Anthropic looks capital-light on paper and is one of the most operating-levered companies ever to file for a public listing. The prospectus filed on October 26 shows compute costs falling from more than 250% of quarterly revenue to 48%. That drop reads like an efficiency story. It is mostly a fixed-cost story. Anthropic has moved its largest expense from the variable column to the committed column, through compute contracts reported at more than $600 billion. A fixed cost always shrinks as a percentage when revenue grows fourteenfold. The same ratio climbs back just as fast if growth stalls, and the third quarter of 2026 already showed growth slowing.
The independent valuation that prompted this note lands at $1.475 trillion for the equity, just under the $1.5 to $2 trillion offer range. Its most useful output is not that figure. It shows which inputs carry the value, and the cost of capital is not one of them.
Where the Capex Went
The intuitive mistake starts with the cash flow statement. A frontier model company should be a capex monster, yet the filing shows no large capital expenditure line and no balance sheet loaded with compute assets. A casual reader concludes that Anthropic rents its way to scale and keeps its capital needs small.
The capital is there. It sits on someone else's balance sheet. The hyperscalers and data center operators that build the clusters book the capex. Anthropic books a promise to pay for them. Economically, that promise behaves like debt with an asset attached: payments are contractual, they do not flex with demand, and they expose the company to default if revenue falls short. Credit markets have priced this. The company's BB+ rating is a speculative-grade grade on a firm holding $114 billion in cash before IPO proceeds. Bondholders are looking straight past the empty capex line to the commitment schedule.
The valuation handles this correctly by folding the present value of compute commitments into debt. That is the right treatment, and it changes how the company should be described. Anthropic is not an asset-light software firm. It is a lessee carrying hundreds of billions in fixed obligations, and it reports high gross margins because the lease cost is spread over a fast-growing revenue base.
The 48% Ratio Cuts Both Ways
The mechanism is simple. Treat compute as roughly fixed over a contract window. When revenue in the first half of 2026 runs about fourteen times the prior-year half, a fixed cost collapses as a share of sales, and the income statement swings toward profit. The company reported an operating profit in its most recent quarter. For the first six months it still posted an adjusted operating loss of more than $8 billion.
The headline loss was $80 billion. Of that, $72 billion was a non-cash charge for expensing the conversion option on a convertible bond. That item is real accounting but carries no operating meaning, and anyone using the $80 billion figure is arguing in bad faith in one direction or the other.
Run the mechanism in reverse and the risk is plain. The prospectus names three forces behind slowing third-quarter growth: OpenAI gaining share, compute capacity acting as a constraint, and some enterprises moving to cheaper open-source models. None of those forces lowers Anthropic's committed compute bill. If revenue growth decelerates while commitments keep stepping up, the 48% ratio does not hold. It rises, because the denominator is growing more slowly than the contract schedule. This is the textbook operating-leverage trade, and it is why a quarter of reported profit tells you less than it seems. It is a reading taken at the point of highest leverage, on the way up.
A second-order wrinkle: compute does not stay bought. Chips age fast and need replacing, and data centers are drawing community opposition that could raise their cost even if chip prices fall. The valuation assumes Anthropic eventually needs about $1 of new capital for every $3 of new revenue. That sales-to-capital ratio of three matches what technology companies earn on invested capital today. Applied to a revenue path toward $1 trillion, it still means very large annual reinvestment, funded either by operating cash or by new commitments of the same fixed kind.
What Actually Moves the Valuation
The valuation rests on four judgments, and they do not carry equal weight.
- Revenue in 2036. The base case is roughly $1 trillion. No company has that much revenue today; Amazon and Walmart are around $700 to $750 billion. The path relies on Anthropic leading the premium segment of an AI market estimated at $5 trillion or more, driven by displacement of highly paid white-collar work. It also leans partly on an explicit move into businesses its agents disrupt, with pharmaceuticals named as the first candidate.
- Target margin. A 75/25 mix of premium and mass-market revenue, at roughly 40% and 10% operating margins respectively, gives 32.5% before regulation. A 5-point regulatory drag brings it to 27.5%.
- Reinvestment efficiency. The sales-to-capital ratio of three described above.
- Risk. A 9.74% cost of capital against a September 2026 US median of 9.16% and a third quartile of 9.88%, plus a 10% probability of failure with half the value lost in that case.
The range of outcomes shows where the sensitivity lies. Revenue of $200 billion in 2036, mass-market margins, a sales-to-capital ratio of one, and harsh regulation produce equity worth negative $373 billion: the company does not survive as a going concern. Revenue of $1.5 trillion, 40% margins, a ratio of four, a first-quartile 7.79% cost of capital, and no regulatory drag produce $4.04 trillion. The swing from the base case to the bad case comes almost entirely from revenue, margin, and reinvestment. The discount rate moves by about two percentage points across the whole range, and the author explicitly calls it among the least important inputs.
The margin assumption is the one most easily checked against public data. The 40% premium-segment figure is taken from the top quartile of enterprise software. Incumbent software margins are the closest observable ceiling for what a closed frontier model, trained on client data, could earn at maturity. The difference is that those incumbents reached their margins without a committed compute schedule running into the hundreds of billions. Anthropic would need software-like margins while carrying something closer to a utility's fixed obligations. That combination has no clean precedent, which makes it a hypothesis rather than a calibration.
The Case That the Leverage Is a Feature
The bull argument deserves a fair hearing, because part of it is right.
Locking in compute is what lets Anthropic keep pace with OpenAI, which held the early lead in capacity. In a market where capacity is the binding constraint, fixed commitments are a moat as much as a burden. A rival who cannot secure chips cannot serve demand. The customer data supports real traction: enterprises spending more than $1 million a year on Claude rose to 1,900 from fewer than 100 a year earlier. About 80% of revenue comes from enterprises paying by usage, the segment where switching costs grow as models are built into workflows. And liquidity is large: $114 billion in cash, plus roughly $100 billion expected from the offering, enough to cover several years of commitments without new money.
The weak link in that case is stickiness, and it has not been tested. Nobody yet knows whether enterprise spending on AI agents holds through a recession or gets cut like any discretionary IT budget. Concentration makes the exposure sharper: the top five customers supply 25% of revenue. If one large customer delays, while the compute schedule does not, the leverage mechanism described above starts working against the company. The bull case also requires the cash cushion and capital-market access to last. Both shrink together in the scenario where they would be needed, because a stalling revenue line burns cash and cuts the market value that would back any new raise.
Governance Removes the Brake
The governance terms matter here because of the leverage, not apart from it. In a company with large fixed commitments, the main protection for shareholders is the ability to force cuts when revenue disappoints: stop renewing, stop expanding, preserve cash. Anthropic's structure takes that protection away.
There are five share classes. A benefit trust selects the board. The founders hold 50.1% of voting power regardless of how many shares are issued later. On top of that, the company is a public benefit corporation, and the prospectus states that it may make decisions that "negatively impact the short-, medium-, or long-term performance of our business or the market price." Put those together and the people deciding whether to sign the next round of compute contracts cannot be overruled by the people funding them. Alphabet and Meta already have dual-class structures. This one goes further.
The lock-up schedule matters less than the coverage suggests. Insiders other than founders can sell up to 10% on day one, further tranches unlock over the first 180 days, and founders and key named figures can trade after 365 days. With deep institutional ownership and plenty of liquidity, unlock dates are more likely to drive headlines than prices.
The Condition That Decides It
The offer range sits right at what an optimistic story is worth. At $1.5 trillion or more, buyers are paying for roughly $1 trillion of 2036 revenue at software-grade margins, with no allowance for the leverage running in reverse.
The variable to track after listing is compute cost as a share of revenue, quarter by quarter. If it keeps falling from 48% while revenue growth re-accelerates past the third-quarter slowdown, the fixed-cost bet is working and the premium segment is real. If the ratio flattens or rises while new commitments are announced, the same mechanism that produced the recent quarter's profit is turning against it. In that case, no shareholder has the votes to stop the next contract from being signed.
