Barclays breaks down the AI profit structure: For every $100 in revenue of model companies, $35-40 flows to cloud providers, bringing them $10-20 in operating profit

Author | Kozmon

      Editor | Hard AI

For every $100 earned by an AI model company, about $35-40 flows to AWS, Azure, and GCP in the form of inference computing costs. Of this revenue, cloud vendors earn $10-20 in operating profit, corresponding to an operating margin of about 35%-45%.

This is the core finding from Barclays’ unit economics research report on the AI industry released on August 28.

Meanwhile, the paid inference profit margin of AI labs has soared from the teens in 2025 to over 50-65% in 2026, with adjusted gross margin up 30-50 percentage points year-on-year. The driving force comes from enterprise customers and agentic workflows becoming "must-have" products in the market.

Barclays analyst Ross Sandler believes that current actual profit margins may be even higher than the report’s estimates, but with intensifying frontier competition and increased compute supply, they are expected to gradually decline.

One industry, two financial pictures

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

APIs inherently have higher inference profit margins than subscriptions. Coupled with differences in training cost allocation and partner revenue sharing, the adjusted gross margins of the two differ by 17 percentage points—around 55% for Lab A vs 38% for Lab B.

The revenue recognition method further exaggerates this “distortion.” Lab A recognizes indirect API revenue on a gross basis, while Lab B uses a net basis, or does not recognize indirect API revenue from strategic partner-operated projects at all. Barclays compares this to UBER and LYFT: with the same core business, differences in accounting standards make reported numbers vastly different. As AI labs begin to disclose GAAP reports, investors need to adjust for these accounting differences when comparing across companies.

Inference profit margins soar

By individual product line—

Subscription products (such as Claude Code, Codex) have an estimated inference profit margin of around 70%, which is the lowest among the three product lines. The reason is that AI labs are willing to subsidize token costs to retain users. Subscriptions are usually monthly with usage caps, and lately, the frequency of usage resets has increased significantly, likely due to pressure to retain users and model efficiency improvements.

Direct APIs are the earliest business model of AI labs and also the most profitable. Developers like Cursor and Figma pay based on token consumption. Barclays estimates that the current API inference margin has surpassed 80%. Improved model token efficiency (fewer tokens needed for the same output), increases in nominal API pricing, and infrastructure-level optimization of inference services—quantization, speculator technology, and new generation compute—all continue to enhance profit potential. Barclays comments that Q2 2026 API inference margins are much higher than shown in reports, but are expected to decline at some point in the future.

Indirect API has a user experience consistent with direct API, but billing is between users and cloud providers. As indirect API accounts for a higher share of revenue, differences in revenue recognition methods at various labs will result in less comparability in financials.

How much do cloud vendors earn?

For every $100 in AI lab revenue, Lab A accounts for $35 for the cloud vendor, with $11.8 in profit after deducting infrastructure costs, equating to a 34% operating margin. Lab B, due to strategic partner revenue sharing arrangements (20% of revenue up to a cumulative cap), lets cloud vendors take more—$41 in revenue and $19.1 in profit, with a 47% operating margin.

Barclays points out that the revenue shares increase the apparent profit margin for the cloud provider, but stripping out the share, actual profit per token remains unchanged. This arrangement is expected to end after 2028.

Agentic subscription products create extra value for cloud vendors. Such stateful runtime products often require access to upper-layer software like databases, making each dollar earned more valuable. In some cases, there is also revenue sharing between the cloud vendor and the AI lab.

From training-led to inference-led

Barclays estimates that overall AI lab revenue will rise from $7 billion in 2024 to $137 billion in 2026, reaching $690 billion in 2028. The year-end ARR is even more aggressive: around $200 billion by the end of 2026, and $782 billion by the end of 2028.

Currently, training expenditure still accounts for about 48% of AI lab revenue, meaning nearly every dollar of AI lab revenue is mirrored by about a dollar in cloud vendor revenue. But the share of training cost is falling fast—from 96% in 2024, projected to fall to 35% in 2027, and 30% in 2028. And as inference profits surpass training spend, the overall profitability of AI labs will continue to improve.

The ratio of cloud vendor AI income to AI lab revenue is also declining: from 153% in 2024 to 90% in 2026, and projected to hit 73% in 2028. Barclays projects that AWS, Azure, and GCP will maintain their current shares of AI lab compute spending for another two years, but starting in 2028, guaranteed AI infrastructure investments will come online and become AI labs’ first choice—at which point the three major cloud vendors may be squeezed out of market share in both training and inference.