AI's Debt-Funded Buildout Becomes a Credit Market Test

The Bank for International Settlements now counts roughly $1 trillion riding on AI infrastructure, and its flagship annual report warns the boom is headed for a reckoning, according to Fortune. That number deserves attention for a reason most AI coverage skips: the money is increasingly borrowed, not earned. Venture capital and retained corporate cash built the early phases of this buildout. The next phases are being financed by lenders who must now decide how much future AI revenue can support today's debt.

From Equity to Credit

Seeking Alpha frames the shift plainly: data centers, specialized chips, power contracts and networking infrastructure demand enormous upfront spending while most eventual returns remain uncertain or years away. That mismatch between when money goes out and when money comes in is the classic opening for debt financing, and Wall Street has moved to fill it. Bloomberg's markets team has tracked how AI debt is reshaping Wall Street itself, with underwriting desks building new structures for a borrower profile that barely existed five years ago.

The scale is easier to see in the capex numbers than in the bond prospectuses. Amazon, Alphabet, Meta and Microsoft burned roughly $416 billion on AI infrastructure in 2025, and free cash flow is shrinking as the bills arrive, as we laid out in Big Tech's AI capex tab. Nikkei's analysis of the same five companies plus Oracle found $1.65 trillion in off-balance-sheet AI commitments, more than their reported debt combined; the details are in Big Tech's $1.65 trillion off-balance-sheet AI debt. When obligations that large sit outside the balance sheet, the credit market becomes the place where the real risk assessment happens, whether or not investors are watching.

Why This Cycle Might Be Different

The strongest argument against an easy comparison to the dot-com bust is physical. Fiber-optic overbuild in the late 1990s left creditors holding assets with little resale value and cash flows that never materialized. This time, the spending is tied to data centers with long-term leases, power purchase agreements, and chips that someone has contracted to buy. Seeking Alpha's argument is that these physical assets and contracts give creditors collateral and recurring cash-flow claims that software promises never offered. A data center with a fifteen-year lease from an investment-grade tenant can be underwritten like real estate. A startup's burned equity cannot.

That argument holds up to a point, and then it runs into three specific exposures.

First, power. Data centers are only as good as their electricity supply, and power costs are volatile while lease rates are fixed. A borrower who signed a long-term lease and a variable-price power contract has a claim on future revenue and a mismatch on the expense side.

Second, obsolescence. Chips depreciate faster than buildings. A GPU that anchors a financing model today may be superseded in three years, which shortens the useful life of the collateral compared with the life of the loan.

Third, concentration. The tenant base for hyperscale data centers is a handful of companies. If a lender's collateral is a building leased to two cloud providers, the credit analysis is really an analysis of those two companies' AI revenue, however diversified the loan documentation claims to be.

What the Bond Market is Saying

CNBC's take on the shift is that tech investors now have new reasons to watch the bond market, because financing conditions, not earnings guidance, will determine whether projects move forward. The mechanics are simple: if yields stay low, leverage accelerates capacity; if yields rise, the same leverage magnifies losses and pushes companies to delay projects. Investors trained to read earnings calls now need to read credit spreads.

The open question is where the losses land if demand disappoints. Seeking Alpha's framing is that the repayment risk could sit with chip buyers, infrastructure landlords, cloud providers, lenders or customers, and the answer changes who feels a bust. A slowdown that forces landlords to renegotiate leases lands on real estate investors and, through them, on the pension funds and insurers holding those bonds. A slowdown that squeezes cloud providers' margins lands on their shareholders and employees. A slowdown that hits customers through AI service repricing lands on businesses that restructured operations around a tool they can no longer afford at scale.

Federal Reserve officials are already watching the boom for financial stability risk, as our earlier reporting on Fed officials scrutinizing AI investment risks noted. Their concern maps onto the same question: the people most exposed to an AI credit event will not be the ones who chose the exposure. A teacher's pension is upstream of a private credit fund that holds a data center loan, and the teacher cannot audit the lease terms.

The Opaque Middle

Barry Eichengreen's September commentary for Project Syndicate identifies the structural gap: much of this lending is flowing through private credit markets, where terms are disclosed to a small set of investors rather than marked publicly. Public bond markets force repricing daily. Private credit funds mark quarterly, sometimes on their own models, sometimes with valuation agents they select. Eichengreen's concern is that opacity in private debt markets will delay the discovery of losses until they are larger, spreading stress through banks and funds rather than concentrating it where it can be resolved early.

That critique deserves its strongest counterargument. Private credit's defenders argue that negotiated loan terms, tighter covenants, and direct lender relationships reduce the reflexive selling that turns stress into crisis in public markets. A syndicated bondholder sells first and asks questions later; a direct lender can restructure. Both descriptions can be true depending on the loan, which is precisely the problem: without public marks, nobody outside the deal can tell which loans are the resilient kind.

History's Unclear Verdict

The BIS has compared the current boom to historical investment busts, including telecom and railway manias that ended in widespread default, according to International Business Times. The historical record, read carefully, cuts both ways. Railway investment in the 1840s destroyed many investors and left Britain with a rail network it still uses. The fiber-optic overbuild bankrupted WorldCom and Global Crossing but bequeathed cheap bandwidth that made the modern internet possible. Busts following genuine productivity revolutions have historically produced social gains even as individual creditors took losses. The BIS is not claiming the technology is worthless; it is claiming the financing structure can fail before the technology succeeds.

That distinction is where the honest uncertainty sits. If AI revenue grows into the capacity, today's debt will look like the financing behind the railways, and the lenders who extended it will have funded something durable. If demand plateaus at levels well below current buildout, the debt becomes the story, because physical collateral with fast-depreciating equipment and concentrated tenants is worth far less in a downturn than its underwriting assumed.

The practical test to apply to any AI financing is the one Seeking Alpha suggests: ignore the headline capex and ask who bears the repayment risk, and whether that party has priced the downside. When the answer runs through private credit funds nobody can audit, held on behalf of pensioners nobody told, backed by chips that lose value faster than the loans amortize, the test is still being taken, and the grade arrives only when yields rise or demand does not.

The $1 trillion question, in other words, is not whether AI works. It is whether the people who borrowed to build it can pay back the people who lent to fund it, on schedules written for a future neither has seen yet.

By Raj Mehta, Buzzrag Global Markets and International Finance Reporter