GPU Finder Research
GPU Cloud Availability Index
Monthly, citation-ready GPU cloud availability research with provider reliability rankings, coverage, current stock signals and downloadable data.
Published by GPU Finder; authored by Alexander Salikov. New editions are frozen monthly after coverage review.
Latest edition
GPU Cloud Availability Index — August 2026
Covers 1 August 2026 through 24 August 2026. This first edition is a frozen month-to-date release because no complete earlier calendar month remains in the retained history.
What the index measures
The index measures how much of known, provider-reported time a tracked GPU configuration was marked available. It keeps current availability, historical reliability and data coverage separate: a current signal is a point-in-time report, while reliability is a duration-weighted period measure.
Covered GPU models
H100, H200, B200, A100. Results appear only where retained provider signals support them. The live catalogue remains available on the GPU model index.
Methodology
Version 1.0 treats unknown as missing data, caps a stored signal at 30 hours, and excludes rows below 48 covered hours or 50% coverage from rankings. Current signals older than six hours are labelled stale. available and limited are provider-reported signals, not launch guarantees.
Data limitations
Availability APIs differ by provider, region and GPU count. Providers without a usable signal remain unknown. High-churn sources can have shorter retention and may be marked insufficient. The export contains availability only—no prices or internal database identifiers.
Published editions
Frozen 24 August 2026 · partial-period edition
Data dictionary
| Field | Meaning |
|---|---|
| reportingMonth | Edition key in YYYY-MM form. |
| reportingPeriodStart | Inclusive UTC start of the measured period. |
| reportingPeriodEnd | Exclusive UTC end of a complete period, or the freeze timestamp for a partial period. |
| generatedAt | UTC timestamp when the immutable edition was frozen. |
| gpuModel | GPU Finder canonical GPU model. |
| providerName | Visible provider name. |
| providerSlug | Stable GPU Finder provider slug. |
| gpuCount | Physical GPUs in the listed instance or node. |
| aggregationScope | all_tracked_regions: provider/configuration results are pooled across tracked regions. |
| regionCount | Number of tracked region labels contributing records. |
| currentAvailabilityStatus | Frozen current signal: available, limited, unavailable, unknown, or stale. |
| availabilityCheckedAt | Most recent contributing availability check, or null. |
| reliabilityPct | Duration-weighted available hours / known covered hours; null below the coverage threshold. |
| coverageHours | Known provider-signal hours summed across contributing configurations. |
| expectedCoverageHours | Reporting-period hours multiplied by contributing configurations. |
| coveragePct | coverageHours / expectedCoverageHours × 100. |
| observationCount | Stored change or heartbeat records inside the reporting period; not the reliability denominator. |
| dataCompleteness | complete, partial, insufficient_coverage, or unknown. |
| methodologyVersion | Version of the Availability Index calculation contract. |
| sourceUrl | Canonical HTML research edition for citation. |
Citation and reuse
Cite the canonical HTML edition for conclusions and the CSV or JSON distribution for record-level reuse. Do not use provider action redirects as research citations.
Suggested hub citation: GPU Finder, “GPU Cloud Availability Index,” methodology version 1.0, https://gpufinder.dev/research/gpu-cloud-availability-index.
Availability-only exports are governed by the Terms of Service. A separate explicit open-data license still requires owner approval.