AI profit structure: For every $100 in revenue generated by model companies, $35-40 flows to cloud providers, bringing them $10-20 in operating profit.

AI profit structure: For every $100 in revenue generated by model companies, $35-40 flows to cloud providers, bringing them $10-20 in operating profit.

For every $100 earned by an AI model company, approximately $35-40 flows to the three major cloud providers—AWS, Azure, and GCP—in the form of inference computing fees. These cloud providers then earn $10-20 in operating profit from this revenue, corresponding to an operating profit margin of approximately 35%-45%.

This is a key finding from Barclays' AI industry unit economics research report released on August 28.

Meanwhile, the AI Lab's paid inference profit margin has surged from a dozen percentage points in 2025 to over 50%-65% in 2026, with adjusted gross margin increasing by 30-50 percentage points year-on-year. This is driven by enterprise clients and the Agentic workflow becoming a "must-buy" product in the market.

Barclays analyst Ross Sandler believes that current actual profit margins may be higher than the report estimates, but are expected to gradually decline as frontier competition intensifies and computing power supply increases.

Two different financial profiles within the same industry

Barclays constructed two hypothetical frontier lab models to break down the profit differences. "Lab A" derives approximately 70% of its revenue from APIs and 30% from subscriptions; "Lab B" is the opposite, with 80% coming from subscriptions and only 20% from APIs.

APIs inherently offer higher inference profit margins than subscriptions. Adding to this the differences in training cost amortization and partner revenue sharing, the adjusted gross margins differ by 17 percentage points – approximately 55% for Lab A and approximately 38% for Lab B.

The revenue recognition methods further amplify this "distortion." Lab A recognizes indirect API revenue using the gross method, while Lab B uses the net method, or even completely omits indirect API revenue from strategic partners' operations. Barclays draws a parallel between Uber and Lyft: the same core business, but drastically different reported figures due to different financial accounting methods. Once AI labs begin disclosing GAAP statements, investors will need to disentangle these differences in accounting methods when making cross-company comparisons.

Inference profit margins soar

Specifically, for each product line—

The inference profit margin for subscription products (such as Claude Code and Codex) is estimated to be around 70%, the lowest among the three product lines. This is because the AI Labs is willing to subsidize token costs to retain users. Subscriptions are typically based on monthly fees and have usage caps. Recently, the frequency of cap resets has increased significantly, likely due to a combination of retention pressures and improvements in model efficiency.

Direct APIs were the earliest and most profitable business model for the AI Labs. Developers like Cursor and Figma pay based on token consumption. Barclays estimates that current API inference profit margins exceed 80%. Improved model token efficiency (reduced tokens required to complete the same task), increased nominal API pricing, and infrastructure optimizations for inference services—quantization, speculator technology, and next-generation computing power—are all continuously releasing profit potential. Barclays states that API inference profit margins in Q2 2026 are significantly higher than those shown in the report's charts, but expects them to revert downwards at some point in the future.

The end-user experience of indirect APIs is the same as that of direct APIs, but the billing relationship lies between the user and the cloud provider. As the proportion of indirect APIs in revenue increases, the differences in revenue recognition methods among different labs will further widen the comparability of financial statements.

How much do cloud vendors earn?

Of every $100 in revenue from the AI labs, Lab A receives $35 from cloud providers, contributing approximately $11.80 in profit after deducting infrastructure costs, resulting in an operating profit margin of about 34%. Lab B, due to revenue sharing with strategic partners (accounting for 20% of revenue, with a cumulative cap), receives more from the cloud provider—$41 in revenue, $19.10 in profit, and an operating profit margin of 47%.

Barclays points out that the revenue sharing inflates the apparent profit margins of cloud vendors; after removing the revenue sharing, the actual profit per token remains the same. This revenue sharing is expected to return to zero after 2028.

Agentic subscription products create additional value for cloud vendors. These stateful runtime products often need to access upper-layer software resources such as databases, resulting in higher value per unit of revenue. In some scenarios, there are even revenue-sharing arrangements between cloud vendors and AI labs.

From training-driven to reasoning-driven

Barclays estimates that total revenue from its AI labs will grow from $7 billion in 2024 to $137 billion in 2026, reaching $690 billion in 2028. The year-end ARR figures are even more aggressive: approximately $200 billion by the end of 2026 and approximately $782 billion by the end of 2028.

Currently, training expenses still account for about 48% of AI lab revenue, meaning that almost every dollar of lab revenue corresponds to nearly one dollar of cloud provider revenue. However, the proportion of training costs is rapidly declining—from 96% in 2024 to an estimated 35% in 2027 and 30% in 2028. As inference profits gradually surpass training expenses, the overall profitability of AI labs will continue to improve.

The proportion of AI revenue from cloud vendors in their AI lab revenue is also declining: from 153% in 2024 to 90% in 2026, and is projected to drop to 73% in 2028. Barclays predicts that AWS, Azure, and GCP will maintain their respective shares of computing power spending for AI labs over the next two years, but starting in 2028, guaranteed AI infrastructure projects will become operational and will be the preferred choice for AI labs—as a result, the three major cloud vendors may gradually lose market share in training and inference.

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.