Barclays Sees Big Three Clouds Taking $35–$40 per $100 of AI-Lab Revenue

AI model companies may be improving the economics of serving customers, but the largest cloud platforms still sit deep in the revenue stream. Barclays estimates AWS, Microsoft Azure and Google Cloud receive $35 to $40 in inference-compute fees for every $100 AI labs generate—and retain $10 to $20 in operating profit.

A growing pool, with less spent on training

Barclays projects total AI-lab revenue will rise from $7 billion in 2024 to $137 billion in 2026 and $690 billion in 2028. Those are forecasts, but their scale makes the allocation of revenue between model companies and infrastructure suppliers commercially significant.

The bank expects training expenditure to fall from 96% of AI-lab revenue in 2024 to 35% in 2027 and 30% in 2028. Barclays expects labs’ overall profitability to improve as inference profit overtakes training expenditure, but serving those workloads would continue to send substantial spending to the cloud platforms.

Product mix changes what a margin means

The cloud figures do not translate into a single uniform lab margin. Barclays estimates paid inference margins rose from low double digits in 2025 to 50%–65% or higher in 2026, driven primarily by enterprise customers and agentic workflows.

Three reasons comparisons can diverge

• Barclays estimates direct API inference margins exceed 80%, while subscription products carry margins of about 70%.
• Its hypothetical API-heavy lab posts an adjusted gross margin of about 55%, versus about 38% for a subscription-heavy lab—a 17-percentage-point gap.
• Revenue-recognition choices can widen the reported difference when indirect API sales are recorded on a gross basis, net basis, or not recognized by a lab.

Owned capacity is the later constraint

Barclays expects AWS, Azure and Google Cloud to hold their respective shares of AI-lab compute spending over the next two years. From 2028, it expects secured self-built AI infrastructure projects to come online and potentially reduce the providers’ share of both training and inference spending. The potential challenge is therefore a forecast tied to future capacity, not a market-share shift already recorded.