Research note · AI infrastructure · August 12, 2026
A dollar of capex is not a dollar of growth.
AI spending headlines are supposed to make the argument feel finished: a giant number means giant demand. But capex is an input, not a result. Before it becomes durable earnings, a dollar spent on a data center has to be deployed, utilized, priced correctly, depreciated honestly, and—where debt is involved—financed at a rate that does not eat the return.
Figure 1 · The capex relay race
A dollar of hardware creates a chain of claims. Breaking any link can break the economics.
A high capex figure tells me something about ambition. It does not settle utilization, pricing power, or return on invested capital.
Two models can both be “AI exposure” and still have opposite economics
A hyperscaler may fund capacity with operating cash flow and serve a broad set of customers. A specialist lessor may fund a smaller fleet with debt and depend on a few large counterparties. Both can report strong demand. The cash-flow consequence is not remotely the same.
| Question after the spending headline | Why it matters |
|---|---|
| Is capacity already contracted? | Demand visibility is not the same as realized revenue. |
| Who funds the hardware? | Customer prepayments and provider debt have different risk. |
| What is the useful life? | A longer life lowers near-term depreciation but does not extend the cash outlay. |
| Can power and construction keep up? | A GPU without power is an expensive inventory item. |
| What is the exit price of compute? | Supply growth can matter as much as demand growth. |
My simplest test: show me the bridge
I do not need a company to predict the future perfectly. I need a bridge: capex to available capacity, capacity to utilization, utilization to revenue, and revenue to cash after depreciation and interest. When one of those steps is missing, the headline should become a question mark rather than an arrow pointing upward.
ModelTry the stress test
The on-site AI Capex Reality Check lets a reader vary utilization, depreciation, pricing, and financing assumptions. It is a sensitivity tool, not a forecast—and it is more useful when the inputs are taken from filings rather than optimism.