The math is brutal. Barclays ran the numbers on AI model revenue. Their conclusion: for every $100 of AI income generated by companies like OpenAI and Anthropic, the cloud providers take $35 to $40. The model companies keep $65. Then they pay their own costs. Net profit? $10 to $20. Maybe.
NVIDIA takes its cut on the chips. The hyperscalers take theirs on the infrastructure. The model labs — the ones doing the actual innovation, the ones absorbing the public scrutiny, the ones facing existential legal battles over training data — they get the residual. This isn't a partnership. It's a toll road.
Data over drama. Let's break down who actually owns the revenue stream.
If we take Barclays' $35 cut and assume a $15 profit for the cloud provider, their operating cost is around $20. That maps to a 57% gross margin. AWS and Azure have run at 55% to 65% gross margins for years. This isn't an anomaly. This is industrial-scale rent collection dressed up as a technology service.
The structure of modern AI delivery is not a market. It's a feudal system with NVIDIA as the landlord of the raw materials, hyperscalers as the regional governors, and model labs as the serfs working the land.
The 35% figure reveals something deeper about AI commercialization. In my years running arbitrage strategies and managing infrastructure risk, I've learned that whoever controls the settlement layer controls the economics. In crypto, that's the exchange. In AI, that's the cloud.
First, the cost structure. A single 8-card H100 node runs $15,000 to $30,000 per month in a managed environment. Production-grade API workloads require multiple nodes. NVIDIA dominates the silicon. Power draw is extreme. Cooling is a second-order infrastructure problem that consumes both electricity and engineering hours.
The cloud providers' real advantage isn't technical brilliance in AI. It's capital expenditure at a scale that creates a moat. They negotiated volume discounts on GPU procurement — 20% to 30% below market rates. They operate utilization strategies that amortize hardware costs across thousands of tenants. They have the balance sheet to hold hardware through market cycles. That's their edge. It's also their vulnerability.
Core order flow analysis follows a simple rule: track where the money must pass before it reaches its final destination. In AI, the flow is unambiguous. Enterprise customer → API subscription → model company → cloud provider → NVIDIA. The choke point is the second-to-last hop. It's also the most predictable profit center in the stack.
Cloud providers collect their 35% regardless of whether the model company is profitable. OpenAI can burn billions on R&D. Anthropic can spend a dollar to earn a dollar. The cloud provider's invoice still gets paid. That's a structural certainty. It resembles a hedge — a short volatility position on AI innovation. The cloud is short model failure risk and long compute consumption. It profits from the process of AI experimentation, not the outcome.
But let me introduce a nuance that Barclays left out. The 35% figure — that's the headline. The true extraction rate is likely higher. AWS Bedrock and Azure AI aren't simple compute farms anymore. They layer model hosting fees, orchestration fees, API gateway charges, fine-tuning computation costs. When a model company deploys on a hyperscaler platform, it pays not only for compute but for the privilege of being within the ecosystem. The actual total extraction rate can exceed 40%.
Liquidity vanishes. Lessons remain.
Now let's flip the contrarian lens. The conventional narrative is that cloud providers are untouchable. I disagree.
Consider the migration economics. Companies that rely on AI are increasingly priced-sensitive. They see the 35% markup on their API bills. They understand that the cloud provider is a counterparty with market power. For some workloads, it's shifting to on-premise or colocation. The data residency and compliance requirements of regulated industries — banking, healthcare, government — already push toward private deployment. The hyperscalers have a pricing ceiling here.
Then there's the open-source threat. Llama and DeepSeek produce competitive models without the API toll booth. We're seeing price wars on token costs. DeepSeek priced its API at levels that forced incumbents to re-evaluate. This compression flows directly into the cloud provider's absolute revenue per token, even if the percentage split remains intact.
I've verified the math with real-world benchmarks, not just Barclays' models. In 2024, I ran a private AI infrastructure deployment for a research fund. We had the choice of running inference on Azure or spinning up a dedicated GPU cluster using a neutral provider. The hyperscaler pricing was more convenient. But conductivity and data transfer costs — the egress fees, the storage overhead, the host of small charges that never appear in headline pricing — pushed the cost delta past 25%. We went the private route. The savings were real. The operational burden was also real.
That's a clue about the market's direction. The deal between Microsoft and OpenAI is famous. What's less understood is that OpenAI's massive Azure credits are a strategic necessity, not a preference. They were burning cash on compute. The cloud credits preserved their runway. But this debt — the 35% extraction — is now embedded in their cost structure. They can't easily leave. And it's not just OpenAI. Anthropic signed a multi-billion dollar deal with Amazon that included a substantial GPU cluster commitment. They're locked in.
Meanwhile, the counter-movement is forming. There's a growing role for neutral compute providers, GPU marketplaces that operate independently of the hyperscaler stack. I've tracked the arbitrage: using independent GPU rental markets, you can often get the same NVIDIA H100 capacity for 20% to 30% less than the standard hyperscaler rate. The trade-off is less managed service. The trade-off is more operational complexity. But for sophisticated teams that understand infrastructure — and I count myself in that group — the cost advantage justifies the burden.
A subtle conflict is emerging. NVIDIA wants to sell GPUs at high margins. Hyperscalers want to drive down their procurement costs and retain pricing power. They're building custom ASICs — Trainium, TPU. If these gain traction, the cloud's effective cost per token declines significantly. That means the cloud's 35% take could widen its profit margin to $20 to $25 per $100, without changing the headline percentage.
NVIDIA knows this. That's partly why they're diversifying into networking and software. The question is whether they can maintain their monopoly position when their own customers are building alternative silicon.
There's a parallel to the DeFi ecosystem. In the early days, arbitrage opportunities existed because power was fragmented. As markets mature, the infrastructure layer extracts its rent. The DEX aggregators take fees. The MEV extractors take their share. The validator networks take their cut. Every layer of the stack that sits between the user and the underlying asset is a toll booth. The AI stack is no different.
Let's sum up the asymmetry. Cloud providers are the only participants in the AI ecosystem whose margin is structurally protected. NVIDIA is exposed to business cycle risk. Model companies are exposed to competition risk. But the cloud is exposed to neither, as long as AI workloads continue to grow. The only real threat is an application-level breakthrough that renders the current infrastructure obsolete or a macro event that triggers a capital expenditure freeze.
Watch the key signals: supply of Blackfinn Blackwell chips. If NVIDIA's Blackwell pricing drops more than 30% from H100 levels, infrastructure costs decline for everyone. Watch for capacity utilization rates reported by cloud providers. Watch deployment announcements for AWS Trainium2 and Google TPU v5. If those ASIC deployment rates jump past 40%, the NVIDIA narrative starts to crack.
The cloud extraction is a long-latency market signal. It tells you that AI model companies are not as profitable as their headline revenue suggests. It tells you that the capital base of the AI industry is being transferred upstream. And it tells you that the only hedge for an AI model company is to own its compute.
That's the playbook. The model companies may be the faces of AI, but the cloud is the body. The smart money is already positioning on the infrastructure side. So the question is not whether AI will change the world. The question is who gets paid for the privilege.