The AI capital cycle and how it is financed

Arguments about an AI bubble usually collapse into arguments about whether the technology is real. That is the wrong axis. Railways were real. Fibre was real. Both produced enormous long-run value and destroyed the capital of the people who funded the buildout, because the returns arrived after the debt came due.

The three questions

• What is the depreciation schedule of the asset, and does it match how it is being accounted for? Compute hardware in continuous heavy use has a shorter economic life than a long useful-life assumption implies, and the assumption flows straight into reported earnings.
• Who bears the risk if utilisation falls? Take-or-pay contracts move it to the tenant. Speculative capacity keeps it with the developer and its lenders.
• Is the revenue external? Circular arrangements, in which a supplier invests in a customer that then buys its product, produce revenue that is real accounting and weak evidence of demand.

A capital cycle turns when the marginal project is funded by someone with no operating experience, using debt, on the assumption that prices hold.

Power is the binding constraint

The buildout is limited less by chips than by interconnection: transformers, turbines, transmission and grid queues measured in years. That constraint has two effects. It slows deployment, which pushes revenue further out than the capital was underwritten for. And it puts data centre developers in direct competition with households for electricity, which turns a technology story into a politics story wherever retail rates rise visibly.

Where the leverage actually sits

Hyperscalers largely fund from cash flow, and their balance sheets can absorb a great deal. The fragility is at the tier below: specialist compute providers, developers and leasing structures funded with private credit, asset-backed borrowing and vendor financing, often collateralised by the hardware itself. That collateral is worth what the next buyer pays in a downturn, which is precisely when there is no next buyer.

What a correction looks like

Not a sudden stop. A pause in new orders, a slip in delivery schedules, two quarters of quiet renegotiation, then an abrupt repricing of the smallest listed players and the fastest write-downs in the vehicles that never had to mark anything. The compute keeps running. The equity behind it does not survive intact.

What to watch

• Useful-life assumptions in depreciation disclosures, and any change to them.
• The mix of contracted versus speculative capacity in new builds.
• Vendor financing and equity stakes taken in customers.
• Utility interconnection queues and retail rate cases, which are the earliest public signal of a physical constraint becoming a political one.