AI infrastructure debt risks grow as financing moves beyond corporate bonds
AI’s data-center buildout is drawing on leases, project loans and private credit, making leverage harder to measure.
By Theo Nakamura · Staff Writer
· 4 min read
AI infrastructure debt risks are drawing closer attention as technology companies and their partners fund data centers through more than ordinary corporate borrowing. CNBC reported that rising hyperscaler bond issuance, complex financing structures and the failure of AI-focused hedge fund Situational Awareness have sharpened questions about how much leverage supports the buildout and where it sits.
For everyday investors, the distinction matters. A company’s reported debt captures only one part of the financial commitments connected to AI computing capacity. Leases, project finance and equipment loans can place claims on the same expected AI demand in different corners of the market.
How is AI infrastructure being financed?
Bonds remain one channel, but CNBC reported that hyperscalers and their backers are also using joint ventures, leases and other arrangements. Chicago Booth Review separately described project and construction finance, equipment-backed lending, asset-backed securities, private credit and special-purpose vehicles, or SPVs, as part of the funding mix.
An SPV is a legally separate entity created for a defined project or pool of assets. In an AI infrastructure deal, it can own equipment or a data-center project and borrow against expected lease payments or asset collateral. That can distribute risk among tenants, developers, lenders, insurers, suppliers and investors rather than leaving it solely with the technology company using the capacity.
Corporate bonds: debt issued by a hyperscaler, with repayment backed by the company.
Leases: commitments to pay for facilities or equipment over time.
Project and equipment finance: borrowing tied to a specific data center, server fleet or other asset.
Private credit and structured financing: lending and securities that can spread exposure among different investor types.
The dollar amounts are substantial, though they measure different things. The Federal Reserve said in February that U.S. data-center spending was expected to exceed $500 billion in 2025. Moody’s Ratings forecast that six hyperscalers it rates would spend close to $800 billion in 2026 and close to $1 trillion in 2027. Moody’s said the companies are large, established businesses that often have strong balance sheets, while warning that heavier capital spending and additional borrowing can pressure credit measures and free cash flow.
Why are leases harder to read than debt?
Goldman Sachs analysts estimated that hyperscalers had $1.5 trillion of combined lease commitments for data centers, research facilities, offices and equipment, CNBC reported. About $1 trillion of that estimate was for leases that had not yet commenced. It is an analyst estimate of broad commitments, not a uniform measure of AI-only debt.
An uncommenced lease is a signed arrangement for which the lease period has not started. The related future payments can become a cash-flow demand once the lease begins, even though the commitment is not yet recognized in financial statements in the same way. Deloitte’s summary of lease-accounting requirements says disclosures cover leases that have not commenced, as well as the timing and uncertainty of lease cash flows.
The risk is also about overlap. Chicago Booth Review said legally separate holdings, including technology shares, infrastructure funds, private-credit vehicles and data-center lending, can share exposure to future AI demand. Its author estimated some parts of this financing system, while stressing that private-credit totals are not fully observable.
Equipment-backed lending adds another uncertainty. Chicago Booth Review reported that some specialist computing providers have pledged GPUs as collateral. If a borrower cannot pay, a lender may need to sell the chips, and the resale value of used GPUs is uncertain as newer hardware arrives.
CNBC’s report on Situational Awareness is a separate warning about market leverage rather than evidence of trouble in data-center project debt. The hedge fund could not meet margin calls after losses on concentrated, borrowed equity positions, illustrating how borrowing can magnify swings in AI-linked securities.
A practical way to follow the buildup is to separate reported corporate debt, signed but uncommenced leases, project or SPV and equipment financing, vendor guarantees, and the investors ultimately holding each claim. Some figures are reported obligations, others are analyst estimates, and some parts of the system remain difficult to measure.
This story draws on original reporting from CNBC.
