Anthropic’s inference business could hit 88% margins, SemiAnalysis estimates

In a report published in early October 2026, SemiAnalysis estimated that Anthropic’s inference business could achieve margins as high as 88% once compute costs are accounted for.

Where the money actually comes from

Roughly 75-85% of Anthropic’s annual recurring revenue comes from API services, which means developers and businesses paying per use. API margins alone are estimated to exceed 80%, according to the research.

Consumer subscriptions make up around 10% of revenue but consume more than 40% of inference compute, per the findings. Yet subscriptions still reportedly yield compute margins of approximately 50%. Subscription users get over four times more compute per dollar than API buyers, and on certain workloads the tiers may deliver up to five times the API value compared to OpenAI.

Subscriptions reduce blended revenue per megawatt by approximately $36 million, according to the analysis, but do not turn negative on a gross-margin basis under realistic utilization conditions.

From deep red to mid-60s

Anthropic’s overall gross margins sat at negative 94% in 2024, per the research. By mid-2026, those margins had climbed into the mid-60% range. The improvement is attributed to optimization in inference operations.

Prior projections also estimated the company was on course for more than $1 billion in quarterly operating profit by Q3 2026.

A different bet than OpenAI

SemiAnalysis’s findings suggest Anthropic may reach sustained profitability earlier than its peers. The reasoning centers on business model design: Anthropic leans API-first and usage-based, while OpenAI relies on a much larger free-user base. Free users bring scale and brand awareness, but they do not pay for the compute they consume, which may limit how efficiently that audience converts into revenue.

What this means for the AI economy

There are caveats worth keeping in view. The 88% figure is an estimate of what the inference business could achieve, and it covers inference specifically rather than the company’s total spending. Training frontier models remains enormously expensive, and post-compute inference margins do not capture that full picture.

The subscription dynamics also bear watching. A product segment that uses more than 40% of compute while bringing in roughly 10% of revenue is sensitive to changes in how people use it. If subscribers start running heavier workloads, that roughly 50% margin could come under pressure, though the analysis indicates it stays positive under realistic utilization.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.