Everything About AI
NextFin News - Yields on CoreWeave's junk bonds surged to 15% this week, a level that usually signals distress, as the AI cloud provider's push to borrow in investment-grade markets collides with a credit market that is starting to freeze out technology startups. The selloff has pushed the company's credit-risk gauge to a record that prices roughly 60% odds of default, and it is raising a wider question for the roughly $570 billion of AI-related debt expected to reach markets in 2026: who keeps funding the buildout when the cheapest money runs out?
The timing is awkward. Just as CoreWeave's existing junk bonds flash distress, co-founder and finance lead Brannin McBee has told investors the company wants investment-grade ratings for more of its chip-backed debt, and eventually for CoreWeave itself, so it can tap the vast pools of capital managed by insurers. "Like, that's all the money in the world that sits in the insurance tranche," McBee said. The plan is sound in theory. The market's reaction suggests it will not be cheap to execute.
The 15% Yield Is a Financing Cost, Not Just a Sentiment Signal
The trigger was simple enough to plot on a chart. On Tuesday, venture investor Chamath Palihapitiya shared a graphic showing the yield on $1.25 billion of CoreWeave bonds due in 2032 touch 15%, then close at a record 14.5%. By Wednesday the bonds had fallen below 80 cents on the dollar, a level widely treated as the threshold for financial strain. CoreWeave shares fell 15.46% that day after closing Tuesday at $77.81, and the company's credit-risk gauge climbed to a record that prices about a 60% chance of default.
For a company that is losing money while spending aggressively, a bond yield is not a mood — it is a price. It is the rate CoreWeave would pay for the next dollar of expansion capital, and at 15% it prices a default probability that makes new high-yield issuance nearly impossible. That matters because CoreWeave raised more than $35 billion in 2026 through chip-backed loans, high-yield bonds, convertible debt and equity sales, and it lifted its capital-spending forecast in August by $4 billion to as much as $39 billion for the year. Analysts estimate revenue of nearly $13 billion in 2026 and a loss of $2.13 billion.
The arithmetic that worries bears is stark. D.A. Davidson analyst Gil Luria said CoreWeave is borrowing at around 9% while earning returns of just 1% on its investments, and that even the 4% to 5% the company targets would fall short of its borrowing costs.
"The equity value is going to be zero," Luria said on the Prof G Markets podcast, if the AI investment cycle turns.His wipeout warning describes a downside scenario rather than a formal forecast, and D.A. Davidson's rating on the stock is Neutral after a May downgrade that cut the price target from $175 to $100. But the underlying point is mechanical: if the return on invested capital stays below the cost of capital, growth destroys equity value.
Consider what the 15% yield implies for the next phase of the buildout. CoreWeave's second-quarter revenue was $2.58 billion, up 112% from a year earlier and above the $2.56 billion estimate, and the company guided third-quarter revenue to $3.4 billion to $3.6 billion. Growth is not the problem. The problem is that every incremental data-center rack financed at distressed yields has to clear a much higher hurdle before it adds to equity value rather than subtracting from it. Chief Executive Michael Intrator said in an internal memo that the market is acting "irrationally." Elon Musk responded to the bond chart with a single word: "Yikes."
The Structural Shift: AI Outgrew the Balance Sheets That Built It
The immediate selloff has cyclical drivers. A hot inflation reading revived rate-hike fears, AI jitters spread through chip stocks, and Wall Street recorded its worst week since June as the Nasdaq Composite tumbled 4.5% on Friday. Sentiment can mean-revert. But the financing problem underneath is structural, and it will not fix itself.
The AI data-center buildout carries a price tag of about $3 trillion, and not even Amazon, Microsoft, Meta Platforms or Alphabet can foot that bill with cash alone. Global AI-related debt issuance is projected to hit roughly $570 billion in 2026, close to double last year, with about $236 billion raised by mid-June. Goldman Sachs strategist Amanda Lynam estimated that $489 billion of AI-related debt has been issued this year, already above the $322 billion that came to market in 2025, with about 40% issued directly by hyperscalers. Amazon alone has raised about $53 billion in debt this year.
Because the hyperscalers cannot absorb the whole buildout, the industry has migrated toward project finance: asset-backed loans collateralized by the GPUs themselves rather than by the corporate balance sheet. CoreWeave's precedent is an $8.5 billion loan backed by computing infrastructure and a contract with Meta Platforms, which received investment-grade ratings from Moody's and DBRS. That deal brought CoreWeave's total equity and debt financing commitments over the past year to $28 billion. The model lets a loss-making company borrow cheaply against contracts with creditworthy customers.
Think of it as the difference between lending to a startup and lending to a toll road. The toll road may be owned by a risky company, but the cash flows are contracted, the asset is specific, and the lender has collateral. That is the structure CoreWeave is selling, and McBee now wants to extend it to debt backed by contracts with lower-rated AI labs. Nearly all of CoreWeave's 2027 capacity is already sold, he said, and the company has signed contracts extending into 2028 and beyond. Its $104 billion backlog represents contracted future revenue, however, not profit — and not cash. Visible Alpha consensus forecasts cited in reporting project $71 billion of cumulative negative free cash flow between 2026 and 2028.
McBee argued the spending reflects expansion rather than weakness.
"If we stop building today, [we'd be] immensely cash-flow positive," he said.
That sentence captures the whole tension. The losses are a choice to expand, not proof of a broken business — but the choice only works if capital stays available at a cost below the return the assets produce. The buildout has shifted from a demand problem to a financing problem, and financing is priced by the market, not by management. This is the structural leg of the story: the AI boom has moved from a world where a handful of investment-grade balance sheets funded everything to a world where dozens of speculative-grade builders must roll debt over at market rates. That regime does not revert on its own.
If CoreWeave Is Shut Out, the Whole Segment Reprices
The second-order risk is not that CoreWeave defaults. It is that the marquee name in AI cloud infrastructure cannot issue junk bonds at a survivable cost, and the entire segment reprices with it. More than $26 billion of AI data-center issuance has already come to the U.S. high-yield market, and European high-yield is getting its first taste of the boom. If that door closes for CoreWeave, it closes for every smaller player without an investment-grade balance sheet.
The spillover is already visible. Shares of other neocloud operators, including Nebius and IREN, extended their declines alongside CoreWeave, and a poll of retail traders on a social investing platform showed a split between those buying the dip and those waiting for further falls. Research firm Gartner has said about 80% of neoclouds could shut down by 2030, and that more than $3.5 trillion of AI infrastructure spending could face impairment if providers keep up their current investment pace. Gartner calculated that data-center owners would need to generate a cumulative $8.5 trillion in revenue through 2030 to cover capital spending, financing costs and a 12% return on invested capital.
The credit market is starting to price similar concerns. On prediction markets, traders assigned a 13% probability to CoreWeave announcing bankruptcy by December 2027, rising to 27% by 2029 — lower than the 60% implied by the credit gauge, but bankruptcy is distinct from an equity wipeout, and prediction markets are not a primary pricing source for institutional credit. Even so, the direction is clear: a company that was a market darling in March is now being underwritten as a potential bankruptcy candidate.
There are also warning signs that the collateral underpinning the project-finance model is not as risk-free as the structure assumes. Oracle recently issued a force-majeure citation to Blue Owl over a massive data center in New Mexico, over concerns the site would not be finished when scheduled to come online in 2028 — a sign that customers may have the legal ability to walk away from commitments. And Meta Platforms, one of CoreWeave's largest customers, is building a cloud business to sell its spare AI computing power, which could reduce its need for third-party capacity if internal utilization falls.
The crux of McBee's plan is the rating itself. Insurers typically prefer investment-grade assets, which is why the insurance tranche is so large — regulatory capital rules make high-yield bonds expensive for them to hold, while investment-grade paper fits their mandates. But if the debt is backed by contracts with lower-rated AI labs, the paper can only be investment-grade if the collateral and structure are strong enough to stand apart from the borrower. That is the test the market will apply to the next deal, and it is why a 15% yield on the unsecured 2032 bonds is a warning about the pricing of the secured paper to come.
The Counter-Thesis: Demand Is Real, and the Backlog Is a Moat
The strongest case against the bear thesis is that demand is not the problem. Nearly all of 2027 capacity is sold. Contracts extend into 2028. Revenue grew 112% in the second quarter and beat estimates. The $104 billion backlog is real contracted revenue. And the $8.5 billion investment-grade, GPU-backed loan proves the asset-backed model works when the collateral is a top-tier customer contract. Some investors are acting on that view: ARK Investment Management, led by Cathie Wood, bought about $19 million of CoreWeave shares on Wednesday, treating the weakness as an entry point rather than a warning.
In this reading, the 15% yield and the 60% default odds are a cyclical overshoot — the same kind of repricing that followed every rate scare in the past decade — and Intrator's "irrational" label will look prescient once spreads normalize. The bull case is a demand case: as long as hyperscalers and large AI labs keep signing multi-year contracts, CoreWeave's assets produce the cash flows that service the debt.
The answer to that case is that both stories can be true at once. Demand can be real and financing can still be the binding constraint. The bull thesis depends on three assumptions holding together: customers remain creditworthy, financing stays available, and utilization stays high. Break any one and the arithmetic Luria describes kicks in. The bear thesis is a financing thesis, not a demand thesis, and it does not require the AI boom to end — only to slow enough that refinancing becomes expensive. A boom financed with debt that must be rolled over at market rates is exposed to the market's mood in a way a cash-funded boom is not.
What Comes Next: Three Horizons and a Falsifying Signal
In the short term, measured in weeks, this is a sentiment and liquidity move. The selloff may have overshot, and a successful investment-grade placement or stabilizing commentary from management could produce a rebound. The 13% bankruptcy odds on prediction markets suggest traders are not pricing a near-term collapse, and CoreWeave's $104 billion backlog gives it time to prove the model.
Over the medium term, six to eighteen months, the binding constraint is the refinancing cost and the utilization rate. The key data points are the next CoreWeave debt placement and its all-in cost, the level of the 2032 bonds relative to 80 cents on the dollar, and quarterly free cash flow against the projected $71 billion of negative free cash flow for 2026 through 2028. If the company can place new investment-grade, asset-backed debt at high-single-digit cost while junk stays shut, the squeeze eases. If it cannot, equity dilution or restructuring moves onto the table.
Structurally, the industry is moving toward project finance with investment-grade tranches. The survivors will be the builders who can collateralize capacity with contracts from creditworthy customers and borrow at single-digit costs. Hyperscalers with investment-grade balance sheets — Amazon, Microsoft, Meta and Alphabet — can borrow more cheaply than anyone and may end up consolidating more of the buildout. Nvidia, which sells the GPUs regardless of who finances them, is exposed to demand rather than to the credit structure. The most exposed are the neoclouds without investment-grade access and the high-yield funds concentrated in AI data-center paper.
Three scenarios frame the path. In the base case, CoreWeave places investment-grade, asset-backed debt at a high-single-digit all-in cost, the high-yield market stays shut for the segment, smaller neoclouds consolidate or fail, and the AI buildout continues at a slower, more expensive pace. In the upside case, AI demand accelerates, utilization stays above 90%, CoreWeave refinances cheaply, spreads compress, and the "irrational market" read is vindicated. In the downside case, a hyperscaler customer delays or walks from a commitment, the junk market closes for the segment, and CoreWeave is forced to dilute equity or restructure.
The falsifying signal is specific. If CoreWeave prices a new investment-grade, GPU-collateralized facility at an all-in cost below 9% within the next two quarters, the financing-squeeze thesis is wrong. If instead the 2032 bonds trade above a 15% yield for a month and any planned high-yield deal is pulled or priced above 12%, the squeeze is confirmed.
The AI boom was never likely to be stopped by a lack of demand. If it stops, it will stop at the price of the money that builds it.
Explore more exclusive insights at nextfin.ai.
