Everything About AI
NextFin News - CoreWeave's shares slid roughly 25% this week, and the yield on its 2032 bonds jumped to nearly 15%, as investors turned on the debt-fueled AI build-out and forced one of the market's most visible AI winners to defend its balance sheet. The selloff is not just about one company. It is a test of whether credit markets will keep funding an AI infrastructure boom that has leaned on borrowed money at a pace no previous cycle has seen.
CoreWeave's bonds traded below 80 cents on the dollar this week, a level that typically signals distress, after the company issued the 9.625% notes in June. The stock, which reached a 52-week high of $141.79, has lost more than half its value over the past year and closed Friday at $61.39. The pressure spread beyond CoreWeave: the Nasdaq Composite fell about 1.9% on Wednesday, while Oracle dropped nearly 9%, neocloud peer Nebius declined about 10%, and Dell fell roughly 11%, as the market repriced the risk of AI capital spending financed with debt.
The central question is no longer whether AI demand is real. It is whether the returns on the capital being poured into AI infrastructure can service the debt that is paying for it - and what happens to the lenders, funds, and investors holding that paper if they cannot.
The Credit Market Is Asking a Question the Equity Market Ignored
CoreWeave's stock had gained 8.7% this year through Tuesday's close. By Wednesday it had fallen as much as 17%, to $64.41, before closing down around 15% for the day. The move was not triggered by a missed earnings print or a lost customer. It was triggered by a bearish research note arguing that enterprise and consumer AI revenue will be insufficient to support profitable returns on all planned AI capital expenditures, and by a chart of CoreWeave's bond yield that drew a one-word response - "yikes" - from Elon Musk after investor Chamath Palihapitiya posted it.
That sequence matters. In the first nine months of 2026, the equity market rewarded AI infrastructure stories on the promise of future demand. The credit market, which has to be repaid in cash on fixed dates, is now asking a different question: what does the collateral look like if the story slows down?
The answer, so far, is making lenders nervous. CoreWeave's 9.625% bond due in 2032 carried a yield to maturity of about 14.8% this week, while its 9.75% note due in 2031 yielded roughly 14.9%. Yields move inversely to prices, so a jump to that level means the market is pricing a materially higher probability of stress than it was when those bonds were issued at par just months ago. In late July, during the last scare, CoreWeave's credit default swaps blew out to roughly 855 basis points, a level that on standard pricing models implied about a 50% probability of default within five years. The stock bottomed near $60, then nearly doubled to around $117 by mid-August after second-quarter revenue rose more than 200% year over year and the company reported a backlog of roughly $104 billion.
The equity market forgave the scare in three weeks. The credit market did not. CoreWeave's most recent financing still carries the terms lenders demanded when they were frightened.
In mid-July, the company went to the leveraged-loan market to raise $2.6 billion through a delayed-draw term loan, the fifth-and-a-half such facility in its series. The deal launched at a spread of SOFR plus 425 to 450 basis points, offered at 99 cents on the dollar. It stalled. To get it across the line, CoreWeave accepted a spread of roughly SOFR plus 550 basis points, priced the loan at about 96 cents, and agreed to a cash lockbox that directs customer payments toward debt service before the company can touch them, plus a 1.35x debt-service-coverage covenant and full amortization. Rating agencies Fitch and Moody's scored the loan BB+ and Ba2, respectively - one notch below investment grade. Loan-market coverage pegged the all-in cost at roughly 9%.
The concessions are the point. When a trade is genuinely bulletproof, financing is supposed to get cheaper. This one got more expensive, more restrictive, and harder to place, all at once. Investors only lined up after CoreWeave caved to the lenders' demands, and the company described the closed deal as "meaningfully oversubscribed." The demand appeared after the terms moved in the lenders' favor - not before.
"The era of easy money, which we experienced in the first nine months of the year, is starting to disappear and dissipate," said Luis Alvarado, co-head of global fixed income strategy at Wells Fargo Investment Institute. "Now the extra dollar or the marginal dollar is going to be more looked at with a microscope into the concessions that the issuer is giving you back."
Not every borrower is being treated the same way. Lambda, another neocloud, secured a $1 billion senior secured, fixed-rate term loan on October 1 - its third GPU-backed facility of 2026 - at 6.78%, rated A-low by Morningstar DBRS and Baa1 by Moody's, and tied to an infrastructure partnership with Microsoft. The loan was oversubscribed. That bifurcation is the clearest signal in the sector: capital is still available for AI infrastructure, but it is migrating toward borrowers with the strongest counterparties and the cleanest structures, while everyone else pays up.
Why the GPU-Collateralized SPV Model Worries on the Way Down
To understand why the credit market is re-examining CoreWeave, you have to look at how the company financed its growth. CoreWeave did not simply borrow against its balance sheet. It built a series of delayed-draw term loan facilities, each housed in a separate, bankruptcy-remote special purpose vehicle created to own the GPUs and serve as the borrower for the loan. The company's most recent $8.5 billion facility is secured by substantially all the assets of CoreWeave Compute Acquisition Co. VIII, LLC, a subsidiary rather than the parent, and allows multiple draws until June 2027.
On the way up, this structure was a feature. It let CoreWeave raise billions without diluting shareholders or bloating the parent balance sheet, and it gave lenders assets they could point to plus contracted cash flows from specific customer contracts. Each tranche's credit story rests on a single counterparty. According to loan-market reporting, the largest counterparty behind the $2.6 billion DDTL 5.5 facility is Anthropic, at roughly 40%, followed by the quant trading firm Jane Street at about 35%, with Midjourney, Hudson River Trading, and Anysphere making up the remainder.
On the way down, the same structure raises three uncomfortable questions.
First, the collateral depreciates fast. Nvidia releases new chip architectures on a cadence that can leave today's flagship GPU two generations behind by the time a five-year loan matures. The secondary market for seized GPU collateral is thin; there is no deep, liquid marketplace where a lender can quickly sell thousands of high-end accelerators at a predictable price. If a borrower defaults, the asset a lender takes back is worth less than the loan, and worth less every quarter.
Second, the debt load has grown faster than the revenue base that must service it. CoreWeave's total debt exceeded $21 billion as of May, up from under $8 billion in 2024. Its 2026 capital spending guidance calls for $31 billion to $35 billion, against roughly $6.2 billion of trailing-twelve-month revenue. Research from MLQ.ai put the company's debt-to-equity ratio above 4.8x, with more recent estimates near 8.9 and a current ratio of about 0.5. The company is paying roughly 25% of its revenue just to cover interest on its debt.
Third, the complexity itself is a risk factor. S&P Global Ratings wrote in a December 2025 report that "CoreWeave is rapidly evolving its complex debt capital structure as it seeks to finance its significant capital expenditure needs." Complexity is manageable in good times. It becomes a liability when creditors need to move quickly.
There is also a distribution question that most equity investors never have to confront. Some of this paper ends up in broad-based corporate bond funds that sit inside target-date retirement products, according to a credit strategist interviewed on the structure. The retail holder has no idea why they own it, no framework for evaluating it, and no ability to exit it independently of the fund. That is not a prediction of default. It is a warning that the investors holding this risk may not be the investors who understood it when they bought it.
The Second-Order Read: It Is the Marginal Dollar of AI Capex That Is Repricing
The first-order effect of CoreWeave's slide is obvious: the company's cost of capital rises, and its stock falls. The second-order effect is larger, and it reaches every company racing to build AI infrastructure.
Hyperscalers are on track to spend about $800 billion on AI infrastructure this year, according to JPMorgan, and that figure is expected to exceed $1 trillion annually by 2027. For most of this cycle, that spending was funded with cheap equity and cheap debt. If the marginal dollar of AI capital expenditure now costs 9% to 15% instead of the near-zero real rates of the early build-out, the economics of the entire GPU-rental arbitrage compress. A project that clears its hurdle rate when capital costs 4% does not clear it when capital costs 12%.
That is why the selloff spread to Oracle, Dell, and Nebius even though their businesses are not CoreWeave's. The market is not asking whether each company will default tomorrow. It is asking whether the return on invested capital across the AI infrastructure complex can keep up with a cost of capital that is resetting higher. The debt concern is hitting the shakier stocks outside the "Magnificent Seven" the hardest, precisely because those are the names whose financing depends on the story continuing to work.
This creates an asymmetry that benefits the strongest balance sheets. If weaker neoclouds lose access to cheap financing, the hyperscalers and top-tier infrastructure owners can consolidate capacity, pick up customers whose providers stumble, and negotiate better terms with chip suppliers. The end of easy money does not kill AI infrastructure spending. It redistributes it - from the borrowers who needed cheap money to survive to the owners who can fund growth from their own cash flows.
The bear case goes further. Gil Luria, a veteran technology analyst at D.A. Davidson, speculated on a podcast that CoreWeave's stock could fall to zero if the AI spending boom wanes. The specific risk he identified is not default in the next quarter; it is customer concentration and renewal. If the industry overbuilds and customers fail to renew their contracts when they expire, the backlog that supports today's debt simply does not convert into tomorrow's revenue. With Microsoft accounting for 62% of CoreWeave's 2024 revenue, and the top two customers representing roughly 65% of first-quarter 2026 revenue, that concentration is the single most important variable in the credit story.
The Counter-Thesis: The Backlog Is Real, and the Market May Just Be Early
The strongest argument against the bear case is straightforward: CoreWeave's backlog is real, contracted, and large enough to cover its obligations. Chief Executive Mike Intrator told employees in an internal memo this week that the company's deal backlog is large enough to more than support its debt obligations, and that he is confident in the business model, customers, and execution. The company reported revenue backlog of $104.2 billion in the second quarter, up from $99.4 billion in the first, and said it had secured more than $25 billion of net new customer commitments in the third quarter.
Intrator also argued that CoreWeave is being punished for other companies' problems. Weakness and delays around the attempted IPOs and debt financings of several rivals "are bleeding into the narrative around CoreWeave," he wrote, naming Firmus, AirTrunk, Lambda, and Nscale. "These companies are trying to raise enormous amounts of capital without the execution, revenue, scale, or software solutions that CoreWeave delivers to its customers," he said. Firmus scrapped a listing that had aimed to raise up to $5.5 billion after investors rejected its $30 billion price tag. AirTrunk pushed back a planned Singapore real estate investment trust IPO until next year. JPMorgan and Goldman Sachs pulled back from fully underwriting a $5 billion loan package for neocloud Volta Infrastructure Holdings, with commitments due October 14.
The bankruptcy-remote SPV structure is also a genuine protection for lenders: each facility is tied to a specific customer contract, so the credit of each tranche rests on a contracted counterparty rather than on the parent company's unsecured promise. And demand for AI compute remains strong - Oppenheimer has argued that AI demand runs about four times available capacity, which is why the company kept an Outperform rating and a $150 price target even as the sector wobbled.
These points are real. But they answer a question the market is no longer asking. The market is not debating whether CoreWeave has a backlog. It is debating whether that backlog converts into revenue and cash flow fast enough to service debt that carries covenants, amortization, and collateral that depreciates every quarter. A backlog is a promise of future revenue. Debt service is a claim on cash today. The gap between the two is where the risk lives, and no amount of reassurance about the size of the backlog closes that gap by itself.
"When you are building a new type of technology company, there will be periods of time when the market misunderstands and acts irrationally," Intrator wrote in the memo, which a company spokeswoman confirmed. He added that he looked forward to "having the opportunity to put the facts in front of the market" on the company's next earnings call, with quarterly results expected in November.
The market may well be early. It may also be right. The two are not mutually exclusive, and the cost of being early is very different for a bondholder with a fixed maturity date than for an equity holder with a long time horizon.
What to Watch: The Signals That Will Settle the Argument
The argument between the bull case and the bear case will be settled by three observable signals over the coming months.
First, backlog conversion. In the November quarterly results, the market will look for evidence that the $104.2 billion backlog and the $25 billion of new third-quarter commitments are turning into recognized revenue and operating cash flow at the pace management has guided. If conversion lags, the "irrational market" defense fails, because the credit thesis rests on cash flow, not contracted promises.
Second, the bond market itself. If the 2032 note's yield stays above 12% for two consecutive weeks, or if the notes trade back below 80 cents on the dollar, the credit market is telling you it does not believe the equity story. Credit investors are slower to move than equity traders but harder to bluff. Their pricing is the most honest signal in the trade.
Third, refinancing. CoreWeave's latest facility allows draws until June 2027. Watch whether the company can access new capital on terms that do not keep ratcheting tighter. If each new facility demands a higher spread, a deeper discount, and more restrictive covenants, the squeeze compounds even if the operating business holds up.
The base case is continued volatility with the backlog ultimately supporting the debt, but at a higher cost of capital that compresses margins. The upside case is that November's results reassure investors, bond yields compress back toward the low double digits, and the stock recovers a meaningful portion of its losses as the sector's bifurcation favors the strongest operators. The downside case is that backlog conversion disappoints, covenant pressure builds, and the contagion that hit Oracle, Nebius, and Dell spreads to other leveraged AI infrastructure names.
Short term, this is a sentiment and liquidity story: the stock will trade on headlines about bonds, rival financings, and the November earnings date. Medium term, it is a fundamentals story: revenue growth, margin, and cash conversion against a $31 billion to $35 billion capital spending plan. Long term, it is a structural story about the end of the easy-money regime for AI capital expenditure - and that regime change does not reverse on its own.
This is not a verdict on AI demand. It is the credit market charging the AI build-out its first real risk premium, and the companies that built their growth on the assumption that the premium would never arrive are the ones now paying for it.
Explore more exclusive insights at nextfin.ai.
