> ## Content Index > Fetch the complete content index at: https://capital-matter.ghost.io/llms.txt > Use this file to discover other available public pages before exploring further. # When the collateral is a chip: how GPU-backed loans really work - URL: https://capital-matter.ghost.io/when-the-collateral-is-a-chip-how-gpu-backed-loans-really-work/ - Published: 2026-10-26T11:08:38.000Z - Updated: 2026-10-26T11:08:38.000Z - Author: Kaviya Gopal When the collateral is a chip: how GPU-backed loans really work A note before we start: the worked example below is illustrative. It uses simple assumptions to show the mechanics, not the terms of any real loan. Issues 03 to 05 were about buildings. A data center is collateral with an economic life measured in decades, and even where the lease is short, as at Hyperion, a 16-year value guarantee sits behind it. A GPU is a different kind of collateral. Accounting schedules for these chips run roughly four to six years, and the people arguing about the right number are not close to agreeing. That raises an obvious question: how do you lend against something that may be worth a fraction of its cost before the loan is repaid? In Issue 02, I flagged GPU-collateralized financings starting to show up in DACH term sheets. This is the instrument behind them. ## The template: chips plus contracts in one vehicle The structure was established by CoreWeave. Its first delayed-draw term loan, in August 2023, put a fleet of H100 chips and the customer contracts attached to them inside a bankruptcy-remote vehicle, with a reported advance rate of about 70 cents on the dollar. It reportedly priced near 15%. Three years later, the same structure is being syndicated to institutional investors, and the pricing varies a lot: | Facility | Date | Size | Pricing | Rating | Who pays the rent | | ------------------ | -------- | ------ | ------------------------------ | --------------------------- | ------------------------------------ | | CoreWeave DDTL 4.0 | Mar 2026 | $8.5bn | SOFR + 225 bps | A3 (Moody's), A(low) (DBRS) | Reportedly backed by a Meta contract | | Lambda term loan B | Aug 2026 | $926M | SOFR + 300 bps, issued at 99.5 | Baa2 (Moody's) | An unnamed investment-grade customer | | CoreWeave DDTL 5.0 | May 2026 | $3.1bn | SOFR + 450 bps | Ba2 (Moody's), BB+ (Fitch) | Two non-investment-grade customers | The collateral is the same kind of asset in all three. The price differs by more than two percentage points. What changed was the customer. A GPU loan is priced mostly on the credit of whoever has promised to pay for the chips, and the hardware is the secondary backstop. That is tenant-credit substitution again, from Issue 01, applied to a shorter-lived asset. DDTL 5.0 also shows how lenders try to line the loan up with the chips. CoreWeave describes it as built to "align funding with the deployment schedule and useful life" of the hardware, with a maturity of about 5.5 years. Lambda's loan is fully amortizing and runs to December 2030. ## The problem in numbers Take an illustrative fleet of GPUs bought for $1,000M, financed with $750M of debt (a 75% advance rate) and $250M of equity. Assume an 8% interest rate, level annual payments, and a four-year take-or-pay contract with a customer. There are two ways to structure the repayment: | | **A. Repaid inside the contract** | **B. Loan outlives the contract** | | -------------------------------------------- | --------------------------------- | --------------------------------- | | Amortization period | 4 years | 6 years | | Annual debt service | $226M | $162M | | Contracted cash flow available for debt | $306M | $306M | | DSCR during the contract | 1.35x | 1.88x | | Loan balance when the contract ends (year 4) | About $0 | About $289M | Structure B looks more comfortable on the day it is signed. It has a higher DSCR and more cash left over for equity. But at year four, the customer's obligation ends and $289M of debt is still outstanding. Whether that debt gets repaid now depends on one of two things: re-contracting the chips at a good price, or selling them. Here is what the chips would need to be worth at year four to cover the balance: | GPU value at year 4 (% of cost) | Value | Cover on the $289M balance | Result | | ------------------------------- | ----- | -------------------------- | -------------------- | | 50% | $500M | 1.73x | Covered | | 30% | $300M | 1.04x | Barely covered | | 20% | $200M | 0.69x | About $90M shortfall | Compare this with Issue 04\. For the building, the facility's value had to fall about 45% before debt exceeded it. For the chips, a similar loss can happen inside four years. That is why lenders like Structure A: if the loan is repaid out of the contract itself, the resale value of the chips mostly stops mattering. ## What do GPUs actually retain? This is where nobody has a clean answer. - **Accounting lives disagree.** One analysis notes that CoreWeave depreciates its chips over six years and Nebius over four, a gap of about 50% on similar hardware. The general market assumption is five to six years. - **Rental prices have swung.** The same analysis puts H100 rental rates at about $8 an hour in early 2024, $1.70 in October 2025 and about $2.35 by March 2026\. A separate report put the fall at roughly 70%, to about $2.50 an hour, by December 2025. - **The sceptics point to obsolescence.** New generations arrive roughly yearly, and Blackwell is reported at about 25 times the efficiency of the previous generation. - **The optimists point to demand.** The CEO of a used-GPU marketplace, who has a commercial interest in saying so, reports H100 server prices up more than 10% year on year in 2026 and active demand for older chips. Another dataset shows used H100 cards well below their 2024 peak. Both can be true. - **Resale tracking** for H100s reports retention of roughly 60 to 75% at two to three years, 45 to 60% at three to four years, and 25 to 35% after five. These are tracker estimates, not audited prices. Notice what is missing: a full cycle of observed recoveries. As one analysis puts it, no GPU generation has yet been through a complete depreciation cycle at hyperscale. It is the same gap KBRA flagged for data center buildings in Issue 04, but here the clock runs in years, not decades. ## The risk that does not show up in the model Analysts who follow these structures have pointed to a wrong-way risk. If a customer signed a contract at prices well above today's market rate, it has a growing incentive to renegotiate or leave, just as the lender's collateral value and replacement revenue are falling. The two shocks arrive together. A related concern is correlation. Customers, lenders, chip vendors and, in some deals, vendor backstops all depend on continued AI demand. Spreading exposure across different names does not spread that underlying driver. That is the subject of the next issue. ## Three questions to ask of any GPU-backed loan **Does the loan repay inside the contract?** If yes, the chip's resale value is a footnote. If no, the balance at contract end is the real credit risk. **Who is the customer, and what happens at renewal?** The rating of the debt follows the customer, not the hardware. A contract struck above market is a risk, not just a comfort. **Whose depreciation schedule is the loan measured against?** If two lenders can look at identical chips and see lives that differ by 50%, the schedule is a choice, not a fact. A building usually outlives its loan. A GPU may not. The structures that work are the ones that make that matter less. --- Next issue: circular financing. What happens when the customer, the chip supplier, the landlord and the lender are the same few names.