Behind Surging GPU Spot Rental Rates: Credit Tightening Drives Inversion of Computing Power Term Structure
nashnova research
Nvidia B200 spot rent rose 23.5% in three months to $6.94/GPU-hour; legacy H100 is up 67% since July — but the driver is not just demand. Loan-to-value ratios for smaller cloud providers collapsed from 80% to 50%, forcing buyers into the spot market.
How much have GPU spot rents actually risen?
B200 spot climbed from $5.62/GPU-hour in late August to $6.94 by November 19, peaking at $7.10 — a 23.5% gain in three months.
The older H100 is up 67% since July 2025, from $1.96 to $3.10/hour.
This means → the repricing spans the entire compute supply curve, not just the latest chip.
Why isn't this just a demand story?
Stanley Lee, Chief Product Officer at Liquid Compute, told Bloomberg's *Odd Lots* podcast that high spot prices reflect rising capital costs for buyers, not just underlying demand.
Many AI inference and application companies want long-term compute contracts, but smaller cloud providers — the "new clouds" — cannot get the loans to underwrite those deals.
In plain terms = it is not that buyers lack appetite for long-term compute. Sellers cannot borrow, so contracts fall through, and buyers are pushed into the spot market at higher prices.
How badly have lending terms deteriorated?
The standard deal: a multi-hundred-million-dollar three-year amortizing loan, collateralized by GPUs, repaid from AI customers' take-or-pay contracts — agreements that obligate the buyer to pay whether or not they use the capacity.
Loan-to-value (LTV) stood at 80% this summer, dropped to 70% after Labor Day, and now sits at 50%.
This means → new-cloud providers must put up nearly twice the equity they needed months ago. Many deals that would have closed simply cannot.
What does "term-structure inversion" tell us?
Short-term spot compute is priced at roughly $40–50 billion per gigawatt; long-term contracts sit around $20 billion — buyers pay a near-100% premium to avoid long-term commitments.
In plain terms = normally, locking in a long-term contract costs more — just as fixing a long-term interest rate usually carries a premium. Here the relationship has flipped: spot is far more expensive than forward.
This reflects a market where pricing is no longer driven by pure supply and demand — credit withdrawal has distorted the term structure.
Does Dell's "GPUs hold their value" argument hold up?
Dell cited H100 rental data last week to argue GPUs retain value: a 58% increase in one year, which he said ends the depreciation debate.
But if the primary driver of high spot prices is credit contraction rather than a demand surge, rising rents cannot simply be equated with asset appreciation.
This means → if financing conditions ease and the long-term contract market restarts, spot rents could fall back — and the "value retention" thesis would lose its foundation.
What is the next thing to watch?
The key test: whether financing conditions for new-cloud providers improve and long-term contract markets reopen.
If LTV ratios keep falling, the term-structure inversion will persist, and high spot prices will increasingly signal financing distress, not a boom cycle.
In plain terms = the direction of spot rents depends less on AI demand and more on whether banks are willing to lend again.
市场有风险,内容仅供研究参考,不构成投资建议。
