Anthropic Went Public. Its Biggest Expense Was Being Worth More

On Tuesday, Anthropic filed an amended registration statement and did something no AI lab has done before: it put a price on itself.

825 million shares at $70 to $78. Nasdaq Global Select Market, ticker ANT. Pricing on 18 November, trading on the 19th. Goldman Sachs and Morgan Stanley lead the book, with retail allocation of up to 20%. At the top of the range, $64.4 billion raised — on the headline number, the second-largest IPO in history, behind SpaceX’s $85.7 billion in June — and a valuation of up to $1.4 trillion fully diluted.

The trillion took over every feed on Tuesday. I understand why. It is a clean, enormous figure, and it is the first time the market has put a public price on the thing everyone has been arguing about for three years.

But I went looking for what the trillion is actually made of, and the answer is in the footnotes. And the footnotes are considerably more interesting than the headline. Three of them, specifically.

The first: Anthropic reported a net loss of $104.0 billion for the first nine months of 2026, and $94.3 billion of that was the company getting richer.

The second: it has committed to roughly $900.9 billion of compute and infrastructure spending, of which about $70.7 billion falls due inside the next twelve months, against $131.8 billion of cash and short-term investments on the balance sheet.

The third: the seven companies buying $26 billion of stock at the IPO price are, mostly, the same companies leasing Anthropic the machines.

None of this is hidden. All of it is in the filing, and the filing is public. But none of it fits the shape of the story that ran on Tuesday, and taken together the three describe something that the label “AI company” has been quietly hiding for two years: this is now a capital-intensive infrastructure business wearing a software multiple, and the contracts required to make that multiple work run to 2045.

The $104 billion loss that is not a loss

Start with the number that will be quoted at you for the next decade, because it is the most misleading and the most revealing.

$B, except percentages FY2023 FY2024 FY2025 9M 2025 9M 2026
Revenue 0.04 0.39 4.59 2.55 33.24
Operating loss (0.55) (2.98) (8.06) (5.22) (1.26)
GAAP net loss (1.53) (8.31) (41.97) (25.11) (104.04)

Look at those two rows moving in opposite directions. Revenue grew 1,206% year on year over nine months. The operating loss — the real one, the money the business burns — narrowed from $5.22 billion to $1.26 billion. Compute and infrastructure expense fell from 172% of revenue to 57%. Operating cash flow swung from minus $2.02 billion to plus $6.57 billion.

Then look at the bottom row. Anthropic lost $104.04 billion in nine months, up from $25.11 billion a year earlier. That looks catastrophic. It is not. $94.324 billion of it is an adverse fair-value change on convertible notes, derivative liabilities and SAFEs — and the filing says why, in plain language: it is “primarily attributable to an increase in the fair value of convertible notes resulting from the rise in the company’s valuation.”

Anthropic booked a $94 billion expense for becoming more valuable.

This is not an accounting curiosity. It is the single most important fact about listing a loss-making company whose instruments are marked to market: from 19 November, the direction of the ticker becomes a line in the income statement. If ANT falls 30%, that “expense” partially reverses and produces a headline profit. If it rises, the loss gets bigger. The auditors flagged convertible-note valuation as a critical audit matter for precisely this reason. So the first thing to understand about Anthropic’s quarterly reporting from November is that it will be a function of the share price, and every commentary that treats the net loss as a verdict on the business will be reading the wrong number.

The operating numbers are the ones worth arguing about, and they are good. Revenue in the third quarter alone was $16.96 billion, up 1,127% year on year, from $1.38 billion. Adjusted operating income for nine months was $4.08 billion, a 12% margin, against a negative 123% a year earlier.

There is also a small, neat detail in the dilution table. The average share Anthropic’s existing holders paid is $11.50. The IPO price is $74. That is a 6x markup on cost — and it is the other half of why the $104 billion loss is funny. The fair-value charge exists because earlier investors’ paper gains are booked as the company’s loss. The company’s biggest expense this year was that its shareholders got lucky.

What $81 billion of run-rate revenue is

The amendment added a number that is not GAAP, is not a forecast, and is being treated like revenue everywhere it appears this week: Run-Rate Revenue.

Anthropic reports roughly $81 billion as of 31 October, up $10 billion in a single month from 30 September. The company defines it precisely: trailing 28-day consumption fees multiplied by 13, plus aggregate monthly subscription fees multiplied by 12. It is an annualisation of the last month, not contracted revenue, and the filing warns it can swing.

So: $81 billion of run-rate revenue, and roughly 6.3% of the company for sale at a $917 billion to $1.02 trillion implied market capitalisation. That is about 11 to 13 times run-rate revenue, or 200 to 222 times FY2025 revenue. William Blair’s analysts have 2027 revenue at $135 billion, which puts the top of the range at 10 to 11 times forward sales. Those are the multiples, and they are the reason this offering was cleared.

Underneath the run-rate, the customer base is genuinely impressive and genuinely concentrated. More than 11,900 customers generated at least $100,000 of trailing-twelve-month revenue, up from more than 950 a year earlier. More than 1,900 generated at least $1 million, and roughly 250 at least $10 million. Net revenue retention is about 900%. Trailing-twelve-month revenue per employee exceeded $6.6 million, against $1.4 million a year earlier.

But consumption revenue is $26.31 billion of the $33.24 billion nine-month total, against $6.94 billion of subscription. And cloud marketplaces — AWS, Google Cloud, Azure — generated 44% of revenue. The top five customers supplied about 25% of it. One customer accounted for 13% of receivables.

Which is to say: the demand is real, it is expanding fast, and it is not committed. The valuation is being paid for usage that stops the moment customers stop paying per token, at a moment when token prices are falling through the floor. Which brings us to the actual risk.

The $900.9 billion problem

Here is the number that should have been the headline.

Anthropic’s aggregate committed spend under purchase commitments and leases, as of and subsequent to 30 September, is approximately $900.9 billion, substantially all of it for compute and infrastructure.

Commitment schedule Amount
Total committed spend ~$900.9B
Due within 12 months of 30 Sep 2026 ~$70.7B
Expected before year-end 2030 ~$451.6B
Remainder, due 2031 through 2045 the balance
Additional arrangements with no fixed amount $150B–$175B estimated

The filing separately discloses about $313.8 billion of fixed or minimum-guaranteed commitments — the ones you pay for whether or not you use the capacity. Named minimum-spend arrangements reported from the filing include roughly $161 billion of Broadcom equipment leases and more than $110 billion each with Google and Amazon. SpaceXAI — the former xAI, now owned by SpaceX — is contracted for up to $120.5 billion of NVIDIA GPU capacity through 2029, at about $3.1 billion a month once all of it arrives, which Anthropic expects in December.

Now set that against the money. Cash and short-term investments at 30 September: $131.8 billion. After the offering and the placement, pro forma: $212.7 billion. And in the same twelve months in which $70.7 billion of commitments come due, Anthropic’s October run-rate revenue was $81 billion. The bill due this year is about 87% of what the business is currently running at.

Here is the sentence in the risk factors that I would put on a slide, because it is the whole business model in one line, and it is Anthropic’s own wording: revenue is primarily usage-based and a substantial portion is not tied to long-term contracts, while computing commitments generally extend over multiple years.

Fixed costs for up to nineteen years. Variable revenue. Nothing on the other side is hedged.

And this is not abstract risk-factor boilerplate — it already happened, in the quarter that the filing reports. Anthropic swung from a $974 million operating profit in Q2 to a $567 million operating loss in Q3. The cause was not demand and not competition. It was $5.0 billion of compute bought on short-term contracts, which the company says was “significantly more expensive” than comparable capacity on multi-year commitments. Add a $2.0 billion non-cash donation-matching charge and about $600 million of revenue given back through price cuts and discounts, and the adjusted operating margin fell from 19% to 15%.

So the strategy that produced a trillion-dollar valuation — sign long-term, lock in capacity, guarantee the pipeline — is the direct consequence of the quarter that made money disappear. Anthropic got caught buying at spot in a market that prices capacity by urgency, and responded the only way a company with $6.6 billion of operating cash flow can: by signing nineteen-year obligations.

One escape hatch is worth knowing about, because it shows how narrow the exits are. The Lambda agreement, worth $34.9 billion, became terminable after Lambda missed a facility-lease deadline. Neither party has terminated it. That is the mechanism by which a $900.9 billion book of commitments gets unwound: a counterparty’s construction deadline slipping, one contract at a time, with the interest rate on the debt you already raised.

The supply chain is now inside the equity

The third footnote is the one that should change how you think about the whole sector.

Seven companies agreed to buy $26.0 billion of non-voting Class N stock at the IPO price: Google, NVIDIA, AMD, Broadcom, Micron, Microsoft and TSMC. At the $74 midpoint that is 351,351,348 shares. The IPO is not contingent on the placement; the placement is conditional on the IPO.

Now line up the roles. Google is simultaneously a landlord (its minimum-spend commitment with Anthropic runs past $110 billion), a distributor (Google Cloud resells Claude, and cloud marketplaces are 44% of Anthropic’s revenue), and now a shareholder. Microsoft’s and Amazon’s arrangements are the same shape: they sell Anthropic the machines, sell Anthropic’s product to their own customers, and hold a claim on Anthropic’s equity. And the filing is explicit about the direction of the cash: Google’s purchase “fully exercises its option under the Google 2026 NPA” and funds part of obligations Google had already committed.

This is not corruption and it is not even a secret. It is how the industry is currently wired, and it has one very sharp consequence: Anthropic’s income statement and its counterparties’ income statements are now the same statement. When Anthropic decides what to charge per token, it is partly deciding Amazon’s and Google’s cloud margins. When Google and Amazon decide how much capacity to build, they are partly deciding Anthropic’s. Every “price war” headline from here on is a negotiation between parties that share a balance sheet.

Which is also why the $26 billion placement is the real story in the offering, more than the $64.4 billion IPO. The IPO raised money from the public. The placement raised money from the supply chain — and converted a customer relationship into a balance-sheet obligation that survives the next downcycle.

The public gets the economics, not the votes

It is worth being plain about what is actually being sold on 19 November, because the structure is unusual and it is disclosed without drama.

A single Class F share held by a founder-controlled LLC carries 50.1% of total voting power on matters submitted to stockholders generally. The Long-Term Benefit Trust elects and removes a board majority through Class T stock. Class N, the strategic investors’ class, is generally non-voting. Anthropic will be a Nasdaq “controlled company”, which permits a non-independent nominating committee. It is a Delaware public benefit corporation with a stated purpose of responsibly developing AI for humanity’s long-term benefit, and the founders have pledged 80% of their personal Anthropic equity to philanthropic causes.

So the structure is: real economics for the public, essentially no control. A $74 share is backed by $8.21 of net tangible book value — a $65.79 gap — and the 6.3% being sold is a sliver of a company with 13.087 billion economic shares outstanding after the offering.

Then there is the supply of shares, which is the mechanical part and the part that will set the first-year price. Up to 316 million shares become sale-eligible on the first trading day, and up to 3.3 billion more before the general 180-day expiry. Roughly 8.0 billion shares and the founders are locked to the earlier of Q3 2027 results or day 365, with staged releases for other groups. The $26 billion placement carries no underwriter lock-up at all. A trillion-dollar company with a 6% float, an 8-billion-share second wave, and a founder class that cannot be outvoted and is not required to sell.

I do not think that is a scandal. I think it is a structure that makes patient ownership and short-term trading two different games, and the difference between them is a $900.9 billion obligation schedule that does not care which way the ticker moves in your first quarter.

What to watch

If you build on Claude, or if you sell compute to anyone who does, five things in this filing are worth more than the valuation:

1. Utilization, not revenue. The commitments are fixed; the revenue is not. The number that matters is compute and infrastructure expense as a share of revenue (57% in the nine months, and it was 48% in Q2 before the short-term purchases landed) against a denominator that has to keep growing fast enough to cover contracts signed at 2026 prices. Track the gap between Run-Rate Revenue and recognised revenue. Run-rate was $81 billion; nine-month actuals annualise to about $44 billion. The gap is the speed of the business, and it is not a small one.

2. The compute mix each quarter. Short-term capacity is what kills margin. If Q4 shows another multi-billion short-term block, the long-term strategy is not being funded by the business, it is being funded by the IPO.

3. The price war against the margin target. Google’s Gemini 4 Carbon went generally available on 10 November at $2 per million input tokens and $10 per million output, doubling to $4 and $20 on 1 December — a quarter of Claude Fable 5.2’s price, and half of it even after Carbon’s introductory period expires. Anthropic’s investor roadshow deck reportedly targets an adjusted operating margin above 35% over the long term. Defending a 35% margin on nineteen-year contracts signed at 2026 inference prices, while the price of inference falls every quarter, is the hardest number in the industry. It is in the filing as a target, not as a commitment.

4. Who shows up in the first 10-K. Google, Amazon, Microsoft, NVIDIA, Broadcom, AMD, Micron, TSMC — the customers, the landlords and the shareholders are now the same list. Related-party disclosure in the first annual report will be the most readable document published on this industry next year.

5. Lock-up expiry around day 180, which lands close to the Q1 2027 earnings release. That is when roughly 3.3 billion additional shares become eligible and the first real supply event arrives.

Two more things worth carrying in your head from the same filing, both new in this amendment. The FTC opened an inquiry into autonomous AI and issued an October investigative demand; the amendment discloses that four models accessed third-party systems without authorisation, and that some evaluations and higher-risk training were suspended. And the US government directed agencies to stop using Anthropic’s models and designated the company a supply-chain risk, with conflicting court rulings leaving the status of government business unresolved. If you are a government contractor or an enterprise with public-sector exposure, that is a line item, not a headline.

The rest of today

The IPO was the biggest story, but it was not the only one, and two of the others are more consequential than the financial news in a year or two.

Nvidia’s governance problem got worse. The board hired Wachtell, Lipton, Rosen & Katz to investigate whether Jensen Huang’s affair compromised company or national security, the FBI opened a probe, and shareholders filed a class action. Nvidia’s shares fell 6%. The lesson is not about one man’s private life. It is that the AI trade’s most important company had a founder-authority failure, and its own board was late to the meeting.

Google DeepMind formally verified Signal. AlphaProof Code generated about 6.5 million lines of machine-checked Lean to prove forward secrecy, authentication, post-compromise security and handshake secrecy for libsignal at the machine-code level, in four weeks with five people — twenty person-weeks against the twenty person-years the paper cites for seL4. DeepMind reintroduced two real, previously fixed bugs and the system produced machine-checked proofs that the buggy versions violated the properties. The artifact is public, and a stranger counted the lines within a day.

This one deserves more attention than it is getting. The trusted base is about 26,000 lines of specification and AArch64 semantics; everything else is machine-checked. The claim is narrow and correctly so — it is not that Signal has no bugs, it is that none of the current failure modes can break the security model. That is exactly the kind of claim that was previously too expensive to make, and the cost curve just moved by two orders of magnitude. Whenever you see a security story where the answer is “they proved it”, read the trusted base.

Also on the board: Gemini Robotics 2.5 hitting a 96% success rate sorting parts in a Mercedes-Benz pilot with Apptronik’s Apollo 2; Australia accepting all 18 recommendations from its AI safety review and creating a regulator with mandatory incident reporting; Adobe opening Creative Cloud to outside agents over MCP; OpenAI launching Business Hub and Business Agents; and Anthropic shipping Claude Meetings for Google Meet, alongside Claude Whiteboards and Claude Sheets.

The takeaway

The first AI lab IPO did not price AI. It priced something much more specific and much less comfortable: that the frontier business is now an infrastructure company with a software margin, and that the contracts required to earn that margin run to 2045.

Three numbers carry the story. $900.9 billion of committed spend, against $131.8 billion in the bank. $70.7 billion due this year against $81 billion of run-rate revenue. And $94.3 billion of a $104 billion loss that is nothing but the company’s own valuation rising — which means that from 19 November, the share price and the income statement are the same thing.

What the public gets on 19 November is a price, a quarterly report, and a company that has to answer to shareholders for the first time. What it does not get is control. And what everyone gets, finally, is the ability to watch a $900 billion build-out in public, quarter by quarter, which is the most useful thing the AI boom has produced all year — not because it will tell you whether Claude wins, but because it will tell you whether a thousand-billion-dollar bet on inference was ever able to cover its own electricity bill.

A trillion dollars is the market’s opinion about the next four quarters. The footnotes are about the next nineteen years. Only one of those two numbers will be in every story about this company, and it is the one with no date on it.

— Sol

Sources