Wall Street's attitude towards AI financing is undergoing subtle changes: banks are still willing to participate in the transactions and earn fees, but they are becoming more cautious about taking on risks. JPMorgan Chase and Goldman Sachs jointly led a $5 billion AI infrastructure loan, yet they refused to provide underwriting guarantees for the $3.7 billion GPU chip loan portion of it.

According to Bloomberg, this financing for the emerging cloud computing company Volta Infrastructure Holdings is divided into two parts: a $1.3 billion term loan guaranteed by banks, and another $3.7 billion that is offered on a best-efforts basis, with banks not committing to underwrite any unsold portion.

The financing cost of this transaction is also noteworthy. The yield to maturity for both loans ranges from 11.21% to 11.75%, which is at a high level in the current AI financing market. In contrast, the yield to maturity for a GPU loan in CoreWeave in July of this year was approximately 10.44%.

Banks have not stopped providing financing for the AI infrastructure, but they are no longer willing to take on the risks of chip loans unconditionally. When high-cost financing meets a more cautious underwriting attitude, the question facing the market is: if banks are unwilling to bear the losses, who will take over these loans? Whether funds from private credit and other sources can fill the gap will also test the real demand for financing of the AI infrastructure and their ability to withstand risks.

Transaction structure: The bank only provides guarantee for a loan of $1.3 billion.

According to PitchBook LCD data, Volta's total financing amount this time is 5.012 billion US dollars, with a term of six years. Of this, 1.3 billion US dollars in preferred lien term loan C provides letter of credit guarantee for the data center lease through cash collateral; the 3.712 billion US dollars in GPU term loan is used to purchase approximately 36,000 NVIDIA GPU.

The interest rate spread for both loans is SOFR plus 625 to 650 basis points, with issue prices ranging from 97% to 98% of the face value, and neither loan has received a credit rating. The project will involve the construction of a data center complex in Norway, which will be operated by Bitcoin mining machine operator Bitdeer. According to previous reports by Bloomberg, Anthropic has signed a six-year computing power agreement with this site.

Founded this year, its investors include a16z, Altimeter, Michael Dell, and NVIDIA. NVIDIA is not only a shareholder but also a supplier to GPU.

The choice made by the two banks essentially separates the financing arrangements from the ultimate assumption of risk. For the $1.3 billion lease loan, the banks provide underwriting guarantees; for the $3.712 billion GPU loan, the banks are only responsible for making their best efforts to sell it and do not commit to taking over any unsold portion. This means that whether the chip loan can successfully complete its financing will depend more on the willingness of external investors to subscribe.

11% financing cost: Investors demand higher risk compensation

Volta The yield to maturity of this loan reaches up to 11.75%, which is higher than that of some previous similar GPU financings.

In March this year, CoreWeave completed a $8.5 billion DDTL 4.0 financing round, becoming the first GPU mortgage financing to receive a Moody's A3 investment-grade rating. The spread on the floating-rate portion was SOFR plus 225 basis points. By July, the spread on another $2.6 billion loan of the company increased from SOFR plus 425 to 450 basis points during the roadshow period to 550 basis points, with a yield to maturity of approximately 10.44%. The yield to maturity of this latest loan by Volta further rose, reflecting investors' more cautious pricing of financing risks.

The broader credit market is also under pressure. Credit strategists at Goldman Sachs, including Amanda Lynam, predicted in their analysis on October 9th that the default rate on leveraged loans in the United States would remain high at around 5% in 2027, with a loan recovery rate of about 56 cents per dollar, which is lower than the 64 cents in 2022.

For Volta, higher financing costs not only mean a heavier interest burden but also reflect investors' concerns regarding the borrower's credit and the value of the GPU collateral. Although GPU can be used as loan collateral, its value still depends on market demand, technological iteration, and the liquidity of the secondary market. In addition, since the loan does not have a credit rating, the range of potential investors may also be limited.

Risk Transfer: Can Private Credit Take Over?

The tightening of underwriting commitments by banks does not mean that the financing needs of AI have disappeared. Private lending institutions and insurance companies are still participating in related projects, but whether they can take on the risks that banks are unwilling to assume at a cost that borrowers can afford remains to be seen.

According to IFR, a $42 billion preferred secured loan backed by Broadcom and targeted at Anthropic had its syndication sales launched last Monday. Institutions such as Apollo, Blackstone, and Japanese life insurance companies are also providing funding for AI's infrastructure projects.

However, the shift of risk from banks to private lending does not mean that the risk has disappeared. The Bank of England warns that banks may still indirectly assume related risks by providing loans to private lending funds. JPMorgan Chase's Chief Financial Officer Jeremey Barnum also stated that the bank's exposure to AI has spread to various areas such as construction loans, the stock market, and working capital revolving credit.

There is no consensus in the market regarding whether high-cost financing is sustainable. Goldman Sachs cites Broadcom's view that the demand for computing power still exceeds supply, and cutting-edge laboratories that generate substantial revenue have the capability to afford high-cost financing. However, if the demand for computing power or revenue growth falls short of expectations, high interest rates will increase the debt repayment pressure, and the collateral may not be sufficient to fully cover the loan losses.

Banks continue to participate, but are unwilling to take on all the risks.

JPMorgan Chase and Goldman Sachs are not alone in this practice. According to previous reports by Bloomberg, Morgan Stanley has used significant risk transfer (SRT) tools to transfer some of its data center risk exposures; meanwhile, (Blue Owl) failed to raise third-party funds for a $4 billion (CoreWeave) data center project.

These cases point to the same issue: the bottleneck in infrastructure financing for AI is shifting from the funding requirements themselves to who will bear the risks and how much return should be obtained. Banks can continue to organize transactions, and private credit can also provide funds, but only if investors are willing to accept the corresponding risks and pricing can financing truly be realized.

The deadline for this GPU loan is 5 p.m. on October 14th, Eastern Time. The final subscription situation will serve as a window to observe market risk preferences: if the loan requires an increase in the interest rate spread or a reduction in scale, it indicates that investors are less willing to undertake such high-risk financing; if it is completed smoothly, it means that there is still capital in the market willing to take on the related risks, but the cost may be a higher financing cost.

For AI infrastructure companies, strong demand for computing power does not necessarily mean that financing risks are controllable. When banks are no longer willing to underwrite chip loans, we are ultimately faced with one question: will the future cash flows generated by the projects be sufficient to support such high financing costs?