The Bank for International Settlements published Bulletin No 140, “Building AI on credit: structures and spillovers”, on 27 October. Its authors, Mathias Drehmann and Egemen Eren, find that capital expenditure on artificial-intelligence infrastructure “is growing fast, and a rising share is financed with long-term debt and through complex off-balance sheet arrangements”. They give no single figure for the debt-financed share, but they size the pieces.
Capital expenditure by the five largest US hyperscalers is expected to exceed $800bn by the end of 2026, around 2.5% of US GDP. Their gross bond issuance reached $222bn by September. Market sources cited by the authors put bonds and asset-backed securities linked to off-balance-sheet vehicles at $75bn to $100bn so far this year. Moody’s Ratings, also cited, puts the largest hyperscalers’ data-centre lease commitments at $1.2trn by mid-2026, of which around $820bn relate to leases that have not yet started.
How the structures work
A special purpose vehicle owns a data centre and earns income from its contracts. Its debt, a bond or asset-backed security, is often placed privately with several institutional investors. Investment-grade hyperscalers lease capacity from the vehicle, sometimes on short initial terms with options to renew, and in some cases take a minority stake or guarantee the assets’ residual value.
The Bulletin’s example is Hyperion, a joint venture between Meta and funds managed by Blue Owl, financed partly with $27.3bn of privately placed, A+ rated bonds. Meta’s lease commitments begin in 2029 and initially run for four years, with options to renew for up to 20 years in total, and Meta does not consolidate the venture. The bonds priced at 225 basis points over Treasuries in October 2025; two weeks later Meta’s own 20- and 30-year bonds priced at 88 and 98 basis points. Hyperscalers gain capacity “without adding debt directly to their balance sheets”.
Where the risk sits in insurer and fund portfolios
The authors identify four weaknesses. Exposures are concentrated on a few firms whose off-balance-sheet commitments already exceed their reported debt, according to Moody’s. Links are dense, since a chip vendor may sell to, invest in and guarantee the same customer. Residual value guarantees are “typically capped”. And leases that have not started are disclosed but not recorded as debt, while guarantees are often recorded at little or no value until payments become likely.
The authors write: “An insurer that holds a hyperscaler’s bonds and the debt of several vehicles leased to it, for example, could be more exposed to that hyperscaler than its bond holdings alone suggest.” By the same logic, a pension fund holding a hyperscaler’s shares and bonds alongside private credit funds that lend to its vehicles could be similarly exposed; the Bulletin’s own example is an insurer. On that logic, issuer-by-issuer limits would understate the exposure unless they are rebuilt around the ultimate tenant or guarantor. Losses, the paper says, could fall on insurers, funds and their bank lenders.
Market spillovers are small so far
On 18 days in 2026 when AI-financing news reached markets, the 10-year Treasury term premium rose by around 1.5 basis points after controls, and the increase reversed within two trading days. Investment-grade spreads did not change significantly. Since most of the debt is long term, a run on short-term funding also looks less likely, the authors judge.
The caveat is scale. In 2026 hyperscaler bonds of ten years or more were about 24% of corporate issuance at those maturities and 13% of gross Treasury issuance, up from about 3% and 1.5% in 2024. If supply keeps growing at this rate, a concentrated group of duration investors will have to absorb more claims with correlated exposures. A higher term premium would lower the present value of pension liabilities but mark down long bonds and raise borrowing costs beyond the AI sector.
What to watch
The authors argue that “Closer monitoring and fuller disclosure of these commitments would help investors, rating agencies and supervisors assess the risks.” Microsoft’s quarterly filing of 28 October showed $364.6bn of leases, primarily for data centres, that had not yet commenced. Watch whether rating agencies and supervisors start to treat such commitments more like debt.
Sources
- Building AI on credit: structures and spillovers (BIS Bulletin No 140, PDF), Bank for International Settlements, 27 October 2026
- BIS Bulletin No 140: Building AI on credit: structures and spillovers, Bank for International Settlements, 27 October 2026
- Form 10-Q for the Quarterly Period Ended September 30, 2026, Microsoft Corp. (SEC), 28 October 2026