Key Takeaways
- • Hyperscalers (Alphabet, Amazon, Meta, Microsoft, Oracle) spent approximately $380 billion on capital expenditures in 2025 and are expected to roughly double that in 2026, putting capex on course to overtake operating cash flow.
- • The hyperscalers issued approximately $120 billion of corporate bonds in 2025, versus an average of about $28 billion annually from 2020 to 2024, and first-half 2026 issuance has already surpassed the 2025 total.
- • Author's research maps about $520 billion of senior hyperscaler bonds, roughly $250 billion of project and data-center finance, about $60 billion of infrastructure and ABS, and an estimated $200 billion of private credit, with about $35 billion of loans secured directly against GPUs.
- • A stress test estimates realized AI-related credit losses of $60 billion to $140 billion, with first-loss positions sitting mostly outside the regulated banking system, principally with private-credit funds, their LPs (pensions and endowments), and PE-owned insurers.
- • Investors should map shared cash-flow dependencies across public equities, infrastructure funds, private credit and insurer holdings, since legally separate AI exposures may share a common demand-driven risk.
Summary
Concern about AI-related debt has largely centered on software, where private-credit lending to software companies exceeded $500 billion by the end of 2025, according to the University of Chicago Booth School of Business article. The author argues that the broader AI infrastructure build-out, covering chips, data centers, power, servers and compute capacity, has largely been financed off the largest technology companies' balance sheets through project finance, equipment-backed lending, leases, asset-backed securities and private credit. Some loans are secured by GPUs, whose useful economic life is shortened by generational upgrades every two or three years, leaving used-chip values uncertain.
The author states that the hyperscalers are not in financial difficulty, but that the marginal financing of AI infrastructure now extends well beyond their balance sheets. S&P Global Ratings has identified about $675 billion of signed but not yet commenced hyperscaler lease obligations, which are excluded from the funded-debt figures in the analysis. The research identifies about $520 billion of senior unsecured hyperscaler bonds, $250 billion or so of project and data-center finance, and roughly $60 billion of infrastructure and asset-backed securities. The roughly $200 billion private-credit figure is described as an estimate rather than an observed total, and specialist compute providers have borrowed about $35 billion secured against GPUs and other assets.
Vendor financing and credit enhancement are described as among the least transparent parts of the market. The article cites Nvidia's September 2025 agreement to buy $6.3 billion of unsold CoreWeave data-center capacity through early 2032. It also describes a June 2026 structure in which Apollo and Blackstone created a special purpose vehicle to raise $35 billion of debt for Anthropic's computing capacity; the SPV purchases chips and leases them back, and Broadcom has promised to cover any shortfall if chips sell for less than the outstanding debt. The author warns that Broadcom's guarantee is weakest in the conditions that would trigger it.
The article argues that asset-allocation categories obscure common economic exposures. An institution holding hyperscaler stock, a data-center fund, a private-credit vehicle financing compute providers and insurer-held infrastructure securities may still depend on a shared set of variables, including AI demand, long-term capacity contracts, counterparty creditworthiness and refinancing ability.
The author's stress calibration suggests a severe debt re-rating could reduce AI-linked equity values by roughly $10 trillion to $14 trillion, which the author describes as illustrative rather than a forecast. Realized credit losses are estimated at $60 billion to $140 billion. The author expects permanent losses to fall first on private-credit funds and dedicated AI-infrastructure credit vehicles, ultimately resting with their limited partners, principally pension plans and endowments, as well as PE-owned insurance platforms and, to a lesser extent, BDC investors. Banks are described as sitting higher in the capital structure, and the author concludes the outcome would differ from 2008 with smaller aggregate credit losses than equity losses. The article closes with five questions for investors: the source of cash flow, durability of commitments, refinancing at maturity, position in the capital structure, and shared dependencies across the portfolio.
The author states that the hyperscalers are not in financial difficulty, but that the marginal financing of AI infrastructure now extends well beyond their balance sheets. S&P Global Ratings has identified about $675 billion of signed but not yet commenced hyperscaler lease obligations, which are excluded from the funded-debt figures in the analysis. The research identifies about $520 billion of senior unsecured hyperscaler bonds, $250 billion or so of project and data-center finance, and roughly $60 billion of infrastructure and asset-backed securities. The roughly $200 billion private-credit figure is described as an estimate rather than an observed total, and specialist compute providers have borrowed about $35 billion secured against GPUs and other assets.
Vendor financing and credit enhancement are described as among the least transparent parts of the market. The article cites Nvidia's September 2025 agreement to buy $6.3 billion of unsold CoreWeave data-center capacity through early 2032. It also describes a June 2026 structure in which Apollo and Blackstone created a special purpose vehicle to raise $35 billion of debt for Anthropic's computing capacity; the SPV purchases chips and leases them back, and Broadcom has promised to cover any shortfall if chips sell for less than the outstanding debt. The author warns that Broadcom's guarantee is weakest in the conditions that would trigger it.
The article argues that asset-allocation categories obscure common economic exposures. An institution holding hyperscaler stock, a data-center fund, a private-credit vehicle financing compute providers and insurer-held infrastructure securities may still depend on a shared set of variables, including AI demand, long-term capacity contracts, counterparty creditworthiness and refinancing ability.
The author's stress calibration suggests a severe debt re-rating could reduce AI-linked equity values by roughly $10 trillion to $14 trillion, which the author describes as illustrative rather than a forecast. Realized credit losses are estimated at $60 billion to $140 billion. The author expects permanent losses to fall first on private-credit funds and dedicated AI-infrastructure credit vehicles, ultimately resting with their limited partners, principally pension plans and endowments, as well as PE-owned insurance platforms and, to a lesser extent, BDC investors. Banks are described as sitting higher in the capital structure, and the author concludes the outcome would differ from 2008 with smaller aggregate credit losses than equity losses. The article closes with five questions for investors: the source of cash flow, durability of commitments, refinancing at maturity, position in the capital structure, and shared dependencies across the portfolio.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.