Anthropic is asking the global market for capital on unaudited figures and on compute contracts payable even when unused, and the stock on offer will not buy control.
30 September 2026 · The 2025 figures come from a confidential registration statement described by Reuters and the Financial Times. The company declined to comment.
Anthropic is preparing a stock-market debut after the US midterm elections at a valuation of about $2tn and an offering of about $100bn. According to the confidential filing described by Reuters, recognised revenue for 2025 was $4.59bn. The listing is not a sale of that result, and a large economic stake in class A still does not obtain control. The incoming shareholder is buying into a company that Dario Amodei and the long-term benefit trust will run, priced on more than $100bn of revenue annualised from the latest month and still without an audited account. This autumn Anthropic is the only large model company coming to market: OpenAI has pushed a listing toward 2027, and SpaceX already completed one in June.
In 1602 the Amsterdam investor in the Dutch East India Company bought a share with no vote at the general meeting, together with a monopoly that paid a dividend. The difference now is that there is no monopoly here. The closer comparison is with the railways of the 1840s. Freight is growing, but the rent on the track is due even when the revenue is not enough to cover it. The price of the listing assumes that future traffic will be enough to pay for track already leased.
Three prices for a share that is not yet public
The confidential filing was made on 1 June. The Reuters and Financial Times accounts omit the offer price, the number of shares at listing, the lock-up period, the fully diluted capital table after options and convertibles are exercised, a product split of revenue, and the payment schedule to cloud providers for 2027 and 2028. They also omit whether the $100bn is new stock to fund compute, or a sale by the funds that came in during the May round. Some of the summer 2026 figures come from conversations with investors.
Table 1. Three valuations
| Valuation | $bn |
|---|---|
| Round, May 2026 | 965 |
| Secondary market, thin | ≈ 1,330 |
| Banks’ target | ≈ 2,000 |
The middle row is a thin secondary market, not a price from regular trading in completed deals. Sources: company announcement of Series H; Nasdaq Private Market and other venues cited in late September; bankers’ comments on indicated demand.
The company has warned investors not to buy its shares on some private secondary markets. Bankers who spoke to Reuters about the mandated offering put it at about $100bn on a valuation of about $2tn. If orders value the company near $1.3tn, the price merely repeats the thin secondary market: holders willing to sell are scarce. If they value it near $2tn, the new buyer pays above that market in order to obtain stock at the listing.
What the 2025 accounts show
Table 2. 2025 accounts
| Item | 2024 | 2025 |
|---|---|---|
| Recognised revenue | $386m | $4.59bn |
| of which billed tokens | - | $3.8bn |
| of which subscriptions | - | $789m |
| Operating expenses | - | $12.65bn |
| of which compute and infrastructure | ≈ $2.5bn | $7.33bn |
| Operating loss | $2.98bn | $8.06bn |
| Revaluation of financial instruments | - | ≈ $34bn |
| Net loss after the revaluation | $8.31bn | $41.97bn |
| Cash and short-term investments, 31 Dec | - | $20.28bn |
From the draft seen by Reuters. The revaluation is not a cash outflow. It records a change in the fair value of financial instruments. The public filing will have to show the fully diluted count.
The United States accounts for nearly two-thirds of sales. Two unnamed customers each hold 12 per cent. Some $2.16bn - 47 per cent - passed through Amazon and Google platforms, against about $351m of commission, or about 16 cents on the dollar, according to the same account.
Table 3. What the two 12 per cent customers do to the valuation
| If the two 12 per cent customers are | What that does to $2tn |
|---|---|
| Amazon and Google | Down. Related-party sales are stripped out before the price-to-revenue ratio is read. |
| A lender and an integrator | Slightly down. The risk is commercial and eases if the customer stays. |
| Internal or contractual load | Sharply down. The top line and the 2026 annualised figure are re-cut. |
| Outside customers on multi-year spend | Almost no effect on price. Concentration is milder than it looks. |
Recognised revenue and annualised revenue are not the same number
Table 4. Recognised revenue - the accounts
| Period | $bn |
|---|---|
| 2025, full year | 4.6 |
| First quarter 2026 | 4.7 |
| Second quarter 2026, preliminary | 11.5+ |
Table 5. Revenue annualised from the latest month. This is not an audited line.
| As of | $bn |
|---|---|
| End-2025 | ≈ 9 |
| May 2026 | 47 |
| July 2026 | 65 |
| September 2026 | 100+ |
The second table is a message to prospective investors, not an audited statement.
The Financial Times writes that the third quarter is also expected to show a positive adjusted operating profit. Against the 2025 loss that is a turn, but it is not profit under US generally accepted accounting principles.
The coding agent is growing. Customer retention is not visible.
The valuation does not rest on consumer Claude. It rests on the corporate channel and the coding agent. The known path of Claude Code revenue runs from zero in 2025 to about $1bn annualised in November and $2.5bn in February 2026. By May about $8bn is cited. Estimates toward August speak of about $15bn, which the company has not confirmed. Menlo Ventures, itself an investor, assigns Anthropic about 40 per cent of corporate spending on large models and about 54 per cent in coding. The trend is indicative, but it is not enough to ground the valuation.
The journalists who saw the draft have not published three things: what share of second-quarter revenue came from customers who also paid in the first; whether the same customer is spending more; and whether the contract is for seats, metered use or a committed spend. Two possibilities remain: either those figures are not in the document, or the publications did not extract them. The banks running the offer have not said which of the two is true. Until they do, retention is asserted, not shown.
The customer pays for tokens directly to Anthropic or through Amazon and Google. Anthropic books the full amount as revenue and remits about 16 cents of every dollar to the platform. The same customer often already pays AWS or Google Cloud for the machines on which the model runs. According to the draft, Alphabet holds about 14 per cent of the equity with no vote. Amazon is the largest strategic shareholder. Nvidia is mentioned as a possible large investor in the offering. The margin stays with Anthropic only if the customer is paying for the model itself. If the customer is paying mainly for cloud and chips, the profit goes to Amazon, Google and Nvidia.
Contracts that are due even when unused
According to the draft seen by Reuters, Anthropic has signed at least $518bn of cloud and infrastructure commitments, of which about 80 per cent is due even if unused. Cash at the end of 2025 was $20.28bn. A credit line of about $15bn provided liquidity before the listing. Once drawn, the loan is debt and is repaid before shareholders’ claims.
Table 6. Three kinds of obligation
| Item | $bn |
|---|---|
| Cash, end-2025 | 20 |
| Google, take-or-pay minimum | ≥ 111 |
| Amazon, take-or-pay minimum | 110 |
| Microsoft, take-or-pay minimum | 31 |
| Broadcom, lease | ≈ 161 |
| Akamai and option | 12 + 9 |
The table shows that against $20bn of cash at the end of 2025 stand at least $252bn of guaranteed payments to Amazon, Google and Microsoft and about $161bn of Broadcom lease. The annual burden of about $32bn is an even split of the three cloud contracts until a payment schedule appears.
If the three guaranteed payments to Amazon, Google and Microsoft are spread evenly over the lives of the contracts - which the public document does not yet show - the result is about $32bn a year, before the Broadcom lease and the Akamai contract. That is an estimate, not a schedule, and it must be replaced by the actual instalments for 2027 and 2028. At full load on the stated 3.5 gigawatts the sum enters the cost of the service. At light load the company pays for capacity it does not use. The same names to which the minimum is owed also hold equity in Anthropic, a sales channel, and a model of their own.
Control does not follow capital
Dario Amodei holds about 2 per cent of the equity. According to the draft, control runs through a founder vehicle and a single class F share with 51 per cent of the votes, and the long-term benefit trust elects a majority of the board. A class A holder cannot force a sale, a dividend or a change of mission even if he assembles a large economic stake. Anthropic is a Delaware public-benefit corporation, so the board may set the mission alongside the shareholder’s economic interest. A suit against such a decision is hard once the arrangement is written into the charter. This is not a control stock. Anyone who accepts the arrangement should demand a lower price-to-revenue ratio than at a company where capital appoints the board. The discount used here is 15 to 25 per cent. It is the author’s judgment, not a discount the market has been seen to apply.
What to look for in the public document
The target valuation is about two trillion. What matters more is this: what recognised revenue is for the last twelve months; what the guaranteed payments to Google, Amazon and Microsoft for 2027 and 2028 are relative to the margin after platform commission; and how many shares remain after options and convertibles are exercised.
For two trillion the investor is not buying the 2025 accounts
He is accepting four things at once: that the corporate channel and the coding agent will hold high revenue; that scarce compute will support the price of the token; that through 2027 there will be no second large public company of the same profile; and that recognised revenue for twelve months can reach $80-100bn at a margin after platform commission sufficient to cover the capacity contracts and training without a new issue.
Table 7. A working case for 2028
| 2028, working case | Share of corporate model spend | Margin after the platform versus the compute burden | Implication for $2tn |
|---|---|---|---|
| Upper | Holds near today’s reading (about 40 per cent on Menlo, an investor) | Covers the minima and training | The ratio to audited revenue can be defended near the May valuation |
| Mid | The share is split with OpenAI and Gemini | Covers the minima; training requires a further raise | $2tn is dear; the valuation shifts toward $1.0-1.4tn |
| Lower | The coding agent loses the development environment | Part of the minimum sits unused | The stated price does not hold even on a $1.3tn issue |
The table is a frame, not a forecast. Menlo is a shareholder. The $32bn burden is an even split, not a schedule.
Table 8. Price-to-revenue
| Denominator | Price-to-revenue | Reading |
|---|---|---|
| 2025 accounts: $4.59bn | 436 to 1 | Refuse, if that is the whole story. |
| $40-55bn recognised | 36-50 to 1 | Dear even after a discount for the lack of control. The range $1.0-1.4tn is still not a fair price on recognised revenue. |
| $100bn annualised | ≈ 20 to 1 | A case for a purchase. It holds only with an auditor and a covered compute burden. |
The May round paid about 21 times the then annualised revenue at a $965bn valuation. Repeating that multiple on a larger annualised figure without an audited account is not a new valuation.
How much of the equity will trade on day one
At a stated valuation of two trillion, if all $100bn is new stock, about 5 per cent of the equity trades on day one. If part of the offering is a sale by existing holders, the tradable stake on day one is larger and the cash that enters the company for capacity is smaller.
Table 9. What follows for the tradable stake
| Moment | What follows |
|---|---|
| New stock only | About $100bn tradable. The public funds growth. The S&P 500 does not admit the name: one year of seasoning, four quarters of profit, at least 10 per cent free float. |
| Mixed offering | More stock on day one, less cash for the minima. The May funds exit through the public. |
| Around day 15 | Nasdaq-100 can add a name in the top 40. Russell and MSCI do not wait a year. |
| Tax on employee stock | Double-trigger awards are taxed at a US listing even if sale is locked. Standard withholding often fails to cover the top rate. On day one the offering measures the tax, not belief in Claude. |
SpaceX raised about $75bn in June at a valuation of about $1.8tn. At the debut it traded above the offer price; later part of that premium faded. Anthropic wants to raise a similar or larger sum months later, with a smaller tradable stake on day one. Strong demand before listing and a fall after the open can occur in the same issue.
Anyone who already holds Nvidia, Amazon and Alphabet and then buys Anthropic adds the same exposure to the model, the chip and cloud services. The only incremental exposure is the corporate channel not already in those three names. Replacing Nvidia with Anthropic at this valuation is a bet on first-day premium, not on the 2025 accounts.
As of 30 September 2026 there is no case for taking part until a public prospectus is out. Two trillion sits well above the thin secondary market, where the valuation was recently $1.3tn, and above the May round at $965bn. The compute contracts are signed forward and are payable whether the capacity is used or not. It is still unknown whether the $100bn is new stock, a sale by existing holders, or a mix. Audited figures for a trailing twelve months have not been seen.
In 1602 buyers in Amsterdam acquired shares with no right to set the company’s course; in return the company held a monopoly that paid a dividend. Nothing of that kind exists here - neither monopoly nor dividend. The working comparison remains the railway: freight may grow, but rent on the track is due even when revenue does not cover it. Until a public prospectus shows what traffic is already on the books and what take-or-pay cloud bills fall due in 2027 and 2028, the two trillion is a price of track, not of freight.
Frequently Asked Questions (FAQ)
Can the stock be bought before the listing?
Not on an exchange. On private markets qualified buyers swap stakes amid scarce supply and often with company consent. The price in late September is quoted from $965bn to about $1.5tn, which points to thin supply. For most accounts through a broker outside the United States that market is closed.
What happens if orders come in at $1.3tn rather than $2tn?
If orders value the company at about $1.3tn, the market pays that valuation, not the $2tn being sought. The price-to-revenue ratio on annualised figures without an audited account is lower. The premium to the last private price shrinks. The remaining question is where the shares on offer come from: if a material part is a sale by existing holders, even the lower valuation can finance their exit rather than the company’s needs. The revenue split, the schedule on the compute contracts and the share count after options and convertibles still have to be seen.
What changes if the offering slips past November?
If the listing takes place after November, the third quarter may enter the public prospectus, the post-midterm Congress will already be known, and OpenAI will still not be a public alternative. Delay makes the deal dearer if annualised revenue slows while the issue waits, or if the 2027 minima start before proceeds from the offering arrive.
What is the largest red flag in the leaked information on the IPO?
The $2tn being sought sits on more than $100bn of annualised revenue that still has no audited account, while recognised revenue for 2025 is $4.59bn. Against $20.28bn of cash at the end of 2025 stand at least $252bn of guaranteed payments to Amazon, Google and Microsoft and about $161bn of Broadcom lease. The filing is not public, so neither the revenue split, nor the 2027 and 2028 instalment schedule, nor the share count after options are exercised can be checked. The control arrangement is a separate question and is compensated with a valuation discount. That does not change the main point: the price treats as given revenue the accounts have not yet recognised, against commitments already taken that far exceed cash on hand.
Is SpaceX a warning or a counter-example?
SpaceX raised about $75bn in June. The first day traded above the offer price; later part of the premium faded. The warning is that a large issue with a small tradable stake can be left short of buyers after the open. The opposite lesson is that waiting has a price too if the valuation stays high after listing. For Anthropic, waiting makes sense until a prospectus appears in which recognised revenue supports a valuation between $1.0tn and $1.4tn and the guaranteed payments are covered; or a first public account in the same range; or the lock-up is near its end and more old stock reaches the market than the secondary market had allowed for.
Important notice
This article is for information and education only. It is not financial advice, an investment recommendation or professional advice of any kind. The views expressed are the author’s personal opinions and are not recommendations to buy, sell or hold securities, assets or investments. Investing involves substantial risk, including the possible loss of all capital. Past performance is not a guarantee of future results. Each reader should conduct independent research, consult a qualified financial adviser and weigh financial circumstances, risk tolerance and investment objectives before taking any decision. The author and borsite.net accept no liability for any loss arising from the use of the information in this article.