Neoclouds (CoreWeave, Nebius, Iren and hundreds of lesser-known ones) buy Nvidia chips with borrowed money and rent out computing power to AI companies. Elon Musk said that over the past few weeks, debt markets have almost completely cut off funding to anyone who doesn’t earn money outside AI data centers, and he promised neoclouds a “mass die-off.” He stressed, however, that this doesn’t apply to SpaceX:
“We closed a $40B credit facility to buy GPUs, but we have an investment-grade rating and revenue from Starlink and launches,” he wrote.
So far, we only have Musk’s word that the deal has closed. Earlier this month, the FT reported on talks expected to wrap up in 2027, but on Friday, according to Bloomberg, the deal had stalled amid turmoil in the credit market.
Musk isn’t alone in his forecast. Gartner, for one, expects more than 80% of independent neoclouds to cease to exist as standalone companies by 2030.
On Monday morning, in premarket trading, shares of Nebius, CoreWeave and Iren were already down more than 4%. Still, the market had been on edge before that: earlier in the month, insurance against a possible CoreWeave bankruptcy became so expensive that investors read it as a distress signal.
The Problem Isn’t Demand, It’s the Cost of Money
A few days ago, Nvidia CEO Jensen Huang said demand for neoclouds is “off the charts,” but the companies are short on working capital.
Neoclouds borrow to buy GPUs, expecting to recoup the cost by renting out computing power. Companies like these die the day they have to refinance. The GPUs pledged as collateral lose value faster than the loans against them are paid down, since if a site isn’t ready on time because of a power shortage or delivery delays, interest starts accruing right away anyway.
Overall, the industry is gradually turning into a very expensive infrastructure business. The economics of these projects will likely need a serious rethink: even with huge demand, a startup may not survive unless it can optimize its borrowing costs and bring capacity online on schedule.
On top of that, infrastructure that isn’t ready for today’s demand could leave a huge number of data centers unfinished, either because there simply won’t be enough electricity or because deliveries of those same graphics cards get delayed. In effect, we’re watching this market take shape in real time.
SpaceX Looks Like a Neocloud Too
Investor James Chanos calls SpaceX the world’s largest neocloud: in 2025, its capital expenditures came to $20.9B against operating cash flow of $6.8B. It rents out computing power, including to Anthropic and Google.
In other words, SpaceX is also building extremely expensive infrastructure and needs financing. The difference is that it has other sources of revenue. Rockets and Starlink now serve as a credit history for chips: according to Musk, the investment-grade rating and revenue from those businesses let SpaceX raise money even when others in the market can’t.
The situation looks a lot like a carve-up of the market for computing power. Musk didn’t raise that $40B for no reason. We suspect he did more than predict bankruptcies and takeovers among neoclouds: he has decided to lead the race himself.
Bottom line: AI is turning into capital-heavy infrastructure, where the winner isn’t the best idea but the cheapest money. When assets are sold off, they get snapped up by whoever has the lowest cost of capital. Instead of healthy competition, we get a club of companies that can afford debt, and that club will buy up the capacity of smaller players who simply get knocked out of the market.