CITIC Securities: The market's concerns about Coreweave are excessive, and the short-term funding gap for new cloud vendors "can still be supported."

The rise in U.S. Treasury rates, combined with CoreWeave CDS briefly exceeding 1000 basis points, has led to growing market concerns about the debt sustainability of Neocloud.

However, a report from CITIC Securities on October 19 titled "How to View Neocloud's Debt Issues" argues that this panic misjudges the timeline: the short-term financing gap for new cloud vendors can still be covered by existing channels, and the real stress test will not arrive until 2028—when the ability of AI demand to continue translating into rental income will determine the quality of this round of high-leverage computing power expansion.

The calculations provided in the report show that capital expenditures related to new cloud vendors will approach $500 billion by 2027, corresponding to an annual new financing demand of about $310 billion. This scale far exceeds the pressure that can be captured by merely looking at the balance sheets of publicly listed companies, but CITIC Securities believes that public market financing, GPU collateral financing supported by customer contracts, and customer prepayments are sufficient to cover this gap in the short term.

Signs of easing have already appeared on the financing cost side. CoreWeave's latest loan backed by investment-grade customer contracts (DDTL 4.0) has seen its spread drop to SOFR + 2.25%, serving as evidence that short-term financing channels still have the capacity to absorb this demand.

$830 billion in off-balance-sheet leases determines the true investment intensity of new clouds

Cloud service providers (CSPs) transfer computing power demand to new cloud vendors (such as Neocloud) through long-term leases, requiring new clouds to complete data center construction and GPU procurement before the lease starts. Therefore, the scale of off-balance-sheet leases directly determines the investment intensity of new clouds. As of the end of the second quarter of 2026, the four major CSPs (Amazon, Google, Microsoft, Meta) have uncommitted lease commitments reaching $830.5 billion, a year-on-year increase of 249%.

Based on this lease scale, CITIC Securities estimates that total capital expenditures for new clouds will approach $500 billion by 2027. This figure is significantly higher than the consensus expectation of about $300 billion in capital expenditures for the top five new clouds, with the difference mainly coming from expansion targets like xAI and investments from unlisted new cloud vendors.

Financing gaps under term mismatches are dispersed across the industry chain

The business model of new clouds inherently involves a term mismatch: upfront investments rely on debt financing, with rental income only gradually recovered after the lease starts. Based on approximately $830 billion in leases, and calculating with a 2.5-year cash recovery cycle, a 4-year delivery cycle, and a 60% leverage ratio, CITIC Securities estimates an annual new financing demand of about $310 billion.

The key point is that this portion of debt does not solely fall on the balance sheets of publicly listed companies, but is dispersed across various segments including new cloud vendors, data center developers, and private credit. Looking only at the debts of publicly listed companies would systematically underestimate the overall funding pressure on the industry chain.

Replacing self-credit with customer credit, short-term financing can still be absorbed

CITIC Securities believes that the supply of short-term financing has the capacity to absorb demand. The core logic lies in new clouds leveraging small equity to drive large investments—using customer credit to replace their own credit: GPU collateral financing supported by investment-grade long-term contracts shifts the lender's risk exposure from the residual value of GPUs to the cash flow from contracts themselves. The spread of CoreWeave's DDTL 4.0 dropping to SOFR + 2.25% is evidence of this mechanism in operation.

Customer prepayments effectively serve as interest-free financing. According to disclosures from Nebius's shareholders, about 70% of its new contracts signed in the second quarter included prepayment terms; in the context of continued tight supply of computing power, this proportion is likely to remain high in the short term.

2028 is the turning point: computing power prices are more important than interest rates

CITIC Securities assesses that the core variable for new clouds' debt repayment ability is not interest rates, but computing power leasing prices. According to data from Silicon Data, the leasing price index for H100 has risen by 31% year-on-year; CITIC Securities estimates that for every 10% increase in computing power prices, CoreWeave and Nebius's net profits in 2027 will increase by approximately 62% and 69%, respectively. However, even in this price environment, CoreWeave's free cash flow remains negative—price increases improve profit elasticity but do not absolutely fill the funding gap.

This means that starting in the second half of 2027, delivery progress, model capabilities, and commercialization advancements will become key tracking indicators. Power constraints will support rental income while also delaying revenue recognition; meanwhile, interest rates are unlikely to materially decline before 2027, further compressing the space for improvements in financing costs.

CITIC Securities predicts that 2028 is expected to be a turning point for cash flow and financing costs, provided that AI demand continues to materialize during this window; if commercialization falls short of expectations, the term mismatch issues currently masked by short-term financing channels will be concentrated and exposed in 2028. In the medium to short term, given the continued tight supply of computing power and resilient prices, overly extreme pricing expectations in the new cloud sector are likely to be corrected, but the sustainability of this correction ultimately depends on the demand quality that begins to be gradually validated in the second half of 2027.