CoreWeave Credit Crisis Resembles the 2023 SVB Panic More Than the 2008 Subprime Mortgage Crisis?

The yield on CoreWeave’s unsecured bonds has soared to around 15%, with CDS spreads surpassing 1,000 basis points, raising strong doubts in the market about systemic risks to the entire AI capital expenditure cycle.

However, leading Wall Street institutions are collectively speaking out: this is not a 2007 New Century Financial-like systemic collapse, but more akin to the phase-specific panic triggered by Silicon Valley Bank (SVB) in 2023—a very small-scale, structurally fragile instrument is being treated by the market as a barometer for a trillion-dollar AI capex cycle.

Ajay Rajadhyaksha, Global Head of Macro Research at Barclays, pointed out that CoreWeave’s approximately $1.5 billion in unsecured bonds represent a structurally subordinated position within a total debt of about $35 billion. Using this instrument’s price signal to negate the entire AI cycle is highly reminiscent of using a $4 billion regional bank ETF in 2023 to predict a U.S. recession. Goldman Sachs believes the market has chosen the wrong benchmark—the unsecured bond market does not reflect CoreWeave’s true incremental financing cost.

Historical Precedent: SVB or the 2007 Financial Crisis?

The core of this debate is the choice of historical analogies.

In April 2007, the largest U.S. subprime mortgage lender, New Century Financial, filed for bankruptcy. Subsequently, the entire credit market gradually collapsed, with the S&P 500 index falling about 40% over the next two years. In March 2023, Silicon Valley Bank was shut down by regulators, immediately sparking a narrative: regional banks are the lifeline of SME credit and a credit freeze would trigger a recession. The bond market rapidly priced in a 75 to 100 basis point rate cut for that year, and the VIX panic index soared even more dramatically than in 2007.

In the end, the Federal Reserve did not cut rates at all in 2023—instead, it kept hiking. The U.S. economy grew far beyond expectations that year, and a recession never materialized. The regional bank ETF, KRE, recovered all its losses, and the S&P 500 surged about 40% in the following two years.

Bloomberg columnist John Authers noted that both crises appeared severe in real time, but ultimately went in opposite directions. SVB was an instance of panic that was over-extrapolated, not a precursor to systemic risk; the bankruptcy of New Century Financial revealed a structural flaw that spanned the entire financial system. From a structural standpoint, CoreWeave’s current predicament more closely resembles SVB.

Rajadhyaksha summarized the core error in the SVB panic: “The mistake was not in recognizing that Silicon Valley Bank had collapsed. It actually did. The error was in extrapolating from a small, highly volatile, structurally weak instrument to the entire U.S. economy.”

Rajadhyaksha directly likens the current market panic’s logical structure to the SVB episode.

The “tail” in 2023 was the KRE, an equally weighted regional bank ETF with about $4 billion in assets under management, and many of its constituents were thinly traded. The decline of this small ETF led to the repricing of trillions of dollars in rates, investment-grade credit, and equity positions, fueling a self-reinforcing panic cycle.

The “tail” in 2026 is CoreWeave’s roughly $1.5 billion structurally subordinated unsecured bond.

Rajadhyaksha outlined CoreWeave’s three-layer debt structure: The primary component is secured delayed draw term loans (DDTLs), held within independent special purpose vehicles (SPVs), each backed by specific GPU hardware, customer contracts, and datacenter leases—when Meta signs a long-term “pay whether you use it or not” contract and assigns it to an SPV, the real collateral for the loan is Meta’s payment obligation; second, there are about $9 billion in convertible bonds with coupons between 1.75% and 2.875%; at the very bottom are unsecured notes, with tens of billions of dollars in senior claims ahead of them in liquidation.

“A 15% yield tells you something about that piece of paper, but it doesn’t tell you that the company is about to stop operating,” Rajadhyaksha wrote, using an analogy: “If someone has a $35 million mortgage and a small credit card, the high credit card interest shows a credit card issue, not that every mortgage borrower in the city is in trouble.”

The numeric comparison is equally intuitive: In March 2023, the market priced in a 100 basis-point rate cut due to a $4 billion ETF’s decline; in 2026, the market begins to question a trillion-dollar capex cycle because a roughly $1.5 billion structurally subordinated bond trades at 80 cents on the dollar.

The main argument from Goldman Sachs analysts Gabriela Borges and Maura Hager also focuses on misreading the debt structure.

Goldman notes that CoreWeave’s main financing tool is the DDTL—asset-backed project financing collateralized by specific customer contracts, GPU infrastructure, and related cash flows. Lenders underwrite based on the quality of specific contract cash flows, not overall CoreWeave credit, shifting financing risk off the company’s balance sheet and reducing refinancing risk. The unsecured bond market for new CoreWeave issues has been closed for months, so using its yields to measure financing costs is relying on an obsolete price signal.

Based on Q2 interest expense annualized to $2.56 billion over roughly $30 billion in total debt, Goldman estimates CoreWeave’s current weighted average debt cost at about 800–900 basis points. For incremental financing, Goldman assumes that investment-grade clients (e.g., Meta) correspond to DDTL pricing around SOFR+300 basis points, high-yield clients around SOFR+600, with blended costs between 750 and 1050 basis points—using a conservative base of 1,000, this means incremental financing costs would only rise 100–200 basis points.

Goldman also cited market data: DDTL 5.0 and 5.5 both trade above issue price discounts, showing that the asset-backed financing market’s confidence in CoreWeave is intact.

Of note, Goldman disclosed in a report appendix that it currently serves as bookrunner or co-manager for a pending CoreWeave underwriting, holds over 1% of CoreWeave shares, received investment banking compensation from CoreWeave in the past 12 months, and anticipates further related business in the next three months.

The other half of Goldman’s bullish thesis rests on the expectation that CoreWeave’s compute monetization capacity will rise.

Goldman estimates CoreWeave’s current blended monetization level at about $10 million in annual recurring revenue (ARR) per MW. Supporting the upside: Short-term (3-6 month) contract signings have reached $40 million/MW; CoreWeave raised prices 25% in July and then another 10% in October; contracts signed in Q2 showed profit margins 5–10 percentage points higher than recent deals. Goldman uses $13 million/MW as the new “reasonable base case” for long-term contracts.

Plugging into the example model (five-year, 200MW contract, 80% debt financing, 10% financing cost, zero residual value) yields a 15% unlevered IRR—a positive spread of about 3 percentage points over an 11.8% weighted average cost of capital.

However, Goldman’s sensitivity analysis also reveals a vulnerability: If the actual blended rate remains at $10 million/MW, the unlevered IRR drops to 6.3%, a negative spread of 6 percentage points at the same debt cost. The entire bullish case rests on CoreWeave raising per-megawatt revenue about 30% above current levels.

Goldman’s own model data also reveals how long this path may take: free cash flow remains negative $28.2 billion each in 2026 and 2027, negative $15.8 billion in 2028, with long-term debt rising from about $44 billion this year to around $91 billion in 2028. Free cash flow is only expected to turn positive by 2030.

Rajadhyaksha shifts focus deep into the supply chain, using hard data to counter the systemic collapse narrative.

TSMC recently reported revenue exceeding its own guidance by $1 billion, notching record profits for the sixth straight quarter. High-performance computing’s revenue share rose to 68%, and the company raised its Q4 growth forecast. ASML also posted record sales and more than doubled operating profits; CEO Christophe Fouquet said plans for 2027 and 2028 output will fall short of demand, and some customers are already trying to book systems for 2029 and 2030.

The five hyperscalers (Microsoft, Meta, Alphabet, Amazon, Oracle) are expected to spend $725–800 billion on AI infrastructure in 2026, and over $1 trillion in 2027. CoreWeave’s 2026 capex guidance is about $35–39 billion—just 4%–5% of the total above.

“Even if CoreWeave completely stops its capex, we would lose just 4%–5% of total AI capex,” Rajadhyaksha noted. “Hyperscalers’ spending mostly relies on free cash flow, not the unsecured bond market. Extrapolating from an emerging cloud provider to some of the world’s most profitable companies stretches the logic too far.”

Barclays concedes that a key difference from 2023 is that there’s no equivalent central bank tool as a safety net. In 2023, the Fed’s Bank Term Funding Program (BTFP) allowed banks to pledge underwater bonds at par for liquidity, crucially breaking the cycle of panic. There’s no similar policy tool for AI infrastructure financing. But Rajadhyaksha argues that the hyperscalers’ real demand for compute power itself forms the backstop for this cycle.

Rajadhyaksha admitted, “The market tends to find the smallest, most volatile instruments in complex systems and treat them as a barometer for the whole. In 2023 it was KRE; in 2026, CoreWeave’s unsecured bonds. Both signals are real. The 2023 extrapolation was wrong, and we believe the 2026 extrapolation will be wrong, too.”

Despite the collective “damage control” from leading Wall Street institutions, all sides clearly identify one common risk trigger.

Rajadhyaksha emphasized that whether the AI capex cycle will repeat the subprime bust depends on whether compute rental rates stay high enough to cover capital costs. “If rentals don’t remain as high as expected, a true crisis could ensue. CoreWeave is an alarming case, but it doesn’t offer evidence on this front.”

Goldman’s sensitivity analysis points to the same core vulnerability: If per-MW ARR fails to rise from $10 million to $13 million, the entire positive spread logic unravels completely. And if supply catches up with demand and pricing falls, CoreWeave will still be locked into data center leases lasting seven to sixteen years, making exit extremely costly.

Brian Stutland of Equity Armor Investments argued that CoreWeave’s core issue is its own business model, not the future of AI as a whole, though he admitted this is “a two-sided coin”—whether it is an isolated company-level problem or reveals a broader structural fragility is currently “about a 50–50 probability.”

For investors, the current key is not whether CoreWeave will fail, but whether compute pricing can sustain the financing logic for this cycle. The answer to that question will determine where this panic ultimately leads.