Try to find the resale value of a robot. Not an arm on a machine-shop floor — those never had a public price to depreciate from. A humanoid, a quadruped, something on the fast frontier where the whole capital case rests on what the machine is worth in three years. You won't find a used market with public prices, because one doesn't exist yet.
So if you're financing a fast-frontier robot today, you are modeling its depreciation off nothing observed. The nearest asset that does have a public resale market — one that shares the defining trait of robotics, a performance frontier that moves in hard generational steps — is the data-center GPU. We maintain an archive of used-GPU asking prices for exactly this reason: it is the closest thing to a live preview of how fast-frontier hardware loses value. And the preview says the thing almost every financing model gets wrong.
It does not decay smoothly.
The plateau that isn't safety
Take the workhorse data-center card of its generation. For more than a year after its successor became available — from roughly nineteen to thirty-one months past that point — its used asking price held flat at its full prior-year baseline. Zero observed decay. A residual model checked at that stage would have looked not just fine but conservative.
Then it cliffed. By month thirty-seven the same card was asking about 43% of that baseline. By month forty-six, roughly 37.5%. More than half the value came off not gradually across the plateau but in a step, and the step landed when the successor's successor reached the market and released the supply that had been holding the older card up.
That is the shape that breaks amortization schedules. A loan structured against smooth exponential decay is over-collateralized straight through the plateau — the borrower and lender both feel safe — and then suddenly, deeply underwater at the cliff, at a date neither the calendar nor a decay curve predicted. The plateau is not evidence of durable value. It is the quiet before a repricing.
The opposite regime is real too
Depreciation is not even the only direction. In the same archive, a prior-generation enthusiast card was seen asking 2.67× its $1,499 launch price — thirty-one to thirty-seven months after its own successor had shipped. Nearly three years into the next generation's life, sellers wanted more than double the original sticker. When demand outruns supply, "residual value" stops being a fraction of cost and becomes a multiple of it.
That figure is a single observed listing — one seller, n=1 — so treat the magnitude as unconfirmed and the direction as the point: residual paths are regime-dependent, not monotonic. A scarcity shock can send used prices above new-unit list, for years, in a market that a smooth-decay model assumes only ever falls.
And the floor is a real floor
Decline, when it finally sets in, also settles rather than vanishing. A deep-vintage data-center card from an earlier generation was still asking 25% to 34% of its baseline at fifty-four to sixty-nine months past its successor — years out, well below cost, but holding a nonzero floor rather than trending to zero. The old card keeps a residual because it still does real work; the collapse is the mid-life cliff, not the long tail.
What a robot buyer should take from this
We do not forecast robot residuals. We don't forecast anything — this publication reports observed market state only, and there is no observed robot resale market to report. But the shape of the risk is already visible in the one fast-frontier asset that has receipts, and three lessons carry over cleanly.
Expect event-driven repricing, not a curve. The thing that moves a fast-frontier residual is a supply release from a later generation — a discrete event — not the steady passage of time. Financing built around a smooth decay assumption is mispriced against the wrong process.
Early "residuals are holding up" proves nothing. The workhorse card held its full value for over a year past successor availability before the cliff. A young robot category showing firm resale prices in its first years is not evidence that a cliff won't come; the plateau is exactly what the pre-cliff period looks like.
Watch successor availability, not the asset's age. The useful clock is not months since purchase. It is where the next generation — and the one after it — sits in the supply pipeline. That is a thing a buyer or lender can actually watch, which makes cliff risk manageable in a way curve-fitting never will be.
Every figure above is a used asking price drawn from public listings and archived with a checksum at the moment of capture — asking, not cleared, so real transaction prices sit somewhere below. We publish the observation and its date, and nothing we can't show you the receipt for. The robot version of this market doesn't exist yet. When it does, we'll be measuring it the same way — and the buyers who studied the preview will not be the ones surprised by the first cliff.
The Robot Ledger is published by Pexara AI LLC. Figures may be cited with attribution; underlying archive available for verification inquiries.