Five Questions for the AI Capex Cycle
The AI buildout is the largest capital-spending story in markets, and most commentary on it collapses into one of two lazy positions: “it’s the new electricity” or “it’s the new fiber bubble.” Both are conclusions. What an investor needs is a framework — a set of questions that stay useful whichever way the cycle breaks. Here are the five I keep returning to.
1. Who actually bears the capex?
“AI spending” is not one thing. A dollar of compute can sit on the balance sheet of a cash-rich platform company funding it from operating profits, a specialist cloud borrowing against GPU collateral, a startup renting by the hour, or an infrastructure fund that securitized a data center lease.
These are radically different risk positions attached to the same physical asset. Capex funded from a fortress balance sheet can be wrong for years without forcing a sale. Capex funded with leverage against the asset itself turns a demand wobble into a refinancing event. When you read any headline spending number, the first question is not “how big?” but “on whose balance sheet, funded how, with what obligations attached?” Cycles rarely break where the spending is biggest; they break where the funding is most fragile.
2. What is the real depreciation clock?
Accountants assign hardware a useful life; the market assigns it a competitive one, and the two need not agree. The economic life of an accelerator is set by the release cadence of better chips, the efficiency gains in models themselves, and the price at which older hardware can still serve workloads profitably.
The framework question: for each player, what happens to their economics if the fleet earns for meaningfully less time than the depreciation schedule assumes — and, symmetrically, who quietly wins if older chips remain economically useful for longer than the pessimists assume? Depreciation assumptions are one of the few places where a spreadsheet quietly encodes an entire worldview. Read them.
3. Is the revenue external, or circular?
Every infrastructure boom develops circularity in its later stages: vendors financing customers, ecosystem players buying from each other, investment dressed as revenue. The pattern is old — it decorated the telecom buildout — and it is not automatically fraudulent or even irrational. But it changes what a revenue number means.
So trace the dollars. How much of a company’s AI revenue is paid by end customers with non-AI income — enterprises, consumers, advertisers — versus paid by other members of the same buildout, funded by capital raised on the strength of the buildout? External revenue tests demand. Circular revenue tests fundraising. A boom is maturing well when the external share grows; it is hollowing when reported growth increasingly depends on money that entered the loop as investment.
4. What is the utilization story?
Between “we bought the compute” and “the compute earns money” sits utilization, and it is the least-disclosed number in the entire story. A data center is a factory; a factory’s economics are set by how full it runs and at what price. Publicly, you mostly get proxies: waiting lists, rental prices for compute, commentary about capacity constraints, the speed at which new capacity gets contracted.
The proxies are still worth watching systematically, because turning points show up there first. Scarcity pricing for compute is the boom working. Softening rental prices while record capacity comes online is the single most important warning the cycle can give — it is what overcapacity looks like before it appears in anyone’s income statement.
5. What breaks first if the money gets expensive?
Every long-duration buildout embeds an assumption about the cost and availability of capital, because the payback stretches years into the future. The stress-test question: if funding conditions tighten — rates, credit spreads, equity appetite, any of it — which participants must return to the market soonest, and what do they have to sell or cancel to avoid it?
Order matters. Leveraged single-asset players feel it before diversified platforms; companies pre-selling capacity feel it before companies renting month-to-month; and cancelled orders flow up the supply chain with a lag, hitting the suppliers everyone considered the safest way to play the theme. Mapping that sequence in advance is worth more than any forecast of whether tightening comes.
Using the framework
Notice what these five questions have in common: none requires predicting whether AI “works,” and none produces a price target. They locate where the risk sits and what evidence would move it — which is what analysis is actually for. Demand can be genuinely transformative and specific balance sheets can still be ruinously wrong; the railway mania built the railways and bankrupted their financiers. The technology succeeding and your investment succeeding are separate questions. The framework exists to keep them separate.