Anthropic's 77% margin bets on enterprises fixing their own governance

Of the two frontier AI labs racing to the stock market, Anthropic is the one investors have cast as the calm bet. It sells mostly to enterprises. It brands itself around safety. And when analysts model its valuation, they give it a lower multiple than OpenAI, roughly 20 times run-rate revenue against OpenAI’s 34, which is the market’s way of saying an enterprise-heavy, recurring-revenue business looks less risky than a consumer subscription base [reported]. Anthropic is about to become the first pure AI-safety company to go public [verified]. The story writes itself: the grown-up in the room is cashing out.

Here’s what actually happened, for anyone who skipped the filings. A company that wants to go public files a document called an S-1 that lays its finances bare for the first time. Anthropic’s is in process, targeting a Nasdaq listing in October 2026 [reported]. The number everyone was waiting for is gross margin, the share of revenue left after the direct cost of serving each customer. For a software company you’d want that north of 70%. Anthropic disclosed 40% today, with a plan to reach 77% by 2028 [verified]. Revenue is running at a $65B annual rate, up roughly sevenfold in a year [verified]. So the whole valuation now hangs on one forward number: that 40 climbs to 77 in about two years.

That climb is the entire bet, and it needs something the safety brand can’t deliver. Anthropic’s safety and governance edge is aimed at exactly the barrier that caps its own enterprise demand, because that barrier lives in the customer’s org chart, not in the model. What it sells as its moat sits on the wall that blocks its own growth.

Walk the margin math and you’ll see why the two connect. At Anthropic’s scale, gross margin on AI is mostly a utilization story. Margins expand when expensive computers run hot against steady, high-value work that keeps coming. The work that does that is agents in production: AI systems running continuously inside a company, on real contracts, at volume. A 40%-to-77% path assumes that kind of enterprise usage compounds fast. I haven’t seen Anthropic’s own breakdown of how much of the jump comes from usage growth versus falling cost per query, so treat the utilization framing as my read, not a disclosed fact. But the volume has to come from somewhere, and enterprise agents in production are where the bull case puts it.

Now the problem. That exact usage is stuck, for reasons no model vendor can fix. McKinsey surveyed more than 10,000 executives and found 88% of companies have deployed AI somewhere while only 39% report a real hit to profit, a 49-point gap between using it and getting value from it [reported]. Deloitte’s survey of 3,235 leaders found 21% have mature governance for AI agents against 74% who expect to deploy them widely by 2027, a 53-point readiness gap [verified]. Researchers at Harvard Business Review have a name for the pattern: pilot-rich, transformation-poor. Companies run dozens of trials and almost none scale into the operating model [reported]. Their diagnosis is blunt. This is a coordination failure inside the buyer, and it needs organizational change, not a better tool.

Read that against the margin bet and the collision is exact. The demand Anthropic’s trajectory needs exists as intent and stalls as deployment. Enterprises want the agents. They can’t operationalize them past the pilot, because the gate is on their side of the contract: who’s allowed to let an agent act, and who’s accountable when it does. A safety-first model doesn’t close a customer’s governance gap. Anthropic can ship the most auditable, most aligned frontier model on the market and still watch its enterprise revenue pool in pilots that never compound.

The honest objection is that the regulators are finally showing up, and that could be the tailwind that breaks the ceiling. Singapore’s IMDA published the first full governance framework for autonomous AI agents in January 2026. NIST, the US standards body, started an agent-standards effort in February. Microsoft shipped a runtime governance toolkit in April [verified]. Maybe the plumbing arrives and the deployments follow. But today’s evidence points the other way. Even under the new Singapore rules, the routine layer of agent decisions, the ordinary sub-threshold calls where most of the actual work happens, sits in a governance vacuum that neither the regulation nor the runtime tooling closes [reported]. The rules are landing on the boundary cases and leaving the everyday middle uncovered, and the everyday middle is where production volume lives. The ceiling isn’t lifting on the schedule the margin needs.

There’s a fair version of the bull case that survives all this, and I want to be straight about it. Gross margin can climb on the cost side alone: cheaper hardware over time, more efficient models, better pricing. If the 77% is mostly a falling-cost story, the demand ceiling barely matters and there’s no collision. That’s the real question, and I can’t settle it from the outside. The one clue we have is that Anthropic’s own investor case names “enterprise governance readiness” as a variable its trajectory depends on. If the company is leaning on the buyer getting its house in order, then it’s betting on the one thing its product can’t move.

Which flips the safe-bet story on its head. The market prices Anthropic’s 85% enterprise concentration as the reason it’s lower-risk than OpenAI [reported]. But if the ceiling holds, that concentration is maximum exposure to the single variable the vendor can’t fix. OpenAI’s crowd of consumer subscribers doesn’t need a governance regime to keep paying. Anthropic’s enterprise-and-safety purity, the thing that reads as caution, is what binds it most tightly to a ceiling on the buyer’s side.

So if you’re holding this IPO as an input and not a headline, put the 40% aside. It’s real, and the company is right that it’s survivable. The number that decides everything is the 77%. When the full S-1 is public, look at the margin bridge and ask one thing: does the jump lean on usage and volume, or on cost per query falling? If it’s cost, this is a footnote and you can watch the demand line at your leisure. If it leans on volume, then the safety company’s valuation is a wager that enterprises fix their own governance by 2028. And the first quarter enterprise agent revenue shows up as pilots that never scaled, the 77% stops being a plan and turns back into the open question the 40% was supposed to settle.