AI Infrastructure Debt Reaches $1 Trillion, Hidden Risks Emerge

The financing system supporting artificial intelligence infrastructure has grown to exceed $1 trillion, but most of this debt sits outside the consolidated balance sheets of major technology companies, creating exposure that traditional financial metrics fail to capture.

While attention has focused on software lending, the physical build-out of AI—chips, data centers, power connections, and specialized compute capacity—has been financed through a complex web of project loans, equipment-backed lending, asset-backed securities, and even loans secured directly by GPUs. Stefan Hepp, adjunct assistant professor at Chicago Booth, mapped this financing system in a new working paper and found that losses could reach $140 billion in a stress scenario.

The hyperscalers' expanding footprint

The five largest cloud computing companies—Alphabet, Amazon, Meta, Microsoft, and Oracle—spent approximately $380 billion on capital expenditures in 2025 and are projected to spend roughly double that in 2026. This marks a shift: capital spending is now on track to exceed operating cash flow for the first time.

Corporate bond issuance reflects this acceleration. The hyperscalers issued roughly $120 billion in bonds last year, compared to an average of $28 billion annually between 2020 and 2024. Issuance in the first half of 2026 has already surpassed the entire 2025 total.

Yet conventional debt figures tell only part of the story. S&P Global Ratings identified approximately $675 billion in signed but uncommenced lease obligations across these companies—commitments excluded from funded debt totals.

Where the financing actually sits

Hepp's analysis identifies several distinct channels. Investment-grade corporate bonds account for about $520 billion, held primarily by bond funds, insurers, and pension funds. Project and data-center finance adds another $250 billion, typically tied to specific facilities with shorter maturities. Infrastructure and asset-backed securities represent roughly $60 billion.

Private credit contributes an estimated $200 billion, though this figure is difficult to measure precisely. Specialist compute providers have borrowed about $35 billion secured directly against GPUs—a structure carrying significant risk given that these chips are displaced every two to three years as faster models emerge.

Vendor financing adds another layer of complexity. In September 2025, Nvidia agreed to a $6.3 billion deal to purchase CoreWeave's unsold data-center capacity through early 2032, effectively underwriting the residual value of its own hardware.

A June 2026 transaction illustrates how intricate these structures have become. Apollo and Blackstone created a special purpose vehicle to raise $35 billion for Anthropic's computing capacity. The debt went not to Anthropic but to purchase chips and lease them back. Senior lenders priced the debt close to investment-grade rates because Broadcom guaranteed to cover any shortfall if the chips sold for less than outstanding debt. Yet Broadcom's guarantee is weakest precisely when it would be triggered—if multiple AI labs simultaneously failed to honor commitments and flooded the secondary market with used chips.

The diversification illusion

Investors may hold positions across public tech equities, infrastructure funds, private credit vehicles, and real estate debt, viewing these as diversified. Legally they are separate, but economically they may share common dependencies: demand for AI computing, the durability of long-term capacity contracts, the creditworthiness of a few counterparties, and the residual value of specialized infrastructure.

Hepp's stress test suggests that a severe debt re-rating could reduce AI-linked equity values by $10 trillion to $14 trillion. While this wouldn't translate directly to credit losses—a market capitalization decline doesn't cause bonds to default—realized credit losses could still reach $60 billion to $140 billion.

Crucially, first-loss positions sit mainly outside the regulated banking system. Permanent losses would fall first on private credit funds and AI infrastructure vehicles, ultimately landing with pension plans, endowments, and private equity-owned insurance platforms.

Why it matters

Unlike the 2008 financial crisis, AI infrastructure debt is less concentrated in highly leveraged banks and relies less on short-term funding. This structure may reduce systemic contagion risk but makes the accumulation and distribution of risk harder to assess in real time. Losses would likely emerge as a sequence of refinancing difficulties, weaker fundraising, and gradual portfolio markdowns rather than a single visible crisis. For institutional investors, understanding which cash flows ultimately support seemingly diversified holdings has become essential—the same contractual relationship can appear differently depending on where it sits in a portfolio.

The analysis was first reported by Chicago Booth Review, based on Hepp's working paper examining AI infrastructure financing structures.