CoreWeave's unsecured bond yields have surged to around 15% and its CDS spreads have broken past 1,000 basis points, fueling intense market doubts about systemic risk across the entire AI capital expenditure cycle.
Yet leading Wall Street institutions are speaking with one voice: this is not a 2007-style systemic collapse like New Century Financial, but rather a temporary panic akin to the one triggered by Silicon Valley Bank in 2023. A very small, structurally fragile instrument is being treated by the market as a barometer for a trillion-dollar AI capital spending cycle.
Ajay Rajadhyaksha, head of global macro research at Barclays, noted that CoreWeave, Inc. (CRWV) has roughly US$1.5 billion in unsecured bonds that sit in a structurally subordinated position within total debt of about US$35 billion. Using the price signal from this instrument to reject the entire AI cycle is the same as using a roughly US$4 billion regional bank ETF in 2023 to predict a US recession. Goldman Sachs, meanwhile, argues the market has picked the wrong reference point, since the unsecured bond market does not represent CoreWeave's true incremental cost of financing.
Historical precedent: SVB or the 2007 crisis?
At the heart of this debate is the choice of historical analogy.
In April 2007, New Century Financial, the largest US subprime mortgage lender, filed for bankruptcy, and the broader credit market subsequently unraveled step by step, with the S&P 500 falling about 40% over the following two years. In March 2023, Silicon Valley Bank was closed by regulators, and the market quickly formed a narrative: regional banks are the lifeline of small and medium-sized business credit, and a credit crunch would trigger a recession. The bond market rapidly priced in 75 to 100 basis points of rate cuts for that year, and the VIX fear index spiked even more than it had in 2007.
In the end, the Federal Reserve did not cut rates once in 2023, but continued raising them instead, US economic growth for the full year far exceeded expectations, and the recession never arrived. The regional bank ETF KRE recovered all its losses, and the S&P 500 rose about 40% over the following two years.
Bloomberg columnist John Authers pointed out that both crises looked severe in real time, yet ended up in completely opposite directions. SVB was a panic that was extrapolated too far, not a precursor to systemic risk; the bankruptcy of New Century Financial exposed a structural flaw running through the entire financial system. In terms of shape, CoreWeave's current situation is closer to SVB.
Rajadhyaksha summed up the core mistake of the SVB panic: "The error was not in recognizing that Silicon Valley Bank had failed. It did fail. The error was extrapolating from a small, volatile, structurally fragile instrument to the entire US economy."
Barclays: a US$1.5 billion 'tail' is wagging the whole 'dog'
Rajadhyaksha directly compares the logical structure of the current market panic with the SVB period.
The "tail" in 2023 was an equal-weight regional bank ETF with roughly US$4 billion in assets under management, KRE, many of whose constituents traded thinly. It was the decline of this small ETF that drove the repricing of trillions of dollars of rates, investment-grade credit, and equity positions, creating a self-reinforcing cycle of panic.
The "tail" in 2026 is CoreWeave's roughly US$1.5 billion of structurally subordinated unsecured bonds.
Rajadhyaksha laid out CoreWeave's three-layer debt structure: the most important is the secured delayed draw term loan, held in separate special purpose vehicles, each tied to specific GPU hardware, customer contracts, and data center leases. When Meta signs a multi-year "pay whether you use it or not" contract and assigns it to an SPV, the loan's substantive collateral is Meta's payment commitment. Next is about US$9 billion of convertible bonds with coupons between 1.75% and 2.875%. At the very bottom are the unsecured notes, behind which tens of billions of dollars of claims queue up in a liquidation.
"A 15% yield tells you something about that piece of paper, but it cannot tell you that this company is about to cease operations," Rajadhyaksha wrote, offering an analogy: "If a person has a US$35 million mortgage plus a small credit card, a high credit card rate tells you about the credit card problem, not that every mortgage borrower in the city is in trouble."
The numerical contrast is equally stark: in March 2023, the market priced in 100 basis points of rate cuts because of a US$4 billion ETF decline; in 2026, with roughly US$1.5 billion of structurally subordinated bonds trading at 80 cents, the market has begun questioning a trillion-dollar capital expenditure cycle.
Goldman Sachs: the market has chosen the wrong benchmark for financing costs
The core argument of Goldman Sachs analysts Gabriela Borges and Maura Hager also rests on a misreading of the debt structure.
Goldman points out that CoreWeave's main financing tool is the DDTL, an asset-backed project financing secured by specific customer contracts, GPU infrastructure, and related cash flows. Lenders underwrite the cash flow quality of specific contracts rather than CoreWeave's overall creditworthiness. This structure shifts financing risk from the company's balance sheet to the underlying assets and reduces refinancing risk. The unsecured bond market has been closed to new CoreWeave issuance for months, so using yields in that market to gauge financing costs is equivalent to referencing a price signal that has ceased to function.
Based on annualized second-quarter interest expense of US$2.56 billion divided by about US$30 billion of total debt, Goldman estimates CoreWeave's current weighted average cost of debt at roughly 800 to 900 basis points. For new financing, Goldman assumes investment-grade customers such as Meta correspond to DDTL pricing of about SOFR plus 300 basis points, while high-yield customers are around SOFR plus 600 basis points, giving a blended cost range of 750 to 1,050 basis points. Using a conservative baseline of 1,000 basis points implies new financing costs would rise only 100 to 200 basis points from current levels.
Goldman also cited market data: DDTL 5.0 and 5.5 are both trading at no less than their original issue discount, indicating that confidence in CoreWeave in the asset-backed financing market has not collapsed.
It is worth noting that Goldman disclosed in an appendix to the report that it is acting as lead or joint bookrunner on a pending underwriting mandate for CoreWeave, holds more than 1% of CoreWeave shares, has received investment banking compensation from CoreWeave over the past 12 months, and expects to continue seeking related business over the next three months.
The pivot and fragility of Goldman's bullish case: revenue per megawatt must rise 30%
The other half of Goldman's bullish argument rests on expectations that CoreWeave's compute pricing power will increase.
Goldman estimates that CoreWeave's current blended monetization level is about US$10 million in annual recurring revenue per megawatt. Supporting the upside are several factors: short-term contracts of three to six months are already being signed at US$40 million per megawatt; CoreWeave raised prices 25% in July this year and another 10% in October; and contracts signed in the second quarter carry contribution margins 5 to 10 percentage points higher than recent contracts. Goldman uses US$13 million per megawatt as a "reasonable starting assumption" for newly signed long-term contracts.
Plugging this into an illustrative model of a five-year, 200-megawatt contract with 80% debt financing, a 10% financing cost, and zero residual value produces an unlevered internal rate of return of 15%, a positive spread of about 3 percentage points over an 11.8% weighted average cost of capital.
However, Goldman's sensitivity analysis also reveals the fragility of this logic: using the current actual blended level of US$10 million per megawatt, the unlevered internal rate of return falls to 6.3%, turning the spread into negative 6 percentage points at the same financing cost. The entire bullish case depends on CoreWeave raising revenue per megawatt by about 30% from current levels.
Goldman's own model data also shows how long that path is: free cash flow is negative US$28.2 billion in both 2026 and 2027, negative US$15.8 billion in 2028, and long-term debt will climb from about US$44 billion this year to about US$91 billion in 2028, with free cash flow not expected to turn positive until 2030.
Supply chain data: demand reality is the true backstop for the AI cycle
Rajadhyaksha turns his gaze to the deepest part of the supply chain, using hard data to rebut the systemic collapse narrative.
TSMC recently reported revenue that beat the upper end of its own guidance by US$1 billion, posted a sixth consecutive quarter of record profit, saw high-performance computing rise to 68% of revenue, and raised its fourth-quarter growth guidance. ASML recorded record sales over the same period, with operating profit more than doubling. CEO Christophe Fouquet said planned supply in 2027 and 2028 will not fully meet customer demand, and some customers have begun trying to reserve equipment for 2029 and 2030.
The five largest hyperscale cloud providers, Microsoft, Meta, Alphabet, Amazon, and Oracle, will spend a combined US$725 billion to US$800 billion on AI infrastructure in 2026, with estimates above US$1 trillion in 2027. CoreWeave's 2026 capital expenditure guidance of about US$35 billion to US$39 billion represents roughly 4% to 5% of that total.
"Even if CoreWeave stopped spending entirely, what we would lose is only 4% to 5% of total AI capital expenditure," Rajadhyaksha said. "Hyperscaler spending is funded mainly by free cash flow, not the unsecured bond market. Extrapolating from one emerging cloud provider to one of the most profitable companies in the world stretches the logical chain too far."
Barclays also acknowledges one key difference from 2023: there is no equivalent central bank tool to serve as a safety net. In 2023, the Fed launched the Bank Term Funding Program, allowing banks to pledge underwater bonds at par for liquidity, which became the key to breaking the panic cycle. AI infrastructure financing has no comparable policy tool. But Rajadhyaksha argues that hyperscalers' genuine demand for compute itself constitutes this cycle's own "backstop logic."
Rajadhyaksha admitted: "The market has a tendency to always find the smallest, most volatile instrument in a complex system and treat it as a barometer for the whole. In 2023 it was KRE, in 2026 it is CoreWeave's unsecured bonds. Both signals are real. The extrapolation in 2023 was wrong, and we believe the extrapolation in 2026 is wrong as well."
The real risk trigger: whether compute rents can hold
Although leading Wall Street institutions are collectively trying to put out the fire, all parties have clearly identified one common risk trigger.
Rajadhyaksha points out that whether the AI capital expenditure cycle repeats the subprime path depends on whether compute rents can remain high enough to cover the cost of capital. "If rents fail to be as high as expected, it could trigger a real crisis. CoreWeave is an alarming situation, but it cannot provide evidence on this front."
Goldman's sensitivity analysis also points to this core vulnerability: if revenue per megawatt cannot rise from the current US$10 million to US$13 million, the entire positive-spread logic reverses completely. And once supply catches up with demand and pricing falls back, CoreWeave remains locked into data center leases of seven to 16 years, making exit costly.
Brian Stutland of Equity Armor Investments believes the core of CoreWeave's problem lies in its own business model rather than the overall outlook for AI, but he also concedes it is a "two-sided coin," with the question of whether this is an isolated company-level issue or reveals broader structural fragility currently sitting at "roughly a 50-50 probability."
For investors, the key question now is not whether CoreWeave will collapse, but whether compute pricing can support the financing logic of this cycle. The answer to that question will determine where this panic ultimately ends.
Comments