The number, and what it covers
Investment in American AI infrastructure will total $10.3 trillion between 2025 and 2032, an average of 3.63 per cent of US gross domestic product a year, according to a paper published by the Brookings Institution and presented at its fall economics conference on Friday.
The author is Stijn Van Nieuwerburgh of Columbia University, and the projection covers more than servers. It counts data centre buildings, power systems, networking infrastructure, and specialised chips and other equipment — the whole physical apparatus rather than the computing hardware alone.
His comparison is historical. “The projected buildout would be larger relative to the economy than the major U.S. canal, railroad, electrification, highway, and telecommunications investment booms,” Van Nieuwerburgh writes. That is a claim about share of output, not absolute dollars, and it is what makes the figure worth arguing over: the country has built expensive things before, but not many that took this much of it at once.
Where the paper’s worry actually sits
The projection is not the paper’s argument. Its concern is how the money is raised.
Van Nieuwerburgh writes that the risks are migrating from transparent on-balance-sheet financing by large corporations to opaque off-balance-sheet financing — joint ventures, private credit, securitisation, special-purpose vehicles, lease commitments, loan guarantees and similar structures. Those arrangements depend on AI companies’ cash flows and on collateral values that are subject to, in his words, “uncertain AI demand, rapid technological change, timely access to power and hardware, and the continued credit quality of a small number of [data center] tenants.”
That last clause is the sharpest part. A great deal of this capital is secured against leases signed by a short list of tenants. If one of them slows down, the exposure is not confined to one building.
What he stops short of saying
The paper does not forecast a crash. “It would be premature to conclude that AI infrastructure already poses systemic risk comparable to earlier credit booms,” Van Nieuwerburgh writes. His point is narrower and harder to dismiss: “off-balance sheet structures matter … because they may make correlated exposures hard to observe before a downturn.”
His recommendation follows from that. “The most important policy contribution at this stage may therefore be to improve measurement and transparency while the capital structure of the industry is still evolving.”
What to watch
The paper is a conference draft in the Brookings Papers on Economic Activity, the fall edition of which put three AI papers on the programme for 25 September. Drafts are discussed before they are finalised, and the projection is a projection: it assumes a pace of construction that power availability and chip supply may not deliver.
The testable part is the disclosure question. Off-balance-sheet AI financing either becomes easier to see in company filings and regulatory data over the next year, or it does not — and Van Nieuwerburgh’s argument is that the second outcome is the one that costs something later.
