Anthropic: a $1'475 billion value that hangs on a line missing from the balance sheet

On 26 October 2026, Anthropic made its IPO prospectus public, a 416-page document. On 28 October, Aswath Damodaran, professor of finance at New York University’s Stern School of Business, drew a valuation from it: $1’475 billion for the equity, at the low end of the $1’500 to $2’000 billion range mentioned for the offering. The Finance team at Meotis has gone back over that calculation and the prospectus. More telling than the number is where the value sits: more than three quarters of it comes after 2036, and what weighs most against it is a commitment the balance sheet does not show. On 2 November, the amended prospectus put it at $895.4 billion.

Fourteen times the revenue in a year

Over the first two quarters of 2026, Anthropic’s revenue reached about fourteen times the level of the same period in 2025. The number of companies spending more than a million dollars a year with it went from fewer than 100 to 1’900. Growth slowed in the third quarter, under competition from OpenAI, insufficient compute capacity and customers moving to cheaper open models. And five customers account for 25% of revenue.

Costs tell the same story of acceleration. Compute, the largest expense, stood at 251% of revenue in the first quarter of 2025: for every dollar it took in, Anthropic spent two and a half on computing power. By the second quarter of 2026 the share had fallen to 48%, and the operating margin, what is left of revenue once running costs are paid, moved from minus 249% to plus 8.4%. The company says it made an operating profit in that quarter. For the half year, the reported loss reached $80 billion, of which $72 billion was non-cash: the accounting for the conversion option of a convertible bond. Excluding that charge and one-off items, the operating loss still exceeds $8 billion.

Optimists and pessimists both find support in these numbers. The fall in compute relative to revenue is the optimists’ best argument. It is also where the pessimists’ case begins.

A debt under another name

A frontier AI company consumes a lot of computing power. Anthropic buys it from third parties under multi-year contracts and discloses those commitments in the notes to its accounts. They appear neither in its investment spending nor in its debt. In the first filing they came to more than $650 billion, led by Broadcom, SpaceX, Google and Amazon.

These contracts have moved the largest expense from variable costs, which follow activity, to fixed costs, which are paid whatever happens. They largely explain the 77.5% gross margin of the second quarter: once capacity is paid for in advance, each extra customer costs little. They also make the company sensitive to the pace of its revenue, since a fixed cost is repaid by growth.

The valuation therefore treats them as debt, and so do we. Brought back to today’s value at the rate a BB+ borrower would pay, Anthropic’s current credit rating and the top step of the speculative grade, that is 6.28%, the commitments scheduled year by year are worth $363 billion. They are part of the $472 billion of debt subtracted from the value of the business.

On 2 November, the amended prospectus filled in the box left blank in the first filing: $895.4 billion of commitments in total, almost all of it compute. We have not redone the whole valuation, because the schedule matters as much as the total; with the same method, the share of value absorbed by this debt goes up. We hold this reading until the prospectus shows that a large part of these contracts can be cancelled or reduced at low cost. We note that some of them allow it, and how far that goes is the first thing to read in the next version.

The story behind the $1’475 billion

To get from accounts to a value, the valuation tells the future in four choices, each set out as an assumption readers can replace with their own.

The first choice is about size. AI replaces part of skilled, well-paid work in some industries, which gives an end market of at least $5’000 billion, with Anthropic aiming at the premium end. The prospectus adds the ambition of running some of the businesses its agents are transforming itself, starting with pharmaceutical research. Hence revenue of about $1’011 billion in 2036, against the $700 to $750 billion of Amazon and Walmart, the two largest revenue lines in the world today.

The second is about margin. The market would split into a mass segment, where often open models compete on price at margins of about 10%, and a premium segment, where closed models fed with their customers’ data reach close to 40%, the margin of the best enterprise software companies. With three quarters of its revenue at the premium end, Anthropic would reach 32.5%, brought down to 27.5% by the cost of regulation.

The third is about what growth requires in investment: one dollar for every three dollars of additional revenue, with the compute shortage lasting a while before it eases. The fourth is about risk: a cost of capital, the return an investor demands to carry that risk, of 9.74%, close to the most demanding quarter of US companies, and a one-in-ten chance of failure, with half the value lost in that case.

The result, $1’475 billion for the equity, includes $212 billion of cash, of which the roughly $100 billion expected from the offering. Two lines in the calculation say more than the total. Of the value of the business, 77% comes from the terminal value, the value of everything after 2036. And the first year assumes $103 billion of revenue, five times the $20.6 billion of the last twelve reported months.

The two extreme stories measure the uncertainty. With $200 billion of revenue in 2036, a mass-market margin, scarce and costly compute and heavy regulation, the equity is worth minus $373 billion: the company does not survive. With $1’500 billion of revenue, a premium margin and cheap compute, it is worth $4’040 billion. More than $4’400 billion separates the two, for the same company and the same accounts.

Premium prices will hold

To be worth what the valuation gives it, Anthropic must hold a margin of about 27.5% on a trillion dollars of revenue while paying for compute capacity already committed for hundreds of billions. The price of its best models must therefore stay above what it costs to serve them. We wrote on 2 November, from the same prospectus, that cost of revenue is about 14% of revenue over nine months, and compute as a whole 57%. The margin on serving exists; it pays for capacity and the next model.

For a Swiss company that buys AI, the price war we follow at the low and middle end of the market will stay there for as long as the supplier’s economics rest on the premium end. And a supplier carrying fixed costs has to grow: when its growth slows, its prices, usage caps or priorities can change. We saw it in July, when subscription caps were lowered on the very day the investor roadshow opened. Keeping the ability to switch models without rebuilding your processes protects your price.

What Meotis chose

The prospectus brings three weaknesses to light: compute costs committed in advance, growth that is obliged to cover them, and customers who depend on a single supplier. Meotis built its model to avoid them.

Meotis orchestrates the best models on the market and pays for them as they are used. The cost of compute follows our customers’ activity, and our products rest on contracts with no multi-year capacity commitment.

Our customers pay for set-up, then a monthly subscription per organisation. Because Meotis buys the models it uses, every fall in their price lowers our cost: the price war we follow in these pages works in favour of our margin.

Comoto Enterprise OS orchestrates these models from a sovereign core hosted in Switzerland, routes each task to the model with the right cost-to-capability ratio, and leaves the customer free to change models. Our applications already spread their tasks across several suppliers, and some move on their own to another when one does not respond. If a supplier changes its prices or its caps, we change models, and our customers keep their processes.

Five share classes

The public will only buy class A shares, with ordinary voting rights and little influence over the board. A public-benefit trust holds the class T shares, with no vote, which elect the majority of the board. A single class F share, held by the seven co-founders, carries 50.1% of the voting rights, however many shares are issued later. Alphabet and Amazon hold non-voting class N shares. The public benefit corporation status also allows management to take decisions that hurt the share price in the name of its mission.

Alphabet and Meta have long lived with multiple-vote structures. Anthropic’s structure goes further, and in our view it calls for an additional discount, which the valuation does not quantify. Existing shareholders other than the founders and a few named individuals will be able to sell up to 10% of their shares on the first day, the rest in tranches over 180 days; the founders, after 365 days.

What to watch

The next version of the prospectus will set the number of shares and the price range, still blank. It will show whether the market lines up with a value close to $1’475 billion or with the top of the range. If the stock is priced well above its value, the first weeks of trading will be a bet on price, and buying value will come later. Facebook showed it: floated at $38 in May 2012, the stock was worth less than $18 in early September, three and a half months later.

Three lines in that next version will show whether the story holds: fourth-quarter growth, the schedule of the $895.4 billion of commitments and how much of it can be cancelled, and the credit rating.

This is markets research and commentary written for a general readership. It is not personalised investment advice and does not take account of any individual’s circumstances.