have paid nothing.
At the IPO the notes convert into shares and the effect largely disappears. Real cash is positive: operating cash flow in Q1 and Q2 2026 was $3.9 billion [6]. From 2023 to 2025 it was negative. A warning for the future: employee pay in shares that the company has not yet booked amounts to $26.6 billion. A large part will be recognised in the quarter of the IPO.
Cash and compute
Since it was founded, Anthropic has raised $141.1 billion of capital, most of it this year:
| Funding round (2026) | Raised | Price per share |
|---|---|---|
| Series G (Q1) | $30.2bn | $25.91 |
| Series G-2, Google and Amazon (Q2) | $15.0bn | $25.91 |
| Series H (Q2) | $65.2bn | $58.90 |
The prices are adjusted for the 10-for-1 stock split of September 2026. On 30 June 2026 the company had $131.7 billion of cash and short-term investments. Its commitments are even larger.
Non-cancelable commitments are contracts under which Anthropic, as the buyer, has committed to pay suppliers for compute and cannot cancel them [7]. At the end of 2025 they stood at $54.6 billion, and on 30 June 2026 at about $265.6 billion. That is what the company would earn in almost six years at its Q2 2026 pace. After 30 June Anthropic, again as the buyer, signed further contracts. For long-term ones (more than one year and up to 20 years) it committed to pay at least about $245 billion. For short-term contracts it committed to about $21 billion more of non-cancelable payments.
With Google and Amazon the commitment is a minimum amount of spending. It must pay Google at least about $114.5 billion and Amazon at least $110 billion. If it uses less compute than agreed, it still has to pay the difference up to that amount.
More and more compute will be leased directly, in data centres operated by others on its behalf (Anthropic Direct). It secures the space early and picks the expensive chips later. It is similar to an airline that orders aircraft years in advance: if the passengers come, it earns a lot; if they don’t, it still has to pay for the aircraft. Google is at once a supplier, an investor and a competitor; what this means for Google is covered in our analysis of Alphabet.
Who is in control
A buyer of shares on the stock market will get a share of the profits, but not control of the company.
- Public benefit corporation (PBC). The board must balance the interests of shareholders, of those affected by the company, and the public benefit: the responsible development of advanced AI.
- Co-founders. The seven co-founders will hold 50.1% of the votes at shareholder meetings through a special entity.
- Independent trust. The Long-Term Benefit Trust, with five trustees who have no financial stake, elects a majority of the board.
- Public benefit fund. It received 846.8 million shares, 5% of the total, worth about $50 billion at the Series H price.
- Lock-up. Larger owners may sell after the Q3 2027 results are published, the others after the Q1 2027 results (or earlier, after 365 or 180 days).
Main risks
| Risk | What the filing says |
|---|---|
| Compute commitments | Mostly non-cancelable; they must be paid regardless of revenue. |
| Competitors | OpenAI, Google and open models; more than 80% of revenue comes from models less than six months old. |
| US government | The Department of War designated it a supply-chain risk in February; an appeals court upheld the designation 2 to 1 in September, and proceedings continue. |
| Export restrictions | In June 2026, restrictions on the Fable 5 and Mythos 5 models, which were switched off for all customers until they were lifted. |
| Dependence on clouds | Almost half of sales go through cloud providers that have their own models. |
| Copyright | Bartz settlement of $1.5bn; music publishers and Reddit are also suing. |
| Dilution | $26.6bn of unbooked share-based pay, extra shares for the fund. |
What it could be worth
There is no IPO price yet. The Series H price of $58.90 multiplied by about 16.9 billion total shares gives about $1,000 billion. Media report that the company is aiming for $1,500 to $2,000 billion; this is not company data. For annual revenue we take revenue in Q2 2026 multiplied by four ($46.2 billion).
| Company value | Price per share | Multiple of annual revenue |
|---|---|---|
| $1,000bn | $59.2 | 21.6 times |
| $1,500bn | $88.8 | 32.5 times |
| $2,000bn | $118.3 | 43.3 times |
A multiple does not say how fast revenue has to grow. A reverse DCF shows that: it takes the company’s value and calculates what future it already includes. The required return here is 10% a year, roughly what the US S&P 500 index has returned over the long term, including dividends.
At $1,000 billion, Anthropic would need about $473 billion of annual revenue in ten years, in 2036. That is roughly what Alphabet, Google’s parent company, has today. Revenue growth would have to be about 26% a year for ten years in a row. At $2,000 billion it would need about $964 billion of revenue, twice what Alphabet has today. Growth would have to be about 36% a year. The calculation assumes the operating margin rises to 35% over ten years.
What the filing says about the future of AI
The filing is about one company, but it is the first audited look at the finances of a developer of the most advanced models. It shows seven directions that apply to the whole industry.
- From chat to digital colleagues. The money no longer comes only from answering questions, but from agents that do the work themselves. Claude Code plans, writes, tests and fixes software code. Claude Cowork does the same with office work: it reads files and works in applications. Claude Enterprise users without Claude Code spent less than $75 a month in Q3 2026; the most active 5% with Claude Code spent about 20 times that.
- Tasks keep getting cheaper, usage keeps growing. In September 2026 Opus 5.5 cut the cost per task by about 40% compared with Opus 5, and Fable 5.1 by about 25% compared with Fable 5. When Opus 4.5 brought a similar price cut, revenue growth on the previous quarter jumped from 48% in Q4 2025 to 132% in Q1 2026.
- Businesses are only at the start. Customers paying at least $100,000 make up more than half of the Fortune 500. At these companies, Anthropic receives on average less than 0.1% of their total expenses.
- The bottleneck is physical. According to the company, demand for compute has long exceeded the capacity available to it. The constraints are chips and memory made by few manufacturers in few places, power and data centres. Hence commitments of hundreds of billions.
- AI helps develop AI. More than 90% of contributions to the company’s code in September 2026 were written by Claude. The company says its models are speeding up its research, and lists as a risk that progress could outpace its ability to check safety.
- The most capable models will not be available to everyone. In April 2026, the Mythos model was first given only to a few selected organisations for cybersecurity, because it can find vulnerabilities in software. The company expects similar access limits for future models, and the state adds its own through export restrictions.
- The lead is short, and the giants are intertwined. Customers move quickly to a newer model. Some cloud providers are at once Anthropic’s compute suppliers, a sales channel, customers, investors and competitors.
All seven directions lead to the same question: will demand for AI grow at least as fast as the compute commitments? If it does, the commitments are a source of advantage. If it doesn’t, they become a burden that still has to be paid.
What to watch
- Price range and number of shares in the amendment to the filing (S-1/A).
- Revenue in Q3 2026, which the filing does not yet include.
- The margin when new compute arrives, for example about $3.1 billion a month at SpaceXAI from December 2026.
- The share of policy violators and the number of customers above $100,000.
- The dispute with the US government and new export restrictions.
- The cost of share-based pay in the quarter of the IPO and the end of the lock-up in 2027.
Key takeaways
- Anthropic filed an S-1 on 26 Oct 2026 to list on Nasdaq (ticker ANT); it does not yet include the share price or the number of shares.
- Revenue: $41m in 2023, $386m in 2024, $4.6bn in 2025; $11.6bn in Q2 2026, 14.7 times more than in Q2 2025.
- Revenue growth on the previous quarter: 48% in Q4 2025, 132% in Q1 2026, 144% in Q2 2026; the company attributes it to new models.
- The cost of compute fell from 251% of revenue in Q1 2025 to 48% in Q2 2026; the operating result turned positive for the first time in Q2 2026 (margin 8%, adjusted 19%).
- The adjusted margin leaves out pay and donations in shares and the Bartz settlement, so it is always higher than the ordinary margin.
- The margin will fluctuate: when a large block of new compute arrives, the cost comes before the revenue.
- Net loss in Q1 and Q2 2026: $80.7bn, almost all on paper because of the revaluation of convertible notes; operating cash flow was +$3.9bn.
- More than 11,900 customers pay at least $100,000 a year (a year earlier more than 950); in Q1 and Q2 2026 they brought in 74% of revenue.
- $141.1bn of capital raised since founding; $131.7bn of cash and short-term investments on 30 Jun 2026.
- Non-cancelable compute commitments: about $265.6bn (30 Jun 2026), followed by about $245bn long-term and $21bn short-term; with Google and Amazon it must pay the difference up to the minimum spend.
- Control: the co-founders hold 50.1% of the votes, an independent trust elects a majority of the board, and the company is a public benefit corporation.
- Valuation: $1,000bn is 21.6 times annual revenue; the reverse DCF requires about $473bn of revenue in 2036 (roughly today’s Alphabet), and at $2,000bn about $964bn.
- The main question: will demand for AI grow at least as fast as the compute commitments?
Key metric charts
The charts that the numbers in the text point to. All data is from the S-1.
Notes on the data: Data for 2026 is unaudited. Shares and per-share prices are adjusted for the 10-for-1 split (17 Sep 2026); the Series G and H prices per preferred share are $259.1364 and $589.0095. The cost of compute is the “compute and infrastructure” line of the income statement. Revenue growth on the previous quarter is calculated from the quarterly revenue in the filing. The adjusted margin is a company measure (operating result excluding share-based pay, donations in shares and the Bartz settlement). The number of customers is the count at quarter end, based on spending over the trailing 12 months. Bridge to net loss (Q1 and Q2 2026, $bn): operating result −0.69, interest income 1.11, change in fair value of convertible notes, derivative liabilities and SAFEs −72.39, other expense (mainly share purchase agreements with Google and Amazon) −8.76, taxes −0.02. Commitments after 30 Jun 2026 are minimum payments under new and amended contracts: the long-term ones have terms of more than one year and up to 20 years and start between 2026 and 2036; the short-term ones are, according to the filing, short term in nature with non-cancelable minimum payments. For Google ($114.5bn, April 2026 to July 2033) and Amazon ($110.0bn, May 2026 to April 2036) the filing states that if Anthropic spends less, it must pay the difference. The total number of shares (about 16.9bn) is derived from the fact that 846,840,960 Class N shares equal 5% of fully diluted shares on 30 Jun 2026. Annual revenue ($46.2bn) is revenue in Q2 2026 multiplied by four; the company’s Run-Rate Revenue measure is still blank in the filing. Commitments in years of revenue: 265.6 divided by $46.2bn. The commitments by partner cannot be added up: some are minimum commitments, some upper limits, and for AMD and TPU 2028 to 2031 they are company estimates; commitments after 30 Jun 2026 also include amendments to existing contracts. Reverse DCF: starting revenue $46.2bn and margin 8.4% (Q2 2026), rising evenly to 35% over ten years, required return 10%, tax 21%, ROIC 25% (an assumption, as the company has no history of profits), ten years of the solved growth, then five years of fade to 2.5%, net cash $131.7bn. The first version of this article (30 Oct 2026) solved for growth over five years; this one looks at ten. The long-term S&P 500 return is about 10% a year (1928 to 2025, including dividends, compounded year by year; Engineering Investor calculation from Damodaran’s data). Alphabet’s revenue for the last four quarters (Q4 2025 to Q3 2026) is $471.1bn. The cost reductions with Opus 5.5 and Fable 5.1 and the share of Fortune 500 companies’ expenses are as stated by the company.
Educational content, not financial advice. At the time of publication the author is not recommending the purchase or sale of any security. Data is taken from the sources above as of 30 Oct 2026 and may change.