How getflops.ai normalizes GPU cloud pricing
The site compares cloud GPU rental markets by reducing provider-specific catalogs into per-GPU hourly snapshots, then grouping those snapshots by GPU, provider, and pricing type.
Provider data sources
Collectors use public catalogs where available and provider APIs where credentials are required. Each run stores a fresh timestamped snapshot instead of mutating old price rows, so trends can be reconstructed from historical snapshots.
Hyperscaler and OCI list pricing
AWS, Azure, GCP, and Oracle-style rows are normalized from published pricing sources into comparable hourly GPU rows.
Marketplace and dedicated GPU APIs
Vast.ai, RunPod, and Lambda rows are collected from provider APIs and reduced to current GPU/pricing-type snapshots.
Daily cadence
Provider collectors run independently so one slow or failing source does not block every other provider from updating.
What the price columns mean
The comparison tables focus on dollars per GPU hour. When a provider lists multi-GPU instances, the total hourly instance price is divided by the number of GPUs so it can be compared against single-GPU marketplace listings.
| Field | Meaning |
|---|---|
| min | Lowest observed per-GPU hourly row in the latest provider/GPU/pricing-type snapshot. |
| median | The primary comparison value, chosen because marketplace outliers can distort means. |
| mean | Average per-GPU hourly price across rows in the same snapshot. |
| max | Highest observed per-GPU hourly row in the latest snapshot. |
| num_offers | Number of rows contributing to the current aggregate. For list-price providers this may be one. |
How to read the numbers
Price is not total cost
Networking, storage, commitment discounts, support, quotas, and deployment friction can change the real cost of a workload.
Availability moves quickly
Marketplace prices can change between collection runs. Treat low rows as current signals, not guaranteed reservations.
Pricing types are separate
On-demand, spot, community, and reserved-style rows should be compared within their own risk and commitment model.