The AI Crash Made Computing More Expensive, Not Cheaper
The crash made the chip cheap as collateral and dear as a rental — and the two prices can't meet, because the shortage underneath them is electricity, not money.
The credit market has already written CoreWeave's ending. Insurance against the company defaulting within five years now costs 1,232.6 basis points — one basis point being a hundredth of a percentage point, so a buyer pays roughly 12% a year to insure its debt, a price implying the market puts better-than-even odds on the firm not surviving five years [1]. Turn the page and the same company's own paperwork says something else. CoreWeave carries $47 billion in debt, lost $709 million last quarter, and has more than $10 billion maturing next year [2]. Yet its filing shows customers now signing multiyear commitments for earlier-generation GPUs at prices above their original contracts — leases that run longer than the accounting life over which a chip's cost is written off [2].
Customers are signing multiyear commitments for earlier-generation GPUs extending beyond their depreciable lives, with pricing above their original contracts. — CoreWeave
Two verdicts, one machine, in the same weeks. The company the credit market has already written off is, in its own filing, the one raising prices. Which of these prices is real? Ask the banks, and the cheap one is. Lenders have cut the advance rate — the share of a chip's value they will lend against it — from 80% to 50% [3]. Nebius shelved a bond offering. Even loans carrying guarantees from Google and Anthropic are being stress-tested and set aside rather than syndicated [4]. In a bank's ledger, a GPU is collateral, and collateral is worth only what someone will pay for it in a fire sale. After CoreWeave's October implosion, they are repricing that fire sale daily. The people renting the chips keep a different ledger. The B200 rental index is up 23.5% in three months, to $6.94 per GPU-hour [3]. A four-year-old H100 rents for 67% more than it did in July 2025 [3]. Spot — paying as you go, no long contract — has run to between $40 and $61 billion per gigawatt-year, the cost of a gigawatt of computing capacity for a year, against long-term contract floors near $20 billion and a hurdle rate of $11.6 billion, the price at which a hyperscaler earns a 15% return [5]. The marginal buyer at those spot prices is a venture-backed startup paying to train a model, not a profitable business serving one [5]. In a renter's ledger, a chip is a working machine, worth whatever it earns before the next generation renders it obsolete. The reason the two prices cannot meet is that only one of them is set by money. The shortage behind the rental spike is electricity. Hyperscalers are holding hundreds of billions of dollars in GPUs in warehouses — bought under irrevocable take-or-pay orders with TSMC, meaning they must pay whether or not the chips ever get switched on — because a shortage of rhenium, a metal used in gas-turbine alloys, has left them without generators to power the hardware [6]. Goldman Sachs estimates only half of the planned data-center capacity will actually be built near-term [7]. Oracle declared force majeure — a clause freeing a party from a contract when circumstances beyond its control intervene — on its New Mexico project after regulators blocked a gas pipeline, and Microsoft faces permitting delays in Michigan [7]. Goldman itself calls the spot premium unsustainable, and projects capacity multiplying through 2028 [5]. That sharpens the point rather than softening it: the closing date on this gap is set by pipelines and turbines, not credit desks. A rate cut, a refinancing, a fresh rescue — none of them shorten the lead time on a gas turbine. The bill travels outward from there, and it lands first on the labs. Anthropic has committed between $518 billion and $895 billion to future compute and infrastructure [8]. The thing it sells — tokens — fell more than 40% in price between June and October, even as usage quadrupled [9]. The inputs got more expensive: DRAM contract prices rose 80%, and AWS raised Nvidia Capacity Block reservation prices at least 15% [9]. In the third quarter Anthropic booked $16.96 billion in revenue and still lost $567 million on operations [9]. The borrowing market has thinned to the giants. Goldman still projects hyperscaler debt issuance climbing to $420 billion by 2027, while neocloud paper sits shelved [4]. The crash did not close the capital tap; it moved it, concentrating it in the handful of balance sheets the market still considers safe. Third, to everyone else. AI hardware demand, by the IMF's count, now exceeds 10% of world goods trade [10]. The Fed is widely expected to raise rates a quarter point at its December 8–9 meeting [10] — the policy answer to the inflation the buildout itself generates, landing on the debt stack that funds it.
This calls for a tightening of monetary policy in many economies. — Kristalina Georgieva
The machine the credit market has already buried is the same machine renters are lining up to overpay for. Both prices are real; they are just set by different fears — one fears the loan, the other fears the dark warehouse. And the second fear has the longer horizon. Somewhere a four-year-old H100 sits under a lease longer than its accounting life, rented at more than its original contracts [2]. The Fed's answer to the whole thing is due December 8 [10].
- 1. AI Stock Selloff Spikes Credit Risk for Hyperscalers
- 2. CoreWeave Faces Financial Pressure After Nvidia CEO Removal
- 3. GPU Rental Prices Surge Amid AI Market Volatility
- 4. Banks Tighten Lending Standards for AI Debt Deals
- 5. AI Compute Spot Prices Reach $61 Billion Per Gigawatt
- 6. Hyperscalers Warehouse Billions in GPUs Amid Power Shortages
- 7. AI Infrastructure Faces Trillion-Dollar Risks and Physical Bottlenecks
- 8. Anthropic Files for Record-Breaking IPO Amid Existential AI Warnings
- 9. AI Token Prices Drop 40% Amid Rising Infrastructure Costs
- 10. Federal Reserve Signals Rate Hikes as AI Drives Inflation