← Back to feed

Anthropic posted $16.8 billion in third-quarter revenue, according to reporting from the Financial Times. The number is large. The trajectory behind it is more complicated.

Quarter on quarter, revenue grew approximately 45 percent — from $11.5 billion in Q2 to $16.8 billion in Q3. That same metric came in above 100 percent between Q1 and Q2, when revenue more than doubled from roughly $4.8 billion. The deceleration is significant enough that investors who had projected the company would pass $100 billion in annualized revenue by year-end are now looking at SemiAnalysis estimates closer to $74 billion annualized.

The distillation cutback

One reason for the deceleration is voluntary. Competing AI labs had been training their own models on Claude’s outputs, a practice called distillation, which Anthropic allowed under prior terms. That activity had accounted for 5 to 7 percent of Anthropic’s total revenue. The company cut off that stream materially during Q3 — trading short-term revenue for a competitive moat on its training data.

The strategic logic is straightforward: letting rivals buy their way to a better model at cost is a bad deal once you realize how much you are subsidizing them. The revenue hit is real, but so is the rationale.

Compute economics

Anthropic told investors it earns $1.70 for every $1 spent on compute. At face value, that implies compute costs are roughly 59 percent of revenue. The company reports an 85 percent adjusted gross margin, but executives have flagged that figure excludes a significant portion of compute spending depending on accounting treatment. The gap between 59 percent compute share and an 85 percent gross margin is where the definitional choices live.

The company reported adjusted profitability for a second consecutive quarter with an 11 percent adjusted operating margin. Executives warned investors that heavy spending on compute expansion and biomedical research could change that trajectory quickly.

Customer pressure

Two of Anthropic’s major enterprise customers, Meta and Microsoft, have moved toward building and deploying their own in-house AI tooling, reducing reliance on third-party model APIs. Price competition with OpenAI and cheaper Chinese model providers — particularly DeepSeek — has also compressed what existing customers pay per token. Neither development is terminal, but both put pressure on the revenue-per-customer line.

The IPO window

Bankers have reportedly floated a public offering that would raise more than $100 billion at a valuation of approximately $2 trillion. That would surpass SpaceX’s June 2026 listing, which set the record for the largest offering ever at $1.77 trillion. An early S-1 draft devotes almost a third of its length to risk factors, including AI models behaving unpredictably or maliciously and what Anthropic itself describes as possible existential risks to humanity — language that is unusual even by modern technology IPO standards.

Whether the market will price that risk premium in or ignore it the way it has ignored similar disclosures from other frontier labs remains to be seen.