Neocloud Selloff: CoreWeave's Bonds Hit 15% While Its Meta-Backed Box Borrows at 6%
CoreWeave, Nebius and IREN had their worst week in nearly a year after a CoreWeave bond touched a 15% yield. Inside Michael Intrator's "box," the best CoreWeave loan costs about 6%. The neocloud trade now turns on which rate wins.
CoreWeave's bonds hit a 15% yield and the neoclouds had their worst week in nearly a year. Inside Michael Intrator's "box," the best CoreWeave loan still costs about 6%. Which rate wins decides the neocloud trade.
Updated Oct. 18, 2026. This piece first ran on Sept. 15 after the Amodei slowdown selloff; it now covers the CoreWeave bond rout of Oct. 13 to 16.
Chamath Palihapitiya posted a chart of one CoreWeave bond on Tuesday night, Elon Musk replied "Yikes," and the neoclouds spent the rest of the week falling. The bond was CoreWeave's 9.625% note due 2032, and its yield touched 15% on Oct. 13. CoreWeave ($CRWV) closed Friday at $61.39, down about 25% on the week, which Reuters called its worst week in more than 11 months. Nebius ($NBIS) finished at $165.47 and IREN ($IREN) at $26.99, a one-year low, after each lost about 12% on Friday alone.
A neocloud is a company that buys Nvidia GPUs, puts them in powered data centers and rents the computing to AI labs and big tech, mostly on multi-year contracts. All three run the same machine. They differ in what they own and in what their money costs, and this week the bond market started charging for that difference.
How Michael Intrator's box works
CoreWeave's chief executive explained the machine to Jason Calacanis on the All-In podcast in March, in an episode recorded at Nvidia's GTC conference. He calls it "the box." CoreWeave signs a customer (his example was Microsoft), buys the GPUs, signs a data center lease, and puts all three into a separate company. The customer pays that company, not CoreWeave.
The money then falls through a waterfall. Co-founder Brian Venturo laid out the order on the Big Technology Podcast in January: power and the data center first, then interest to the lenders, then principal, and whatever is left goes up to CoreWeave. Intrator's claim on All-In was that "within two and a half years of a five-year deal, we have paid for everything." (That's his number from an interview, and no filing breaks it out.)
The contracts carry the revenue. Committed contracts made up 98% of CoreWeave's revenue in the second quarter, and the backlog was $103.7 billion at June 30, before another $25 billion of commitments signed in early Q3. Short three-to-six-month deals signed since June went for about $40 million per MW, so demand was the one thing nobody argued about this week.
CoreWeave borrows at two very different prices
Lenders loved the box at first, and you can read it in the loan pricing. CoreWeave's first GPU loan in 2023 cost around 15% a year. By March 2026, an $8.5 billion box backed by Meta's contract was rated A3 by Moody's, and Intrator said on the Q1 2026 call it priced "at a level implying a cost of less than 6%." That loan is non-recourse, so if the Meta box failed, its lenders couldn't come after CoreWeave itself.
The other boxes don't work that way. The 10-Q says every delayed-draw term loan except the Meta one is "unconditionally guaranteed" by CoreWeave, and $31.4 billion of the company's $35.1 billion of June debt is recourse. Pricing also turned before the bond scare did: the $2.6 billion loan CoreWeave closed in August priced at SOFR plus 5.5 points, wider than it was marketed, against customer contracts averaging about three years.
The 15% is what a buyer earns on that 2032 bond after it fell to around 79.5 cents on the dollar on Oct. 14, so CoreWeave still pays only its 9.625% coupon on the bonds it already sold. But the next bond prices off that 15%, and the cost of insuring CoreWeave's debt against default hit a record 1,194.5 basis points on Oct. 15.
What the All-In podcast got right, and the number it got wrong
The All-In hosts spent half an hour on CoreWeave in the episode published Oct. 16, with Ben Horowitz sitting in. Chamath's bear case runs through the box. The contracts are fixed price, so the cash flowing into each box can't rise. By his team's math, CoreWeave still has to borrow about $35 billion more to fund contracts it already signed, at an average cost of about 8.5% today. If that new money costs 10% to 15%, he puts CoreWeave within about $5 billion of break-even, and he called the cycle of credit insurance, bond yields and borrowing costs feeding on each other "the negative reflexive loop." He also said, "I don't think that they're in real risk" (while reading the insurance market as pricing roughly coin-flip odds of default over five years).
Two of his numbers need fixing. He said a GPU loan, DDTL 1.0, had "spiked to 15%," but that loan priced near 15% back in 2023 and has no public trading price. The bond that hit 15% this week was the unsecured 2032 note. He also cited credit insurance at about 953 basis points, and the record print came in higher.
David Sacks argued the other side. The box lenders hold an investment-grade tenant's contract plus the GPUs as collateral, he said, and what repriced was CoreWeave's unsecured paper, which "does not increase the cost that CoreWeave is paying" on bonds already sold. His north-star metric is payback period, and a box that earns back its cost in half the contract is hard to lose money on. Horowitz said a16z owns none of the neoclouds because his old company, Loudcloud, once borrowed to hold bandwidth that fell from $1,500 to $50 a megabit within two years (his figures). He also said GPU rental prices haven't followed token prices down, and this month's rental indexes back him up.
The macro tape helped Chamath. September inflation printed 3.7% on Oct. 14, with the 10-year Treasury near 5.2%, just off its Oct. 5 high of 5.31%, and the Fed debating a hike. Gartner picked the same morning to predict that more than 80% of neocloud providers won't operate independently by 2030. Jensen Huang defended CoreWeave from Stockholm the next day, saying its cost of capital is "actually decreasing," and the stock bounced 5.9% before giving it back Friday. Bloomberg's tally of $3.3 billion in founder stock sales since the IPO didn't help (the company says every sale was pre-scheduled).
Nebius and IREN built different boxes
Nebius makes its customers carry more of the build. Prepayments covered about 50% to 60% of capital spending on its four newest core contracts, per the Q2 call, and it funds the rest with equity and convertibles, including a $5.75 billion convert that closed in August. Its first GPU-backed loan, $775 million in July, priced at SOFR plus 2.5 points. Nebius also raised on-demand GPU prices 17% to 21% on Oct. 1, its second increase in about three months. Morningstar called it the least levered of the three, and it still fell 12% on Friday (correlation doesn't read balance sheets).
IREN owns the part CoreWeave rents. The former Bitcoin miner holds land and grid connections at Childress and Sweetwater in Texas and in British Columbia. On its $9.7 billion Microsoft contract, a 20% prepayment plus a $3.6 billion investment-grade loan at 6% funds 96% of the GPU cost, and Microsoft accepted the first phase in August. The rest of IREN's customers are smaller, and their GPUs sit on a $2.4 billion facility at a fixed 9%. Morningstar cut its fair value on IREN from $43 to $37 on Friday, citing financing risk.
Which rate wins
The selloff priced CoreWeave's unsecured layer, and Sacks is right that the coupons on existing bonds don't move. But the boxes are only as walled off as their guarantees, and CoreWeave guarantees all of them except Meta's. Every new box that lacks an investment-grade tenant now prices off a company whose bonds yield 15%, while the Fed is debating a hike. Co-founder Brannin McBee told the FT he wants more of the debt rated investment grade so insurers will lend, and that is the right fix, because it turns more boxes into the Meta box.
The lenders and the tenants come out ahead either way. Bond buyers now earn about 15% on CoreWeave paper issued at 9.6%, and Meta and Microsoft hold fixed-price capacity while CoreWeave sells new short contracts at $40 million per MW. My view: Nebius, with customers funding its builds, is the neocloud the credit market should worry about least. CoreWeave is a rate trade until more of its boxes look like the Meta one, and IREN owns its power but rents its money.
Inside the box, CoreWeave's best money costs about 6%. Outside it, the market wants 15%.