The most misleading number in AI finance is the $6.3 trillion "gap." It reads like a shortage of capital. Most of it is better understood as a shortage of balance-sheet room. The first wave of financial engineering around AI data centers does not mainly draw in new kinds of risk-takers. It repackages the credit of a handful of very strong companies so that the debt sits somewhere other than on their books. Intuition says off-balance-sheet structures move risk away from the sponsor. Often the credit stays with the sponsor, and what moves is the visibility. The cycle turns more dangerous at a later stage, when the collateral stops being a hyperscaler's promise and starts being the asset itself.
Where the $6.3 Trillion Gap Comes From
The arithmetic in Marc Rubinstein's Net Interest column is simple. Estimates for AI infrastructure capex through 2030 range from $7.6 trillion to $8.0 trillion. Analysts at KKR estimate that traditional US public investment-grade investors can absorb about $1.7 trillion of that. The remaining $6.3 trillion, on the top end of the range, is expected to come from private investment-grade debt, asset-backed finance, infrastructure debt, structured solutions and, in KKR's phrase, "financing structures that may not even exist yet."
Look at the list. Its first item is still investment grade. The same insurers, pension funds and asset managers who buy public corporate bonds are the natural buyers of private IG paper, infrastructure debt and the senior tranches of asset-backed deals. The KKR split does not say that the world lacks investment-grade savings. It says that public IG issuance by the borrowers themselves cannot carry the load. Those two statements have very different implications.
If the binding constraint were the pool of savers, financing costs would have to rise sharply to pull capital away from other uses. If the binding constraint is how much debt a company is willing to report, the solution is structural. You build a vehicle that borrows, owns the asset, and leases it back. Bond investors get paper backed by a strong tenant. The company gets the asset without the leverage appearing as its own debt. That is what most of the gap is: a classification problem with a capital-markets fee attached.
Inside a Hyperscaler Off-Balance-Sheet Build
Meta's Hyperion campus in Richland Parish, Louisiana, illustrates the pressure. Ground was broken in December 2024 on a 2,250-acre site that Mark Zuckerberg said "will be able to scale up to 5GW over several years." Meta put the initial cost at $10 billion and began funding it from its own resources. Per the column, the scale of the project threatened to overwhelm the balance sheet, so Meta turned to Morgan Stanley. The detailed terms of that structure sit behind the column's paywall and are not reproduced here. What follows is the general architecture of this class of deal, not a description of Hyperion's specific terms.
A sponsor-backed data center financing usually has four moving parts:
- A special-purpose vehicle that owns the land, buildings and often the power infrastructure, financed mostly with long-dated bonds and a slice of equity from an outside investor.
- A lease from the vehicle to the hyperscaler, whose rent is sized to cover debt service.
- Some form of support from the hyperscaler beyond the lease itself, such as a residual value guarantee or a make-whole obligation if it walks away, which is what lets rating agencies rate the bonds close to the tenant's own credit.
- Accounting and rating treatment that keeps the vehicle's debt off the tenant's consolidated balance sheet, or classifies the obligation as something other than funded debt.
Point three is where the intuition fails. Bondholders in a structure like this are not mainly underwriting the data center. They are underwriting the tenant, with the building as secondary protection. If the vehicle's bonds trade close to the sponsor's own spreads, the market is telling you exactly that. The risk has not left the sponsor. It has moved from a line called "debt" to a set of lease commitments and guarantees that analysts have to dig out of footnotes.
That has two consequences. Reported leverage understates economic leverage, which means the headline balance sheet becomes a worse guide to a company's true commitments the longer the build continues. And the vehicle's own risk profile depends on a mismatch that is easy to overlook: the bonds can be long-dated, while the most valuable equipment inside the building depreciates over a few years, and the tenant's appetite for that specific site depends on technology choices no one can fix today.
Reading the Railroad Ladder Correctly
Rubinstein's historical parallel is the American rail boom. Between 1870 and 1890, rail capex ran at 2.24% of GDP. AI capex is projected to reach 3.63% of GDP out to 2032. The comparison needs care: one figure is a realized twenty-year average, the other a forward projection over a shorter window. On its face, though, the AI build would be roughly 1.6 times as intense relative to the economy.
The richer lesson is in the passage he quotes from Richard White's Railroaded. Investors moved from government bonds, to government-secured railroad bonds, to convertibles, to mortgage bonds, and on to "first mortgage bonds and second mortgage bonds," "mortgages on the trunk line and mortgages on the branches," land grant bonds, income bonds, and finally "bonds on anything and everything that investors might accept as collateral."
Read that as a sequence, not a list. Each rung moved further from the strongest credit and closer to the specific asset. The early paper leaned on sovereign backing. The late paper leaned on a branch line's ability to earn its keep, or on land whose value depended on the line being built.
Map it onto AI finance. Today's flagship structures sit near the top of the ladder: the asset is new, but the credit behind it is a company that generates enormous operating cash flow. The rungs below are visible already in outline. Debt secured on GPUs themselves. Data centers leased to tenants without investment-grade ratings. Financing for compute providers whose revenue depends on contracts with a few AI labs, which in turn depend on continued fundraising. Each step down substitutes asset value, or a weaker counterparty, for sponsor credit. The column's framing, that financial engineering "tends to flourish when conventional balance sheets can no longer keep pace" and can "obscure where its risks are accumulating," fits this pattern precisely. Obscured risk at the top of the ladder is mostly a disclosure issue. Obscured risk at the bottom is a solvency issue.
The Case That This Is Just Good Plumbing
The strongest counterargument deserves a fair hearing. The structures can be read as efficient rather than evasive. A hyperscaler that preserves its credit rating keeps its own funding cost low, which benefits every project it sponsors. Insurers and pension funds want long-dated, highly rated assets that match their liabilities, and few things fit that brief better than a lease from a company of Meta's scale. Outside equity investors get exposure to infrastructure returns they cannot easily buy elsewhere. Rubinstein frames the engineer's job as allocating "cash flows and risk among the parties most willing and able to bear them," and on that definition these deals are doing their job.
The railroad analogy also cuts the other way. Many nineteenth-century lines were built ahead of any traffic, financed by promoters with little income of their own. Today's largest AI spenders fund much of their capex from existing businesses that already generate cash. A build financed partly by advertising and cloud revenue is a sturdier foundation than a branch line financed by land grants.
All of that is correct for the top rungs. It does not settle the question for the lower ones, and it leaves one specific weakness at the top: guarantees and lease obligations are contingent claims. They look small in normal times and become very large precisely when a site is no longer wanted, which is also when several sites might be unwanted at once. The protection for bondholders is real. The question is how quickly it converts into balance-sheet debt for the sponsor if the demand forecast behind 5GW campuses proves too high.
The Rung to Watch
The decision variable is not the total volume of AI financing. That will be large regardless, because the capex estimates require it. The variable is the mix of collateral. As long as the marginal dollar is lent against hyperscaler credit, the boom's financing is opaque but sound, and the main cost is that headline leverage ratios mislead.
The view changes when the marginal structure leans on the asset rather than the tenant. Specifically, it becomes materially riskier when three things show up together: debt tenors that run well beyond the useful life of the equipment securing them, leases that tenants can exit without guaranteeing the residual, and a growing share of issuance from borrowers whose own credit is below investment grade. That combination would mark the move from government-secured bonds toward "bonds on anything and everything." KKR's phrase about structures "that may not even exist yet" is a forecast that the market will eventually reach those rungs. Where new AI deals sit on that ladder, rung by rung, is the cleanest available read on where this cycle stands.
