Hook
Barclays recently dropped a number that should unsettle anyone building on top of the AI stack. For every $100 of AI model revenue generated, cloud providers extract $35 to $40. The model companies—OpenAI, Anthropic, the names we all know—keep the rest. But here's the part the press release doesn't emphasize: after paying for compute, the cloud provider's profit sits at $10 to $20. The model company's margin? Often negative.
The code does not lie; only the founders do. And in this case, the code is written in capital expenditure, depreciation schedules, and utilization rates.
I've spent the last decade auditing smart contracts and incentive structures. This is the same pattern I saw in DeFi summer, in the NFT minting fiasco, in the Terra collapse. When someone controls the underlying infrastructure, they don't need to win the application layer. They just need to collect rent.
Context
The AI industry has settled into a three-tier hierarchy. At the bottom sits NVIDIA, selling the picks and shovels. In the middle, hyperscale cloud providers—AWS, Azure, GCP—operate the GPU clusters that make modern AI possible. At the top, model companies like OpenAI and Anthropic build the algorithms and products that generate revenue.
The Barclays analysis reveals the profit distribution across these layers. For every $100 in AI model revenue, cloud providers take $35 to $40. Their operating costs run roughly $17 to $25, leaving $10 to $20 in pure profit. That's a 28% to 50% net margin on AI-related revenue.
Compare that to the model companies. They collect the $65 to $60 that remains, but they're paying for research, development, and their own compute costs. Most are operating at a loss. OpenAI reportedly spends more on inference and training than it generates in revenue.
This is not a partnership. It's a toll booth.
Core: The Systematic Teardown
Let me break down what this 35% extraction actually means, layer by layer.
The Cost Structure Behind the Toll
Based on my audit experience and industry financial models, the $35 to $40 cloud take breaks down roughly as follows: $12 to $15 covers hardware depreciation (GPUs on a four-year cycle), $8 covers electricity and cooling, $5 covers networking and operations, and the remaining $7 to $10 is profit. Barclays' $10 to $20 profit range aligns with this structure.
The cloud providers' margin—around 57% gross—matches the standard gross margin range for AWS and Azure core services. This isn't a premium for innovation. It's a toll for access to infrastructure that model companies cannot realistically replicate.
I don't trust the audit; I trust the gas fees. In this case, the "gas fees" are the capital expenditure barriers that keep new entrants out of the compute market.
The Utilization Game
Here's what the Barclays report doesn't tell you. AI inference demand is spiky. Nighttime utilization drops. Idle GPUs still consume electricity and accrue depreciation. The cloud provider's profit margin depends on peak-hour pricing covering the cost of idle capacity.
This is the same dynamic I identified in the Compound protocol's interest rate models back in 2020. The system works during normal conditions, but the rounding errors and edge cases only surface under stress. For cloud providers, the stress case is simple: if AI application revenue grows slower than capital expenditure, the 35% extraction becomes insufficient to cover costs.
The rug was pulled before the mint even finished. In this case, the "rug" is the capital expenditure bubble that could deflate when enterprise AI spending inevitably slows.
The Model Companies' Trap
Model companies face a structural disadvantage. They generate revenue through API calls, priced per token. But their cost structure includes both the cloud extraction and their own research and development expenses.
Consider the math. A model company generates $100 in revenue. The cloud takes $35 to $40. The company retains $60 to $65. But their R&D costs—researchers, data, training runs—often exceed that amount. The result: high revenue, negative margins.
This is the valuation paradox. Investors value model companies on revenue multiples, but the revenue is subsidized by cloud providers who extract guaranteed profits regardless of whether the model company succeeds.
The Vertical Integration Response
Some players are fighting back. Meta has invested heavily in its own data centers. xAI is building a massive supercomputer in Memphis. OpenAI has reportedly explored designing its own chips.
The logic is straightforward: if you can't avoid the toll, build your own road. But this strategy requires massive capital expenditure and years of lead time. Most model companies don't have that luxury.
The alternative is the open-source route. Models like Llama and Mistral can run on commodity hardware or through neutral cloud providers like CoreWeave. This avoids the hyperscaler toll but sacrifices the integrated services and optimization that AWS or Azure provide.
The Hidden Leverage
The cloud providers' power extends beyond compute. Microsoft's $13 billion investment in OpenAI creates a circular relationship. OpenAI pays Azure for compute. Microsoft records that as Azure revenue. A portion of that revenue flows back to Microsoft as profit. The extraction is effectively an internal transfer price.
This creates a transparency problem. When the cloud provider is also an equity holder, the 35% extraction becomes a mechanism for profit shifting rather than a market-based price. Regulators should be paying attention to this structure.
Contrarian: What the Bulls Got Right
I've spent this article dissecting the cloud providers' extraction. But let me be precise: the bulls have a point.
The 35% extraction isn't pure rent-seeking. It reflects real costs. Building and operating GPU clusters at scale requires enormous capital. The hyperscalers have spent billions on data centers, cooling systems, and networking infrastructure. They've optimized power usage, negotiated favorable electricity rates, and developed custom silicon.
The margin they earn is compensation for bearing the capital expenditure risk. If AI demand collapses, the cloud providers absorb the depreciation losses. The model companies can walk away. The cloud providers are stuck with billions in sunk costs.
This is the "toll booth" argument from the other side. The toll isn't just for access. It's for risk absorption.
The bulls also correctly note that cloud providers are investing in AI-specific infrastructure. AWS Trainium, Google TPU, and Azure's Maia chips are designed to reduce inference costs. These investments could eventually lower the extraction rate while maintaining margins.
Reentrancy is not a bug; it is a feature of trust. The same applies to cloud extraction. It's a feature of the capital-intensive nature of AI infrastructure.
Takeaway
The 35% extraction rate is not sustainable in its current form. The pressure points are already visible: OpenAI's chip ambitions, Meta's data center buildout, the rise of neutral cloud providers, and the growing regulatory scrutiny of Microsoft-OpenAI and Amazon-Anthropic relationships.
The question isn't whether the extraction rate will change. It's whether the change comes through competition, vertical integration, or regulatory intervention.
For investors, the signal is clear. Cloud providers are the safest bet in the AI value chain. They earn guaranteed profits regardless of which model company wins. But that safety comes with a ceiling. The extraction rate caps their upside.
For model companies, the path forward is either scale to the point where self-built infrastructure becomes viable, or accept the toll and focus on applications that can generate enough revenue to absorb it.
The code does not lie. The capital expenditure schedules don't lie. The utilization rates don't lie. The only question is who gets to write the next version of the code.
I'm watching the capital expenditure to revenue ratios. I'm watching the OCID settlement between Microsoft and OpenAI. I'm watching the deployment rates of Trainium and TPU. The signals will tell us which direction the extraction rate moves.
Until then, the toll booth stands. And it's collecting.