The optimistic pitch for GPU lending says the chips hold value like aircraft. They do not, and the pitch is not needed. The defensible claim is narrower: current-generation GPUs have a liquid global secondary market, and a lender survives depreciation if the outstanding loan balance stays below a conservative estimate of what the fleet would fetch at every point in the loan's life. Everything in GPU collateral practice follows from that one discipline.
The published curves are steep and sharply generation-dependent. The largest secondary-market dataset in circulation, drawn from over 600,000 units and 77,000 transactions since 2023, puts H100 residuals at 27–53% of list price after one year and just 5–14% after three. H200 holds 45–74% at year one and 24–51% at year three. B200 holds 59–80%, then 30–56%. B300, the newest, holds 78–89% at year one and 37–64% at year three. Two caveats travel with those numbers. The publisher sells residual-value insurance, so the data is directionally credible but not neutral. And the bands are wide because the best-informed buyers genuinely disagree; when sophisticated desks differ by twenty points on depreciation, collateral value is uncertain in the technical sense, not just risky.
The mechanism behind the steepness is Nvidia's release cadence. Each generation arrives with a multiple of the prior generation's performance and knocks 30–60% off prior-generation resale within twelve to eighteen months. This is not gradual wear; it is scheduled obsolescence with a public calendar. A lender can read the calendar as easily as a borrower, which is why loan tenor and amortization speed, not loan-to-value at signing, are the real credit decisions. A 65% advance that amortizes over three years can be safer than a 50% advance over five, because the second spends its later years above the curve.
The most useful nuance in the data is that rental rates and resale prices move separately. Through 2025, H100 rental rates fell while H100 resale values rose, because datacenter power and space constraints slowed Blackwell deployment and kept the installed base earning. The two prices answer different questions — what the machine earns versus what the machine fetches — and they must be underwritten separately. Our forward curve models the first. The residual curve is the second. A credit that leans on either one to fix a hole in the other has confused the box with its cash flows.
For India, the standing objection is that the domestic secondary market is thin, so collateral is untestable. The objection dissolves on inspection. GPUs enter India at zero basic customs duty, are air-freightable, and clear through the same global broker and IT-asset-disposition channel as everyone else's fleets. Recovery from an Indian default is an export transaction: global broker bid, less 15–25% for logistics, warranty transfer and compliance. Indian collateral tracks global curves precisely because nothing anchors it locally. The honest caveats sit elsewhere: portability cuts both ways, so serial-number tracking, geofenced telemetry and colocation access control are collateral hygiene rather than paranoia; and no lender has yet repossessed a GPU fleet in India, so enforcement is untested, which argues for title-retention lease structures over hypothecation in early transactions.
What does not work is treating the residual curve as a substitute for revenue underwriting. By year three an H100 fleet is worth 5–14% of list; no recovery at that level rescues a loan sized against the machines alone. The residual curve's proper role is to set the amortization boundary and the advance rate, while the offtake contract carries the repayment. One sentence covers the whole doctrine, and it is the sentence we would want every credit committee looking at Indian compute to hear: amortize below the residual curve, lend against the contract rather than the chip, and verify the utilization that connects them.