The global buildout of AI infrastructure is consuming capital at an unprecedented pace, yet the debt engine powering this "super-cycle" has quietly expanded to levels that are raising market concerns. As an increasing number of tech giants slide into negative free cash flow, the demand for external financing is set to climb relentlessly.

According to the latest projections from Goldman Sachs, direct debt issuance by global hyperscalers is expected to reach $420 billion in 2027, marking a surge of more than 60% compared to the estimated $250 billion for 2026. Meanwhile, Bank of America has calculated in its latest report that, by 2030, the combined external financing needs of core AI infrastructure players—including Microsoft, Amazon, Alphabet, Meta, Oracle, SpaceX, CoreWeave, and Nebius—will total between $1.2 trillion and $1.5 trillion. If debt serves as the primary financing channel, this is projected to inject over $300 billion in incremental supply into credit markets annually, with the potential to spike as high as $500 billion in the near term.

This wave of debt expansion is not confined to the investment-grade bond market; it is also seeping through off-balance-sheet special purpose vehicles (SPVs), infrastructure financing, and private credit channels.

Hyperscaler Free Cash Flow Squeeze Makes Debt the Only Way Out

The most immediate financial consequence of the AI investment frenzy is the sharp deterioration in cash flow positions among major hyperscalers. Excluding Microsoft, all other major hyperscalers have already fallen into negative free cash flow. This means that future capital expenditure expansion relies almost entirely on external financing—with debt being the preferred route.

Bank of America estimates that from 2026 to 2028, the combined operating cash flow of the eight core companies will be approximately $3.3 trillion, while total capital expenditures will reach $3.7 trillion (including $2.9 trillion for AI-related capex), leaving a net financing gap of roughly $400 billion over the three-year period. Extending the forecast horizon to 2030, operating cash flow rises to $6.6 trillion and total capex to $6.9 trillion, with the net gap still hovering around $300 billion. When factoring in capital deployment needs such as dividends, buybacks, and mergers and acquisitions, the total funding shortfall expands to $800 billion.

Bank of America emphasizes that debt holds a significant cost advantage over equity—the after-tax cost of debt for hyperscalers is approximately 5%, whereas equity financing costs exceed 11%, a gap of more than double. This makes debt financing the dominant choice for the foreseeable future. However, for companies like Oracle, where borrowing costs have recently climbed sharply, this logic is beginning to lose its applicability.

Off-Balance-Sheet "Time Bomb" Keeps Growing

Beyond direct bond issuance, another risk exposure in the AI financing system that cannot be overlooked stems from off-balance-sheet arrangements. Year-to-date, total AI-related dollar debt issuance across credit markets has reached $568 billion, comprising $259 billion in investment-grade bonds, $256 billion in private credit, direct loans, and other bilateral or non-public financing, $40 billion in high-yield bonds, and $11 billion in institutional loans.

On top of the disclosed on-balance-sheet debt, off-balance-sheet commitments and obligations totaling as much as $3.1 trillion—primarily consisting of $1.1 trillion in undiscounted lease payments and $1.7 trillion in purchase commitments—have grown by $1.3 trillion in just three months. These off-balance-sheet exposures, channeled through SPVs and infrastructure financing structures, map directly onto the private credit market.

Vendor Financing Unlocks the Door for "Unfinanceable" Assets

Among external financing channels, the "vendor financing" mechanism led by chip suppliers is playing an increasingly pivotal role—its core function is to convert assets that were previously difficult to finance into standardized bonds that can access the bank credit market.

Take Nvidia as an example: through "take-or-pay" structures, it provides six-year minimum revenue guarantees for certain Neocloud customers, setting a contractual revenue floor for lenders. This substitutes for hyperscaler offtake endorsements, allowing the related financing to gain market acceptance.

Broadcom has gone even further through its AI XPV Platform. In a debut $35 billion debt package jointly issued with Apollo and Blackstone, Broadcom provides direct guarantees on $31 billion of senior notes and chip residual value (representing 87% of total debt). The value of these guarantees is clearly reflected in pricing differentials: guaranteed A2-rated notes carry an interest rate of 5.75%, while unguaranteed second-lien notes are priced at 8.5%. Bank of America analysts estimate that if the platform expands to 20GW in scale, Broadcom's peak residual value guarantee (RVG) exposure could reach approximately $370 billion by mid-2029.

Bank of America's report points out that the essence of these supplier credit mechanisms is to shift the determination of liquidation rates from capital users to suppliers, thereby converting exposures that previously could not enter conventional credit markets into financeable assets. But at the same time, this also means that a substantial volume of contingent liabilities is quietly accumulating on chip suppliers' balance sheets.

Credit Feedback Loop: The Self-Correcting Limits of the Debt Bubble

Although debt expansion continues, Bank of America also cautions that the market possesses inherent risk-correction mechanisms. The sharp volatility in AI credit spreads in July 2026 provided a preview—spread widening forced some issuers to pause bond issuance plans, confirming that when the cost of debt exceeds the economic rationality threshold for capital allocation, issuers will pivot to equity, convertible bonds, or even proactively cut capital expenditures.

The report also warns that the main constraints on the pace of AI infrastructure development may not be capital itself, but rather physical and institutional bottlenecks such as power supply, regulatory approvals, and construction timelines.

The rapid rise of open-source models constitutes another layer of potential pressure. As open-source solutions capable of delivering near-frontier model performance at minimal cost gain market penetration, token prices for leading frontier models have fallen by more than 50% since the June peak, hitting historic lows. This not only undermines the commercial viability of frontier models but also puts pressure on their trillion-dollar-level IPO valuations. If the profitability inflection point for the token economy continues to be delayed, the revenue assumptions underpinning the entire AI debt ecosystem will face a fundamental reassessment.