CICC: The Financial Cracks of AI "Subprime"

CICC released a research report stating that since October 2026, the yield to maturity on multiple senior unsecured bonds of CoreWeave, a representative Neocloud company, has risen to around 15%, and the five-year credit default swap (CDS) spread has breached 1,100bp. CICC believes this round of AI financing risk is a localized credit contraction for "subprime"-style weak-qualification issuers under a high interest rate environment compounded by AI slowdown concerns. The cost increase for secured DDTLs may be smaller than the rise in unsecured yields, but deterioration in parent company credit could further transmit to new project loans, leading to further tightening of financing conditions, and liquidity pressure and debt repayment risk warrant vigilance. CICC's main views are as follows: The "subprime" cracks: alarms from financial markets. CoreWeave's credit bond yields have risen from around 10% in the first half to 15%, and the five-year CDS spread has breached 1,100bp, implying a five-year cumulative risk-neutral default probability of about 60%. Neoclouds occupy a fragile position in the credit spectrum characterized by high leverage and weak cash flow, akin to "subprime"-style weak-qualification issuers in the AI financing system: the four listed Neocloud companies had 2026 net debt/EBITDA of about 7.5x, significantly higher than the roughly 0.1x level of a sample of six large technology companies. In terms of debt-to-asset ratio, the Neocloud sample was about 84% in 2025, higher than the roughly 45% of the large technology company sample. In terms of cash flow, due to the timing mismatch between capital investment and cash flow recovery, Neocloud free cash flow has long remained in negative territory. How large is the financing pressure? As of the second quarter of 2026, CoreWeave already had about US$35 billion in interest-bearing debt, of which 29% is the unsecured credit bonds that suffered this decline and 39% is secured loans (DDTL) borne by SPVs, but 90% of total debt carries parent company recourse. Rising credit bond yields do not directly change the contractual cost of all existing debt, but may gradually affect subsequent project construction and refinancing of existing debt: based on 2.7GW of signed capacity pending construction, US$300 million/GW unit investment cost, and a 20% order prepayment ratio, static estimates indicate it still needs about US$27 billion in additional financing, with another roughly US$15 billion of debt maturing in concentrated fashion in 2031-2032. Where is the debt servicing pressure threshold? CICC established a five-year GPU project financing model. Under base-case assumptions, unlevered IRR of about 15% exceeds the 10% financing cost, and the debt service coverage ratio is about 1.44x, but the first-year debt service coverage ratio is close to the warning zone of 1.2x. Due to Neoclouds' characteristics of high leverage, high interest rates, and timing mismatch between capital expenditure and cash flow, both rising loan costs and project cash flow volatility will exert notable pressure on debt servicing. Under four scenarios, the minimum debt service coverage ratio will all approach the stress threshold of 1.0x: loan costs rising from 10% to 15%, unit annual revenue falling from US$12 million/MW to below US$10 million/MW, EBITDA margin falling below 60%, or more than 50% of cash payments concentrated in the final year. With debt servicing capacity near the stress threshold, borrowers will face pressure for additional financing. How do the financial "cracks" transmit? Currently, the secondary market yield to maturity on CoreWeave's unsecured bonds has risen to around 15%, but existing DDTLs mainly accrue interest at contractually agreed spreads, and financing costs do not reprice in sync with unsecured bond yields. Compared with parent company credit bonds, DDTLs also have GPU asset collateral and long-term contract protection, so financing costs are relatively lower. However, about 90% of CoreWeave's interest-bearing debt carries recourse, some of which is still guaranteed by the parent company. As parent company credit deteriorates, financing conditions for new DDTLs may tighten. Is it a systemic risk? 1) CICC believes this round of AI financing risk is a localized credit contraction for "subprime"-style weak-qualification issuers under a high interest rate environment compounded by AI slowdown concerns. The cost increase for secured DDTLs may be smaller than the rise in unsecured yields, but deterioration in parent company credit could further transmit to new project loans, leading to further tightening of financing conditions, and liquidity pressure and debt repayment risk warrant vigilance. 2) The sample Neoclouds' 2026 capital expenditure is expected to be only one-twelfth that of the six large technology companies, and the credit quality of large technology companies remains relatively sound. AI debt credit quality is tiered, with project loans protected by long-term customer contracts, asset collateral, and independent cash flow; bank loan exposure to capital shocks is also limited, and static estimates show the direct impact of single-entity exposure risk on the CET-1 adequacy ratio of relevant large banks is about 1-5bp. 3) Although systemic risk is temporarily limited, if AI sentiment continues to be suppressed under a high interest rate environment, it could trigger concerns about the economic cycle, and subsequent focus should be on whether credit risk spreads along Neocloud debt to broader AI debt categories such as private credit. Risks External financing conditions tightening beyond expectations; uncertainty in estimates of new debt financing demand; AI capital expenditure and project delivery falling short of expectations; GPU collateral value declining and financial institution credit losses exceeding expectations.