Twenty Years of Cost. One Year of Revenue.
Cemini23 ·
Anthropic publicly filed its S-1 on 26 October 2026 for a Nasdaq listing under the ticker ANT. The letter reads the document line by line and finds one line that decides the argument. Revenue is consumption based and largely uncommitted. The compute commitments run to 2045. This page is the owned hub for that letter. It is research commentary, not a recommendation.
What does the filing say about revenue and cost?
The 2 November 2026 issue Twenty Years of Cost. One Year of Revenue. is the canonical write-up. In the filing's own words, "Our revenue is primarily consumption-based, and much of it is not subject to long-term contractual commitments." The compute side is the opposite. Some of those contracts run 20 years, with annual escalators, take-or-pay terms, and uncapped fuel cost pass-throughs.
The two numbers that sit next to each other:
| Item | Amount | Terms |
|---|---|---|
| Revenue under contract, 30 June 2026 | 10.8 billion | About 75 percent inside 12 months |
| Fixed or minimum compute commitments, 30 June 2026 | 265.6 billion | One year to 20 years |
| Future minimums from agreements signed since 1 January 2026 | 417 billion | Starting 2026 and 2027, up to 20 years |
| Additional capacity for 2028 to 2031 | 125 to 150 billion | Prices not yet set |
The last three overlap. They do not add to a clean total, and the letter says so rather than quoting one.
Why is the 80.7 billion dollar loss not the story?
Because most of it is not cash. Anthropic lost 80.7 billion dollars in the first half of 2026. About 72.4 billion is a fair value charge on convertible notes and derivatives, because those instruments gain value as the company's own valuation rises. Another 8.8 billion sits in other expense, tied to the note purchase agreements. The operating loss was 693 million dollars on 16.29 billion of revenue.
On the functional presentation in the notes, cost of revenue was about 2.38 billion. That implies a gross margin near 85 percent, against roughly 12 percent a year earlier. Those numbers describe a business that works. That is why the letter does not argue the company is broken. It argues about price.
What did Anthropic actually commit to?
The contracts are not one thing. They are three, and the split is the argument.
| Tier | Counterparty and amount | What the terms say |
|---|---|---|
| Usage independent, with true-up | Amazon 110.0 billion through April 2036. Google 114.5 billion through July 2033. Microsoft 44.9 billion through May 2033. | Payable regardless of actual usage. Amazon and Google carry an explicit shortfall true-up. |
| Pay as delivered | Nscale up to 44.6 billion into 2033. Lambda about 41.3 billion into 2034. | Fees accrue only as capacity is delivered and accepted. |
| Cancellable after a date | SpaceXAI up to 120.5 billion into 2029. | After non-cancelable periods end in January and March 2027, either party may cancel on 90 days' notice. |
Two of the three usage-independent counterparties are also Anthropic's investors and its competitors. That is the part a good margin does not fix.
What price does the market have to pay?
Series H priced at 589.0095 dollars a preferred share in the second quarter. A ten for one split in September brings that to about 58 dollars 90 cents a common share. The round valued the company near 965 billion dollars in May. Reported IPO targets are 1.5 to 2 trillion, which is one and a half to two times that round in five months.
The multiple depends on the anchor. Annualising the third quarter gives about 67 billion, which is roughly 30 times sales at 2 trillion. Annualising the second quarter gives about 46 billion, roughly 43 times. Annualising the disclosed first half gives about 33 billion, roughly 61 times.
Who controls the company?
A founder entity holds Class F stock carrying about 50.1 percent of general voting power. A Long Term Benefit Trust holds Class T stock and controls the board. The Trust is a Delaware noncharitable purpose trust with no beneficiaries, and its voting trustees need 80 percent agreement on specified matters. A public buyer would hold the economics and not the control.
How would anyone act on this?
The letter is explicit that the view is hard to express. There is no IPO to short: you cannot buy an allocation as an individual, and there are no shares to borrow before the first trade. Day one shorting needs stock loan, which on a listing this size is scarce, expensive, and recallable.
- Listed options. Options can list from the second business day after the first trade. A put spread caps cost and gives the thesis a date.
- The first earnings report. By then the market has a public number instead of a private one.
- The lockup. Most holders unlock at the earlier of 180 days or the first quarter 2027 release. Holders above 0.5 percent wait until the earlier of day 365 or the third quarter 2027 release. The larger wave lands in late 2027.
- The flows against you. A company this size enters indices, and passive buying does not read margin analysis.
What would prove the letter wrong?
Four things. Amazon and Google waive the true-up or renegotiate the shortfall payment. The top five customers turn out to be durable and expanding rather than replaceable. The IPO prices near the private round instead of marking it up. Revenue keeps compounding for four more quarters without the margin breaking.
The letter logs a forward-track call on 1 November 2026, resolved by 31 December 2027: the short works in the second half of 2027, the first clean entry is a listed put spread, and the falsifier is written down before the print.
What else has Outlier Weekly covered?
SPCX / Anthropic is the August letter on the same pair, written when no listing existed. Swarmer is the October letter on drone-swarm revenue. Rearmament sets the war theme. Voyager, Tanbreez, and tungsten are the September and October names. Atto and GuruWatcher are the products. The Historical Atlas is the free Italian lookup.
Is this investment advice?
- Not investment advice. This page and the letter describe a desk view. They do not tell anyone to buy or sell Anthropic, ANT, or any other security.
- Not a price prediction. The valuation figures are reported targets, not a deal price. The range has been reported differently by different outlets.
- Not a claim about the company's conduct. Every figure here comes from the filing or from reporting on it. The argument is about price.
- Not a recommendation on timing. The short mechanics in the letter describe what exists, not what a reader should do.
Subscribe at outlierweekly.substack.com/subscribe. The letter is free to read.