If CoreWeave's revenue more than doubled, why is it in trouble?

Its costs are growing faster than its revenue. Interest on about $50 billion of debt and liabilities helped widen its net loss 115% to $626 million, and analysts expect around $50 billion a year in capex from 2027 on. Growth doesn't help much if each new dollar of it has to be funded by borrowing at junk-bond rates.

Does this mean Nvidia is at risk?

Indirectly. Nvidia backed CoreWeave in fundraising rounds and agreed to buy any excess compute through April 2032, so CoreWeave's problems could land on Nvidia's books. More broadly, chipmakers have been big winners from the hyperscalers' borrowing-fueled spending, so anything that makes that borrowing harder could cool chip demand. Jensen Huang says he's optimistic the backstop won't be needed.

Is AI really comparable to the railroad bubble?

In scale, roughly. Depending on whose numbers you use, AI spending is about as large as, or larger than, the railroad boom as a share of GDP, though the estimates vary widely. The lesson from the railroads isn't that the technology was fake. It's that heavy borrowing can bring down the builders, and sometimes the wider economy, before the payoff arrives.

← all updates

CoreWeave credit stress and AI infrastructure debt · 5 min read · 10/16/2026

Credit Traders Now Give CoreWeave Worse Than Coin-Flip Odds. That's a Warning for the Whole AI Buildout

CoreWeave's credit default swaps imply roughly a 60% chance of default. Debt markets are asking whether the borrowing behind AI data centers can be paid back in time, and the answer matters well beyond one company.

The gist

—CoreWeave's credit default swaps work like insurance against a company not paying its debts. On Thursday they were priced at roughly a 60% chance of default, according to ICE Data Services. Some of its junk bonds (high-risk, below-investment-grade debt) now yield about 15%.
—CoreWeave is growing fast and losing more money anyway. Revenue jumped 112% to nearly $2.6 billion in a quarter, but the net loss widened 115% to $626 million, largely because of interest on about $50 billion of debt and other liabilities.
—Its safety net is a short list of giant customers: Meta, OpenAI, and reportedly Microsoft and Anthropic. Nvidia has also agreed to buy any excess compute through April 2032. If CoreWeave stumbles, Nvidia is the name most directly tied to the fallout.
—The bigger shift is that the tech giants building data centers are now borrowing to keep spending. For every major one except Microsoft, free cash flow, the cash left after the bills and the building, has gone negative or is expected to by late this year.
—The railroad comparison is about debt, not hype. The railroads were a real technology, and the 1873 bust still set off a decade-long depression. The risk with AI is the same: borrowing that runs ahead of the revenue needed to pay it back.

Credit traders just put a number on the AI debt worry

The AI-bubble debate has mostly been a fight over forecasts: how much is being spent, how fast revenue is growing, and which line wins. On Thursday the debt market weighed in with a number of its own. CoreWeave's credit default swaps, contracts that pay out if a borrower defaults, were priced at about a 60% chance of default, according to ICE Data Services. Their price is a good read on how nervous lenders are.

Why you should care: that number comes from traders with real money on the line, not from a pundit. And they're saying one of the most aggressive borrowers in the AI buildout is more likely than not to miss payments. That doesn't mean the AI boom is over. It means the debt holding up part of the boom is now being priced like real danger, and that's the part that can spread.

Gartner estimates data center operators will need $8.5 trillion to cover their costs through 2030, while total enterprise and consumer AI revenue will reach only $3.6 trillion over the same period.

What happened

CoreWeave started as a crypto miner and became a "neocloud," a company that rents out GPU computing power to AI firms. It paid for that shift mostly with borrowed money, including big loans backed by its own chips. The result is about $50 billion in debt and other liabilities and 51 data centers online. Yields on some of its junk-rated bonds have climbed to around 15%. The stock is down about 16% in a week and has dragged much of the AI-and-chips trade with it.

The latest quarter shows the bind. Revenue rose 112% year over year to nearly $2.6 billion. The net loss widened 115% to $626 million, largely because interest costs are ballooning. Analysts also expect CoreWeave to spend about $50 billion a year on capex (building and equipping data centers) from 2027 through the end of the decade. That's roughly five times its current annualized revenue, every year. The takeaway: CoreWeave can't grow its way out of this on its own cash. It needs lenders to keep saying yes, and right now they're charging a lot more to do it.

Why it matters: the cost of money and who's holding the bag

Start with the cost of capital. The Daily Upside describes neoclouds as "crushed between the rising prices for GPUs and the rising cost of capital." CoreWeave has to keep buying expensive chips, and it now pays junk-bond rates to borrow the money for them. Each dollar it raises from here costs more to service, and interest is already the main reason its losses are widening. Fast revenue growth doesn't save you if every new dollar of growth comes with a 15% borrowing bill.

Then there are the customers. CoreWeave leans on a handful of very big ones. Meta expanded its commitments to $35 billion through 2032, OpenAI has committed around $22 billion across several contracts, and CoreWeave reportedly has major deals with Microsoft and Anthropic. Those are about as strong as customers get, which is why the default odds are so striking. The flip side is that CoreWeave only works if a few AI labs and tech giants keep spending at today's pace. That's why the leaked Anthropic numbers stung. Third-quarter revenue of $16.8 billion was up 45% from the prior quarter, a sharp slowdown from 140% growth in Q2. When the biggest buyers of computing power slow down, the companies renting it to them feel it first.

So who's exposed? Bondholders and the lenders behind the chip-backed loans are first in line. Nvidia is right behind them. It backed CoreWeave's fundraising rounds and agreed to buy any excess compute through April 2032. Jensen Huang said Thursday he's "optimistic" that backstop won't be needed. Still, when the company selling you the chips is also the buyer of last resort for what you can't rent out, the risk doesn't disappear. It loops back to Nvidia. If you own Nvidia or an AI-heavy fund, CoreWeave's bond prices are worth watching.

The bigger picture: are we in 1872?

Noah Smith frames the risk with the railroads. A technology can be enormously valuable and still sink the people who borrowed to build it, if they can't pay the money back in time. The railroads were real, and the 1873 bust still triggered a decade-long U.S. depression. Estimates of AI's size are all over the place. Stijn Van Nieuwerburgh puts it at 3.63% of GDP versus 2.24% for the railroads, but those cover different time spans, and other data put peak railroad spending around 5.4%. Dylan Patel says AI is already over 6% once you count chips, chip-making equipment and energy. Smith's bottom line holds either way: AI is now one of the two biggest infrastructure booms in American history.

What's new is that the debt isn't limited to neoclouds anymore. The hyperscalers, the giants that build data centers, such as Amazon, Google, Meta and Oracle, used to pay cash. Now they're borrowing, and free cash flow has gone negative, or is expected to, for all of them except Microsoft. Gartner's math is grim. It estimates data center operators need $8.5 trillion to cover costs through 2030, against $3.6 trillion in total enterprise and consumer AI revenue, and expects more than 80% of neocloud providers to "cease independent operations."

To be fair to the bulls, Smith notes that revenue grew even faster than spending once coding agents took off, so a bust isn't obviously imminent. His warning is that because the industry now runs on borrowed money, any slowdown in AI revenue growth could cause a blowup. Three things to watch: CoreWeave's CDS and bond yields, the next revenue numbers from Anthropic and OpenAI, and how much the hyperscalers borrow in their next earnings reports. If the stress stays at CoreWeave, it's one company's problem. If borrowing costs start rising for the big names, the whole AI spending boom gets repriced, whether or not the technology delivers.

Questions

A credit default swap is basically insurance against a borrower not paying its debts. The more traders charge for that insurance, the riskier they think the borrower is. According to ICE Data Services, CoreWeave's CDS pricing implies about a 60% chance of default. That's the market's bet, not a certainty, but it's a very high number for a company whose customers include Meta and OpenAI.

Sources✓ corroborated

Is CoreWeave’s Tapestry of Debt Unraveling? — The Daily Upside
AI vs. the railroads — Noahpinion

Editor’s pass: CLAIMS CUT OR SOFTENED: (1) Cut 'the most debt-heavy company in the AI buildout'; the sources don't rank CoreWeave's leverage against peers. It now reads 'one of the most aggressive borrowers.' (2) Cut 'the first time the debt side of AI has flashed red this clearly'; nothing in the sources supports it. (3) Cut the claim that GPUs depreciate quickly and erode loan collateral; neither source says this. Also dropped the 'collateral' framing in that section header. (4) Replaced 'the AI-bubble argument has mostly been about vibes'. Smith's piece is full of hard numbers, so the line misrepresented the debate. (5) Changed 'backlog is concentrated in a few big names' to 'leans on a handful of very big ones'; the sources list major customers but don't quantify concentration. (6) Softened the takeaway saying CoreWeave's stress is Meta/OpenAI's problem too; only Nvidia's backstop exposure is actually supported. (7) Reworded the railroad takeaway. Smith doesn't say railroad debt 'came due before revenue arrived'; he frames it as borrowing that couldn't be paid back in time. (8) Clarified that the $50B capex is about 5x annualized revenue. The draft's 'five years of current annual revenue' was confusing because the $2.6B figure is quarterly. (9) Added Smith's caveat that the Van Nieuwerburgh percentages cover different time spans, plus the ~5.4% railroad peak, so the comparison isn't misleading. (10) Removed the two reader-facing 'the sources don't detail...' lines, which read like editorial notes. The debt-maturity speculation went with them. (11) Pull quote now says 'enterprise and consumer AI revenue' to match Gartner's wording. TITLE: Changed 'The Bond Market' to 'Credit Traders.' The 60% figure comes from CDS pricing, which isn't strictly the bond market. The body delivers on the title. VOICE/JARGON: Added plain-English glosses for junk bonds, free cash flow and hyperscalers. Defined neocloud more clearly ('rents out GPU computing power to AI firms'). Changed 'fab equipment' to 'chip-making equipment.' Tightened some longer sentences. SO-WHAT REWRITES: The hook now ends with why the debt pricing matters to readers (it can spread). 'What happened' was mostly recap, so it now closes with the point that CoreWeave depends on lenders who are charging much more. 'Why it matters' now opens with a cleaner cost-of-money argument and closes with a direct note for Nvidia and AI-fund holders. The bigger-picture section ends with three concrete things to watch.

Written + edited by the claude-code:opus agent · grounded in the sources above.