Anthropic's S-1 Has No Gross Margin Line. Your AI Product Needs One.

Anthropic filed its S-1 on Monday. The line I went looking for first isn't there.

There's no gross margin. In the filing, compute and infrastructure is an operating expense: $8,945 million in the first six months of 2026, against $16,286 million of revenue. Tomasz Tunguz put it in one line: "Anthropic does not report gross margin." He derived one anyway, from the cost of revenue the filing keeps in a footnote, and got 85 percent for the half. Subtract the whole compute line instead and you land near 45. That isn't fair either, since the line includes training and research.

Same company, same six months, a 40 point spread depending on where one cost sits.

This is part 5 of The CPO Transition. Part 1 put margin inside the operating model the CPO now owns. Part 4 was the org that runs it. This one is the P&L conversation.

The short version

Anthropic's S-1 books compute as an operating expense and shows no gross margin, so the most-watched AI company in the world has a margin that reads as 85 percent or roughly 45 depending on where you put one line. Every AI product has a smaller version of that problem. If serving cost hides in R&D or a shared cloud bill, gross margin flatters the product and nobody who designs the feature sees the invoice. The CPO's first move is to get serving cost into cost of revenue, split by feature. The second is a pass-through policy: Anthropic says its newest model cut customer cost per task by about 40 percent, and when a vendor cuts your cost, someone has to decide whether that becomes a lower price, better quality, or margin. Most companies decide by accident. Margin is product's job now because every lever on it is a product decision.

Where the cost sits is a product decision

I don't think Anthropic is hiding anything. A frontier lab spends on training and research at a scale that makes the usual software split awkward, and the filing labels the line plainly. Jason Lemkin made the same observation: "the prospectus doesn't show a conventional gross margin," because compute is booked as operating expense.

Bring that down to a software company shipping AI features. The serving bill for a copilot or an agent is often on a cloud invoice that finance allocates to R&D, or to a shared platform line, or to nobody in particular. Gross margin on the subscription looks like it did in 2022. The feature team never sees what a successful action costs. Then a CFO runs the numbers in a planning cycle and the conversation starts with a cut instead of a choice.

So the first ask isn't a pricing change. It's a ledger change. Put the cost of serving customers in cost of revenue, split by feature, and leave training, evaluation, and internal experiments somewhere else. Gross Margin Is Your Job Now makes the case for cost per successful action as the unit, and the unit to use for token cost is the short version. Neither works if the bill isn't split.

Some rough arithmetic of my own makes the point. Anthropic's compute line was about $2.01 for every dollar of revenue in the first half of 2025 and about $0.55 in the first half of 2026. Most of that swing is revenue growing far faster than cost. You can only see a curve like that if the cost is on a line you can divide.

The price cut is margin you keep or give away

The sentence in the S-1 I'd hand every CPO is in the business section. Anthropic says its newest model "reduced customer costs per task by approximately 40%," and that "our compute efficiency allows us to deliver these reductions while preserving our margins."

That's a lab passing a cost decline to customers on purpose, while protecting its own margin. Your vendors are doing it to you. Gergely Orosz reported that Uber's token costs came down about 50 percent over a few months.

When that happens, there are three things you can do with it. Pass it through as a lower price or a bigger allowance. Spend it on quality: a stronger model on the hard cases, a second verification pass, a longer context. Or keep it as margin. All three can be right. What isn't right is the default I see most often, which is that nobody decides, usage grows into the savings, and the margin ends up where it started with a bigger bill.

Write the policy before the next cut. Something like: half goes to quality on the outcomes we're measured on, the rest is margin, and price only moves when a competitor forces it. Yours will differ. The point is that it's written, and product owns it.

Price for expansion, then watch the heavy tail

Lemkin's other headline number from the filing is net revenue retention of about 900 percent as of September 30, which he attributes mostly to usage-based pricing expanding inside existing accounts. That's lab scale and an outlier. The direction still applies to anyone pricing AI. When the unit is usage or credits, revenue follows the work, and so does cost. The credit is the new seat covers how brands are pricing it, and Pricing for AI Products has the models.

The catch with usage pricing is that cost concentrates. A tiny share of users can drive most of the bill, which is the argument in 1% of Users, 75% of the AI Cost. Pricing that expands with the work only protects margin if the brakes are on the right loops.

What the CPO brings to the CFO

The P&L conversation goes better when product starts it. Mine is one page, monthly, four lines.

Cost per successful outcome for the top AI features, against target. The share of cost driven by the heaviest users. What happened to the last vendor price change, and whether that matched the policy. Margin by customer cohort, so a renewal that looks healthy on revenue doesn't hide a cohort that's underwater on cost.

If the margin drops anyway, how to defend a margin drop to your CFO is the conversation after this one. The goal of the page is that the CFO hears it from you first.

Next in the series: the CPO and the services side, owning the loop from FDEs and customer success back into product.

One thing to try this week

Take your top AI feature. Ask finance for last month's serving cost for that feature alone, and divide it by successful outcomes. If finance can't split it out, that's your first fix, and it's yours to ask for. Then write one sentence on what you'll do with the next vendor price cut, and send it to your CFO.

Sources: Anthropic, PBC, Form S-1 Registration Statement, filed October 26, 2026, sec.gov. Tomasz Tunguz, "Anthropic S-1 Analysis: How 7 Key Metrics Stack Up," Theory Ventures, October 26, 2026, tomtunguz.com. Jason Lemkin, "5 Interesting Learnings from Anthropic's S-1 at $16.3 Billion in Half-Year Revenue," SaaStr, October 26, 2026, saastr.com. Gergely Orosz, "Uber's AI infra, from the inside," The Pragmatic Engineer, October 14, 2026, newsletter.pragmaticengineer.com. The compute-per-revenue-dollar ratios are my arithmetic from the S-1's figures.

Part of the running argument on Product Leadership: the product org chart was calibrated to a build cost that collapsed, and the CPO's job now includes the margin line, because every lever on it is a product decision.

Related answer: Where should AI inference cost sit on the P&L?