AI's Credit Markets Are Starting to Split Chips From Contracts
On October 1, GPU-cloud provider Lambda closed a $1.008 billion fixed-rate term loan priced at 6.78%, carrying investment-grade ratings from Moody's (Baa1) and Morningstar DBRS (A low). A day later, Reuters and the Financial Times reported that Amazon was in talks to move roughly $8 billion of Nvidia Grace Blackwell systems — installed across more than a dozen U.S. data centers in five states — into a special-purpose vehicle, leasing the chips back from outside investors who would be offered up to a 10% equity stake in exchange for taking the hardware off Amazon's balance sheet. Same underlying asset, same week, two opposite instincts: one buyer borrowing directly against GPUs, one trying to get out from under them.
Both moves are reactions to a problem Nvidia created for itself two months earlier. In August, Nvidia signed preliminary agreements with six of Wall Street's largest pools of capital — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR — to build "compute financing platforms" that would lend against GPUs as collateral, with Nvidia backstopping up to 25% of a given deal's residual value, a guarantee that could total as much as $125 billion against the full $500 billion target. The pitch: treat racks of accelerators like aircraft or office towers, assets with a known resale value that can be pledged against a loan. As of early October, by most accounts, not a single transaction under that framework had actually closed.
Why Bankers Don't Trust a GPU's Resale Value
The holdup is depreciation. Banks that finance physical infrastructure are used to underwriting assets whose value erodes on a predictable, decades-long curve with an established secondary market. A GPU's resale value, by contrast, resets every time Nvidia ships a new architecture — and Nvidia controls that release schedule. Amazon has already cut the assumed useful life of some of its own servers from six years to five, citing the pace of AI hardware change, and analysts have flagged that AI-linked financing demand is unusually sensitive to swings in credit conditions — precisely the kind of correlated risk a lender doesn't want sitting underneath thousands of near-identical loans at once. Without a track record for what a three-generation-old accelerator is actually worth on the used market, a 25% guarantee from the chipmaker itself reads as reassurance, not proof.
Contracts Are Becoming the Better Collateral
Lambda's loan shows where lenders are gravitating instead: back the contract, not the chip. The $1.008 billion facility is secured by committed deployments with two investment-grade customers, including Microsoft, rather than by the GPUs' resale value — and it priced accordingly, with a fully amortizing structure running to 2033 and a rate well inside what GPU-collateral deals have typically commanded.
"Building on our investment-grade Term Loan B and bank lending facility, this is the third new credit market Lambda has opened in the last 18 months." — Michel Combes, CEO of Lambda
The same divide is visible inside a single balance sheet. Meta's $27 billion Hyperion bond, sold in October 2025 to fund its Louisiana data-center campus, has been sliding ever since: traded around 96 cents on the dollar in July, down to roughly 94 cents by late September, pushing yields well above issuance even as Meta's stock and the broader Nasdaq kept climbing. When Meta returned to the market for a follow-on data-center financing vehicle in July, it paid roughly 40 basis points more than the original Hyperion notes carried — a small but telling markup showing fixed-income investors, unlike equity investors, are pricing in growing caution about AI payback timelines the longer these structures sit outstanding.
For the vendors actually building this infrastructure — foundries, server integrators, data-center developers, power suppliers — the split matters more than any single headline number. A financing market that only rewards contracted, investment-grade-backed deployments will keep funding hyperscalers and neoclouds that can show multi-year offtake agreements, while capacity bet purely on GPU resale value gets pricier to finance or doesn't get built. Amazon's SPV maneuver looks like a hedge against exactly that risk: rather than wait for Nvidia's $500 billion platform to prove out a depreciation model, it's trying to hand the residual-value question to investors now, while the hardware is still new. Whether that bet pays off depends on something nobody in this market can yet price with confidence — what a Blackwell-generation GPU is actually worth in 2029.