The AI infrastructure boom is creating a hidden layer of debt, and credit investors may be among the first to feel the consequences if the cycle turns.
The Bank for International Settlements recently published an analysis of how the major technology companies are financing their enormous expansion in data-centre capacity. The headline numbers are already familiar: corporate bond issuance has surged, credit spreads have started to widen for some lower-rated companies, and the largest technology companies are committing hundreds of billions of dollars to infrastructure.
I wrote about the more visible, on-balance-sheet debt in my previous article. What caught my attention in the BIS analysis, however, was the other side of the equation: the financing that does not necessarily appear as conventional corporate debt at the hyperscaler level.
The basic structure is relatively simple. A technology company can create or participate in a joint venture that builds and owns a data center. The joint venture then raises debt from private investors, while the technology company takes an equity stake and signs a long-term lease or capacity agreement.
From an economic perspective, this can look remarkably similar to the technology company borrowing money itself to build the data center.
From an accounting perspective, however, it can look very different. The debt sits at the project or joint-venture level, while the technology company records payments under its contractual arrangements. In other words, some of the financial risk has moved outside the corporate balance sheet.
There are three reasons why I think this deserves more attention.
The first is the sheer scale of the investment cycle. Microsoft, Amazon, Alphabet, Meta and Oracle are collectively spending hundreds of billions of dollars on data centers, chips, networking equipment and other AI infrastructure. Their enormous operating cash flows cover a significant part of this spending, but they cannot cover everything indefinitely as capital expenditure continues to rise. The remaining funding has to come from somewhere, and increasingly it is coming from a combination of corporate debt, private credit and project-level financing.
The second issue is who is providing that capital. Private credit funds, insurance companies and other institutional investors have become increasingly important sources of financing for infrastructure projects. That is not necessarily a problem, but the risks involved in financing a data center are different from those involved in lending to a conventional company. A traditional corporate lender can spend most of its time analyzing revenues, margins, cash flow and leverage. A data-center lender has to make assumptions about construction schedules, electricity availability, equipment deliveries, future demand and the utilization rate of the finished facility. Those assumptions can be difficult to get right, particularly in an industry that is expanding so quickly that there is relatively little historical data on which to base them.
The third issue is the way these structures can make risk appear safer than it really is. A data-center SPV backed by a 15-year agreement with Microsoft or Amazon looks extremely attractive. You have a physical asset, a long-term contract and a counter-party with a very strong balance sheet. But the most important risk may not be that Microsoft or Amazon defaults. The more realistic question is what happens to the project if construction takes significantly longer than expected, the electricity connection is delayed, costs increase sharply or the technology company eventually discovers that it needs less capacity than it originally expected.
How the structure works
The structure varies from transaction to transaction, but the basic mechanics are relatively straightforward.
A group of investors establishes a joint venture to develop a data center, with the vehicle acquiring the land, securing an electricity connection and arranging the financing required to construct the facility.
The joint venture then raises debt, often through private placements rather than through the public bond market. A small number of institutional investors may provide the financing, with the creditworthiness of the project depending on both the completed data center and the contractual commitment from the hyperscaler.
The technology company may own part of the project, but its more important role is often its long-term agreement to use the facility.
That agreement is what creates the expected cash flow needed to repay the debt. For the technology company, this arrangement can be attractive because it provides access to additional capacity without requiring the company to finance the entire project directly. It can also allow some of the borrowing to remain at the project level rather than appearing as conventional corporate debt.
For the investor, the proposition is equally attractive. Instead of lending to a highly leveraged start-up, the investor is financing a physical asset that is expected to generate cash flow under a long-term agreement with one of the world’s largest technology companies.
The problem is that this apparent simplicity can hide a concentration of risks that are difficult to observe from outside the transaction.
Three risks I am watching
1. Refinancing risk
The first risk is refinancing. Data centers can take several years to develop and construct, while the debt used to finance them may have a much shorter maturity. That creates a potential mismatch between the life of the asset and the maturity of the financing. Imagine a project that takes four years to build but has six-year debt. The facility may have only two years of operating history before the debt needs to be refinanced.
If credit markets remain favorable, that may not matter. But if the refinancing takes place during a period when private credit investors have become more cautious, the economics can change very quickly. The new lenders may demand a substantially higher interest rate, additional collateral or tighter financial conditions.
A project that looked highly attractive when the original financing was arranged can therefore become much less attractive when it has to return to the market. This is one reason I am particularly interested in the maturity profile of the debt being raised today.
2. Guarantees can become real liabilities
The second risk is the use of guarantees and other contractual commitments from the hyperscaler. Some structures may include completion guarantees, minimum-payment commitments, capacity agreements or other forms of support designed to give investors additional comfort.
These commitments may not appear as conventional debt on the hyperscaler’s balance sheet, but they still represent economic obligations. That distinction becomes particularly important when something goes wrong.
If a project runs significantly over budget and cannot complete construction without additional funding, a guarantee that looked like a remote contingency when the deal was signed can suddenly become a very real liability for the technology company. At that point, the distinction between on-balance-sheet and off-balance-sheet financing becomes much less meaningful. The obligation still has to be met by someone.
3. Private credit valuations can lag reality
The third risk is more subtle and relates to the way private credit is valued. Unlike public bonds, private loans do not trade continuously in a transparent market. An investor therefore does not receive a new market price every day telling them that the value of their investment has fallen.
That can make private credit appear more stable than it actually is. If the economics of a data-centre project begin to deteriorate, the value of the loan may not immediately reflect that change. The adjustment could come later, when the project needs refinancing, when the assumptions underlying the original valuation are reassessed, or when an investor is forced to sell.
A loan that was valued at par when it was originated could therefore be worth materially less two years later if construction costs have risen, electricity supply has been delayed or expected utilization has fallen. The risk has not necessarily appeared suddenly. The market may simply have taken longer to recognize it.
Scenarios
50%
Projects are completed broadly on schedule, demand continues to grow and contractual cash flows are sufficient to service the debt.
Delayed completion
30%
Construction and power delays push projects beyond their original timelines, forcing some vehicles to seek additional financing or refinance on less favorable terms.
Structural repricing
20%
AI infrastructure spending slows materially, demand for new capacity falls and some projects struggle to refinance their debt at anything close to the original terms.
My base case remains an orderly digestion of the current investment cycle. The major hyperscalers have enormous financial resources, strong access to capital and considerable experience managing large infrastructure projects. It would therefore be wrong to assume that every project financed through an SPV is going to run into trouble.
My greater concern is the middle scenario, where nothing actually breaks but the assumptions made when the projects were financed prove to be too optimistic. Construction delays and electricity constraints are already real issues in parts of the data-center market. A project does not need to fail completely to create problems for its lenders. It simply needs to take longer and cost more than originally expected.
The third scenario is the tail risk. For that to happen, we would need more than a few delayed projects. The growth in AI infrastructure spending itself would have to slow significantly, leaving the industry with more data-center capacity than it needs or can profitably use. I think that is less likely over the next two years, but the probability would increase if capital expenditure continues to grow much faster than the revenues generated by AI services.
What I will be watching
The most useful warning signs may not come from the headline numbers published by the hyperscalers. They may instead appear in the details of individual projects and financing transactions.
At the project level, I will be watching whether data centers are being completed on schedule, whether electricity connections arrive when expected, whether construction costs are increasing and how much new private credit is flowing into data-center financing.
I will also be watching the pricing of new transactions. If investors suddenly demand significantly higher yields to finance similar projects, that would suggest that the market is becoming less comfortable with the underlying risks.
At the hyperscaler level, changes in long-term capacity commitments could be particularly revealing. If companies start reducing, delaying or restructuring their commitments to new data centers, that would be an important signal that the assumptions behind some of these projects are changing.
I would also pay close attention to the footnotes of corporate financial statements. Guarantees and other contractual commitments are often buried there rather than highlighted in the headline financial figures.
Finally, the private credit market itself deserves attention. If private credit funds begin experiencing significant withdrawals, or if institutional investors become less willing to allocate capital to infrastructure debt, refinancing conditions could deteriorate even if the underlying data centres are still operating normally. That is why the financing market may provide an earlier warning than the operating companies themselves.
The credit implication
The on-balance-sheet debt problem I wrote about two weeks ago is relatively easy to see. You can track corporate bond issuance, look at credit spreads and examine the amount of long-term debt being added to hyperscaler balance sheets.
The off-balance-sheet problem is much harder to measure. That does not automatically make it more dangerous, and I would not argue that using SPVs is inherently problematic. In many cases, project financing is simply a sensible way of matching a long-lived physical asset with the investors who are best equipped to finance it.
The problem arises when investors treat the structure itself as a source of safety. Moving debt from a hyperscaler’s balance sheet to a project company does not eliminate construction risk. It does not eliminate electricity risk. It does not eliminate demand risk, and it does not eliminate refinancing risk.
This distinction is particularly important because private credit has grown so rapidly in recent years. The market has attracted enormous amounts of capital, and investors have been willing to accept increasingly complex structures in exchange for relatively attractive yields.
The first warning sign may not be a tech default
A plausible first warning sign would a private credit fund might mark down a data-center investment; or a project might need additional financing because construction costs have increased. A refinancing might take place at a much higher interest rate than the original loan. Or a hyperscaler might have to honor a guarantee that everyone expected to remain a remote contingency.
None of these events would necessarily constitute a crisis. But they would tell us that the economics of the AI infrastructure boom are beginning to change. And because a significant amount of this financing sits outside the traditional banking system and public bond markets, the first losses could appear among private credit funds, insurance companies and other institutional investors rather than among the banks or the technology companies themselves.
Bottom line
The AI infrastructure boom is creating a new layer of financing that is easy to overlook. The major technology companies are increasingly using joint ventures, project companies and other structures to finance data centres without putting all of the borrowing directly onto their own balance sheets. For investors, these structures can offer an attractive combination of yield, physical assets and long-term contracts with some of the world’s strongest companies. But the apparent safety ultimately depends on a series of assumptions. Data centres have to be completed on time. Power has to be available. Construction costs have to remain under control. Demand for computing capacity has to continue growing, and the projects have to be able to refinance their debt when it matures. If those assumptions hold, the structure works very well. If they don’t, the losses may appear first in private credit funds, insurance companies and other non-bank investors.
That is why I think the off-balance-sheet financing of AI infrastructure deserves considerably more attention than it is currently receiving. The most important question is not how much debt the hyperscalers have on their balance sheets. It is how much economic leverage exists around them.
Alessandro Rovelli
Rovelli Research
Independent credit investment research
