A BIS Bulletin Says the Off-Balance Sheet Structures Financing AI Data Centers Obscure the Final Allocation of Risks

The Bank for International Settlements published BIS Bulletin No 140, Building AI on credit: structures and spillovers, dated October 27, 2026. The bulletin says the views "are those of the authors and do not necessarily reflect the views of the BIS or its member central banks." It also says its references to individual firms and projects "are for illustrative purposes." Parts of the bulletin follow, not the whole.

Its first key takeaway is that capital expenditure to build AI infrastructure is growing fast, and that a rising share of it is financed with long-term debt and through complex off-balance sheet arrangements. Its first question, it says, concerns "how emerging financing structures allocate credit risk within the AI ecosystem and who bears losses if returns fall short of expectations."

Emerging Credit Relationships

The bulletin illustrates a stylized version of the structure in a graph. "At its centre is a special purpose vehicle that owns a data centre and earns its income from the contracts it has signed." The vehicle's debt is issued as a bond or an asset-backed security, often placed privately. "Investment-grade hyperscalers lease capacity from these vehicles, sometimes on short initial terms with options to renew." The newer cloud providers the bulletin calls neoclouds, and AI labs, "which are mostly unrated or rated below investment grade," sign similar contracts. A chip vendor sells the hardware and supports its customers' financing. Banks provide construction financing to some vehicles, and may also lend to the private credit funds that hold their debt.

"First, ring-fencing the data centre in a vehicle gives lenders a secured claim on a durable building," the bulletin says. "The building, its power connection and its cooling plant can be rented to other tenants, which makes them relatively easier to put to other uses."

The vehicle's debt is also backed by the data center's chips, which the bulletin calls "arguably less robust as collateral." Chips "have fewer alternative uses and might lose value with each new generation." Therefore, it says, "capacity contracts and initial lease terms typically run for a few years." And because several companies may hold similar hardware from the same vendor, resale values "could fall sharply" if many seek to sell used chips at the same time.

Risks the Bulletin Names

The bulletin says the same structures "might also carry risks to financial stability, especially if their adoption becomes more widespread," and names four.

The first is concentration. "The lease commitments and guarantees rest on a small number of large firms," leaving the holders of the debt dependent on those firms' financial strength. The second is interconnection: a chip supplier may sell hardware to a customer, hold a stake in it, buy its capacity and guarantee some of its payments, so trouble at the customer could reach the supplier through each of these channels. Banks may be exposed directly, through construction loans to vehicles, and indirectly, through loans to the funds that hold vehicle debt.

The third is that the contractual support protects only in part. "If a hyperscaler does not renew a lease or terminates it early, the vehicle may need to find a new tenant or realise the value of the assets, potentially at a time when similar capacity is also being offered to the market." The vehicle can fall back on residual value guarantees, "but they are typically capped."

The fourth is opacity. "Leases that have not yet started are disclosed but not recorded as debt. Guarantees are often recorded at little or no value until payments become likely." Without a clear view of the ultimate guarantor, the bulletin says, lenders "may hold larger exposures to a single firm than they realise." And it says, "Closer monitoring and fuller disclosure of these commitments would help investors, rating agencies and supervisors assess the risks."

Spillovers to Bond Markets

The bulletin's second question is whether the borrowing spills over into wider bond markets. On 18 days in 2026 when news about AI financing reached markets, it finds that "movements in Treasury term premia and investment-grade corporate spreads were small and short-lived." It calls its estimates "event-day associations, not causal effects."

"Immediate spillovers to Treasury and investment-grade corporate bond markets therefore appear limited, but the potential pressure is growing," the bulletin says. If supply expands at this rate, "a concentrated group of duration investors will need to absorb a larger stock of claims with correlated exposures. This could exert greater pressure on long-term yields and, where leverage is shifted into opaque vehicles, add to financial stability risks."