Anthropic's S-1 shows 85% gross margin on running its models

Anthropic's S-1 shows an 85% gross margin in the first half of 2026 from running its models. That line may be the most dazzling number in the whole document.

The number that answers the skeptics

Anthropic, PBC filed its registration statement with the SEC on October 26, 2026, for an offering of Class A common stock. It has applied to list on the Nasdaq Global Select Market under the ticker "ANT." The underwriting bench is a who's who of Wall Street: Goldman Sachs, Morgan Stanley, J.P. Morgan, Citigroup, Barclays, Wells Fargo Securities, BofA Securities, Deutsche Bank Securities, RBC Capital Markets and UBS Investment Bank.

Then there is the margin. The 85% challenges skeptics who said pay-as-you-go model sales wouldn't be profitable. Evercore ISI analyst Mark Mahaney calls Anthropic's customer-serving gross margins "software-like." Selling frontier AI on a pay-as-you-go basis, at scale, with software-grade economics. That is a thesis a lot of people in very expensive suits were waiting to see in print.

How the margin got so fat

The company's own numbers explain the shape of it. Anthropic says revenue in the second quarter of 2026 grew nearly 1,400% year over year, while compute costs grew 300%. It also reports a 5.2x increase in compute efficiency, the processing resources used to train and run models, from the start of 2025 to the midpoint of 2026.

The topline is staggering on its own. First-half revenue came to $16.3 billion, against $1.2 billion in the first half of 2025. The filing also reports net revenue retention of 900%, which means customers who were already paying are, by that measure, paying far more.

The coverage gives neither a product or customer breakdown of that margin nor full-year margins for earlier years. This is a first-half snapshot, not a historical trend.

The part where the money leaves

Now, a glamorous balance sheet always has a back room. Anthropic posted an $80.7 billion net loss in the first half of 2026, up from $13 billion a year earlier. Reporting attributes nearly $9 billion of that to compute and infrastructure and does not explain the rest. Our full breakdown of the S-1's losses, commitments and founder voting control has the table.

Mahaney's own caveat points the same direction: R&D for model training absorbs most of the gross profit. So the 85% is the margin on serving customers, not a promise that the company as a whole is anywhere near black ink. The honest reading is that selling the model looks wonderful and building the next one looks like a bonfire.

The commitments are the size of a small country's budget. The Financial Times reports nearly $800 billion in computing commitments over 10 years, while Axios cites $750 billion. One disclosed piece: up to $120.5 billion through 2029 for Nvidia-GPU compute from SpaceX, around $3.1 billion a month once all promised capacity is delivered. The filing also lists Nvidia-GPU capacity agreements with Microsoft, Nscale, Lambda and other cloud providers.

If the efficiency gains and the margin hold, that spending is an investment. If they slip, it is a very large invoice.

What it means

The skeptical case is simple: expensive models might never pay their way. Anthropic's S-1 challenges the serving-cost half of that argument, in a more serious venue than a keynote slide. The other half, whether training costs ever shrink relative to revenue, stays open.

The governance picture is just as striking. Anthropic is a Delaware public benefit corporation. A single Class F share, held by a Founder LLC whose members are the seven co-founders, carries 50.1% of the voting power, and the company says it is a "controlled company" under Nasdaq rules. Buyers of Class A get one vote per share and a front-row seat.

What to watch

• The terms and the calendar. The coverage cited here supplies neither proposed offering terms nor a roadshow or trading-debut date. Reported valuation talk differs: the Guardian cites investor and analyst figures above $1.5 trillion, while an IFR report cited by Golden Era points to $2 trillion.
• The pushback. New York City Comptroller Mark Levine wrote to CEO Dario Amodei asking Anthropic to pause the IPO until it "fully and transparently addresses" concerns over AI safety, governance and the pace of development. No formal response from Anthropic has surfaced in the reporting.
• Washington. The Trump administration's supply chain risk designation bars military agencies from contracting with Anthropic. The company has challenged it in court; the cited reporting does not give the challenge's current status.
• The use of proceeds. Anthropic says it intends to use net proceeds for "general corporate purposes," including capital expenditures, working capital, operating expenses, hiring and possible acquisitions.

An 85% serving margin, nearly 1,400% year-over-year Q2 revenue growth and a proposed ticker. Everything is historic. Especially today.