Methodology / Research draft
How SBX is calculated
SBX measures quoted rental prices for B200, B300, GB200 and GB300 compute in USD per physical GPU-hour.
Comparable observations
The primary draft uses public on-demand list prices for exclusive instances. The full instance price is divided by its physical GPU count. Region, hardware, bundle contents and commercial terms are preserved. Spot, reserved and account-specific prices are separate observations.
A provider’s published list price does not establish available inventory or the price of a completed transaction. NVL72 and HGX systems are not performance-equivalent units.
Provider and model prices
Independent operators must agree on matching source records. Accepted quotes are summarized within each provider using SKU and regional medians. Model prices use explicit weights for independent provider groups. Adding more nodes or duplicate listings does not increase a provider’s weight.
The composite
SBX is the arithmetic weighted average of all four model prices. Weights are published with each snapshot. A missing model or fixed-weight provider constituent makes the affected index unavailable; the remaining weights are not silently redistributed.
Equal weights, when configured, are a research choice. They do not represent measured market share or certification against an industry standard.
Availability and publication
Prices require fresh observations, sufficient independent reports, comparable terms and approved source rights. A dash means no qualified price is available. An unsigned page capture or a valid node signature alone is not proof that a price is correct.
The public collector network and Pyth publication are separate stages. Pyth publication remains pending until publisher admission, feed identifiers and external readback are verified.
Reproduce the result
Snapshots identify the methodology, provider registry and input batch hashes. The open source node software verifies signed observations and reproduces calculations from the same inputs.