Barclays: Cloud Providers Take $35-41 for Every $100 of AI Model Company Revenue
nashnova research
Barclays' latest report dissects AI inference economics: for every $100 in model-company revenue, $35–41 flows to the big three cloud providers as compute fees, yielding them 35%–45% operating margins — but that high-margin window may be shorter than the market expects.
What exactly do cloud providers earn in the AI value chain?
AI model companies (OpenAI, Anthropic, etc.) sell inference services to customers, but inference runs on cloud infrastructure. Of every $100 in revenue, $35–41 goes to AWS, Azure, and GCP as compute fees.
Cloud providers keep $10–20 in operating profit from that fee, at 35%–45% operating margins.
This means → cloud providers play a "toll booth" role: regardless of which model company wins, rising inference demand keeps delivering high-margin revenue to the clouds.
Are model companies themselves profitable now? Why did margins spike so fast?
Barclays estimates that paid-inference margins at AI labs surged from the low teens in 2025 to 50%–65%+ in 2026 — a 30-to-50-percentage-point year-on-year jump.
Two drivers: enterprise customers buying at scale, and agentic workflows — products that let AI autonomously execute multi-step tasks — becoming a "must-have."
Margins vary sharply by product line: direct API margins top 80%, the most lucrative; subscription products (e.g. Claude Code, Codex) run around 70%, the lowest of the three — because labs subsidize token costs to retain users.
Analyst Ross Sandler notes actual margins may be even higher than estimated, but expects them to decline as frontier competition intensifies and compute supply grows.
Same industry — why can two AI labs' financials look so different?
Barclays built two hypothetical labs to illustrate: Lab A (70% revenue from API) posts an adjusted gross margin of roughly 55%; Lab B (80% from subscriptions) manages only about 38% — a 17-percentage-point gap in the same industry.
The gap widens further due to revenue-recognition methods. Lab A books indirect API revenue on a gross basis, inflating top-line numbers; Lab B uses net recognition — or doesn't book partner-operated indirect API revenue at all.
In plain terms = same core business, different accounting treatments, very different-looking reports. Barclays likens it to Uber vs. Lyft: identical ride-hailing model, but financial-reporting choices make the two look like different industries. Investors will need to strip out these differences once AI labs start filing GAAP statements.
Of the money cloud providers collect, how much is real profit?
For Lab A: of every $100 in model-company revenue, $35 reaches the cloud provider; after infrastructure costs, that yields $11.8 in profit at a 34% operating margin.
Lab B's structure is more favorable for clouds: a strategic-partner revenue share (accounting for 20% of revenue, with a cumulative cap) means the cloud provider collects $41 and keeps $19.1 in profit — a 47% operating margin.
This reflects a structural nuance: the revenue-share mechanism inflates the cloud provider's apparent margin, but per-token actual profit is no different once the share is stripped out — and that share is expected to drop to zero after 2028.
After 2028, does the cloud providers' golden era end?
Cloud-provider AI revenue as a share of AI-lab revenue is falling: from 153% in 2024 to 90% in 2026, and a projected 73% by 2028.
This means → AI-lab revenue is growing faster than the fees labs pay to clouds — model companies are keeping more of each dollar earned.
Barclays expects that from 2028, committed AI infrastructure projects will come online and become labs' preferred option, potentially causing the big three clouds to lose share in both training and inference.
Training-cost share is also dropping fast — from 96% in 2024 to a projected 30% by 2028. In plain terms = the slice of spending that goes to "teaching AI new things" is shrinking; inference revenue is overtaking training costs, steadily improving overall lab profitability — but it also means the cloud providers' current high-margin window may close sooner than the market assumes.
市场有风险，内容仅供研究参考，不构成投资建议。