Thorpe Methodology

Data

Thorpe maintains a proprietary multi-source pipeline for GPU pricing and availability. Inputs are normalized into standardized benchmark groups, retained as append-only observations, and reconciled against raw payloads to support full auditability of every published benchmark value.

Benchmark construction

Thorpe groups observed listings into standardized benchmark families based on GPU type, configuration, and related market attributes. Values are calculated using documented grouping and normalization rules designed to support comparability over time.

Tier 1 — Required match

• ·GPU type must match (handling naming variants like A100_80G, H100_SXM, RTX_A6000)
• ·GPU count must match exactly (1×, 2×, 4×, 8×)
• ·Only on-demand pricing model is included

Tier 2 — Confidence scoring

• ·Exact VRAM match: 100% confidence
• ·Close VRAM match (within 10%): 80% confidence
• ·Unknown VRAM: 70% confidence
• ·Significant VRAM mismatch: excluded

Tier 3 — Context fields (display only)

• ·vCPUs, RAM, storage, boot time, interconnect type
• ·Displayed for reference but not used as hard matching criteria

Update cadence

Thorpe collects and appends observations every 4 hours. Published benchmark views are updated on that cadence unless otherwise stated. Historical data is retained indefinitely.

Caveats and exclusions

Observed listings are normalized into benchmark groups, but GPU markets remain heterogeneous across provider configuration, region, availability, and commercial terms. Users should evaluate benchmark outputs alongside the documented methodology and current coverage.

• ·Coverage is not exhaustive — Thorpe tracks providers available through current data sources, not every GPU cloud provider
• ·On-demand only — spot, reserved, and enterprise contract pricing is not included in benchmark indices
• ·Availability is volatile — a listing marked available at ingestion time may no longer be available minutes later
• ·Same GPU ≠ same machine — two listings in the same benchmark group may differ in vCPUs, RAM, storage, and interconnect
• ·4-hour snapshot granularity — data reflects pricing at ingestion time, not continuous real-time pricing

Change control and incident policy

Material changes to benchmark construction, source coverage, or calculation logic are documented through versioned methodology updates and benchmark notices. If a data or calculation issue is identified that could materially affect published benchmark values, the issue is reviewed, documented, and communicated through a formal notice process.