There is no tape for used GPUs. Assemble the price instead.

The pieces are public but scattered; the fusion is the product. Residual value is modelled, then checked — the rental model fills the gaps the resale data leaves, and the resale data validates the model.

The premise: a used GPU is worth what it can still earn. Sparse resale listings cannot be trusted to tell you the price, but what an hour of each chip rents for is observable right now across dozens of clouds — so the engine values the machine the way any income-producing asset is valued, then applies one economic bound and one reality check.

Vein one, cloud spot rental rates — scraped from provider listings, the asset's earning power observed in near real time; the neocloud composite is the primary input. Vein two, secondary hardware prints — broker and refurbisher listings that validate the model but never set it. Vein three, replacement cost — the frontier's street price per unit of performance is a hard ceiling on every prior generation. Vein four, the GPU-backed debt stack — facility pricing tells us what depreciation schedule the lending market underwrites, and anchors the 12% unlevered discount rate. Vein five, the obsolescence calendar — each successor volume ramp adds a displacement increment to everything behind it.

Worked through the H100: start from the observed neocloud rate of $3.25/hr — the median of each provider's cheapest on-demand per-GPU listing (hyperscaler rates carry a platform premium and marketplace rates are interruptible, so both are recorded as bounds, not inputs). Project that rate decaying at 11.2% a year, steepening by 6 points each time a successor generation reaches volume production (3 ramps are on the calendar). Convert rate to cash: 85% utilisation, minus $0.32/hr of all-in running cost — power at the chip's draw, datacentre overhead, colo, ops — across 730.5 hours a month. Cash flows stop when the decayed rate no longer covers running cost, or at age 7 — one year past the six-year useful life the lending market underwrites. Discount at 12%: above the 6.55% investment-grade and 8.8% sub-IG GPU-backed debt costs, below a neocloud equity hurdle. That sum is the income PV: $41,588. Then the cap: nobody rationally pays more for old silicon than the same performance costs new — the B200-192GB at $47,500 street delivers 6.6 units of performance, so H100-class performance (3.0 units) is bounded at $21,591. The index value is the lower of the two, refereed against the broker band of $15,000-$28,000.

Base rate decay is calibrated at run time from the observed H100 tape: 11.2% a year (median across channels — hyperscaler 11.2%, marketplace 17.3%, neocloud -4.3%). The boom-suppressed single series never sets the base; the cross-channel median does.

Income PV is what the rental stream is worth; the ceiling is what frontier arbitrage allows; COMPUTE-RV is the lower of the two.

The H100 print is the tell. Its rental stream is worth $41,588 at a 12% discount, but nobody pays more for a used chip than frontier performance costs new — the ceiling prices H100-class performance at $21,591, and the secondary market band is $15,000–$28,000. Model, arbitrage bound and observed prints are read against each other every print.

Successor silicon is on the calendar. When the frontier's price-performance steps, every ceiling reprices at once.

The stress assumes the next frontier lands at 1.7x the current frontier's street price-performance — conservative against vendor-claimed uplifts. The lending market is underwriting six-year useful lives against this curve; the stressed column is what those books mark to if the roadmap delivers.

Facts, opinions and judgments kept as distinct types; identical inputs, identical curve.

Every input row carries a source URL and access date; rental rates are scraped observations, not assemblies. The engine is deterministic stdlib code and the output embeds the SHA-256 of its own canonicalised inputs (a9af611f…). Prints are filed into the ANTIKYTHERA chain — append-only, hash-linked, anchored outside our control — so the track record cannot be quietly rewritten. That ledger is the moat: once the curve has run through a cycle and the useful-life calls prove out, these marks become the cited ones.

spec 0.1.0 deterministic monthly print