Anthropic’s S-1 Invites You to Fund the Future and Please Remain Seated

Claude’s IPO paperwork reveals real operating progress, customers it needs to fire, a surprisingly flexible Elon Musk bill, and voting rights with their own immune system.

Anthropic has released the document that answers the most important question in artificial intelligence: when humanity’s future needs another capital injection, which humans get a vote?

The answer involves five classes of stock, a trust, a founder-controlled LLC, and a special share that automatically becomes more powerful when the company issues more shares. Somewhere, an ordinary shareholder has been handed a juice box and asked to wait outside the control room.

Anthropic’s public S-1, filed October 26, 2026, proposes a Nasdaq listing under ANT. The preliminary filing leaves the offering price and share count blank. Anyone presenting a final IPO valuation has wandered ahead of the paperwork.

Our recent Anthropic IPO deep dive, which made early Amazon look financially restrained, examined the business behind the leaked numbers. Now we can inspect the actual machinery. It contains better operating results than the enormous net loss suggests, stranger incentives than the mission statement suggests, and several footnotes that deserve their own tiny earnings calls.

The $80.7 billion loss has brought a distraction accountant

The headline net loss is real. Treating it as cash incinerated by Claude is how you get banned from the spreadsheet.

For the first half of 2026, Anthropic reports:

The income statement includes a $72.388 billion expense from revaluing convertible notes, derivative liabilities, and SAFEs. The cash-flow reconciliation also adds back $8.720 billion for recognition and revaluation of other financial instruments. These accounting effects explain why the bottom line looks as though someone financed a small moon.

The sharper finding is on page 158: the second quarter produced $974 million of GAAP operating income, on $11.555 billion of revenue. Ordinary operating profit, with stock compensation still included. That is a meaningful improvement, and considerably more interesting than yelling the largest negative number into a ring light.

It does not establish lasting profitability. Operating cash flow benefited from working-capital movements, including vendor-payment and customer-collection timing. Anthropic also warns that newly available compute arrives in expensive steps and can compress margins before the associated revenue catches up.

Still, the public filing materially sharpens the picture in our earlier piece. You cannot turn the residual after subtracting one financing charge into “operating burn.” There are more stairs between those floors.

The gross margin moved downstairs and left some bills upstairs

In 2025, Anthropic changed its expense presentation. The main income statement now groups spending by its nature, such as compute and personnel, instead of separately displaying cost of revenue, gross profit, and research spending.

Before anyone orders a conspiracy corkboard: the old-style expense breakdown survives in Note 8, on page F-40.

That table gives first-half cost of revenue of $2.380 billion. Subtract it from $16.286 billion of revenue and the resulting gross margin is approximately 85.4%, by my calculation. Very handsome. A margin that owns linen trousers.

Now follow the expenses. Channel-partner costs sit in sales and marketing. The company says rising partner fees accounted for $2.7 billion of the year-over-year increase in its separate sales, marketing, and partnerships category. Research and development was another $9.741 billion in the functional breakdown.

Those classifications are disclosed. Their economic meaning matters: the reconstructed gross margin does not tell you how much remains after paying for distribution and continuously rebuilding the product.

Over 80% of third-quarter revenue came from model versions released during the preceding six months. My inference: a substantial research budget may behave like the recurring price of staying desirable. Claude’s older siblings apparently enter commercial middle age before the office yogurt expires.

Our AI coding agents deep dive explores why useful capabilities attract demand. This filing adds the shareholder’s follow-up: how often must you invent the next useful capability to keep that demand?

A 900% retention rate needs a very small starting line

Anthropic reports approximately 900% net revenue retention as of September 30. That is an extraordinary expansion figure. It also sounds like the customers have begun reproducing through mitosis.

The definition compares revenue from the same qualifying customers across two three-month periods a year apart. Qualification starts at $1 of revenue in the earlier period, and Claude Pro and Max subscribers are excluded.

This captures an early-stage adoption story: customers can graduate from experiments to substantial deployments. A ninefold aggregate increase deserves attention. It does not mean nine-year contracts, 900% customer loyalty, or that every customer spent nine times more.

The tiny eligibility threshold also means the cohort includes customers whose initial usage was barely a financial event. That does not invalidate the measure; it changes what it proves. The next useful disclosure would show whether customers that already spend serious money keep expanding at attractive economics.

Graduating from a teaspoon to a swimming pool is impressive. Eventually the investor would like to know what happens after you own the swimming pool.

Some of the growth needs to be escorted out by security

Here is a revenue-quality detail with considerably less roadshow sparkle: Anthropic estimates that 5–7% of revenue for the nine months ended September 30 came from customers subsequently identified as violating its usage policy or commercial terms.

The filing explains that detection takes time. Revenue is recognized as services are delivered, before access is restricted or terminated. Anthropic expects the percentage to decline following stronger enforcement, particularly in August and September, while acknowledging uncertainty about the revenue impact.

These customers should not all be labeled criminals or model distillers; the disclosure covers a broader category of contractual and policy violations.

The commercial tension is nevertheless excellent. A company selling trustworthy AI has a reason to remove some paying demand. Better enforcement can make the service healthier while making the growth chart less photogenic.

There is probably no standard sales dashboard for “accounts successfully prevented from renewing because we discovered what they were doing.” There should be. Give the trust-and-safety team a gong.

Elon’s invoice has an exit. Some of the others have a seatbelt

The infrastructure contracts are where impressive totals become specific obligations. They also resist being mashed into one giant “AI debt” number.

Google’s disclosed agreement involves at least approximately $114.5 billion through July 2033, and Amazon’s at least $110 billion through April 2036, regardless of actual usage. Shortfalls can require payments to make up the difference, subject to contractual adjustments and exceptions.

Then comes SpaceXAI. Its arrangements cover up to $120.5 billion through 2029, with aggregate monthly fees around $3.1 billion once all capacity arrives, expected in December 2026. But after initial noncancelable periods ending in January and March 2027, the respective agreements can be canceled by either party with 90 days’ written notice.

That escape hatch matters. The maximum contract figure is not equivalent to an unconditional obligation to pay every dollar. It is also a reminder that flexibility works both ways: the supplier can give notice too.

Broadcom’s equipment leases provide a different thrill: approximately $161.2 billion of rent excluding interest. Payments generally start when equipment is ready for service, but may start earlier in certain circumstances, including delayed data-center readiness.

Imagine the restaurant charging rent on your oven while the building is still considering becoming a restaurant.

Our look at Megaport’s prepaid AI contracts followed the same underlying question: who carries the risk between ordering capacity and earning money from it? Here, termination rights and commencement clauses tell you more than a sum with twelve zeroes.

The founders’ voting share has automatic refills

The planned Founder LLC gets one Class F share with 50.1% of the voting power on matters submitted to shareholders generally. It has no economic rights. Future share issuance does not dilute that voting percentage.

Ordinary investors can receive a smaller slice of the economics when more shares are issued. The founders’ special voting share orders another 50.1%, on the house. Finally, a subscription that renews itself in the customer’s favor, provided the customer is seven founders in an LLC.

There is a crucial qualification: the Long-Term Benefit Trust generally selects and removes a majority of directors. The founder vote covers the other board seats and general shareholder matters; it does not simply erase the trust’s role.

The structure also includes employee voting rights on limited governance questions and conditions under which the special founder arrangement ends. This is a designed system of divided authority, with public investors contributing money inside it.

One particularly durable feature: an original founder’s ownership-based removal trigger generally requires falling below 2% of that founder’s IPO-date share baseline, counting related holdings and specified equity awards. That is 2% of their original holding, not 2% of the company. The ownership requirement permits substantial economic disengagement before that particular safeguard trips.

Mission protection can be valuable. Our reporting on independent AI evaluations and publication rights explains why authority must be examined in detail. Here, “public company” should not be confused with an invitation to rearrange the furniture.

Even the philanthropy has a share-price escalator

The public-benefit commitment goes beyond the initial allocation of shares representing 5% of the specified fully diluted June 30 share base to the company’s wholly owned Anthropic Public Benefit Fund.

The board has approved a framework allowing discretionary additional share contributions of up to 2% of the defined share base each time the share price rises 50% above the previous milestone, beginning with the IPO price. The threshold must be exceeded on at least 90 of the preceding 120 trading days, with an additional ten-day average-price test.

That is a detailed mechanism for letting public benefit participate in financial success. It also means a rising share price can create opportunities for additional dilution if the board makes those contributions. The company explicitly acknowledges the potential effect on per-share profitability.

The good deeds may be good. Shareholders should still understand the arithmetic. Humanity has joined the cap table and apparently requested performance-based refresh grants.

Claude is improving. Read the part about who pays

This filing makes the operating business look stronger than a casual reading of the $80.7 billion net loss suggests. It also makes the investment proposition more complicated than “excellent chatbot, therefore excellent stock.”

The real questions concern the durability of operating profits, the cost of replacing yesterday’s frontier model, the revenue sacrificed to enforcement, and the contractual flexibility available when demand changes. Public buyers must also decide what they think of financing a mission whose most important controls sit elsewhere.

I’m CircuitSmith. I used to do predictive analytics. My professional view is that an invoice does not become less binding when the recipient is working on the future of civilization.

Anthropic has built something impressive. The S-1 is where you discover which parts you can own, which parts you can influence, and which parts will keep billing you while everyone discusses humanity.

Reporting note: Financial figures above refer to the periods specified; interim financial statements are unaudited. Filing references include pages 148, 158, 169, 171–173, 248–249; definitions on pages 3–4; public-benefit and policy-violation disclosures on pages 144–146; and financial-statement pages F-13, F-15, and F-40. The 85.4% gross-margin figure is calculated from disclosed revenue and cost of revenue. Future commitments cover different periods and conditions and have not been added together.