The Debt Machine Behind the AI Boom

How Wall Street is quietly financing trillion-dollar data centers — and why some of that risk is landing in commonplace retirement portfolios

Note: This piece has been fact-checked against reporting from Bloomberg, CNBC, Fortune, Nikkei, Tom’s Hardware, Forbes, and analyst notes from Goldman Sachs and Moody’s.

The artificial intelligence build-out has become one of the largest capital projects in modern history, and it is no longer being funded primarily out of Big Tech’s famously deep cash reserves. In 2025 alone, AI-related companies and projects tapped debt markets for at least $200 billion — likely an undercount, since many deals are private — and Wall Street projects hundreds of billions more in issuance for 2026. Morgan Stanley and JPMorgan analysts estimate the infrastructure push could ultimately drive up to $1.5 trillion in additional tech-sector borrowing, while UBS forecasts as much as $900 billion in new issuance in 2026 alone. Oracle recently completed an $18 billion bond sale that made it, by Citi’s count, the largest investment-grade issuer among non-financial U.S. companies.

Veteran credit bankers say they have never seen anything like it. Matt McQueen, who oversees global credit and securitized products at Bank of America, told Bloomberg the numbers are unlike anything seen in a 25-year career, and that “you have to turn over all avenues to make this work.”

The off-balance-sheet layer – ‘Margin Equity’ Loans

The headline bond sales are only part of the picture. A large and growing share of the financing sits off the hyperscalers’ balance sheets entirely, structured through special purpose vehicles (SPVs) and long-term leases.

The mechanics work like this: a developer or financial sponsor sets up a separate legal entity — the SPV — which borrows the money, builds the data center, and leases it back to a tech giant under a long-term contract. Because the tech company doesn’t control the SPV, the debt isn’t recognized on its balance sheet; the lease obligation appears only in footnotes until the facility comes online. The signature example is “Beignet,” a roughly $30 billion financing for a Meta data center in Louisiana, payable by an SPV rather than by Meta itself.

The aggregate numbers are startling. A Nikkei analysis found that Alphabet, Amazon, Meta, Microsoft, and Oracle carry roughly $1.65 trillion in data center obligations listed off their balance sheets — about 122% of the debt they do report. Goldman Sachs analysts estimate hyperscaler lease commitments have grown from about $200 billion five years ago to $1.5 trillion today, including roughly $1 trillion in “uncommenced” leases not yet reflected in financial statements. Goldman warned this “can understate leverage and future liquidity needs.” Moody’s has cautioned that the six largest hyperscalers — projected to spend $785 billion this year and nearly $1 trillion in 2027 — face a “material shift” in their balance sheets, with capital expenditures now exceeding earnings at several firms. Alphabet went free-cash-flow negative for the first time as quarterly AI capex hit $44.9 billion.

Critics have drawn comparisons to Enron, which collapsed in 2001 partly because of debts hidden in off-balance-sheet entities. The comparison is imperfect — these lease structures are legal, disclosed in footnotes, and represent an accepted accounting practice — but the scale of obligations building outside conventional balance-sheet scrutiny is genuinely unprecedented.

Chips as collateral ‘NVDA’ will ‘Buy Back’ 5 y/o chips from the Margin Equity Loan Holder, if data centers can’t pay back loans.

A second, riskier layer of the machine treats the GPUs themselves as collateral. In GPU-backed SPV structures, a vehicle purchases the chips and leases them to the AI company; the debt is secured by the hardware, not the borrower’s corporate assets. xAI’s roughly $20 billion lease structure for its Memphis “Colossus 2” supercomputer — about $7.5 billion in equity and $12.5 billion in debt — is the flagship example, with a notable twist: Nvidia itself invested up to $2 billion in the SPV’s equity, effectively helping finance demand for its own products.

The obvious problem is depreciation. A data center built today, and the chips inside it, could become obsolete before the debt financing them is repaid. Investor Michael Burry has publicly challenged the depreciation schedules the industry uses; Nvidia CEO Jensen Huang counters that GPUs like the A100 keep working for close to a decade. The gap between those two views is, in effect, the gap between a sound loan and a bad one.

The Nvidia backstop — real, but narrower than rumored

In August 2026, Nvidia announced financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR intended to unlock more than $500 billion in outside capital for data centers, chip fabs, and power plants. To address the depreciation objection head-on, Huang said Nvidia may provide residual-value support covering up to 25% of a given financing opportunity, assessed project by project, if chips fail to hold their expected value at the end of a term.

Several caveats matter. The arrangements are memoranda of understanding, not executed contracts; no rate, tenor, or first-loss terms have been published. Huang has described the support as “supplementary,” suggesting Nvidia does not absorb first losses, and said its exposure runs below other compute-financing arrangements — Meta, for instance, gave its Hyperion joint venture a residual-value guarantee covering sixteen years, which helped carry that paper to an A+ rating from S&P, and Broadcom’s AI financing structure reportedly covers 100% of shortfalls to senior tranches. Nvidia’s most recent 10-Q disclosed existing facility-lease guarantees with maximum gross exposure of just $3.5 billion, so a backstop at the announced scale would be structurally and quantitatively different from anything the company has actually booked. Burry called the plan a “sign of desperation”; Bloomberg has tallied roughly $750 billion in circular Nvidia-linked deals struck in a single summer.

Where the risk lands – ABC Mom Pop Corp. Pensions

The end buyers of this ‘depreciateing assets’ debt are the institutions: insurers, pension funds, and endowments. The logic is structural. Data centers are expensive, long-lived projects; insurers and pension funds hold long-dated liabilities — annuities and retirement benefits payable decades out — and hunt for long-duration assets to match them. Private-credit and infrastructure managers act as middlemen, converting data center projects into debt those institutions can hold. As Bloomberg put it, whether you’re an institutional investor or an individual saver, the fixed-income side of your portfolio is getting more and more AI-heavy — often without the saver knowing it.

If AI revenue fails to materialize on schedule, the failure modes are prosaic rather than cinematic: refinancing becomes difficult as debt comes due, borrowing costs rise just as revenue disappoints, SPV equity holders must inject cash to reduce their own returns, and in the extreme case, bankruptcy — with lenders left to recover value from rapidly depreciating hardware and purpose-built facilities. A Citi credit strategist captured the mood: “There is something inherently uncomfortable as a credit investor about the transformation of the sort we’re facing.”

Even Microsoft CEO Satya Nadella seemed to acknowledge the historical rhyme, recommending on a recent earnings call the book 1873 — a history of railroad-era financial engineering that ended in a crash.

What this means for individual investors

None of this requires believing in secret schemes or fraud. The documented facts — $1.65 trillion in off-balance-sheet obligations, chips as collateral for hundreds of billions in loans, unresolved disputes over how fast that collateral loses value, and pension money absorbing the paper — describe a system with genuine concentration risk, built at historic speed, whose central assumption is that AI revenue arrives before the debt does.

Reasonable steps for individuals are unglamorous: understand what your bond funds and target-date funds actually hold, since AI-linked credit is increasingly embedded in “safe” fixed-income allocations; check whether your equity holdings depend heavily on hyperscaler capex continuing at current rates; and treat any single narrative — bull or bear — with the skepticism the incomplete public record deserves. Market corrections of 20% or more are historically routine events; portfolios built to survive them don’t need to predict when.

Sources

• Bloomberg (via Insurance Journal / EnergyNow): “The $3 Trillion AI Data Center Build-Out Becomes All-Consuming for Debt Markets” (Feb 2026)
• Nikkei (via Tom’s Hardware): “AI tech companies have ‘hidden debt’ worth around $1.65 trillion” (Jul 2026)
• CNBC: “AI infrastructure debt and leverage draw market scrutiny” (Aug 14, 2026); “Dust to data centers” (Dec 31, 2025)
• Forbes: “Big AI Data Center Owners Are Massively Expanding Their Debt” (Jul 23, 2026)
• Fortune: “Nvidia found a new way to keep the AI boom funded: your retirement money” (Aug 12, 2026)
• MLQ News: “Nvidia’s $500 billion financing plan is a framework, not a loan book” (Aug 2026)
• Futurism: “AI Companies Are Trying to Hide a Staggering Amount of Debt” (Jul 2026)
• Roic News / carboncredits.com: xAI $20B GPU lease SPV reporting (Oct 2025)
• Bird & Bird: “GPU-Based Financing in the Global Data Center Market” (2025)