Written by: Rita

Two leading AI laboratories, both with total revenue of 100 dollars in 2026, one ultimately profits 55 dollars, while the other only 38 dollars. This 17 dollar gap does not stem from model capabilities, but from the business structure itself.

On August 27, Barclays released a report on the U.S. internet industry, systematically breaking down the unit economics of AI laboratories and hyper-scale providers. The profit margins of API products far exceed those of subscription products, and differences in business structure explain most of the gross margin gap for AI laboratories. As AI laboratories begin to disclose GAAP financial data, investors need to penetrate the revenue figures to understand these structural differences.

API profit margins exceed 65%, while subscription products are nearly 30 percentage points lower

Barclays estimates that paid inference profit margins for AI laboratories have surged to over 50% to 65% in 2026, significantly up from mid-double-digit levels in 2025. API products have the most substantial profit margins, while subscription products are relatively low, mainly because some users benefit from the subsidies on token costs provided by AI laboratories to maintain competitiveness.

Subscription products are mostly monthly fee-based or usage-based, usually with usage caps, allowing users to switch between different models, with token consumption rates varying greatly. Enterprise subscriptions are priced per license, with higher limits. APIs adopt a pure pay-as-you-go model, allowing developers (e.g., Cursor, Figma) to embed AI functions in their products, calling APIs to generate tokens. As model token efficiency improves and API pricing rises, API profit margins continue to widen.

Barclays lists a set of hypothetical data. In Laboratory A's revenue structure, 30% comes from subscriptions, 45% from direct sales of APIs, and 25% from indirect APIs. Conversely, Laboratory B derives 80% from subscriptions, with direct and indirect APIs accounting for only 10% each. The difference in inference profit margins between the two expanded from 4% and 8% in 2025 to 55% and 38% in 2026. Barclays points out that APIs and paid enterprise accounts have higher profit margins, and as the business structure tilts in this direction, the overall profit margins still have room to increase.

The income recognition method for indirect APIs creates a misleading report

The essence of the business for indirect APIs is the same as that of direct sales APIs, where users call the AI laboratory's API interfaces to generate tokens, but billing and customer relationships are handled by hyper-scale providers, meaning users do not interact directly with the AI laboratory.

The issue lies in the method of income recognition. AI Laboratory A recognizes indirect API income on a gross basis, and this portion continues to expand. Laboratory B has smaller related income, recognizing it on a net basis, and does not recognize income at all when its strategic partner operates indirect APIs.

Barclays compares this to Uber and Lyft. Although the core businesses are the same, the methods of income recognition for total revenue and net revenue differ, leading to significant discrepancies in the companies' reports. As the share of indirect APIs rises in the overall revenue of AI laboratories, this recognition method will produce significant distortions in revenue figures, potentially misleading the market's judgment on the relative progress of various AI laboratories.

Out of every 100 dollars of AI revenue, 35 to 40 dollars flow to cloud vendors

Barclays estimates the flow of AI laboratory income along the industry chain. In 2026, out of every 100 dollars of AI laboratory income, approximately 35 to 40 dollars will ultimately convert into inference revenue for hyper-scale providers, which then generate about 10 to 20 dollars of operating profit, corresponding to an operating profit margin of about 35% to 45%.

The flow varies under different business models. In the direct sales API model, approximately 35 dollars out of every 100 dollars of AI laboratory income flows to hyper-scale providers, generating about 11.8 dollars of operating profit, with a profit margin of about 34%. For Laboratory B, which has a higher proportion of subscriptions, the income flowing to hyper-scale providers due to revenue sharing with strategic partners reaches 41 dollars, with operating profit of about 19.1 dollars and a profit margin of about 47%. Barclays emphasizes that after excluding revenue-sharing factors, the actual profit per token remains consistent across companies.

Inference profits will eventually surpass training costs, with cloud vendor shares expected to decline in 2028

The majority of current income for AI laboratories flows into the revenue statements of hyper-scale providers primarily because training costs constitute a very high proportion of total costs. Barclays predicts that inference profits will gradually surpass training costs, driving up the profit margins of the AI laboratories themselves.

The landscape is changing. Starting in 2028, AI laboratories' self-built backstopped AI infrastructure projects will gradually come online, becoming the preferred source of computing power. Barclays expects that AWS, Azure, and GCP will continue to control the major share of AI laboratories' computing expenditures for the next two years, after which the self-built infrastructure will gradually take precedence.

Barclays' conclusion is clear. The unit economics of AI laboratories are improving rapidly, but factors such as business structure, income recognition methods, and partnership relationships create significant apparent differences. Investors comparing different AI laboratories should not only look at revenue figures but need to penetrate down to the underlying variables such as product structure, the proportion of APIs versus subscriptions, and methods of income recognition. The competition of model capabilities has been replaced by the competition of business models.

Disclaimer

This article is a compilation and interpretation of a third-party brokerage research report (Barclays, August 27, 2026) by Chaos Research, combined with the整理 of publicly available market information. The ratings, target prices, profit forecasts, and related judgments cited in the text are the views of the brokerage analysts and only represent the positions of their respective organizations, do not reflect the views of Chaos Research, and do not constitute any investment advice.

The market carries risks, and decision-making must be independent. This article should not be used as a basis for buying or selling any securities.