For every $100 that AI model companies pull in, somewhere between $35 and $40 ends up in the hands of cloud providers like AWS, Azure, and Google Cloud Platform. That’s the takeaway from a Barclays research report dated August 28, 2026, which digs into the economics of running AI models at scale and finds that the real money in AI isn’t where most people think it is.
The cloud providers aren’t just moving bits around. They’re converting that $35-$40 revenue share into operating profits of $10 to $20 per $100, translating to margins between 35% and 45%.
The inference economy takes center stage
Barclays’ analysis marks a significant pivot in how Wall Street thinks about AI spending. The conversation has long been dominated by training costs, the massive upfront compute bills required to build foundation models. But the report shifts attention to inference, the ongoing cost of actually running those models every time someone asks ChatGPT a question or an enterprise deploys an AI agent.
To illustrate the dynamics, Barclays constructed two hypothetical AI labs with different business models. Lab A, oriented more toward API-based revenue, generates roughly $35 in cloud revenue per $100 earned, with the cloud provider pocketing about $11.80 in profit at a 34% margin. Lab B, leaning toward subscription models, pushes closer to $41 in cloud revenue per $100, yielding approximately $19.10 in cloud provider profit at a 47% margin.
Subscription-based AI products tend to drive more consistent inference loads, which means steadier, higher-margin revenue for the infrastructure layer.
Growth numbers that demand attention
The bank estimates that global AI lab revenue was around $7 billion in 2024. By 2026, that figure has ballooned to an estimated $137 billion. Looking further out, the projection reaches up to $690 billion by 2028.
Profitability has kept pace with the growth. Paid inference services saw their margins climb to between 50% and 65% in 2026, up from low double digits in 2025.
The cloud provider moat, and its cracks
Barclays’ analysts flag a risk that investors should weigh carefully: AI labs may eventually build their own infrastructure. If companies like OpenAI, Anthropic, or other major model providers decide to reduce their dependence on third-party cloud services, the revenue share that currently flows to AWS, Azure, and GCP could shrink meaningfully.
The analysts project that margins could actually exceed current estimates in the near term, but they expect moderation over time as competition intensifies and the infrastructure landscape shifts.
The difference between Lab A and Lab B economics in Barclays’ framework suggests that not all cloud workloads are created equal. Providers that can attract subscription-heavy AI customers stand to earn meaningfully higher margins than those serving primarily API-driven workloads.