A GPU cluster pledged as loan collateral carries three separate numbers, and the credit agreements underpinning the current AI infrastructure boom mostly reference one of them. That gap is the central mispricing sitting inside tens of billions of dollars in debt now secured by chips rather than by a borrower’s broader balance sheet. Untangling it means separating three concepts that loan documents tend to treat as interchangeable.
Face value is the accounting number: purchase price minus depreciation on whatever schedule the borrower picked. CoreWeave depreciates its GPUs over six years. Nebius, running comparable hardware for comparable workloads, depreciates the same chips over four. That divergence alone should concern anyone underwriting loan-to-value covenants, because face value is the figure those covenants are built on, and two operators holding similar fleets can report meaningfully different collateral coverage.
Liquidation value is what a distressed buyer actually pays. Analysis published by Meg McNulty at CipherTalk, drawing on secondary-market data, puts moderately used GPUs two to three years old at roughly half to two-thirds of new-unit cost under ordinary conditions. In a default scenario where several neoclouds come under stress at once, the buyer pool shrinks just as supply floods the market, and recovery can fall to something closer to a third of what the loan assumed. That is the number a fire sale actually clears at, well below what most loan documents model.
Going-concern value is the hardest number to pin down and the one that matters most. It measures what a cluster is worth as a working asset to a new operator, not as a warehouse of silicon. A lender exercising step-in rights after default does not just inherit chips. It inherits a colocation contract it did not negotiate, a rack layout it did not design, and none of the operational knowledge that kept the cluster productive: which nodes run hot, which cooling loops are unreliable, which jobs to reroute before a GPU degrades into a full failure. Meta’s own technical report on training Llama 3 documented over 400 unplanned disruptions across roughly 16,000 H100s in under two months of work. Keeping a fleet at that failure rate productive is a craft, and the people who practice it tend to leave once a borrower defaults.
None of this is priced. GPU-backed loans carry an 8.5-point spread over benchmark rates, by CipherTalk’s accounting, against 1 to 2 points on a typical aircraft loan. That gap exists because aircraft financing has decades of appraisal standards, maintenance logs, and an active secondary market behind it. GPU financing has a rental-rate index that launched in 2024 and one early-stage startup attempting to build a derivatives exchange. Lenders are charging a large premium for uncertainty, which is not the same as hedging a modeled risk.
This is the collateral question sitting underneath the entire neocloud financing wave, from xAI’s special-purpose-vehicle structures to FluidStack’s data-center arrangement with Anthropic. If a cluster is worth far less to its next tenant than its face value implies, loan-to-value ratios calculated off depreciation schedules are systematically overstating what actually backs the debt. The spread between face value and going-concern value is not a rounding error. It is the uncompensated risk sitting inside every GPU-backed credit facility written so far.
Anyone underwriting or buying into GPU-backed debt over the next 90 days should demand two things most current deal structures skip: a documented operational handoff plan that survives a borrower default, and a loan-to-value calculation stress-tested against liquidation value, not face value. Without both, the loan is priced against a number the collateral will never actually deliver.
Analysis published on CipherTalk (Substack) by Meg McNulty, May 2026.
