Ignore the model benchmarks. Watch the invoice.
Barclays dropped a number last week that should disturb anyone building on top of the AI stack: cloud providers are extracting 35-40% of every dollar of AI model revenue. Not as a one-time setup fee. As a standing toll.
For every $100 an AI company books, $35 to $40 flows directly to the infrastructure layer before a single engineer gets paid. The model companies keep the rest on paper, but after compute, labor, and overhead, their practical margin sits somewhere between ten and twenty dollars. Maybe.
The crypto crowd loves to talk about decentralization as an ideological position. It is not. It is an economic survival strategy. And right now, the centralized cloud oligopoly has built a toll booth on the only road to AI scale.
The Infrastructure Tax Structure
Let me break down what that 35% actually buys, because the figure is not arbitrary. It tracks the gross margin profile of mature cloud infrastructure.
If a cloud provider collects $35 and banks $15 as profit, their operating costs hover around $20 per $100 of AI revenue. That works out to roughly 57% gross margins, which sits squarely inside the 55-65% band that AWS and Azure have historically maintained on their core services. This is not price gouging in the traditional sense. This is the economics of a capital-intensive utility with pricing power.
The cloud is not charging for compute. The cloud is charging for the right to exist at scale.
I ran this against my own infrastructure experience. A mid-sized inference cluster, say an 8-card H100 node, runs between $15,000 and $30,000 per month in a managed environment. Production API workloads typically require three to five such nodes. The cost structure aligns almost exactly with Barclays' implied numbers. The report is not modeling hypotheticals; it is describing the current state of the machine.
What the report does not say, and what I have learned from auditing infrastructure deals since 2017, is that the cloud providers are not taking risk. They are collecting rent.
The Rentier's Advantage
This is the part that gets lost in the AI narrative noise.
OpenAI and Anthropic carry the innovation risk. They bet billions on model architectures, training runs, and market adoption. If GPT-6 flops, the equity holders eat the loss. If the regulatory environment tightens, the model companies absorb the compliance burden. If inference costs spiral due to context window expansion, the model companies eat the margin compression.
The cloud providers, by contrast, collect their 35% regardless of whether the underlying models succeed or fail. During the training phase, they sell GPUs. During the inference phase, they sell token throughput. During the inevitable consolidation phase, they sell the exit liquidity to whichever model company survives.
Bets are cheap; exits are expensive. The cloud providers have figured out how to charge for both.
This is the structural insight the bull case for AI misses. The market treats Microsoft, Amazon, and Google as AI winners because they own the distribution channels. True. But they are also the only entities in the stack whose profits do not depend on model quality. They depend solely on model quantity. As long as enterprises keep buying AI services, the toll booth collects.
I have seen this movie before. In DeFi summer, the yield farmers thought they were the innovators. The real winners were the infrastructure providers selling gas and block space. The pattern repeats because the pattern is structural.
Cost Structure Breakdown
Let me get more granular on where that 35% goes, because the composition determines the durability of the toll.
Equipment depreciation is the largest line item, roughly $12 to $15 per $100 of revenue, assuming a four-year GPU depreciation schedule. Power and cooling runs around $8. Networking and operations consume another $5. That leaves $7 to $10 as clear profit, which matches Barclays' $10-20 range once you factor in utilization gains from advanced scheduling.
The optimization levers matter more than most analysts appreciate. Cloud providers have deployed KV cache optimization, continuous batching, and quantization to drive down marginal token costs. These techniques expand the profit zone between what they charge and what they spend. The 35% toll is not static; it is a floor that gets more lucrative as inference efficiency improves.
Meanwhile, model companies face the opposite dynamic. Their costs are not falling at the same rate. Longer context windows and multi-turn agent interactions drive exponential growth in inference requirements. The model company pays for the exponential curve; the cloud provider collects the toll at every inflection point.
Follow the gas, not the hype. The gas here is the growing divergence between model innovation and infrastructure capture.
The Vertical Integration Arms Race
The 35% toll explains the strategic behavior we are seeing across the industry.
Microsoft deepened its entanglement with OpenAI because the toll was not enough. They wanted the equity upside on top of the infrastructure rent. Amazon did the same with Anthropic. Google simply builds everything in-house because they have the capital base to verticalize the entire stack.
This is not partnership. This is the internalization of a toll that was previously paid externally. When Microsoft invests $13 billion in OpenAI and also provides their compute, the 35% is a transfer price, not a market transaction. The public financial statements obscure the true economics.
For independent model companies, the math is brutal. Every dollar of API revenue carries an embedded infrastructure tax that their vertically integrated competitors do not truly pay. Mistral and xAI are building their own supercomputers not because they love hardware, but because the toll is existential.
This is the same logic that drove crypto projects to build their own infrastructure rather than rent from centralized providers. The gas costs were bleeding out the value proposition, so they optimized for self-sovereignty. AI is hitting the same wall, just with more zeros attached.
The Decoupling Delusion
Here is where I diverge from the mainstream take on this data.
The conventional narrative says AI infrastructure is a winner-take-all market where scale begets advantage. Cloud providers hold the capital, the distribution, and the customer relationships. The toll is permanent.
I disagree. The toll is vulnerable, but not from the direction most people expect.
Everyone watches the model layer for disruption. They track GPT releases, open-source alternatives, and benchmark improvements. They expect a DeepSeek moment to upend the economics. That is the wrong frame.
The real disruption will come from the compute layer itself, not the model layer.
The cloud providers are not competing with each other on AI margins. They are competing with NVIDIA over the hardware cost embedded in that 35%. AWS Trainium, Google TPU, and Microsoft Maia are not science projects. They are toll reduction strategies. If custom silicon reaches 40% utilization in production, the internal cost structure shifts dramatically. The profit zone expands from $10 to $20 per $100 of revenue toward $20 to $25.
And then the pricing war begins. Not over models. Over inference cost per token.
The cloud providers will use their hardware advantage to undercut each other on AI API pricing, squeezing the model companies from below while the model companies still cannot escape the infrastructure layer. The toll gets cut, but the toll booth changes ownership.
This is the decoupling thesis the market has not priced. AI's infrastructure cost curve is about to bend, and the beneficiaries will not be the model companies. They will be the hardware players who own the new cost structure.
The Hidden Risks
I am not bullish on this setup. I am mapping the landscape, and the landscape has sinkholes.
The first risk is the capital expenditure cycle. Cloud providers are spending enormous sums on data centers based on projected AI demand. If enterprise AI adoption slows, or if the models fail to deliver on their productivity promises, the toll revenue will not cover the depreciation. The 35% toll becomes a 15% toll, and the profit zone collapses.
Watch the ratio between cloud capex and AI revenue. When capex growth exceeds revenue growth for two consecutive quarters, the toll structure is breaking.
The second risk is energy. Power costs represent over 30% of cloud operating expenses. AI training and inference workloads are energy monsters. A sustained electricity price shock compresses the profit zone directly. The toll would need to rise, but the market will not accept a higher toll during a demand slowdown.
The third risk is regulatory. The entanglements between cloud providers and model companies are drawing antitrust scrutiny. The EU AI Act demands model traceability, which means data flows and logging stay on cloud infrastructure. This gives the cloud providers a compliance choke point over their model company tenants. Regulators may eventually view the 35% toll as an abuse of market power, particularly if the vertical integrations create exclusionary dynamics.
None of these risks are priced into the current valuations. The market sees AI revenue growth and assumes it translates to AI profits. The toll structure says otherwise. The toll structure says the profits are already spoken for.
Positioning for the Inflection
From a capital allocation perspective, the Barclays report confirms what the flow data has been signaling for 18 months.
Own the toll roads, not the cars. The cloud providers are the toll roads. Their AI revenue growth is the toll volume. The model companies are the cars, burning fuel and paying fares.
But the toll road business is about to face competition from a new highway. Neutral compute providers like CoreWeave and Oracle's OCI are offering alternative routes at different price points. They will not replace the hyperscalers, but they will capture a meaningful share of price-sensitive workloads.
The more interesting positioning is in the hardware layer. The cloud providers' push for custom silicon creates a software ecosystem opportunity. Companies building the toolchains, the orchestration layers, and the middleware for Trainium and TPU are positioned to benefit from the toll reduction initiative, regardless of which cloud provider wins the AI revenue battle.
And then there is the decentralized compute angle, which the traditional finance crowd dismisses out of habit. If the toll structure remains persistently high, if the cloud providers maintain their 35% extraction rate, the economics of decentralized GPU networks become compelling by default. Not because of ideology. Because of math.
I have built my career on following the mechanics, not the marketing. The mechanics here are unambiguous. The AI value chain currently transfers value from innovators to infrastructure at a confiscatory rate. That rate will either be disrupted by new infrastructure models, or it will be normalized by regulation.
Either way, the 35% toll is not permanent. The question is who builds the alternative route.
The model companies are betting their futures on escaping the cloud. The cloud providers are betting on vertical integration. NVIDIA is betting on the compute demand curve being unbounded.
One of these bets is wrong. Follow the infrastructure, and you will see which one breaks first.
Bets are cheap. Exits are expensive. The cloud providers built the exit ramp, and they control the toll booth. Smart capital is already mapping the detours.