SpaceX closed $40 billion in debt financing this month to buy GPUs it does not expect to receive until the first half of 2027. The debt is callable, and it sits in a special purpose vehicle that keeps it off SpaceX's balance sheet — a structure Bloomberg reported was chosen specifically to preserve the company's investment-grade rating.
That is the emblematic transaction of this cycle. A rocket company borrowed against chips it hasn't taken delivery of, through a legal entity it doesn't have to consolidate, to protect a credit rating it needs to borrow again. AI capex is no longer a cash-flow story. It is a credit market story, and the semiconductor supply chain has not repriced for it.
Last week this column treated the cloud capex numbers as demand-driven — Meta guiding $135–145 billion for 2026, obligations accumulating toward $825 billion, SK Hyper seeking $350 billion for the first 5GW of a 15GW program. That framing was right for 2024 and 2025, when the money was largely earned. This week the delta is the marginal dollar. It is borrowed, and the borrowing is now the binding constraint on chip orders.
AI capex is no longer a cash-flow story. It is a credit market story, and the semiconductor supply chain has not repriced for it.
The scale is not subtle. US investment-grade issuance is on track for a record $1.76 trillion in 2026, past the 2020 pandemic-era peak of $1.75 trillion, and AI-related IG funding has already cleared $300 billion — nearly three times last year's total. AllianceBernstein's Tim Coffey says borrowing costs aren't constraining activity because the issuers are rate-insensitive. Goldman's Lindsey Rosner calls it a once-in-a-generation capital-intensive investment cycle.
Read the structure underneath the headline, though, and the rate-insensitivity argument gets thinner. The SpaceX deal exists because the debt cannot sit on the balance sheet without threatening the rating. That is the pattern now: special purpose vehicles, equipment-backed loans, lease guarantees, off-balance-sheet commitments that rating agencies bless and that investors price anyway.
The mechanics are rigid in a way the underlying business is not. Physical compute clusters are financed on three-to-five-year debt cycles, as Evan Conrad has pointed out, which means operators must match fixed repayment schedules against multi-year leases. If the leases don't get signed, debt service doesn't flex. The loan is indifferent to utilization.
Depreciation is already eating the revenue. CoreWeave booked roughly $2.6 billion in second-quarter 2026 revenue against about $1.4 billion in depreciation and amortization of intangibles. Nebius: $582 million revenue, $260 million depreciation. Defending profitability is hard even when you win the race to secure GPUs.
The credit market has started to notice. As of October 21, yields on CoreWeave's unsecured debt sat near 15 percent, and the cost of insurance against default jumped 30 percent in a month. CoreWeave shares fell 29 percent in October; Nebius dropped 27; IREN slid 32. Gartner forecasts more than 80 percent of neocloud operators will cease independent operation by 2030, and warns that $3.5 trillion in AI infrastructure investment could be exposed to impairment risk by the end of the decade.
Musk himself said more than 200 neocloud companies could face widespread failures and consolidation as debt financing tightens, and noted that debt markets have largely shut off access for companies without substantial revenue outside AI data centers. The person who just closed the largest GPU debt deal of the year is describing the door closing behind him.
Here is where this stops being a financial-markets story and becomes a semiconductor story. The demand signal for 2027–2028 silicon is not Nvidia's order book. It is the spread on a CoreWeave bond. An order book is a queue of intentions. A bond spread is the price of the money that has to fund those intentions, marked daily by people with capital at risk.
If that spread blows out further, the first thing to die is not the data center. Concrete gets poured, grid interconnects get queued, and a partially built shell is a sunk cost with option value. The GPU order behind it is a purchase order, and a purchase order can be pushed, downsized, or cancelled. So can the HBM stack that goes on the accelerator and the base die underneath it.
SemiAnalysis revised its 2027 data center capacity forecast up from 43 to 50 gigawatts, backed by contracts and satellite imagery. Those gigawatts are real — poured concrete, interconnection queues, verified sites. But 50GW of supply is only locked in if the financing that underwrites the chips arrives on schedule.
If it doesn't, 50GW becomes 50GW of stranded commitments. The memory makers who expanded capacity for HBM stacks and base dies against that forecast hold the inventory. HBM is not a commodity you warehouse and sell into the next cycle at a modest haircut; it is a customized stack qualified to a specific accelerator, and it ages badly.
The BIS laid out this transmission in January: the scale of anticipated AI investment requires firms to shift financing from operating cash flows to debt, with private credit playing a rapidly increasing role. The BIS added that sustainability hinges on AI firms meeting high earnings expectations — and that equity prices have run far ahead of debt market pricing. That gap is the tension. Equity can hold a narrative for years. Credit marks to market weekly.
The demand side is not obviously ready to service the debt. Apollo's Torsten Slok found revenue per worker and profit margins rising for the Magnificent Seven while remaining flat across the other 493 S&P 500 companies, with no clear evidence AI is lifting broader productivity. "The AI story will be decided in the S&P 493, not the Mag 7," he wrote, "and so far, there is no evidence that revenue growth or margins are improving outside the tech sector."
That is the mismatch in one sentence. The debt is being raised against enterprise inference budgets that would have to expand tenfold — a $10 million-revenue company going from $100,000 to $1 million a year in inference spend, 10 percent of top line — to clear the hardware already planned. Consumer subscriptions at $20 a month will not cover it.
So watch the credit, not the order book. Watch whether the next GPU-backed SPV prices tighter or wider than the last one. Watch whether the callable structure in the SpaceX deal gets copied, because callable debt is what you issue when you are not certain you will need the asset for its full life.
If credit holds, the buildout proceeds and the 2027–2028 forecast is roughly right. If it doesn't, the cancellation arrives at the chip and HBM order book before it arrives at the data center — and the fabs, packaging lines and memory expansions built for 50GW discover they were financed by a bond market that changed its mind faster than any process node could ever slip.