On Monday an amended registration statement reached the US securities regulator, and with it came a detailed look inside a frontier AI company. Anthropic, which plans to list on Nasdaq under the ticker ANT, has put its economics and its control structure on paper. Most coverage will quote the headline figures. The more useful exercise is to ask what each figure actually measures.
The first number is growth. Revenue for the nine months to 30 September 2026 was $33.24 billion, roughly twelve times the same period a year earlier, with the third quarter alone near $17 billion. The filing is candid about the cost of that speed: it lists customer concentration as a risk. A buyer should read that as a reminder that the vendor’s own revenue depends on a handful of decisions made in other boardrooms.
The second number is the loss: $104 billion over the same nine months. It looks alarming, but it is two figures welded together. Much of it comes from fair-value changes on convertible notes and derivative liabilities, a largely non-cash effect that typically grows when the value of the company’s equity grows. What is not an accounting effect is the cost of compute, which rose by about $14.4 billion, an increase of 329 percent. The discipline is to separate the loss you can discount from the cash that is really leaving for infrastructure.
The third number is the one to remember: about $895 billion in purchase commitments and leases, almost all of it compute and infrastructure. That is not an expense yet. It is a promise to pay, spread across future years, and it turns the business into a fixed-cost machine. If demand keeps compounding, the commitments will look prescient. If demand flattens, they become a weight. Either way they shape pricing, because a company carrying obligations of that size will tend to price for utilisation rather than discount its way into a market.
Then there is the share. The company is a Delaware public benefit corporation, so its directors must weigh a stated public mission alongside shareholder returns. A Long-Term Benefit Trust elects a majority of the board. And the founders hold a special Class F share carrying 50.1 percent of the voting power but no economic rights. In plain terms, control and ownership have been separated on purpose: the people who steer cannot be paid more for steering, and the people who are paid cannot outvote them. Public investors are being asked to buy into that design.
What should an enterprise buyer do with all this? Treat the filing as a counterparty file, the way a bank treats a borrower. Check three things: how dependent the vendor is on its largest customers, how much of its cost base is committed rather than flexible, and who can change its direction. Then write the answers into your own contracts: portability of prompts and data, the ability to switch models without rebuilding workflows, and pricing protections that outlast a single funding cycle.
The deeper point is that an AI supplier is no longer a software vendor in the old sense. It is closer to a utility with a research lab attached: heavy fixed assets, lumpy demand and unusual governance. A utility is safe to depend on when you understand its economics and have a second source. The filing gives you the first. Building the second is your job.
Think mathematically. Engineer intelligently. Lead strategically. Build responsibly. Scale globally. Create real business value through artificial intelligence.
Think mathematically. Engineer intelligently. Lead strategically. Build responsibly. Scale globally. Create real business value through artificial intelligence.
