Anyone who buys compute at scale already knows the spot rate is the wrong number. Capacity ordered today is delivered in months and paid for over years, so the question that matters is not what an H100 costs this afternoon but what a B300 will cost in the seventh month of a ramp that starts next quarter. Commodity markets answered the equivalent question decades ago with forward curves. GPU rental is now liquid enough for the same treatment: Silicon Data publishes daily forward curves for the A100, H100 and B200 derived from observed term structure, and CME Group is building futures on top of its benchmarks. The mathematics is the standard no-arbitrage construction used in every yield curve since the 1970s; what is new is that rented compute finally has enough observable term pricing to support it.
What the record shows
The H100 is the first datacentre GPU with a full price cycle on the public record, and it teaches three lessons. First, the decline is large. Scarcity pricing in 2023 routinely exceeded $8 per GPU-hour; by late 2025 the same silicon was widely available at $2–4, roughly 64% below peak. Second, the decline is not smooth. One-year contract rates rose about 40% between October 2025 and March 2026, from roughly $1.70 to $2.35 per hour, as financed neocloud demand absorbed float; Silicon Data's B200 rental index jumped 24% in March 2026 alone. A curve fitted as a straight line down would have mispriced every contract signed that quarter. Third, the declines cluster. Prices moved most when the next generation reached volume, not when it was announced.
Underneath the rental market sits a hardware treadmill with a published schedule. Epoch AI's price-performance series shows GPU performance per dollar improving around 30% a year, and Nvidia has committed to an annual architecture cadence: Blackwell now, Rubin in volume in the second half of 2026, Rubin Ultra in the second half of 2027, Feynman in 2028. Each launch reprices its predecessors — but with a lag, because launch and availability are different events. A generation announced at GTC typically takes two to three quarters to reach rentable volume, and its predecessor's rental price holds up until that volume actually lands.
Three forces, one curve
Those observations reduce to a parametric model with three components per GPU class.
A scarcity plateau. While a generation is supply-constrained, rental prices do not decay; allocation, not price, clears the market. The newest class holds a flat plateau measured in months, longest for rack-scale parts (GB300 NVL72) where supply is committed a year ahead, zero for classes already past their scarcity window.
Baseline erosion. Once float builds, rates erode continuously as supply broadens and workloads migrate up-generation. We parameterise this at 1.2–2.2% per month depending on the class's age within the cadence — consistent with the H100's post-scarcity trajectory and with a ~30% annual price-performance drift, but slower than either, because contracted fleets and financing structures resist repricing on the way down.
Launch step-downs. When a successor reaches volume availability, the incumbent steps down 4–8% beyond the baseline, the size falling with distance from the frontier. We date these steps to volume windows, not announcements: Rubin in late 2026 and Rubin Ultra in late 2027 are the two events inside the current 18-month horizon.
The forward mid for class c in month t is then the current reserved rate carried flat through the plateau, decayed at the class's monthly rate thereafter, and multiplied by each step factor whose volume window has passed. Around the mid we carry the low–high band from tracked reserved-term listings, since dispersion between providers is itself persistent — the widest spreads on the board today exceed 4× between hyperscaler list and marketplace clearing prices.
| Class | Plateau, mo | Erosion, %/mo | Rubin step | Rubin Ultra step |
|---|---|---|---|---|
| GB300 NVL72 | 8 | 1.2 | −4% | −8% |
| B300 | 4 | 1.5 | −7% | −6% |
| B200 | 2 | 1.8 | −6% | −5% |
| RTX 6000 Pro | 0 | 1.3 | −2% | −2% |
| H200 | 0 | 2.0 | −5% | −4% |
| H100 | 0 | 2.2 | −4% | −4% |
| MI355X | 3 | 1.9 | −6% | −4% |
Parameters as of 11 July 2026. The MI355X follows AMD's own cadence; we map its step-downs to the nearest Nvidia volume windows pending observable MI400-generation term pricing. RTX 6000 Pro steps are small because workstation-class demand is weakly coupled to the datacentre frontier.
What it is for, and what it is not
This curve prices delivery timing on our request-compute page: move a phase three months right and the estimate reprices every month of the commitment at that month's forward rate, so the cost of waiting — or the saving — is visible before anyone talks to a desk. The same shape is what a lender should hold against a GPU-backed loan, since the collateral's earning power follows the rental curve, not the invoice price.
It is a parametric estimate, not a traded curve. It will be recalibrated as the tracker accrues term-structure observations, and it will be wrong in the specific way the H100 record warns about: demand shocks — a financed offtake wave, an inference capacity squeeze — produce rebounds that no decay model anticipates. We publish the parameters so the disagreement can be precise. Nothing here is investment advice.