AI’s debt boom: where data-centre risk sits off balance sheet

AI data-centre funding does not sit in one capital-expenditure line. A hyperscaler may issue bonds directly, while a joint venture or special-purpose vehicle borrows from private lenders, builds the facility and services that debt with long leases or capacity payments. One site can therefore produce assets, liabilities and contracts across several sets of accounts.

‘Off balance sheet’ does not mean that the debt has vanished, or that every structure conceals distress. A non-consolidated vehicle’s borrowing can remain with that vehicle, while the customer reports leases, purchase commitments, guarantees or an equity stake elsewhere in its disclosures. The investor’s task is to trace who provides more cash and who absorbs losses if construction overruns, utilisation disappoints or refinancing fails.

Evidence cut-off: 23 August 2026.

Announced plans and actual spending are different numbers

A Federal Reserve project-level estimate put annualised US data-centre investment at about US$370 billion in the second quarter of 2026. Its range for 2027 was much wider, from US$360 billion to US$930 billion. The result depends on the flow of new announcements, cancellation rates, the time from announcement to construction and the time required to complete a project.

Federal Reserve Governor Lisa Cook separately noted in May 2026 that announced plans exceeded US$1.5 trillion, but only a small share had been realised. The gap matters more than the headline. A proposed site must obtain power and permits, begin construction, reach completion and attract paid utilisation before it produces operating cash.

Cancellation becomes more expensive as a project advances. A preliminary plan can be dropped; ordered electrical equipment, grid work and minimum-purchase commitments may remain even when demand arrives late. Announced investment should therefore be separated from work in progress, funded completion cost and non-cancellable commitments.

Separate recognised debt from contractual obligations

The Bank for International Settlements describes a common structure. A joint venture or special-purpose vehicle owns or develops the data centre, funded by sponsor equity and private debt. A hyperscaler customer may take a minority interest and sign a long operating lease or capacity offtake agreement, sometimes with a guarantee. Project debt can sit outside the customer’s consolidated balance sheet even though customer payments support repayment.

Not every lease remains off balance sheet. Accounting standards require many commenced leases to be recognised as a right-of-use asset and lease liability. Data-centre leases that have not commenced, long purchase contracts, minimum-capacity commitments and some guarantees may first appear in contractual-obligation or contingent-liability disclosures instead.

Reported net debt is therefore only the first line. At the same reporting date, investors need cash borrowings, lease liabilities, leases not yet commenced, non-cancellable purchase commitments, maximum guarantee exposure, related-entity equity and any commitment to inject more capital. Those items are not one accounting liability, but together they show where future cash calls can arise.

Contracts determine the loss waterfall

In a stand-alone project, rent and capacity payments normally service interest and principal. If the project underperforms, losses do not automatically jump to the hyperscaler. Project cash, subordinated equity and contractual reserves may be used first. Lenders then enforce security and guarantees. The burden returns to the customer where it has promised extra payments, capacity purchases or completion support.

The function of each clause matters more than its label:

• Completion support decides who funds cost overruns and delays.
• Minimum-purchase commitments decide what the customer pays when utilisation is weak.
• Termination rights set the cost of reducing capacity when demand changes.
• Residual-value guarantees allocate losses if buildings, power equipment or servers become obsolete faster than expected.
• Cross-default and ratings clauses can connect one project problem to other funding.

A long contract with one strong customer differs from a facility carrying several shorter reservations. The first can stabilise initial cash flow but creates concentration risk. The second can diversify customer failure but makes utilisation more sensitive to the cycle and competing capacity.

Construction and power delays bring refinancing risk forward

Construction finance often assumes that short or floating-rate funding will be replaced once a completed asset produces stable cash flow. A delayed grid connection, substation, cooling system or chip delivery postpones rent while interest and construction costs continue to accrue. If the loan matures first, the planned long-term refinancing can arrive on worse terms.

That is why total project cost is less informative than the gap between remaining completion cost and committed funding. Investors also need to know whether sponsors can be called for more equity, whether lenders must extend facilities, whether interest is fixed or floating and how long the interest reserve lasts. ‘Still under construction’ does not establish that completion is fully funded.

Operational risk remains after opening. Servers can have a shorter economic life than the building and power connection. A change in chips, cooling or customer specifications may require another round of spending. High current utilisation is not enough if renewal prices fall or power costs rise before the debt is refinanced.

Private credit remains connected to banks

The BIS notes that private-credit funds and insurers can hold data-centre project debt while banks extend facilities to the project, the fund or both. Losses may first sit with non-bank lenders, but bank loans, derivatives, bridge finance and undrawn credit lines keep the system connected.

Stress can travel without an immediate fall in traded bond prices. Weaker fundraising can prevent a private fund from supporting a refinancing. Several projects drawing bank facilities at once can concentrate liquidity demand. An insurer with a large allocation to long project loans can face both credit losses and pressure in managing its own liabilities.

This exposure does not by itself establish a systemic crisis. Governor Cook observed that even ambitious investment and debt projections remained below the peak leverage seen before the global financial crisis. The useful distinction is between a real credit channel and a forecast of systemic failure. Lender concentration, collateral values, maturities, redemption terms and banks’ undrawn commitments decide how far stress can travel.

Cash-rich hyperscalers still have contract risk

Large cloud companies generate substantial operating cash and can access investment-grade bond markets. BIS data show gross bond issuance by large US hyperscalers exceeded US$100 billion in 2025, mostly at maturities beyond five years. Long fixed-rate funding can match the build period better than short debt, but greater issuance and uncertainty about project returns have also appeared in some companies’ credit-default-swap spreads.

Cash is a starting point for repayment capacity, not proof of an adequate investment return. Cash remaining after investment and contractual payments can shrink if cash capital expenditure, lease payments, power costs and interest rise faster than customer receipts. Count each cash payment once. Depreciation is different: it reduces accounting profit but is not a fresh cash payment, and is added back when reconciling profit to operating cash flow, as the SEC’s financial-statement guide explains . Spending to replace equipment is a separate cash outflow. The same financing can remain orderly when utilisation and monetisation produce enough cash.

The distinction also matters when reading cash flow across the AI value chain . A chip supplier’s revenue is its customer’s capital cost. Supplier prosperity and customer cash-flow pressure can coexist until the customer creates paid demand worth more than the full operating and financing cost.

Read five disclosure areas together

Start with cash paid for property, plant and equipment relative to operating cash flow. Then connect balance-sheet borrowings and lease liabilities with notes on leases not yet commenced and purchase commitments, related-party or variable-interest-entity disclosures, maximum loss exposure and management’s construction timetable and utilisation assumptions.

Lay the maturities out by year. A project loan, lease commencement, minimum-purchase payment and major server refresh falling in the same period can create a cash concentration even for a company with current net cash. Match currencies and rate terms as well. An Australian investor faces both the issuer’s US-dollar funding risk and the effect of AUD moves on portfolio returns.

A three-year cash-call example

Consider a fictional project company, not a real issuer. All amounts are US$ millions. It starts Year 1 with 10 of unrestricted cash and a 50 loan. A customer pays 18 each year for capacity; cash operating costs are 6 and interest is 4. The loan principal is due at the end of Year 3. There are no taxes, additional capital costs, distributions or other cash flows in this simplified example, and all surplus cash is retained.

The project generates 8 of cash after operating costs and interest each year, yet it cannot repay the 50 balloon at maturity: 10 + 3 × 8 = 34, leaving 16 to refinance or fund. Positive annual cash generation does not remove the maturity gap.

Now assume the customer separately guarantees up to 10 of that unpaid principal, with no other support obligation. If the guarantee triggers and the customer pays the full 10 directly to the lender, 6 still remains to be refinanced, funded by another party or recovered through enforcement. The customer’s Year 3 outflow becomes 28: its regular 18 payment plus the contingent 10. That regular payment is the project’s inflow, not an additional system-wide obligation to count twice. The project’s 50 loan and the customer’s capped guarantee are not two separate 50 debts.

The figures illustrate cash timing, not an expected loss or a legal conclusion about an actual contract. Real repayment depends on the guarantee’s trigger and enforceability, cash restrictions, other creditors and collateral recoveries. Change any of those assumptions before applying the example to a disclosed project.

Avoid dividing one year’s stated AI revenue by the same year’s capital expenditure and calling the result a return. Data centres support several services, while investment and revenue recognition occur at different times. Contracted capacity, utilisation, customer concentration, unit pricing, power cost, depreciation, interest cover and the direction of free cash flow provide a more useful sequence.

The Australian portfolio channel

Australian exposure is broader than owning a US hyperscaler. Banks can lend to developers or non-bank funds; superannuation and insurance portfolios can hold private debt; listed property and infrastructure vehicles can own powered land or facilities; energy and construction businesses can benefit from the build-out. Each position occupies a different place in the loss waterfall.

For an infrastructure or property vehicle, the central questions are tenant concentration, lease length, power availability, development funding and residual value. For a bank or private-credit exposure, they are seniority, covenants, collateral, maturity and sponsor support. For a US equity holding, the issue is whether incremental paid use covers capital expenditure, leases and operating costs. AUD translation can change the reported return without changing the project’s ability to service US-dollar debt.

This framing prevents a common category error. Strong demand for power, land or construction can support an Australian supplier while a heavily financed project delivers a poor return to its equity or junior creditors. The theme can be economically real without every security linked to it sharing the same risk.

What would change the assessment

The observable conditions that would change the working view begin with conversion and funding. The risk case strengthens when announced capacity keeps rising but construction and completion conversion weaken, power delays increase total cost, long contracts are reduced or renegotiated, and project spreads or refinancing costs rise. If leases not yet commenced, purchase commitments and guarantees grow while hyperscaler free cash flow and interest cover deteriorate, the evidence that external obligations are returning to sponsor cash flow becomes stronger.

The counter-signals are equally observable: funded projects connecting on time; non-cancellable contracts with strong customers; utilisation and paid use generating cash comfortably above financing and maintenance costs; long fixed-rate funding; staggered maturities; limited guarantees; and adequate interest reserves.

Conclusion

Off-balance-sheet AI data-centre debt is not one number. It includes project-company borrowing, hyperscaler bonds and lease liabilities, leases not yet commenced, minimum-purchase commitments, guarantees and private-credit funding recorded in different places.

Follow the facility through its contracts, repayment source, loss waterfall and ultimate guarantor. If completion and utilisation produce cash above financing cost, the structure can accelerate useful investment. If capacity arrives before demand and guarantees activate at refinancing, risk that appeared outside one balance sheet returns inside somebody’s cash flow.

Sources

• Bank for International Settlements, Financing the AI infrastructure boom: on- and off-balance sheet borrowing
• Federal Reserve Board, Estimating Aggregate Data Center Investment with Project-level Data
• Federal Reserve Governor Lisa D. Cook, The Opportunities and Risks AI Presents for the Economy and Financial System
• US Securities and Exchange Commission, disclosure of off-balance-sheet arrangements and contractual obligations
• IFRS Foundation, IFRS 16 Leases