Here is the tell most readers miss: the loans funding AI chips are engineered to expire at exactly the moment the chips they bought stop being worth anything.
Citadel Securities' head investment-grade desk analyst, Jeff Eason, put the number to Bloomberg this week: more than $500 billion in new debt could be issued by 2028 to buy AI chips. Most of that debt runs three to five years — "about as long as the chips themselves stay useful."
Read that again. The tenor, arguably, isn't an accident of the credit cycle. It appears to be matched, deliberately, to the depreciation life of a single generation of silicon. The entire stack is a bet that one cohort of GPUs earns back its purchase price before it becomes economically dead weight.
The entire stack is a bet that one cohort of GPUs earns back its purchase price before it becomes economically dead weight.
And the analyst building that thesis won't vouch for his own figure. Eason called the level "unprecedented" and said it might even be underestimated. He also said it could make up more than 5% of the Bloomberg US high-grade index. Globally, AI-related debt has already reached roughly $570 billion.
That is the contradiction to sit with: the people underwriting this are describing it in the language of things they cannot size.
Last week this thread was about a single guarantee. This week it has metastasized into a quantified warning and a set of copycat structures, and the binding constraint of the whole AI buildout may, arguably, have quietly moved. It may no longer be silicon. What may be shifting, arguably, is solvency.
Start with the plumbing, because the plumbing is the argument.
Any AI compute buildout needs three things at once: capital, an offtake contract, and a datacenter. Lenders won't advance against the first and third without the second — someone investment-grade promising to buy the output. That offtake guarantee is the load-bearing wall.
Nvidia has stepped into that gap directly. Its backstop program offers Neoclouds a take-or-pay commitment — a minimum revenue guarantee on the underlying GPU capacity — typically six years long, with Nvidia buying compute at pre-agreed prices.
Here is what that structure actually does to the math. Neoclouds running under these backstops earn "zero or slightly negative IRRs." The operator makes nothing, or loses a little. And lenders are comfortable anyway, because the backstop covers the debt payments regardless.
Sit with that too. The equity holder's return is near zero, and the deal still closes — because the point isn't the operator's profit. The point is that the credit risk has been lifted off the operator and set down somewhere else.
Naver made the same move explicit this week, and said the quiet part in a headline. Per the disclosed headline, its AI-factory contract structure: take the Nvidia GPUs, hand the financial burden to a Brookfield SPV. In the original — "엔비디아 GPU 받고, 브룩필드 SPV로 재무 부담 던다." Receive the chips; throw the balance-sheet weight into a special-purpose vehicle.
Each of these structures is locally rational. The Neocloud gets financeable. Naver gets compute without the debt on its books. The lender gets a covered payment stream. Everybody de-risks.
But de-risking is not the same as removing risk. It moves risk — and where it moves is the whole story. The demand-collapse risk doesn't vanish when it leaves the operator's balance sheet. It flows up the offtake guarantee and pools on the guarantor: Nvidia, plus a thin layer of investment-grade backstoppers like the Brookfield SPV.
A distributed hardware cycle is being restructured, contract by contract, into a single-point credit exposure.
Now watch the overbuild run in parallel, and watch that nobody is coordinating it.
In Taiwan, three ODM giants committed NT$181.31 billion to expansion in a single stretch. Quanta is reportedly spending NT$19.7 billion on an AUO plant in Taoyuan for new lines; Wistron took $121 million of California real estate; Inventec is expanding across Taiwan, Thailand, and Texas. Wistron and its subsidiary Wiwynn both reportedly hit limit-up on the news.
Dell has raised its AI-server target to roughly $60 billion. It reportedly booked about $16.1 billion of AI-server revenue in a single quarter and sits on $24.4 billion in orders. And its gross margin fell from 21.1% to 17.8% — the dilution is the risk the market is watching.
Alibaba is burning through it. Per its latest report, group free cash flow swung to a negative RMB17.3 billion outflow in one quarter; adjusted EBITDA fell 84% year over year. Management hinted capex would exceed the RMB380 billion already announced, insisting the return over the next three to five years is "very certain."
And China's AI datacenter capex is projected to race from about $140 billion in 2026 to $327 billion by 2030, per Bank of America — roughly 20% of the global AI capex figure the same team now projects above $1.7 trillion for 2030.
None of these actors, arguably, is internalizing the others' capacity. Taiwan's ODMs, on this reading, are building lines for demand Dell is also chasing for orders Alibaba is also funding for a market China is also racing to fill. Each plans against its own order book. The aggregate is a number nobody owns.
That is exactly the condition under which a glut forms — and here is where the two halves of the story collide.
In a normal hardware cycle, overbuild produces a chip glut. Prices fall, inventory clears, marginal operators go bankrupt, and the loss stays roughly where the bad bet was made — on the balance sheet of whoever overbuilt.
The backstop structures break that containment. If utilization disappoints, the Neocloud earning zero IRR doesn't absorb the loss — it was never taking that risk. The loss travels up the take-or-pay guarantee to the party that promised to buy the compute no matter what.
So a correction that the industry's own history would have processed as a distributed inventory washout is now structured to arrive as something else entirely: a credit event, concentrated on a handful of guarantors.
And the concentration keeps tightening as the copycats multiply. Every Brookfield-style SPV, every six-year Nvidia backstop, adds another tributary feeding the same node. The more financeable each deal becomes, the more residual demand risk, arguably, pools on the guarantor of last resort.
This is the part that should unsettle anyone reading the buildout numbers as a growth story. The $327 billion, the $60 billion, the RMB380 billion — treat them not as capacity but as capital at risk, and the aggregate looks less like an industry investing and more like a single balance sheet writing options it has priced at zero.
Nvidia's backstops, arguably, don't eliminate demand risk. On this reading, they socialize it on the way up and concentrate it on the way down. The distributed cycle has, arguably, been folded into an exposure sized like a sovereign's, sitting on top of one company and a thin ring of investment-grade names.
Which leaves the question this thread has been circling, now made sharper by the copies. When the original single guarantee has been replicated into an entire market's plumbing, do those offtake promises still have credible remedies — or has the very structure that made $700 billion financeable also made a correction unsurvivable in its old, containable form?
The tenor tells you the deadline. Three to five years, matched to the chips. The debt, not the silicon, is now the load-bearing wall — and Citadel's own analyst won't tell you how heavy it is.