An AI capex forecast needs more than one financing assumption
A facility debt ratio cannot price the entire AI buildout without separate assumptions for hardware, power and construction. Start with the asset mix before forecasting the debt wave.
I would not put a single loan-to-cost ratio on a spreadsheet cell labeled “AI infrastructure.” The cell contains too many different things.
In “Concrete, Silicon, & Leverage,” Tomasz Tunguz estimates a $4 trillion debt wave, drawing on a roughly $5 trillion buildout and high facility-level leverage. His stated assumptions make the scale argument inspectable. They also expose the question I would ask before using the result: which assets are being financed on those terms?
McKinsey’s $5.2 trillion base-case forecast allocates about 15% to builders, 25% to power and related infrastructure, and 60% to technology developers and designers producing chips and computing hardware. It is a forecast across parts of a value chain, not a disclosed financing plan for one property portfolio.
The debt could still be enormous. The aggregate number needs an asset-level bridge before it becomes a financing estimate.
Concrete and computing equipment need separate rows
Land, buildings, power infrastructure and accelerators can all involve borrowing. That does not make their advance rates, repayment schedules, guarantees or residual-value assumptions interchangeable.
A facility-level ratio describes a particular financing arrangement. Applying it across a broad spending forecast assumes the same financing mix where the evidence may not support one. Some spending may come from operating cash flow, some from corporate borrowing and some from project or equipment finance.
The relevant objection is therefore not that chips cannot secure debt. It is that the way they are financed needs to be specified. A short replacement cycle or a uncertain resale value changes the questions a lender and borrower must answer, even if the equipment sits inside a long-lived building.
I would also check the accounting boundary of the forecast. A value-chain estimate and a buyer’s capital budget may count different activities. Combining them without understanding those boundaries risks counting the same economic demand more than once.
Replace the headline with a sensitivity table
In a hypothetical infrastructure review, I would split a proposed $100 million investment into site work, power systems, computing equipment and other costs. Those amounts would be explicit assumptions, not a miniature claim about the industry.
For each category, I would enter the expected debt share, term, repayment profile and responsible borrower. Then I would test a delayed opening and a faster equipment replacement cycle separately. Those two events stress different parts of the plan.
Only after that would I sum the borrowing and estimate debt service. If a category lacks financing evidence, I would show a range rather than silently importing the building’s ratio.
The revenue requirement deserves the same care. Interest coverage, operating profit and gross margin are different steps in the calculation. A revenue forecast should show the expenses and coverage assumptions connecting them, including repayment and replacement needs where relevant.
This approach may produce a larger or smaller debt estimate. Its advantage is that the reader can see which assumption moves the answer. A trillion-dollar result should become more transparent as it becomes more consequential.
Price the debt against the assets and payment schedules; “AI infrastructure” is not a single collateral class.
