Futures basis
GPU rental prices vs. the futures benchmarks.
CME Group announced H100 and B200 compute futures on NYMEX, settling to Silicon Data's rental indexes, for October 5, 2026, pending regulatory review (press release). This page compares those benchmarks with GridCompute's independent, open hourly prices. It's a cross-check, not a settlement benchmark.
Asking prices for free capacity vs. an assessed benchmark. Not the same thing; the gap is information, not an error.
No benchmark levels are shown yet. Silicon Data's and Ornn's public terms don't allow third parties to store or republish their index values, so levels appear here only once we have a licence or written permission. GridCompute's own prices are always on the GPU pages.
Gaps in the line are hours we didn't collect. Benchmark levels are dots on the day they apply to; nothing is drawn between them.
H100 SXM basis, last 30 days
| Date | Benchmark | Tier | Level | GCX median | Basis | Basis % | Lowest card | Card basis | Source |
|---|---|---|---|---|---|---|---|---|---|
| No benchmark levels for H100 SXM in this window. | |||||||||
Methodology
Basis (market) = GCX median − benchmark. Basis (rate card) = lowest published rate card − benchmark. Basis % = basis ÷ benchmark. The GCX median is the median asking price across available on-demand Vast.ai listings in the hourly snapshot at the benchmark's hour (or up to 3 hours before it); rate cards are RunPod, Verda and Azure on-demand prices from the same snapshot.
Benchmark tiers (neocloud, hyperscaler, composite) are shown separately and never averaged together. Benchmark levels are entered by hand with a link to where they were published, and only where the publisher's terms allow it; they stay under the publisher's terms and are not covered by GridCompute's CC BY licence. Missing hours and days are shown as gaps, never estimated.
No futures settlements are shown: exchange settlement prices are licensed market data, and we don’t scrape exchange sites.
Data: /api/basis/latest · /api/basis/history · see also compute spark spread and the API reference.