Compute Economics · Credit Markets

The Melting Collateral Problem

Someone has to own the risk that a $30,000 GPU is worth $8,000 in three years. Increasingly, it's credit funds — the quiet underwriters of the AI boom.

Watch the rack melt: a live loan-to-value model

$30,000GPU resale value
$21,000Debt outstanding
143%Collateral coverage
$0Cumulative rental revenue

The red column beside the rack is the lender's outstanding debt (amortizing over 5 years); the dark column is what the hardware would fetch today. When red overtakes dark, the loan is underwater — the rack turns red. Drag to rotate.

"Who today carries the capital risk of deploying compute? It's the credit funds that underwrite the GPUs."
— Meltem Demirors, Crucible Capital (paraphrased from a public panel discussion)

Why debt, not equity

A single H100-class server costs $250k–$400k; a modest cluster runs into nine figures. Neoclouds (CoreWeave, Lambda, Crusoe) can't equity-fund that, so they borrow — often in asset-backed facilities where the GPUs themselves are collateral. CoreWeave alone has raised $10B+ in GPU-backed debt.

The blind spot

Traditional asset-backed lenders price risk off decades of data: aircraft, railcars, real estate. GPUs have no such history. Nobody knows the true 5-year residual value of an H100 — it depends on chip release cycles and AI demand, both wildly uncertain. Lenders are underwriting blind.

Depreciation vs. contracts

The standard mitigation: lend only against contracted revenue (a hyperscaler signs a 3-year take-or-pay), not against the metal. Then the credit risk becomes the counterparty's, not the chip's. Merchant (uncontracted) GPU fleets get far harsher terms — or no debt at all.

Why it matters to you

If GPU residuals collapse faster than loans amortize, credit losses ripple into higher compute prices and stalled buildouts. The AI boom's pace is set as much by underwriting committees as by chip fabs.

Worked example: the 3-year cliff

Buy a GPU node for $30,000 with a 70% LTV loan ($21,000 debt, 5-year straight-line amortization). Assume value decays 35%/year and it rents at $2.10/hr at 70% utilization (~$12,900/yr gross).

Year 1: value $19,500 vs. debt $16,800 → coverage 116%, thin but alive. Year 2: value $12,675 vs. debt $12,600 → coverage 100.6%, a coin flip. Year 3: value $8,239 vs. debt $8,400 → underwater. The loan only works if rental cash flow services it before the collateral cliff — which is exactly why utilization, not silicon, is what lenders really underwrite. Recreate this on the sliders above (35% decay, 70% LTV) and scrub the timeline.

Enjoy this tool? Build your own with Super