Cracks in the AI infrastructure financing chain are spreading from the periphery toward the center.
CoreWeave bonds are already deeply distressed, and as a tenant it underpins the repayment capacity of more than $14 billion in data center bonds in the high-yield market. Deutsche Bank's market-wide credit map shows that AI credit divergence is accelerating—AA-rated tech giants still enjoy quasi-sovereign funding costs, but Oracle's CDS at the low end of investment grade has widened sharply. A downgrade would trigger massive forced selling, pushing this "peripheral crisis" into the heart of the market.
CoreWeave's bonds have traded at distressed levels for three consecutive days. Its 9.625% notes due 2032 yield about 14% at the close, and according to Reuters, the yield on its 2031 notes exceeds 16%—a level typically seen as a distress signal. Its five-year CDS closed at 1,117 basis points, implying a roughly 63% probability of default within five years, comparable to CDS levels in the days before Lehman Brothers filed for bankruptcy. The 2032 notes have fallen below 80 cents since issuance in June, entering the distressed range as defined by market convention.
The core of the problem is not limited to CoreWeave itself. CoreWeave is also the sole tenant for seven of the 20 project-financed data center bonds in the high-yield market, involving $14.1 billion in debt, or 31% of that market's total. This means CoreWeave's cost-of-capital crisis directly transmits to the bond repayment capacity of issuers such as Applied Digital, Core Scientific and Galaxy Digital.
At the same time, Deutsche Bank warned that Oracle's five-year CDS surged 36 basis points in two days to 288 basis points. If its BBB rating is downgraded, institutional investors that can only hold investment-grade assets would be forced to sell, triggering a systemic chain reaction.
The Junk Bond Market's AI Exposure: $14.1 Billion Hangs on CoreWeave
According to Morgan Stanley's latest monthly report "The State of AI Financing: Pause, Not Retreat," there are at least 20 project-financed data center bonds in the high-yield market, with total debt of $45.6 billion, financing 4,063 megawatts of critical IT capacity. Each bond is issued by a single special purpose entity, with debt service sourced from rent payments by a single tenant.
CoreWeave is the largest tenant among them, and also the most fragile link. Of the 20 bonds, seven have CoreWeave as tenant, involving $14.1 billion in debt and 1,512 megawatts of capacity, accounting for 31% of the total and 37% of capacity—more than the Google-backed Fluidstack series ($11.9 billion) and Nvidia ($8.4 billion), making it the largest single source of rent in the high-yield data center market.
Specific issuers relying on CoreWeave rent include:
Applied Digital (APLD): two bonds totaling $3.94 billion ($2.35 billion of 9.25% 2030 notes and $1.59 billion of 7% 2031 notes, both for the Polaris Forge project in Ellendale, North Dakota), plus another $2.15 billion Polaris Forge 2 bond with Oracle as tenant;
Core Scientific (CORZ): $3.3 billion, financing 590 megawatts of capacity across Texas, Georgia, Oklahoma and North Carolina, the single largest bond by CoreWeave-leased capacity on the tracking list;
Galaxy Digital (GLXY): $3.51 billion of 9.875% notes for the Helios campus in Lubbock, Texas, with final delivery extending to March 2028;
Edged ($1.3 billion, co-leased by CoreWeave and Alibaba), Elk Grove Village ($900 million) and DDC 01 Propco ($1.12 billion, Richmond, Virginia), the last of which priced on September 23—three weeks before the severe market turmoil.
This is dual exposure, not simple circular financing. Take Galaxy bondholders as an example: they bear both the construction risk in Lubbock and the credit risk of whether CoreWeave can continue paying rent from 2028 onward.
The Market Is Already Pricing by Tenant: Spread Divergence Is Clear
Morgan Stanley's spread data clearly presents the market's judgment. As of September 28, the seven CoreWeave tenant bonds had an average spread of about 400 basis points, double the four Google-backed Fluidstack deals (Cipher, TeraWulf, Flash, Meridian, about 195 basis points), and far above Amazon-leased bonds (about 150 basis points). Same asset class, same structure, similar ratings—the only variables are the name on the lease and whether a hyperscaler co-signs the guarantee.
The widest-spread bonds share common features: Zenith Arc (Jane Street tenant, Oklahoma, final delivery October 2028) at 599 basis points and 11.09% yield; Galaxy at 559 basis points/10.68%; Edged at 498 basis points/9.92% (widened 160 basis points since issuance); and SoftBank-leased SE Cosmos at 451 basis points. Weak tenants, long construction cycles, or both, are the common risk factors.
Notably, the best performer among CoreWeave tenant bonds is the one with the nearest delivery date—Applied Digital's original 9.25% 2030 notes, with final delivery on October 31, 2026, quoted at 104.75 with a spread of 218 basis points. The closer to energized delivery, the less sensitive the market is to tenant credit; the farther the delivery date, the greater the weight of tenant credit.
In addition, Morgan Stanley's scatter plot of loan-to-cost (LTC) versus yield to maturity shows that bonds issued in the second and third quarters of 2026 (the peak of the boom) cluster in the LTC 85%-95% and yield-to-maturity 13%-17% range—meaning the newest issuances have the thinnest equity cushion, the highest required returns and the most concentrated risk exposure.
Deutsche Bank Maps the Whole Market: AI Credit Has Already Split Into Three Tiers
Deutsche Bank Research plotted every bond in the dollar investment-grade and high-yield indices by yield and duration, marking AI-related names, producing the most systematic AI credit panorama to date.
The conclusion: AI financing costs overall are not yet out of control, but divergence at the far end of the curve is widening.
Deutsche Bank divides AI credit into three tiers:
AA-rated AI names (Meta, Alphabet, Nvidia, Amazon): short-end funding costs around 5%-5.5%, at the bottom of the investment-grade distribution, with the market still treating them as quasi-sovereign borrowers;
BBB-rated AI names (Oracle, SpaceX): 10-12 year duration funding costs around 8%-8.5%, at the top of the investment-grade distribution, about 1.5-2 percentage points above bonds of similar duration, forming a clear "AI tax";
High-yield AI names: mostly BB-rated, yielding 7%-11%, broadly in the middle of the high-yield market.
CoreWeave is the sole outlier. Across the entire high-yield index, the highest-yielding bonds (15%-19%) are usually unrelated to AI, but CoreWeave is the exception—its BB+ rated unsecured 2032 notes fell below 80 cents just four months after issuance, with five-year CDS at 1,118 basis points.
Deutsche Bank noted that CoreWeave leases computing power under four- to six-year "take-or-pay" contracts, with a backlog exceeding $100 billion, about two-thirds of which comes from investment-grade customers such as Microsoft, Meta, IBM and Caterpillar. But the bond market clearly cares more about leverage than the customer list.
Deutsche Bank also pointed to a structural contradiction overlooked by the market: advanced GPU chips have a depreciation cycle of about four to six years, matching the contract term, but the debt financing them has a longer maturity.
This means the collateral depreciates before the debt matures, which is precisely the fundamental reason for widening spreads at the far end of the yield curve. Combined with the market repricing Fed rate-hike expectations, term premium will become an additional burden for all borrowers needing to refinance.
Oracle: The Next Domino?
Deutsche Bank explicitly lists Oracle as the next risk node to watch closely.
The bank warned that "this concern could become a self-fulfilling prophecy and risks triggering broader problems for other AI-related names," and singled out Oracle as a "bellwether": its five-year CDS surged 36 basis points in two days to 288 basis points, and Goldman Sachs data showed a single-day widening of 31 basis points on Wednesday.
Oracle's systemic risk lies in its rating position. Oracle currently sits at BBB, the lowest investment-grade tier. A downgrade would force a large number of institutional investors that can only hold investment-grade assets into forced selling—not an orderly repricing, but a forced liquidity shock.
From the bond market's perspective, the spreads on two Oracle tenant bonds (Applied Digital's Polaris Forge 2 and Yondr's Loudoun County project) are already 185-220 basis points wider than Oracle's own unsecured bonds of comparable maturity, with Yondr's spread once widening to nearly 300 basis points in early September. Morgan Stanley warned last week that data center delivery delays could severely hit Oracle's debt position.
Financing Gap and Bull-Bear Divide
CoreWeave's core contradiction lies in a $35 billion financing gap.
Deutsche Bank analyst Brad Zelnick, in a report titled "Dislocation Creates Buying Opportunity," reiterated a buy rating and $150 price target for CoreWeave, noting the company had $15.5 billion of available liquidity at the end of the second quarter, 60%-65% of its backlog from investment-grade customers, and unlevered internal rates of return on new contracts jumping from about 15% to "above about 20%." But he also acknowledged the key contradiction:
CoreWeave will undoubtedly need additional capital over the next 12-18 months to fulfill existing orders, with operating cash flow of about $10 billion over the next 12 months against required capital expenditure of about $45 billion—a gap of about $35 billion.
If even half of that gap needs to be financed at the current unsecured bond rate of about 16%, additional interest expense would reach about $2.8 billion a year, exceeding the $2.8 billion total loss analysts expect CoreWeave to post this year.
Goldman Sachs' Rich Privorotsky leans toward the view that the market is "overreacting locally," arguing that "there are legitimate financing constraints, but this does not negate the underlying demand logic." Morgan Stanley believes that "for high-quality high-yield bonds, risk-reward looks good when rates stabilize."
The key qualifiers in both Goldman Sachs' and Morgan Stanley's views are "high-quality" and "if rates stabilize"—and with the 10-year Treasury yield still above 5% and rate-hike expectations rising, neither condition holds for a BB- rated building shell in Lubbock whose sole tenant's CDS is already priced for default.
Transmission Path: From Periphery to Center
Combining the two reports from Morgan Stanley and Deutsche Bank, the current transmission path of AI credit risk is relatively clear:
The first layer is CoreWeave itself—unsecured bond yields of 16%, CDS implying a default probability above 60%, and a $35 billion financing gap;
The second layer is the $14.1 billion of high-yield data center bonds relying on CoreWeave rent—issuers such as Galaxy, Core Scientific, Applied Digital and Edged, where bonds with longer construction cycles and farther delivery dates carry higher risk;
The third layer is Oracle—BBB rating hanging at the low end of investment grade, CDS continuing to widen, and a downgrade triggering forced selling that spreads the crisis from the high-yield market into the heart of the investment-grade market;
The fourth layer is the institutional investors such as life insurers buying these bonds—they are the largest buyers of such project-financed bonds and are re-examining what kind of credit endorsement the appearance of a hyperscaler name in an offering memorandum actually means.
Deutsche Bank's conclusion aligns closely with Goldman Sachs' Privorotsky:
"As the capex cycle shifts from free cash flow to debt financing, we must increasingly respect the price signals of the credit market. AI financing costs are not extreme overall, but worrying themes have emerged at the periphery."
Credit crises always begin at the periphery. In 2007, subprime was also considered "contained."
Comments