The Yikes Heard Around The World

Chamath Palihapitiya posts a chart. The yield on one of CoreWeave's bonds touched 15%, then closed at a new high. Elon Musk replies with one word: "Yikes."

Then CoreWeave stock falls about 15% in a day. Nebius and IREN, which do the same work, fall hard too. It sounds like CoreWeave is suddenly paying 15% interest.

That never happened. CoreWeave still pays the 9.625% a year it locked in on that bond. That bill did not go up by a dollar. The All-In podcast spent close to half an hour on this in one episode, and the hosts did not agree. Here is what they were fighting about.

Rooms full of chips

GPUs are chips that do thousands of small math problems at once. AI chatbots run on them. CoreWeave is a neocloud, a company that mostly just rents GPUs to AI companies on contracts that run about four to five years.

Those chips cost a fortune, and there are two ways to pay. You can sell slices of the company, which is called equity. Or you can borrow, which is called debt, and you pay that back with interest no matter how business goes.

Jason Calacanis hosts All-In with Chamath and David Sacks. He said CoreWeave, IREN and Nebius were down between 10 and 15% that Wednesday, and Nvidia was down 3%. "CoreWeave is the tip of the spear. They've raised over 35 billion." Counting investors and lenders, CoreWeave said it raised more than $30 billion in 2026, and by the middle of the year it owed about 35.5 billion. It only went public in 2025.

Inside the box

Michael Intrator, CoreWeave's co-founder and CEO, says the way the loans are built makes it safer. "I take my contract with Microsoft and I put it in the box. I go to Jensen and I buy the GPUs. I put it in the box. I take my data center contract. I put it in the box. And now the box governs cash flow."

Bankers call the box an SPV, a special purpose vehicle. It's a separate little company with its own bank account, like a company set up to own one vending machine. Money pours down like a waterfall and fills a row of buckets in order. Say Microsoft pays the box $100. Maybe 20 goes to the data center bucket and 10 to power. The lenders take their interest plus a piece of the loan itself, called principal. Whatever splashes past the last bucket goes back to CoreWeave.

These loans are asset-backed, which means the GPUs are the collateral, like a car loan. Many are delayed draw term loans, pulled out in pieces as CoreWeave builds. Then comes the promise: "within 2 and 1/2 years of a 5-year deal, we have paid for everything." That's CoreWeave's own claim from when it went public. It counts money customers pay upfront, and it counts the money before paying interest. The plan works as long as customers keep paying and borrowing stays cheap.

The seesaw

The parent company also sold bonds backed only by its promise to pay. That is where the 15% comes from.

A bond is an IOU you can sell. You lend a company $1,000. It pays you interest every year, called the coupon, and hands back the $1,000 at maturity. Price and yield sit on a seesaw. Say the bond pays $100 a year. Buy it for $1,000 and you earn 10%. Then people get nervous and the price drops to $800. That same $100 is now 12% of what the new buyer paid, and they still get the full $1,000 back at the end, which is a $200 bonus. Count the bonus with about six years left and you get roughly 15% a year. That number is the yield to maturity. Price down, yield up.

CoreWeave's bond pays 9.625% and comes due in 2032. During the sell-off it traded below 80 cents on the dollar, which pushed its yield to about 15%.

CoreWeave sits below investment grade, in what Wall Street calls junk. Yields get measured against the 10-year Treasury, which is the risk-free rate. The extra on top is the spread. It's counted in basis points, and one basis point is a hundredth of a percent. If the Treasury pays 5% and the bond pays 15%, the spread is 1,000 basis points.

A coin flip on insurance

Chamath says the loudest warning came from a different market, where the traders he calls the sharps buy protection. The smart money, he said, is trading what's called a credit default swap, or CDS. It's insurance against a company not paying its debts, like fire insurance on a house you don't own. The more people think the house might burn, the more the insurance costs.

He said the price "spiked this week to about 953 basis points." By his math that's about a 16% chance of default in any given year and about 55% over 5 years. "So, it's a coin flip whether they default or not. That's not what I'm saying. That's what the insurance market was saying about their debt."

The math goes like this. It costs about $9.50 a year to insure every $100 of debt. Traders usually assume lenders get back about 40 cents on the dollar in a default, so the insurance covers the other 60. Nine and a half divided by 60 is about 16% a year. Stack that over 5 years and you land around 55 to 60%. Bloomberg also reported a record price for this insurance, implying better than a 60% chance of default. That's what the price says. Prices can be wrong.

Why now

Chamath starts with the 10-year Treasury, the tide under every boat. On his chart, over a few months, the extra the riskiest junk bonds pay over Treasuries grew about three points, and CoreWeave's grew about five. During the sell-off, though, the 10-year and junk bonds barely moved. The jump was mostly about CoreWeave.

Then he looks inside the box: "It turns out that these contracts are fixed price." Think of a gym membership locked at $50 a month for 5 years. If the drop-in price doubles to $100, the gym still gets $50 from you. That drop-in price is the spot price, and the hosts say spot has climbed since CoreWeave locked in.

Chamath's firm estimates CoreWeave still has to borrow between 26 and 41 billion, about 35 in the middle, to build what it has signed. It also has to refinance, which means replacing old loans with new ones as they come due. CoreWeave says its average rate is 8.3%. A store that borrows at 8% to stock shelves earning 12% keeps 4 cents on every dollar. If it borrows at 15%, it loses three. That gap is the financing spread, and this whole fight is about it. Chamath adds that about 3/4 of revenue comes from three customers. CoreWeave's own numbers put the top three at about 72%.

He says the problems feed each other. Rising borrowing costs, in his words, "erodes the margin that they have that then brings them close to this break even point that causes the CDS to react which then causes the bond yields to react. That's the negative reflexive loop that they're in." Reflexive means the fear helps cause what it fears, the way a bank run does. Still: "I don't think that they're in real risk but as we all know this is sort of a confidence game."

One thing I'd flag: the order may have run the other way. The bond hit 15% first, and the insurance price set its record after that.

Chamath said his team at Social Capital estimates CoreWeave is within about $5 billion of break even, the point where money coming in just covers what it owes. He said that's also how close it is to breaking a covenant, which is a rule in a loan, like keeping a minimum amount of cash. In its filings, CoreWeave said it was keeping every rule on its box loans.

Rube Goldberg machines

Ben Horowitz, filling in for David Friedberg, co-founded Andreessen Horowitz and before that ran Loudcloud, an early company renting computing online. His view: "They're Rube Goldberg financial machines. They're exceptionally complex." Every step has to go right.

One of those steps is how long a GPU stays useful. Spreading a machine's cost over its working years is called depreciation. CoreWeave spreads its GPUs over 6 years, and Jason says people now talk about 6 to 8. That means the chips could keep earning after a 5-year contract ends.

Ben's firm hasn't invested in any neocloud. When Chamath asked why, Ben went back to 1999, when Loudcloud had to buy bandwidth on long-term contracts. It cost $1,500 a megabit then. By 2001 it was $50, a drop of about 97% in 2 years as Ben remembers it. If you borrow to buy something whose price collapses like that, you still owe the whole loan. That's the fear with GPUs. Ben's counter is that the dot-com bubble had too few people online to use all those lines, while AI has too few chips. He says AI answers keep getting cheaper, but renting a GPU hasn't.

Can CoreWeave go bankrupt?

Jason pulled up a Polymarket market on whether CoreWeave will announce bankruptcy. It showed 15% by the end of 2027 and 23% by the end of 2028. Those were bets at taping, in a small market of about a quarter of a million.

One of the hosts: "this is dumb because I think Michael explained it, you have this opco propco setup. So CoreWeave can't go bankrupt." The opco is the operating company, and the propco is the company that owns the stuff. In that view each box is a propco. If one fails, it goes into receivership, where a court-appointed caretaker can sell its GPUs to repay its lenders.

That view leaves out two things. First, the parent owes its own unsecured bonds, which get paid mostly from whatever spills out of the waterfalls. Second, CoreWeave's filings show the parent promised to pay most box loans if a box can't, so only about $1 in 10 of its debt is truly fenced off. The parent could still go bankrupt if it fell behind on those.

Sacks says the selling hit unsecured debt, a handshake IOU backed only by a promise. "The repricing is of CoreWeave's unsecured obligations. It's true that the market has discounted them by 20%, they're in distress territory. That has brought the implied coupon of those bonds from 9.6% to around 15%. But that does not increase the cost that CoreWeave is paying."

Jason cut in that it changes the cost of future debt, and both of them have a point. A billion dollars costs about $96 million a year in interest at the old rate and about $150 million at roughly 15%. CoreWeave also sold about $14 billion of unsecured bonds in 2026, counting ones that can turn into stock.

Jason also points out that the big customers are investment grade. Microsoft has the top rating there is. But OpenAI has no rating from the major agencies, and CoreWeave said one of its loans backs work for two large customers that aren't investment grade. Sacks says Nvidia chips several generations old still rent at premium prices, and that power and permits are too hard to get for new supply to flood in.

The Achilles heel

Chamath's last warning is that even if every box works, there's one weak spot. In his scenario the 10-year climbs to 5 1/2 or 6% for reasons outside AI, like government debt. At taping it was about 5.1%. "Just like the great warriors, you just slice the Achilles tendon and they just crumble to the knees." The problem is "the fact that you need to finance these things in a variable rate model as a spread to the risk-free rate." He says Microsoft, Google, Meta and Nvidia need to backstop all of it in a worst case, like co-signers who pay if the borrower can't. CoreWeave's box loans float with a short-term rate, but some require it to lock in most of that rate with side deals called hedges.

Sacks pointed to a different number: the payback period, meaning how long a machine takes to pay for itself. "To me, that's like the North Star metric here." CoreWeave says 2.5 years on a typical 5-year deal. He says SpaceX pays back in roughly one year, at about 50 billion per gigawatt versus CoreWeave's 20. His reason is that SpaceX's deals can be cancelled by either side within 90 days, so it basically gets spot. Those per gigawatt numbers are Sacks's estimates. On SpaceX's August earnings call, its finance chief, Bret Johnsen, said its AI compute is getting less than a one year payback.

Chamath came back to the insurance price: "I think the sharps may be wrong, and they usually are wrong when they're buying these credit defaults, but you know, when your expected value is a coin flip in a 5-year window, that's like scary town." His fix is to cancel some contracts and sell at spot, even if that means paying liquidated damages, a penalty agreed in advance, like the fee for ending a phone plan early. He figures CoreWeave has built about half of what it signed up for. Its own count is lower, closer to a third: about 1.5 gigawatts running out of about 4.2 signed.

What I'm watching

So what is actually going on with CoreWeave? The honest read is that the market now charges CoreWeave much more to borrow on its promise alone. A company that grows on borrowed money feels that every time it borrows again. But the boxes it already built are backed by big customers' contracts and the GPUs inside. Chamath worries about the cost of money and the fear loop. Sacks and Ben point to demand that still outruns supply.

CoreWeave is a bet that demand stays strong longer than the debt takes to pay off. Four things will show who's right. First, the payback period. Second, what CoreWeave pays on its next loan, with the 10-year Treasury under it. Third, how much capacity sells at spot. Fourth, whether the boxes keep paying their lenders, because as long as they do, the waterfall works.

Elon's "Yikes" was about a bond price. CoreWeave's bill on that bond is still 9.625%. The one to watch is the next bill.