The Toll Booth at the Gates of Intelligence: Who Really Owns the 35%?

The numbers arrived on a Tuesday, buried in a Barclays research note that most of the crypto Twitterati scrolled past. Cloud providers are taking a 35-40% cut of every dollar of AI model revenue. The model builders—the ones burning billions in compute and talent—are left with a 10-20% profit margin, if they are lucky. I read this in my Lagos apartment, the hum of the generator a constant reminder that infrastructure is never free. While the crowd shouted about token prices, I watched the exit. This is not a story about AI. It is a story about rent, about who owns the land beneath the gold rush. And for anyone who has spent years in crypto, the pattern is hauntingly familiar. The chain remembers what the soul forgets. In crypto, we called it the "Ethereum tax"—the gas fees, the validator rewards, the MEV extracted from every trade. The narrative was always about decentralization, but the economics were about toll booths. Now, the same architecture is being rebuilt in the AI industry, and the toll collectors are not anonymous validators. They are Microsoft, Amazon, and Google. The Barclays data is a rare, clean snapshot of a structural truth: the AI revolution is not being built by the model makers. It is being financed, hosted, and ultimately owned by the cloud oligopoly. To understand the 35%, you have to understand the cost of silence. I spent three months in 2020 manually tracking Uniswap liquidity pools, mapping sentiment against on-chain volume. I learned that the most important data is often the data that is not there. The Barclays report is similar. It tells us the cut, but it does not tell us the cost structure that justifies it. So, I built my own model. Assume $100 of AI model revenue. The cloud takes $35. If the cloud's profit is $15, its operating cost is $20. That implies a gross margin of roughly 57%. This aligns almost perfectly with the 55-65% gross margin range that AWS and Azure have reported for their core cloud businesses. The 35% is not a predatory tax; it is the standard price of admission for enterprise-grade infrastructure. It is the cost of the GPU clusters, the power, the cooling, the redundancy, and the 99.9% uptime SLA that a startup cannot replicate. But here is the insight that the report misses. The cloud provider's profit is not a "share" of the model's success. It is a rent. It is a fixed toll that is paid regardless of whether the model is profitable. OpenAI can burn $5 billion a year, and Microsoft still collects its check. This is the "riskless rent" that I have been tracking since the Terra collapse. In 2022, I watched algorithmic stablecoins fail because they confused trust with collateral. The cloud providers have made no such mistake. They are not betting on the model. They are selling the shovels, and they are charging a premium for the privilege of using them. The model companies are the miners; the cloud is the power plant. And the power plant always gets paid. This creates a profound structural distortion. The model companies are valued on their revenue, but their revenue is a mirage. A $100 of API revenue is really $65 of revenue after the cloud cut, and then you have to subtract the cost of the GPUs, the researchers, and the electricity. The actual net income is often negative. This is why the market is starting to see a "vertical integration" push. Meta is building its own data centers. xAI is building a massive supercomputer in Memphis. OpenAI is reportedly in talks to design its own chips. They are all trying to escape the toll booth. They are trying to build their own road. Noise is the tax we pay for visibility. The noise around AI is deafening, but the signal is in the cost structure. The 35% cut is not static. It is a function of the underlying hardware costs. If NVIDIA's next-generation Blackwell chip is 30% cheaper per FLOP, the cloud's cost basis drops, and the profit margin on that 35% cut expands. This is the hidden lever. The cloud providers are not just rentiers; they are also arbitrageurs. They buy GPUs at wholesale, they optimize utilization with techniques like continuous batching and KV cache reuse, and they sell the output at a price that reflects the scarcity of the moment. The 10-20% profit margin is the floor, not the ceiling. In a period of high demand, that margin can expand significantly. I do not trade tokens; I trade timelines. And the timeline here is clear. The current equilibrium—where model companies are dependent on cloud providers—is unstable. The pressure is building on three fronts. First, the model companies are trying to disintermediate the cloud by building their own infrastructure. Second, the cloud providers are trying to disintermediate NVIDIA by building their own chips (AWS Trainium, Google TPU). Third, a new class of "neutral cloud" providers (CoreWeave, Oracle OCI) is emerging, offering GPU capacity without the strategic entanglements of the hyperscalers. This is the classic crypto pattern: the base layer captures value, then the application layer tries to build its own base, and then a new base layer emerges to serve the disaffected. The contrarian angle is that the cloud providers are not as safe as they look. The 35% cut is a function of scarcity. If the AI bubble deflates—if enterprise adoption stalls, if the cost of inference drops faster than expected—the cloud providers will be left with massive, underutilized data centers. The depreciation on those GPUs is a fixed cost. The electricity bill is a fixed cost. The 35% cut will not save them. They will be forced to cut prices, and the 10-20% profit margin will evaporate. This is the "capex trap" that I have been warning about. The cloud providers are making a massive bet on the future of AI, and if that bet is wrong, they will be the ones holding the bag. The model companies, for all their losses, have the optionality of pivoting. The cloud providers are locked into their physical assets. This is where the crypto lens becomes essential. In crypto, we have a term for this: "the settlement layer." The cloud is the settlement layer for AI. It is the final arbiter of truth, the place where the compute is actually executed. And like any settlement layer, it is subject to the "toll booth problem." The question is not whether the toll will be collected, but who will be allowed to build the road. The current answer is the hyperscalers. But the history of technology suggests that this will not last. The ledger is cold, but the pattern is warm. The pattern is that every centralized bottleneck eventually gets challenged by a decentralized alternative. The challenge to the cloud's AI toll booth will come from two directions: open-source models that can run on commodity hardware, and decentralized compute networks that aggregate idle GPUs from around the world. I have been studying the decentralized compute space for years. The technology is still immature, but the economics are compelling. A decentralized network can offer GPU compute at 50-70% of the cost of a hyperscaler, because it does not have the overhead of data centers, sales teams, and enterprise compliance. The catch is reliability and security. A hyperscaler offers a 99.9% uptime SLA. A decentralized network offers... hope. But as the technology improves—as verifiable computing and trusted execution environments become more robust—the gap will narrow. And when that gap narrows, the 35% cut will come under pressure. The model companies will have a choice: pay the toll to the hyperscaler, or take the risk on the decentralized network. The risk will be worth it, because the savings will be the difference between life and death. To hold is to trust the unseen architecture. This is the core of my thesis. The AI industry is building a new architecture, and the cloud providers are the unseen pillars. They are the ones who make the magic happen, but they are also the ones who extract the rent. The question for investors is not whether AI will be transformative—it will be. The question is who will own the transformation. The Barclays report suggests that the cloud providers are the winners. I am not so sure. The history of technology is a history of disintermediation. The mainframe gave way to the PC. The PC gave way to the cloud. The cloud will give way to something else. The question is what that something else will be. I see three possible futures. In the first, the hyperscalers maintain their grip. They continue to extract 35% of AI revenue, and they become the most valuable companies in history. In the second, the model companies successfully vertically integrate, building their own chips and data centers, and the cloud providers are reduced to commodity utilities. In the third, a new decentralized compute layer emerges, and the 35% cut is compressed to 10-15%, as the market becomes more competitive. I believe the third future is the most likely, but it is also the most uncertain. It will require a leap of faith, a willingness to trust the unseen architecture. We mined the silence in Lagos to find the signal. The signal is that the AI industry is at a critical inflection point. The current value distribution is unsustainable. The model companies cannot continue to lose money forever. The cloud providers cannot continue to extract rent without facing a backlash. Something has to give. The question is when, and how. I am watching the signals. I am watching the capex numbers. I am watching the chip orders. I am watching the open-source community. And I am watching the decentralized compute networks. The next 18 months will be decisive. The toll booth is standing, but the road is being rerouted. The question is not whether the toll will be collected, but who will be on the other side of the gate. While the crowd shouted, I watched the exit. The exit is not a token. It is a structural shift. It is the moment when the model companies realize that they are not in the AI business; they are in the infrastructure business. And the moment they realize that, they will start building. The 35% cut is a tax on their ignorance. The smart ones will pay it for now, but they will be building the alternative in the background. The smart investors will be funding those alternatives. The smart analysts will be tracking the cost curves. The rest will be left holding the bag, wondering why their AI stocks are not going up. The chain remembers what the soul forgets. The chain of value will remember who built the road, and who only collected the toll.