While AI model companies might appear to be raking in massive profits, the real beneficiaries seem to be the cloud service providers powering them. A recent study by Barclays has shed light on the intricate profit distribution within the AI value chain, revealing that for every $100 generated by AI model companies, a significant portion, approximately $35 to $40, flows directly to the major cloud platforms – Amazon AWS, Microsoft Azure, and Google GCP – primarily for inference compute costs.
Cloud Providers See Substantial Margins
These cloud giants are not just facilitating AI innovation; they are handsomely profiting from it. The Barclays report indicates that cloud service providers can achieve operating profit margins of $10 to $20 per $100 of AI model company revenue, with margins potentially soaring as high as 35% to 45%. This insight, derived from Barclays’ unit economics model for the AI industry released on August 28th, offers a clearer picture of where value is accumulating in the burgeoning AI ecosystem.
AI Labs Experience Profit Margin Surge
The study also highlights a dramatic increase in the profitability of AI labs’ paid inference services. Margins have reportedly surged from just over 10% in 2025 to an impressive 50% to 65% or even higher in 2026. This represents a substantial year-over-year increase in adjusted gross margins of 30 to 50 percentage points.
According to Barclays analyst Ross Sandler, the primary drivers behind this significant profit margin expansion are enterprise clients and the rise of what are termed ‘agentic workflows.’ These advanced products have become essential offerings in the market. Sandler suggests that the actual profit margins for AI labs might even exceed these estimates. However, he cautions that as competition among cutting-edge models intensifies and compute supply continues to grow, these margins are expected to gradually normalize over time.
Business Structure Dictates Gross Margin Discrepancies
To illustrate the financial variations among different AI labs, Barclays constructed models for two hypothetical leading AI organizations. ‘Lab A’ derives approximately 70% of its revenue from APIs and 30% from subscriptions, while ‘Lab B’ operates in reverse, with 80% of its income dependent on subscriptions and only 20% from APIs.
Given that API businesses inherently offer higher inference profit margins than subscription services, coupled with differences in how training costs are amortized and revenue-sharing arrangements with partners, these two hypothetical labs exhibit a striking 17 percentage point difference in adjusted gross margins: Lab A boasts around 55%, whereas Lab B is at a lower 38%.
Furthermore, variations in revenue recognition practices amplify these discrepancies. Lab A recognizes indirect API revenue on a gross basis, whereas Lab B uses a net basis, sometimes even excluding indirect API revenue generated by strategic partners entirely. Barclays draws a parallel to the difference between Uber and Lyft, where similar core businesses can present vastly different figures on financial statements due to accounting methods.
The report emphasizes that as AI labs begin reporting under Generally Accepted Accounting Principles (GAAP), investors must first understand these accounting nuances to conduct meaningful cross-company comparisons.
Profitability Across Different Product Lines
A deeper dive into specific product lines reveals that subscription products, such as Claude Code and Codex, have the lowest estimated inference profit margins among the three main product categories, hovering around 70%. This is attributed to AI labs being willing to absorb some token costs to ensure user retention, often offering monthly subscriptions with usage limits. Notably, an increased frequency of usage limit resets may reflect both user retention pressures and ongoing improvements in model efficiency.
Direct API, the earliest commercial model for AI labs, remains the most profitable business line. Tools like Cursor and Figma charge developers based on token consumption. Barclays estimates that current API inference profit margins exceed 80%. Factors contributing to this high margin include increasing model token efficiency (requiring fewer tokens for the same task), rising nominal API prices, and continuous optimization of inference service infrastructure through techniques like quantization, speculative decoding, and next-generation compute hardware.
The report specifically notes that API inference profit margins in the second quarter of 2026 were significantly higher than depicted in charts, but are expected to return to more normalized levels eventually.
Indirect APIs offer a user experience similar to direct APIs, but establish a direct billing relationship between the user and the cloud service provider. As indirect APIs constitute a growing portion of AI labs’ total revenue, differences in revenue recognition practices will further widen the comparability gap in their financial statements.
Unpacking Cloud Service Provider Earnings
Examining the profit structure of cloud service providers reveals that for every $100 in revenue generated by an AI lab, Lab A corresponds to approximately $35 in cloud service revenue. After deducting infrastructure costs, cloud providers can realize a profit of about $11.80, translating to an operating profit margin of roughly 34%.
Lab B proves to be even more lucrative. Due to a revenue-sharing mechanism with strategic partners that involves 20% of revenue with an aggregate cap, cloud providers can gain approximately $41 in revenue for every $100 of AI revenue from Lab B. This results in a profit of $19.10, yielding an operating profit margin as high as 47%.
Key Metrics Comparison:
- Cloud Service Revenue per $100 AI Revenue: Lab A: $35, Lab B: $41
- Cloud Service Provider Operating Profit: Lab A: $11.80, Lab B: $19.10
- Cloud Service Provider Operating Profit Margin: Lab A: 34%, Lab B: 47%
Lab B’s higher profit margins stem from the strategic partner revenue-sharing mechanism. Barclays emphasizes that this arrangement inflates the apparent profit margins for cloud service providers. If this factor were excluded, the actual per-token profit for cloud service providers would be identical for both labs. This revenue-sharing arrangement is projected to diminish to zero after 2028.
Additionally, subscription products for agentic workflows create additional value for cloud service providers. These state-aware runtime products often require the invocation of upper-layer software resources, such as databases, making them more valuable per unit of revenue. In some instances, revenue-sharing agreements between cloud providers and AI labs further enhance the cloud providers’ actual earnings.
AI Industry Shifts from ‘Training’ to ‘Inference’
On an industry-wide scale, Barclays forecasts AI lab revenue to surge from $7 billion in 2024 to $137 billion in 2026, and reach a staggering $690 billion by 2028. The growth trajectory is even more aggressive when measured by annualized recurring revenue (ARR) at year-end, projected to hit approximately $200 billion by the end of 2026 and potentially $782 billion by the end of 2028.
Currently, training expenditure still accounts for about 48% of AI lab revenue, meaning that for nearly every dollar earned by AI labs, a dollar is spent on cloud services. However, the proportion of revenue attributed to training costs is rapidly declining. It is expected to drop from 96% in 2024 to 35% in 2027 and further to 30% in 2028.
The significance of this trend lies in the increasing profitability of AI labs as inference-generated profits begin to outpace training expenditures. This marks a fundamental shift in the industry’s focus, moving from a ‘training-driven’ to an ‘inference-driven’ model.
Concurrently, the ratio of cloud service provider revenue to AI lab revenue is also decreasing, falling from 153% in 2024 to 90% in 2026, and projected to further decline to 73% by 2028.
Barclays anticipates that the market share of AWS, Azure, and GCP in AI lab compute spending will remain largely stable over the next two years. However, by 2028, the industry landscape could undergo a significant transformation. The advent of asset-backed financing for AI infrastructure projects is expected to make them a more attractive option for AI labs. This could lead to the major cloud providers gradually losing market share in both AI training and inference, potentially triggering a reshuffling of current profit structures.
