The AI data center landlords are broke with order books stuffed full
Demand was never the bet; the bet was that borrowed billions could outrun the bill until rent arrived.
The strangest market in America right now sells computing capacity it cannot yet deliver, to customers who have already promised to pay, while its own insurers quietly price its funeral. CoreWeave reported second-quarter revenue of $2.58 billion, up 112 percent from a year earlier, and a contracted backlog that passed $104 billion before another $25 billion of deals landed within weeks after the quarter closed (The Next Platform, Aug 14). In the same quarter it burned $5.7 billion in free cash flow (24/7 Wall St., Aug 12). Both facts are true. Only one of them keeps the lights on.
Start with who these companies are. CoreWeave began life as a New Jersey cryptocurrency mining outfit and pivoted into renting out Nvidia chips; Nebius is a spinout of Russia's Yandex, now based in Amsterdam; Lambda, Crusoe and IREN round out a class of newcomers the industry calls neoclouds, built explicitly to undercut Amazon, Microsoft and Google on GPU rentals (The Next Platform, Aug 14). Their customers are the very giants they undercut: Microsoft, Meta, Google, OpenAI and Anthropic all sign multiyear leases because nobody, not even the hyperscalers, can build fast enough alone. The landlords of the AI boom exist because their tenants are too impatient to wait for their own construction crews.
The trigger this month was Meta. On July 17 the company announced Meta Compute, a service selling raw GPU capacity to outside developers, and CoreWeave's stock fell roughly 35 percent over the following days (TechTimes, Jul 30). Read that again slowly: Meta is simultaneously CoreWeave's largest customer, bound by a $21 billion contract running through 2032, and its newest direct competitor (TechTimes, Jul 30). When your biggest tenant opens a rival letting office across the street, the lease still gets paid. What changes is what anyone will lend against it.
Credit markets repriced before equity analysts did. On July 29, CoreWeave's credit default swap spread topped roughly 855 basis points, the highest among major technology companies and, under standard pricing models, an implied coin-flip chance of default within five years; its shares are down more than 40 percent from a year ago even after earnings beats (TechTimes, Jul 30). Oracle's default insurance climbed to about 215 basis points from 145 at the end of last year, and S&P cut Oracle to BBB-minus, the lowest investment-grade rung, earlier in July (TechTimes, Jul 30). The bond market is not arguing that AI demand is fake. It is arguing that the chain of loans each of these companies must keep refinancing, at worse rates every time, can snap before the rent arrives.
The slow pressure underneath is arithmetic. CoreWeave spent $9.4 billion on capital expenses in the quarter against $626 million of net losses, exiting June with $6.41 billion of cash and 51 data centers holding 1,500 megawatts of active power, with 3,700 megawatts more contracted (The Next Platform, Aug 14). Nebius took in $582 million of quarterly revenue, up 454 percent, while spending $5.7 billion on capital expenses in the same three months, and guides to $20-25 billion of spending for the year against full-year revenue guidance of just $3-3.4 billion (Quasa, Aug 13; 24/7 Wall St., Aug 21). These are companies buying the factory years before the factory earns anything, and every renewal of the debt costs more than the last one.
The pressure reaches well past the neoclouds. The five big hyperscalers issued more than $200 billion in debt in 2026 so far, after averaging under $30 billion a year from 2020 through 2024, and Goldman Sachs pegs global AI-related debt supply at $489 billion this year, ahead of its full-2025 estimate of $322 billion (New York Times, Aug 20; Benzinga, Aug 2026). Alphabet posted its first negative quarterly free cash flow since going public more than two decades ago, and Oracle's long-dated 2054 bonds now yield 7.8 percent (Briefs.co, Jul 29). Borrowing on this scale competes with the United States Treasury for the same bond buyers, which is one reason 30-year yields touched levels last seen in 2007 (New York Times, Aug 20).
History offers one clean model: the telecom fiber buildout of the late 1990s. Global Crossing and WorldCom laid oceans of capacity on junk debt because demand forecasts looked limitless, and demand did arrive eventually; internet traffic kept growing for decades. The builders were simply gone by then, having defaulted into bankruptcy courts that handed the fiber to creditors for pennies. The lesson is not that the demand was imaginary. It is that being early with borrowed money and being wrong look identical from the outside until the refinancing date.
The backlog is real, the demand is real, and neither of them pays the interest bill on time.
The counterexample argues the other way, and it is Microsoft. The company holds AAA ratings from both major agencies, has sold no bonds since 2017, and funds its share of the AI buildout entirely from operating cash flow (Briefs.co, Jul 29). If demand proves real and durable, Microsoft absorbs the eventual profits without ever having paid the toll that CoreWeave and Nebius pay each quarter. The structure of the boom, not its direction, decides who is still standing when the rents come due.
There is a darker wrinkle inside that structure: the money chases its own tail. Bloomberg reported in July that Nvidia is preparing financial commitments potentially exceeding $750 billion, including discussions over a $250 billion guarantee helping OpenAI lease computing capacity, while Nvidia already invests in the very cloud companies that use the proceeds to buy Nvidia's own chips (TechTimes, Jul 30). When the supplier finances the buyers who finance the tenants who pay the supplier, a slowdown anywhere becomes a shortfall everywhere.
Who pays if this breaks? Bondholders first: pension funds and insurance portfolios now hold the GPU-backed paper. Who pays if it works? Electricity customers and ratepayers, as grids strain, and the tenants, whose lease payments fund the whole pyramid either way. The winners in both worlds are already chosen: Nvidia, which collects cash today whether the clouds thrive or collapse, and Microsoft, which never joined the borrowing line in the first place.
Watch two things from here. If the read is right, CoreWeave's next loan or bond prices well above its existing debt, and Nebius returns to shareholders for fresh equity inside six months, because the $20-25 billion spending plan cannot be met otherwise (Motley Fool transcript, Aug 19). If the read is wrong, Meta Compute fails to win outside tenants and CoreWeave prints positive free cash flow within two quarters, proving the backlog converts to cash faster than the debt compounds. Either way, stop watching the backlog announcements. Watch the interest line.