Financing the AI Boom: Where Is the Credit Risk Ending Up?

6

By Shashank Pattekar, International Banker

The Bank for International Settlements (BIS) recently pegged the artificial intelligence (AI)-related capital expenditure of the five largest technology companies across 2025 and 2026 at more than $1 trillion. This investment is being financed across a chain linking hyperscalers, project vehicles, private-credit funds, insurers and banks, meaning that credit risk is potentially borne by multiple parties. For creditors, therefore, it is becoming essential to gain the necessary transparency into who is borrowing, which cash flows service the debt and where losses ultimately land should demand undercut expectations.

Hyperscalers’ balance sheets represent the most visible part of this financing chain, with Alphabet, Amazon, Meta, Microsoft and Oracle all able to use their operating cash flows to fund a substantial share of their investments. PIMCO (Pacific Investment Management Company) reported in May that consensus estimates placed their combined capital expenditure at nearly $690 billion in 2026 and $870 billion in 2027. At those levels, spending would absorb 94 percent of operating cash flow in both years, compared with 40 percent in 2023.

Public-investment-grade bonds provide the most direct source of additional financing, with the issuing company owing the debt and bondholders relying on cash flows from the company’s entire business. The BIS reported that corporate bonds have been hyperscalers’ primary financing source, with gross issuance exceeding $100 billion in 2025. Strong credit ratings and diversified earnings reduce near-term default risk, although additional borrowing raises interest costs and reduces financial flexibility if AI investments fail to generate the anticipated earnings.

“From a credit quality standpoint, the direction of travel seems clear: AI infrastructure is becoming increasingly debt-financed,” PIMCO analysts Lotfi Karoui, Michael Puempel and Amit Arora stated in a May 22 report. “The good news is that the starting point remains strong. Four of the five hyperscalers report net leverage that is barely positive, and even after the recent surge in issuance, the broader technology sector remains the least leveraged sector in the US.”

Tracing the financing becomes more challenging when the infrastructure sits inside a separate entity.

Tracing the financing becomes more challenging when the infrastructure sits inside a separate entity, however. For instance, a hyperscaler can take a minority interest in a joint venture or special purpose vehicle (SPV) that owns or develops a data centre. The vehicle raises debt to finance the data centre’s development, while the technology company supports the project by agreeing to rent the facility or purchase a specified amount of computing capacity. It may also provide a guarantee that gives the project’s lenders additional protection if the asset performs poorly or loses value.

Such commitments can support repayment while most of the project debt remains at the vehicle level. “Alongside traditional bonds, hyperscalers have turned to off-balance sheet arrangements to finance infrastructure expansions, often in partnership with private credit firms,” the BIS explained in its “Quarterly Review”, published in March. “A common structure involves a dedicated vehicle—often a joint venture or special purpose entity—that acquires or develops data centre assets. The vehicle is capitalised with equity from a consortium of sponsors and raises debt through private placements.”

As financing passes through bonds, private-credit funds and project vehicles, the visible sponsor—typically the hyperscaler backing the project—may differ from the legal debtor, even when its leases, purchasing commitments or guarantees support repayment. Meta’s Hyperion data-centre campus in Louisiana provides a useful example of such a structure:

• 80 percent is owned by funds managed by private-credit firm Blue Owl Capital, with Meta retaining the remaining 20 percent.
• Both parties to the joint venture have committed to funding their respective shares of the approximate $27 billion in project-development costs.
• A portion of the capital raised by Blue Owl will be financed through debt placed with PIMCO and other selected bond investors, while Meta will lease all of the completed facilities under agreements carrying four-year initial terms and extension options.
• Meta has also provided a capped residual-value guarantee for the first 16 years of operation, which may require a payment if specified conditions are met following the termination or non-renewal of a lease.

This arrangement reduces Meta’s immediate cash requirement, while the separate project vehicle undertakes much of the borrowing. Nonetheless, Meta’s lease payments and guarantees remain central to creditor protection. As such, investors must examine the legal borrower, security over the assets, lease conditions, guarantee triggers and likely recovery value, as those details will determine how much credit risk has genuinely moved away from Meta versus how much has simply become less visible.

Unlike public bonds, private-credit facilities are negotiated directly, allowing lenders to tailor collateral, covenants and repayment terms to such factors as construction risk, power availability and tenant concentration. This flexibility helps explain private credit’s rapidly growing role in AI financing, with the BIS noting in January that outstanding private-credit loans increased to more than $200 billion by 2025, from near zero just 10 years earlier. The share of private-credit loans to AI-related companies also rose from less than 1 percent of total outstanding loan volumes to almost 8 percent.

Exposures extend beyond private-credit firms. While institutional investors supply capital to private-credit funds, insurers and other investors often directly hold a project’s debt. Banks are also involved via their funding lines to project vehicles.

With privately financed AI infrastructure thus creating potentially sizeable exposures for both banks and nonbank financial institutions (NBFIs), losses could be transmitted deep within the financial system. Lower-than-expected demand could weaken the lease income used to service debt, while rapid improvements in chips could reduce the value of equipment pledged as collateral. Delays in connecting a data centre to sufficient electricity supplies may also postpone revenue. And a retreat by private-credit investors could create refinancing pressure when existing facilities mature.

Such risks are unlikely to impact each borrower equally. While cash-rich hyperscalers possess diversified earnings and relatively low leverage, leveraged data-centre developers, specialised cloud providers and equipment suppliers are more vulnerable; their revenues depend on a small number of customers, and their expansion requires repeated access to external finance. A cancelled capacity agreement, tenant non-renewal or construction delay could impair their ability to service their debts well before the hyperscaler experiences financial distress.

Leases and guarantees provide meaningful protection to creditors, but they also expose the sponsoring technology company to losses. If a project vehicle’s cash flows prove insufficient to service its debt, the sponsor may have to extend financial support when its AI investments are already producing weaker-than-expected returns.

Federal Reserve Bank of Chicago (Chicago Fed) has estimated direct bank exposure to AI-adjacent industries at about 0.8 percent of assets, with delinquency rates close to those of wider portfolios.

For now, the risk is expected to remain contained. On the banking side, for example, the Federal Reserve Bank of Chicago (Chicago Fed) has estimated direct bank exposure to AI-adjacent industries at about 0.8 percent of assets, with delinquency rates close to those of wider portfolios.

That said, the Chicago Fed has also observed notable tail risks emerging, whereby stress in one AI-adjacent industry could spill into multiple interconnected AI-adjacent industries. “For example, if software companies are stressed and unable to maintain their infrastructure spending levels, semiconductor manufacturers, energy companies, and data centers may be affected, impacting their ability to repay their debts to commercial banks,” the bank warned in a February report.

“Banks most likely have additional exposure to AI-adjacent industries through lending to nonbank financial institutions (NBFIs),” the report also acknowledged. “For example, a bank may lend to a private credit institution providing funding for a data center or lend to an investment fund that specializes in AI investments, and stress in the underlying companies may lead to stress in the NBFI borrowers.”

Assuming, therefore, that credit risk ends with the legal borrower or project sponsor provides only partial insight into the overall picture. Determining who is ultimately financing the AI boom requires tracing the exposure through project vehicles to private-credit funds, institutional investors, insurers and the banks supporting them.