Doubts spread over recovering AI infrastructure investments
Caution spreads from stock and bond markets to banks
Higher financing hurdles put pressure on vulnerable companies

Anxiety over recovering investments is growing in the financial markets that funded the race to expand artificial intelligence (AI) data centers. Despite large orders, mounting capital expenditure and interest burdens have led investors to question companies’ ability to repay their debts. With initial public offerings (IPOs) delayed and related bond prices falling, banks that have been unable to sell construction loans to other investors also face pressure over providing additional funding. Across an industry that still needs massive investment, concerns are spreading that some companies may be unable to withstand rising financing costs.

CoreWeave Debt Insurance Costs Hit Record High

According to Bloomberg on the 21st (all dates hereafter are local time), specialized cloud companies that acquire and rent out graphics processing units (GPUs) and data centers needed for AI computing have recently faced financing pressure. At U.S.-based CoreWeave, a leader in the so-called “neocloud” industry, the cost of insurance against default has soared to a record high. As of the 15th, the premium on CoreWeave’s five-year credit default swaps (CDS) reached 1,194.5 basis points intraday (1 basis point = 0.01 percentage point). That implies a market-priced probability of default over the next five years exceeding 60%. The bond market has begun to question the repayment capacity of companies that have expanded AI infrastructure through heavy borrowing.

Massive upfront investment and the increased borrowing used to support it have intensified the funding burden. Neocloud companies must buy expensive chips and secure data centers before providing computing services to customers, so they finance a substantial portion of their initial investment with debt. CoreWeave’s capital expenditure plan for this year, announced in August, was USD 35 billion–USD 39 billion (approximately USD 36.1 billion–USD 40.2 billion), roughly three times the annual revenue of USD 13 billion (approximately USD 13.4 billion) expected by the market. Already elevated financing costs are also weighing on earnings. The company’s second-quarter net interest expense was USD 640 million (approximately USD 660 million), accounting for a quarter of quarterly revenue of USD 2.575 billion (approximately USD 2.66 billion), while its net loss for the same period reached USD 626 million (approximately USD 645 million).

Turning large orders into actual revenue also requires additional funding. CoreWeave reported an order backlog of USD 104 billion (approximately USD 107 billion) as of the end of June, citing it as evidence that it can manage its debt. The figure represents revenue expected to arise from contracts signed with customers and is recognized in earnings only when equipment delivery and service provision requirements are met. With facilities needed to fulfill those contracts yet to be built, rising yields demanded by bond investors are adding to the burden of securing follow-on financing. If the release of new chips reduces the value of existing equipment, further investment may also be needed to replace it.

Investors Raise the Bar for AI Financing

As the funding required to expand facilities continues to grow, investors are becoming more selective about AI companies. SB Energy, a U.S. data center developer backed by SoftBank, recently encountered a setback in its IPO timetable because of disagreements among investors over its valuation. Its proposed valuation of USD 50 billion (approximately USD 51.6 billion), despite not yet having a single operational data center, raised investor concerns. Marketing the offering to investors was also postponed as the company responded to additional questions from the U.S. Securities and Exchange Commission (SEC). Alongside the high valuation, investors cited its dependence on major customer OpenAI as a concern.

Another factor heightening investor caution is the massive funding burden that will continue after the listing. The future revenue from data center contracts secured by SB Energy totals USD 430 billion (approximately USD 443 billion), most of it related to OpenAI. The funding needed just to build the facilities required to fulfill those contracts was estimated at USD 174 billion (approximately USD 179 billion). This means the company will have to keep raising money in bond markets and elsewhere even after boosting its capital through the listing. Investors have already responded lukewarmly to the USD 5 billion (approximately USD 5.16 billion) debt financing being considered alongside the IPO, fueling doubts about whether follow-on funding can be secured as planned.

AI Overinvestment Fears Take Hold in Financial Markets

As markets raise the bar for AI business profitability, concerns about overinvestment voiced since the start of the year are also being more fully reflected in financial markets. In January, the International Monetary Fund (IMF) warned that if AI’s productivity gains fall short of expectations, technology companies could scale back investment and stock market corrections could persist. Recently, concerns about recovering investments have also emerged in data center projects with major technology companies as customers. Bond prices linked to “Hyperion” and “Project Excalibur,” whose tenants include Meta, Nvidia and Oracle, have traded well below face value in recent weeks. This reflects growing investor anxiety over whether future rental income will cover debt accumulated during construction and operation.

As concerns over recovering investments spread, trading in CDS that provide protection against default has also surged. According to Morgan Stanley, weekly average CDS trading volume linked to bonds issued by Oracle, SpaceX, Alphabet, Amazon, Meta, Broadcom and Nvidia rose from less than USD 2 billion (approximately USD 2.06 billion) in July to more than USD 9 billion (approximately USD 9.28 billion) in September. Beyond bond investors, equity and multi-asset investors are also turning to CDS to hedge against losses on AI-related investments. Caution is also being heightened by the difficulty of assessing total debt exposure amid a complex web of intercompany investments and loan guarantees. On the 19th, Mark Habib, head of digital infrastructure at international credit rating agency S&P Global, pointed to a structure in which major technology companies effectively guarantee data center construction and financing for smaller AI companies. His warning was that, with individual transactions insufficiently disclosed, more debt exposure may be hidden than is currently known.

Table 1. Delayed AI Data Center Loan Sales and the Burden on Banks

Banks Even Consider Selling AI Loans at a Discount

The burden of the complex web of AI financing is also building at banks that funded data center construction. Banks recover their funds by first providing loans and then selling portions to other investors, or through operators issuing asset-backed securities (ABS) and other instruments to repay existing loans. They channel the recovered money into new projects, but delays in finding investors to take over the loans have also slowed the circulation of funds. In May, JPMorgan, Mitsubishi UFJ Financial Group (MUFG) and others were reported to have spent more than six months trying to sell investors portions of USD 38 billion (approximately USD 39.2 billion) in construction loans for Oracle-leased data centers. Some banks also pursued plans to sell loans at a discount to nonbank financial institutions.

Banks have sought to sell loans even at a loss because funds tied up in existing loans also constrain financing for subsequent projects. When loans underwritten with the intention of selling them remain on the books, banks must continue maintaining capital against potential losses, while headroom under company-specific and sector-specific lending limits shrinks. Individual banks’ lending limits reportedly also became an obstacle when selling portions of data center loans related to Oracle and OpenAI to other financial institutions. The purchasing institutions likewise faced constraints on increasing lending to the same companies, delaying the sales. Meanwhile, funds remained tied up at the banks that initially underwrote the loans, reducing their capacity to provide additional loans for new data center construction.

Growing Reliance on Private Credit

Data center operators unable to obtain the money they need from banks face the prospect of raising more expensive funding in the private credit market. Private credit managers demand high interest rates and fees in exchange for assuming construction risk or providing funding in line with project schedules. Refinancing options when loans mature are also limited. This means operators may have to offer investors higher returns to secure the funds needed to repay existing loans and invest in facilities. Deals offering double-digit returns to finance AI computing equipment have already emerged. According to the Financial Times (FT), Anthropic recently offered Blackstone participation in USD 9 billion of subordinated loans as part of a chip-financing transaction it is pursuing, with a coupon rate of 10%. After accounting for the issue discount and fees, the return to investors exceeds 10%. The interest rate on subordinated funding Anthropic raised through a similar transaction in June was 8.5%.

Higher financing costs are also increasing the burden of massive capital investments planned for the future. According to a forecast released by market research firm Gartner on the 14th, AI companies’ infrastructure investment will reach USD 1 trillion (approximately USD 1.03e+3 billion) this year, with an additional USD 6 trillion (approximately USD 6.19e+3 billion) expected to be invested in 2027–2030. Gartner concluded that corporate and consumer spending on AI services during this period would be insufficient to generate adequate returns across all planned capital investment. Neocloud companies are particularly vulnerable to cost pressures because they must recoup their investments through computing services to sustain their businesses, according to the analysis. Gartner forecast that more than 80% of neocloud companies would cease operating independently by 2030 because of cost pressures and a failure to differentiate their businesses.