Barclays: For Every $100 AI Model Companies Earn, $35-$40 Flows to the Big Three Cloud Giants

IT Home, August 30 — On the surface, artificial intelligence (AI) model companies appear to be making massive profits, but the ones truly thriving may be the cloud computing service providers behind them.

A latest research report from Barclays breaks down the profit distribution within the AI industry chain. The study found that for every $100 in revenue generated by model companies, $35 to $40 flows to the three major cloud platforms—Amazon AWS, Microsoft Azure, and Google GCP—in the form of inference computing costs.

Cloud service providers can earn $10 to $20 in operating profit from this portion of revenue, corresponding to an operating margin of up to 35% to 45%. This conclusion comes from Barclays' AI industry unit economics research report released on August 28, allowing the public to more clearly observe where value in the AI industry chain ultimately flows.

AI Lab Margins Improve Significantly. The report also shows that the paid inference margins of AI labs have surged from just over 10% in 2025 to 50% to 65% or even higher in 2026; adjusted gross margins increased by 30 to 50 percentage points year-over-year.

Barclays analyst Ross Sandler noted that the primary factors driving this significant margin improvement are enterprise clients and so-called "agentic workflows," which have become "must-buy" products in the market.

He further believes that the actual margins of AI labs may currently be higher than the estimates in the report. However, as competition among frontier models intensifies and computing power supply continues to increase, this metric is expected to eventually decline gradually.

Business Structure Determines Gross Margin Disparity. To help investors better understand the financial differences between various AI labs, Barclays constructed two hypothetical frontier AI labs for comparison.

"Lab A" derives about 70% of its revenue from APIs and 30% from subscriptions; "Lab B" is the exact opposite, relying on subscriptions for 80% of its revenue, with APIs accounting for only 20%.

Because API business inherently possesses higher inference margins than subscription business, coupled with differences in training cost amortization methods and partner revenue sharing, the adjusted gross margins of these two hypothetical labs showed a staggering 17-percentage-point gap: Lab A is approximately 55%, while Lab B is only 38%.

Differences in revenue recognition methods further amplify this "distortion."

Lab A recognizes indirect API revenue on a gross basis, while Lab B uses the net method, or even entirely excludes indirect API revenue operated by strategic partners. Barclays compares this situation to the difference between Uber and Lyft: their core businesses are essentially the same, but due to different accounting treatments, the numbers appearing on their financial statements may be completely different.

The report cautions that as AI labs begin to disclose financial statements under Generally Accepted Accounting Principles (GAAP), investors need to first understand these accounting differences before making cross-company comparisons; otherwise, meaningful comparisons will be difficult.

Margins of Different Product Lines.

Further breaking down specific product lines reveals that subscription products, such as Claude Code and Codex, have an estimated inference margin of about 70%, the lowest among the three main product lines.

The reason is that AI labs are willing to absorb some token costs in exchange for user retention. These subscriptions typically charge monthly while setting usage limits. Notably, the frequency of usage limit resets has clearly increased recently, which may simultaneously reflect the user retention pressure faced by AI labs and the continuous improvement in model efficiency.

Direct API is the earliest business model for AI labs and also the highest-margin business line. Developer tools like Cursor and Figma charge based on the number of tokens consumed by users. Barclays estimates that current API inference margins have exceeded 80%.

As token efficiency continues to improve—meaning fewer tokens are needed to complete the same task—combined with nominal API price increases and continuous optimization of inference service infrastructure, including technologies like quantization, speculative decoding, and next-generation computing power, greater profit margins are being continuously unlocked.

The report specifically notes that the API inference margin in the second quarter of 2026 is actually much higher than the level shown in the charts, but it is expected to return to normal levels at some point in the future.

Indirect API provides end users with essentially the same experience as direct API, but a direct billing relationship is established between the user and the cloud service provider. As the proportion of indirect API in total AI lab revenue continues to rise, the differences in revenue recognition methods among labs will further widen the comparability gap between their financial statements.

How Much Are Cloud Providers Actually Making.

Further breaking down the profit structure of cloud service providers reveals that for every $100 in revenue created by an AI lab, Lab A corresponds to approximately $35 in cloud service revenue. After deducting infrastructure costs, the cloud service provider can obtain about $11.80 in profit, corresponding to an operating margin of about 34%.

Lab B's situation is more lucrative. Due to a strategic partner revenue-sharing mechanism—which involves 20% of revenue and has a cumulative cap—for every $100 in AI revenue corresponding to Lab B, the cloud service provider can earn about $41 in revenue, of which profit reaches $19.10, with an operating margin peaking at 47%.

Metric | Lab A | Lab B Cloud service revenue per $100 of AI revenue | $35 | $41 Cloud provider operating profit | $11.80 | $19.10 Cloud provider operating margin | 34% | 47%

It is worth noting that Lab B's higher margin comes from the strategic partner revenue-sharing mechanism, which involves 20% of revenue and sets a cumulative cap.

Barclays emphasizes that this revenue-sharing mechanism inflates the cloud provider's apparent margin. Excluding this factor, there is no difference in the actual per-token profit for cloud providers between the two labs. This revenue-sharing arrangement is expected to gradually decrease to zero after 2028.

In addition, agentic subscription products can also create extra value for cloud service providers. These stateful runtime products often need to call upon upper-layer software resources like databases, resulting in higher value per unit of revenue. In some cases, revenue-sharing arrangements exist between cloud providers and AI labs, further increasing the cloud provider's actual earnings.

AI Industry Shifting from "Training" to "Inference". Looking at the scale of the entire industry, Barclays expects AI lab revenue to grow from $7 billion in 2024 to $137 billion in 2026, reaching $690 billion in 2028.

Calculated by year-end annualized recurring revenue (ARR), the growth rate is even more aggressive: expected to reach approximately $200 billion by the end of 2026, and potentially $782 billion by the end of 2028.

Currently, training spending still accounts for about 48% of AI lab revenue, meaning that for every $1 earned by an AI lab, it corresponds to nearly $1 in revenue for cloud service providers.

However, the ratio of training costs to revenue is dropping rapidly, from 96% in 2024 to an expected 35% in 2027, and further down to 30% in 2028.

The significance of this trend is that as profits generated by inference business gradually exceed training expenditures, the overall profitability of AI labs will continue to improve, and the industry's focus will gradually shift from "training-driven" to "inference-driven."

At the same time, the ratio of cloud providers' AI revenue to AI lab revenue is also declining, from 153% in 2024 to 90% in 2026, and is expected to further drop to 73% in 2028.

Barclays expects that the share of AWS, Azure, and GCP in AI lab computing expenditures will largely hold steady over the next two years, but by 2028, the industry landscape could reach a turning point.

By then, asset-backed financing AI infrastructure projects will gradually come online and become the preferred choice for AI labs. This means that the current cloud "Big Three" may gradually lose share in both the AI training and inference markets simultaneously, and the current profit landscape will face a reshuffle.