For every $100 an AI model company collects, roughly $35 to $40 flows out the door to Amazon, Microsoft and Google. That is the arithmetic at the center of a new Barclays report on the AI industry’s economics, which lays out in unusually concrete terms how the boom’s profits are distributed.
The three cloud providers, Amazon Web Services, Microsoft’s Azure and Google Cloud Platform, capture the money as inference computing fees, the cost of running the models for customers. Barclays estimates the clouds make $10 to $20 of operating profit on each $100 of AI company revenue, an operating margin of 35 to 45 percent on that business.
Meanwhile, the AI labs’ own paid-inference businesses have become dramatically more profitable. Margins in that business climbed from low double digits in 2025 to 50 to 65 percent or higher in 2026, with adjusted gross margin up 30 to 50 percentage points year over year, the report said. The improvement reflects better model efficiency, higher utilization and pricing power in a market where demand still outstrips supply.
Barclays analysts cautioned that actual margins may run above the report’s estimates, which are based on public disclosures and industry checks. But they expect margins to ease over time, as frontier-model competition intensifies and computing supply keeps growing. The direction of travel, the report argues, is downward from today’s extraordinary levels, toward something closer to normal software economics.
The numbers help explain a divergence that has puzzled investors. Cloud stocks have re-rated upward as AI infrastructure spending flows through their income statements, while model-company valuations rest on the hope that inference margins persist long enough to justify enormous spending on training and data centers. Barclays’ report suggests both can be true at once: the clouds earn reliably on throughput, the labs earn on margin, and the open question is who holds pricing power when supply catches up with demand.
The mechanics matter. A model company charges customers for API usage, and the underlying compute is rented from the clouds. The 35 to 40 percent figure reflects list prices for GPU instances, negotiated discounts and the reserved capacity the labs buy in bulk. The labs’ own gross margins are calculated after that compute cost, which is why their reported margins have swung so sharply as they negotiate better rates and run models more efficiently.
The margin structure also explains the strategic maneuvers of the past year. Labs have signed enormous infrastructure deals to reduce their dependence on the three clouds, clouds have bought into model companies to secure their workloads, and a web of interlocking investments has formed across the industry. Barclays notes that the cloud companies’ AI revenue is partly a function of the labs’ own fundraising: money raised by a lab tends to flow, in part, back to the cloud that hosts it.
Skeptics note that the report is a snapshot, not a forecast. Inference prices have already begun falling as new capacity comes online, and the 50 to 65 percent margins may look exceptional in hindsight. The report’s own conclusion is that margins will be compressed by exactly the forces that created them: competition among model developers and an expanding supply of computing.
The report also quantifies a dynamic that has become central to the AI investment narrative: the circularity of the industry’s money. Much of the capital raised by AI labs is spent, in large part, on cloud computing from the three giants, which in turn invest some of those profits back into AI companies. Barclays does not call the arrangement a bubble, but it notes that the revenue figures on both sides of the ledger are mutually reinforcing, and that a slowdown in fundraising would show up in cloud AI revenue within a quarter or two.
Barclays’ own economists have been among the more cautious voices on AI’s near-term economic impact, and the report sits comfortably with that stance. It describes an industry that is genuinely profitable at the infrastructure layer, genuinely expensive at the model layer, and still searching for the end-customer demand that will justify both. The report’s figures will be cited on both sides of the debate: bulls will point to the clouds’ 40 percent margins, bears to the concentration of revenue in three companies that also finance their own customers.
For investors, the report offers a clearer way to think about the AI value chain. The clouds’ earnings from AI are visible, recurring and growing; the labs’ earnings are real but narrower, and exposed to a pricing war that has not yet arrived. The central finding is that AI revenue is not staying where it is earned: roughly two-fifths of every dollar an AI company collects is already someone else’s revenue, and that share will determine which parts of the industry are actually profitable when the boom matures.
