Torsten Slok at Apollo published a chart last week that is easy to dismiss because it looks too neat. Silicon and equipment sit at a 41% operating margin. Models and applications sit at -59%. Compute and cloud, energy and grid, and the rest of the stack sit between them. Fortune turned the point into a clean headline about AI profits being funded by investors rather than customers.
The chart deserves a slower read. In a normal software story, the layer closest to the customer should have the best chance at durable margin. It owns distribution, usage data, switching costs, and pricing power. The infrastructure suppliers get paid, but the product layer keeps the surplus if the end market is real.
AI is currently paying out in the other direction. The farther a company is from the end user, the cleaner the margin often looks. Nvidia, memory suppliers, equipment vendors, and parts of the power stack are receiving cash from the buildout now. The model and application layer is still trying to prove that end customers will pay enough, often enough, at a gross margin high enough to justify the infrastructure already being ordered.
Start with the steelman for the bulls. Early general-purpose technologies often look wasteful before the applications catch up. Railways, fiber, cloud, and smartphones all had periods where infrastructure arrived ahead of obvious demand. A narrow customer-ROI screen can miss the option value of building capacity before the use cases are fully priced. Goldman Sachs Research now estimates global AI-related investment above $1 trillion in 2026 and argues that the level as a share of GDP still sits inside the historical range for prior technology buildouts. The near-term indicators it tracks also do not show an imminent capex slowdown.
That is a real argument. If AI adoption keeps moving from experiments into everyday workflows, the loss-making model layer may be the temporary price of creating a much larger market. OpenAI and Anthropic can burn capital today if enterprise usage, consumer subscriptions, agentic workflows, and API demand compound fast enough. The chip vendors then look less like bubble beneficiaries and more like toll collectors on a new production function.
The harder question is timing. A capex boom can be right in the long run and still wrong in sequence. Cash leaves first. Revenue arrives later, if the adoption curve is kind. The party that earns a 41% margin today is being paid by a party with a -59% margin today. That does not make the 41% fake. It means the 41% is partly contingent on the -59% remaining financeable.
This is the part of the distribution I would watch. Customer demand and investor funding are both real sources of cash, but they have different reflexes. Customer cash renews when the product keeps producing value. Investor cash renews when the story keeps producing expected value. The first is slower and stickier. The second can be abundant until it becomes suddenly expensive.
Oracle is the clean stress case because it turns an abstract chart into a balance-sheet problem. The AI infrastructure story has pulled Oracle into a much heavier capital-spending posture, with large data-center obligations tied to future compute demand. Fortune cites concerns around negative free cash flow, nearly $130 billion of debt, and roughly $260 billion in lease commitments for AI infrastructure projects. The Register reported earlier this year that Oracle's fiscal 2026 capex reached $55.7 billion, up from $21.2 billion a year earlier, and that management discussed around $70 billion of net cash outlay for fiscal 2027 capital expenditures before some customer prepayment and timing effects.
A project-finance view makes the risk clearer. Data centers can look safe when they are contracted. The contract can look safe when the customer is famous. The famous customer can look safe when capital markets treat AI growth as open-ended. Each link may be rational on its own terms. The chain still depends on end customers eventually producing enough cash to support all the obligations stacked upstream.
That is why the bubble question is too blunt. A bubble frame asks whether AI is real. The better question is where the cash conversion shows up, and when. Models can be useful while model companies remain structurally unprofitable. Chips can be scarce while too many data centers get financed. Enterprise adoption can rise while the revenue per unit of compute disappoints. Several of these statements can be true at the same time.
The open-weight push from Meta points to the same pressure from another angle. Reuters reported that Meta's new Muse Glimmer model is smaller than frontier systems, built for agentic tasks, and designed to run on a Mac or PC with a single graphics card. Zuckerberg's policy argument was about American open-source competitiveness against Chinese labs. The economic argument is sitting underneath it. If customers are worried about ballooning AI bills and security limits around closed systems, cheaper local models become a way to push usage closer to the edge and reduce dependence on metered frontier compute.
That does not kill cloud demand. It changes the shape of demand. Some workloads will keep moving toward frontier models because quality matters more than cost. Some will move toward smaller models because cost and control matter more than the last bit of capability. If that second bucket grows quickly, the upstream margin pool shifts. The owners of scarce frontier compute still win the hardest problems. The owners of distribution, local hardware, and enterprise integration may keep more of the routine work.
The margin chart is useful because it forces a payoff sketch. In the good state, downstream AI revenue catches up before capital markets lose patience. The model layer improves unit economics, applications find repeatable willingness to pay, and infrastructure utilization rises into the buildout. Upstream margins compress from extreme levels but remain healthy because the market is larger.
In the bad state, the application layer keeps growing revenue but not cash flow. Customers use AI, but they bargain down pricing, shift routine workloads to smaller models, or treat features as table stakes inside existing software budgets. Upstream suppliers still report strong demand for a while because orders and leases lag. Then the marginal data-center project starts needing more equity, more debt, or more customer prepayment at worse terms. The buildout does not need to collapse for the equity story to reprice. It only needs the financing chain to demand proof sooner than the applications can provide it.
My prior sits between the cartoon versions. AI demand is not imaginary. The tools are already useful enough that many customers will not go back. The industrial structure around that demand is still young, debt-heavy in places, and unusually dependent on capital markets believing that losses near the user will become profits near the user. That is a narrower claim than calling the whole thing a bubble. It is also the claim that matters for who gets paid.
The 41% can persist for longer than skeptics expect. Scarcity often does. But a margin chain that pays the farthest supplier before the closest customer is a chain with a financing assumption inside it. The assumption may be right. I would still mark it to market often.
Sources: Apollo, Fortune, Goldman Sachs Research, Reuters, The Register.
Originally published at deanlee.info.