According to the Financial Times, Anthropic has reportedly told its investors it is targeting profitability in the second quarter, with a gross margin above 80% on roughly $11.5 billion in revenue. The revenue figure is impressive. The one that matters is the margin, because it answers the question that has been asked for three years: can serving a model be profitable?
We have been following this story since the announcement of the first operating profit in August and the preparations for a stock market listing. Here is the piece of information that was missing.
Why gross margin is the right indicator
Gross margin measures what is left of one euro of revenue once the direct cost of production has been paid. For an AI lab, that direct cost is essentially inference: the electricity and hardware needed to serve each request.
This is exactly the question we have been asking for months and that nobody was answering. We wrote it in our article on the two costs of AI: inference is a charge that recurs with every use, unlike training, which is a one-off investment.
A gross margin above 80% means that serving a request costs less than a fifth of what it brings in. If that figure is confirmed, the argument that these companies are structurally selling at a loss falls apart.
The caveats, which are significant
These are projections communicated to investors. This is neither an observed result nor an audited document. A company preparing a stock market listing presents figures to people it is trying to convince.
Gross margin is not profit. It excludes research, salaries, marketing and, above all, the amortisation of model training. A company can post 80% gross margin and lose money massively.
The scope of the direct cost is an accounting choice. What you put in the cost of production and what you put elsewhere changes the margin without changing anything about the economic reality. This is precisely the kind of point that an audited prospectus clarifies, and that an investor presentation does not.
If the gross margin is high, the question shifts immediately: where does the money go? We documented a twenty-year infrastructure lease worth $9.1 billion. A company can generate an excellent margin on every request and swallow the result in long-term commitments. That is the very model of this industry, and it is what the prospectus will have to clarify. The margin says the product is sound; it says nothing about the soundness of the whole.
What this changes in the debate
This figure, if it holds, settles one controversy and opens another.
It makes it difficult to argue that frontier AI cannot be profitable at the product level. Regarding the absence of measured productivity gains, we noted that the economic case remained to be made. It is advancing on the seller's side; it is still not advancing on the buyer's side.
This is the asymmetry to remember: a company can be highly profitable selling a tool whose return its customers cannot yet measure. That is not contradictory, and it is not sustainable indefinitely.
What to take away
What makes this information useful is not whether it is good or bad, but that it will be verifiable soon.
A company that reports 80% gross margin to investors before a stock market listing will have to publish audited accounts. We wrote that this would be the most illuminating event of the year, and that remains true.
Until then, the right reflex is the one we apply to every announcement: note the figure, note who produced it, and wait for it to be confirmed by someone with no interest in it being true.