Methodology

How the data is built

Every number on the terminal traces back to a timestamped, normalized snapshot. Here is the pipeline.

01

Collect

Connectors pull on-demand and reserved pricing, regional availability, and capacity signals from each provider's public pricing surfaces on a recurring schedule.

02

Normalize

Raw rates are converted to a common unit — USD per GPU-hour — and mapped to a canonical GPU model, provider, and region taxonomy so quotes are directly comparable.

03

Snapshot

Every normalized quote is written as an immutable, timestamped snapshot. Nothing is overwritten, so the full price and capacity history is preserved.

04

Index

Snapshots roll up into market metrics: average rates by model, spot and reserved indices (basis 100), regional spreads, and week-over-week change.

Pricing units

All rates are expressed as USD per GPU-hour. On-demand reflects the spot/hourly list rate; reserved reflects the effective hourly rate of a monthly committed term. The reserved discount is the spread between the two.

Coverage

Flopdesk tracks NVIDIA H100, H200, B200, A100, L40S, and RTX 6000 Ada, plus AMD Instinct MI300X and MI325X, across 10 providers and 12 regions. Coverage expands as new accelerators and providers come to market.

Freshness & limitations

Snapshots are captured daily; the dashboard surfaces the last-capture time on every view. Prices are indicative, aggregated from public sources, and may lag real-time provider changes. They are provided for market intelligence and are not investment advice.