AI in action · Regulation
BIS: AI Debt Is Hidden in Vehicles Few Investors Can Map
AI data centres are increasingly financed with off-balance-sheet vehicles that concentrate risk in a few large firms, a BIS bulletin found. Spillovers to bonds are small so far.
The Bank for International Settlements has drawn a map of how the AI build-out is being financed, and its main finding is that the map is hard to read. In BIS Bulletin No 140, "Building AI on credit: structures and spillovers", published on October 27, economists Mathias Drehmann and Egemen Eren show that a rising share of AI infrastructure spending is funded with long-term debt and complex off-balance-sheet arrangements, and that the ultimate exposures sit with a handful of large companies, more than their balance sheets suggest.
The scale comes first. Capital spending by the five largest US hyperscalers is expected to exceed $800 billion by the end of 2026, around 2.5% of US GDP. Gross bond issuance by those five, Alphabet, Amazon, Meta, Microsoft and Oracle, reached $222 billion by September, mostly at long maturities and equal to more than 10% of long-term US Treasury issuance this year. A further $75 billion to $100 billion of bonds and asset-backed securities linked to off-balance-sheet vehicles has been placed, according to market sources cited by the authors, and Moody's puts the hyperscalers' data centre lease commitments at $1.2 trillion by mid-2026, about $820 billion of it for leases that have not started.
The typical structure is a special purpose vehicle that owns a data centre, borrows against it and leases capacity to a highly rated tech company. The bulletin's example is Hyperion, the joint venture between Meta and funds managed by Blue Owl, financed partly with $27.3 billion of privately placed A+ rated bonds backed by Meta's lease commitments, which Meta does not consolidate. Those bonds priced at 225 basis points over Treasuries in October 2025, while Meta's own 20- and 30-year bonds priced two weeks later at 88 and 98 basis points. Chip vendors add support too: in August 2026 Nvidia signed agreements with six investment firms for platforms aiming to raise more than $500 billion and said it may offer residual value support on up to 25% of a project.
The authors list four risks. Exposures are concentrated, since the hyperscalers' aggregate off-balance-sheet commitments already exceed their reported debt. The links are tangled, because a chip supplier may sell to, invest in, buy capacity from and guarantee the same customer, while banks lend both to the vehicles and to the funds holding their debt. Investors are only partly protected, as chips may lose value with each new generation and residual guarantees are capped. And the structures are opaque: leases not yet started are disclosed but not counted as debt, so an insurer holding a hyperscaler's bonds and the debt of vehicles leased to it may be more exposed than it realises.
For the wider bond market, the news is reassuring for now. On 18 days in 2026 when AI-financing news reached markets, the 10-year Treasury term premium rose by about 1.5 basis points, against a typical daily move of around 3.5, and the effect faded within two trading days. But the bulletin warns that sustained issuance at the current pace could make those spillovers larger and more persistent, and calls for closer monitoring and fuller disclosure of the commitments. With the SEC saying the same week that it is reviewing AI disclosures, the two messages point the same way.
Sources: BIS — Bulletin No 140, Building AI on credit: structures and spillovers, by Mathias Drehmann and Egemen Eren (Oct 27, 2026)