Neoclouds Explained: CoreWeave's Business Works, Despite Risky Borrowing

Neoclouds like CoreWeave, Nebius and IREN have real customers and real demand, and I think the business works. What could break it is the price of the money they borrow to buy the chips.

Goldman Sachs upgraded CoreWeave to Buy on Monday with a $106 target, and on Tuesday HSBC answered with a $39 target, about a third below Monday's close. $CRWV dipped to a new 52-week low of $57.24 on Tuesday before closing at $62.13, after losing about a quarter of its value the week before. IREN fell 4% to a new low of its own the same day, while Nebius edged up.

These three are the biggest public neoclouds, companies that buy Nvidia GPUs and rent the computing power to AI labs and big tech. Nobody on either side of last week's fight thinks demand dried up, so I'd read the selloff as a fight about borrowing costs. CoreWeave can borrow at under 6% when Meta is the customer, while its 2031 bonds yielded almost 16% on Tuesday. The whole business sits in the gap between those two numbers.

How a neocloud makes money

CoreWeave's chief executive, Michael Intrator, explained it to Jason Calacanis on the All-In podcast in March with something he calls "the box." CoreWeave signs a customer (his example was Microsoft), buys the GPUs from Nvidia and leases space in a data center. All three go into a separate company, and the customer pays that company directly.

The money then pays out in a fixed order. Co-founder Brian Venturo described it on the Big Technology Podcast in January: the power bill and the data center get paid first, then the lenders' interest, then the loan itself, and whatever is left goes to CoreWeave. Intrator says a five-year deal pays for everything inside two and a half years, though that's his figure from an interview and no filing breaks it out.

Customers sign up for years at a time and owe the money whether they use the chips or not. That's why 98% of CoreWeave's revenue comes from committed contracts, and why it had $104.2 billion of signed work waiting at the end of June. Meta alone signed a $21 billion deal running through 2032.

Why I think the model works

Lenders have spent three years learning to trust the box, and you can see it in what they charge. CoreWeave's first GPU loan in 2023 cost around 15% a year. By March 2026, a box backed by Meta's contract got an A3 investment-grade rating from Moody's, and Intrator said on the Q1 call it priced "at a level implying a cost of less than 6%." The lender is really betting on Meta paying its bills.

Demand isn't the problem either. GPU rental prices are rising, with an H100 renting for $2.98 an hour on Oct. 19, and CoreWeave has been signing short contracts at about $40 million per MW. Goldman's Monday upgrade made the same case, arguing the market overprices CoreWeave's future borrowing costs and underprices what the borrowing buys.

The companies on the other side of these contracts do well regardless. Nvidia books the GPU sale when it ships, and Microsoft and Meta lock in computing power at fixed prices without putting the hardware on their own balance sheets.

What could break it

The risk is CoreWeave's own borrowing. Its 9.625% bonds due 2032 touched a 15% yield on Oct. 13, Chamath posted the chart and Elon Musk replied "Yikes." CoreWeave still pays the 9.6% it agreed to on those bonds, but every new bond it sells now has to compete with a 15% yield on the old ones. By Tuesday its 9.75% notes due 2031 yielded 15.97%, though Janus Henderson's John Lloyd told CNBC he's keeping them in his fund.

That matters because CoreWeave has to borrow a lot more. On All-In last week, Chamath Palihapitiya said his team figures CoreWeave needs about $35 billion more to fund contracts it already signed, roughly as much as all the debt it had in June. The contracts are fixed price, so the money coming into each box can't rise to cover a higher loan rate. If new money costs 10% to 15% instead of today's average of about 8.5%, he puts CoreWeave within about $5 billion of break-even. He still said he doesn't think CoreWeave is in real danger, and he called it a confidence game.

HSBC put the same worry into a single number on Tuesday. Analyst Abhishek Shukla expects CoreWeave to earn an average 8.9% on its invested capital from 2026 to 2030, which he wrote is "below its cost of borrowing," and he sees its net debt nearly doubling to $92 billion by the end of 2030.

The boxes also protect CoreWeave less than the pitch suggests. Only the Meta box is fully walled off from the parent company. CoreWeave's 10-Q says it guarantees every other GPU loan itself, so a bad box lands on CoreWeave, and $31.4 billion of its $35.1 billion of June debt can reach the parent. Three unnamed customers brought in about 72% of its revenue last quarter, and interest expense hit $640 million, more than the $626 million net loss.

David Sacks pushed back on the same episode. The box lenders hold a strong customer's contract and the chips themselves, he said, so the loans are fine even if the unsecured bonds aren't. His measure is payback: as long as a box earns back its cost in half the contract, it's hard to lose money, and he'd start worrying if that stretches out. Ben Horowitz, sitting in, called neoclouds "Rube Goldberg financial machines" and said a16z owns none of them, though he agreed GPU prices haven't fallen the way his old company's bandwidth costs did in 2001.

Two of Chamath's numbers were off. He said a CoreWeave GPU loan had "spiked to 15%," but that loan cost about 15% back in 2023, and the 15% this month was the unsecured bond. He also cited about 953 basis points for insuring CoreWeave's debt against default, and that price actually hit a record of 1,194.5 on Oct. 15.

CoreWeave, Nebius and IREN aren't equally exposed

How much each neocloud borrows decides how much a rate scare can hurt it.

$NBIS makes its customers pay more of the cost up front. Nebius has more than $40 billion of customer commitments, mostly from Microsoft and Meta, and expects more than $9 billion of prepayments this year, cash for GPUs it never has to borrow. Its first secured loan, in July, was $775 million, while CoreWeave raised more than 45 times that this year.

$IREN sits in the middle. The former Bitcoin miner owns its land and power in Texas and British Columbia, which most neoclouds don't. On its $9.7 billion Microsoft contract, a 20% prepayment plus a 6% loan covers 96% of the GPU cost. Its smaller customers' GPUs are financed at 9%, though, and its total debt reached $7.84 billion, about double what it was a quarter earlier. On Wednesday it announced a planned 300 MW campus in Badajoz, Spain, aiming for first operations in 2028, without saying how it will pay for it.

$CRWV borrows the most and guarantees nearly all of it, and its bonds are the ones trading near 16%. The smaller and private neoclouds are struggling too: Microsoft backed out of Nscale's planned site in England because grid power there won't arrive until the 2030s, and Korea's Elice closed down 32% on its first day of trading on Tuesday.

What to watch

The clearest signal is the price of CoreWeave's next big loan or bond. If it comes in near the Meta box's rate, Sacks was right; if it prices off the 15% bond, Chamath was.

The view

I think neoclouds will mostly work out, because the customers are real and GPU rental prices are still going up. The danger is that the cheap money only comes with Meta- or Microsoft-grade customers, and CoreWeave still has to fund a lot of boxes on its own name at today's rates. Nebius, which gets its customers to pay up front, is the safest of the three. CoreWeave is a bet that its borrowing costs come back down before it needs that next $35 billion.

Goldman thinks they will and HSBC thinks they won't. The bond market, at almost 16%, is siding with HSBC for now.

Tags: neoclouds, pillar, CRWV, NBIS, IREN, debt, gpu