AI Economics
Anthropic's IPO filing: for the first time, frontier models add up
October 26, 2026
The announcement came on 26 October (Fortune): Nasdaq, ticker ANT, USD 16.3bn in revenue for the first half of the year.
Read past the loss
The reported net loss of USD 80.7bn is the number that travels, and it is the least informative one in the document. USD 72.4bn of it is accounting revaluation — not money spent.
What matters sits underneath. The gross margin is around 85 percent, operating cash flow is positive for the first time, and existing customers today spend roughly nine times what they spent a year ago.
Taken together, that is the first documented evidence that frontier models actually pay for themselves. Until now that was an assumption, defended with projections. It is now in a prospectus, where saying it wrongly has consequences.
The counterweight is in the document too
To its credit, the filing states openly that close to half of the revenue flows through the marketplaces of Amazon, Google and Microsoft — which is to say through competitors.
Our take
The economics question is settled, the distribution question is not. For anyone buying AI, the margin figure is the more useful one: it tells you there is room in the price. Frontier models are not being sold at a loss to buy market share, so the price pressure you are seeing comes from competition, not from subsidy — and competition lasts longer than a subsidy does.
Watch the ninefold expansion rather than the revenue. It says that whoever starts, stays and grows. That is the pattern of infrastructure, not of a pilot project. If your own AI spend is drifting upwards without anyone being able to say which use case is driving it, you are part of that number.
And the marketplace dependency is a lesson for your own contracts. A supplier whose access runs through its competitors has a structural weak point that has nothing to do with the quality of the model. Ask yourself the same question about your own supply chain: whose infrastructure does your AI access depend on, and what happens if that relationship changes?
Want to look at your own AI costs and contracts? We are happy to go through them with you.