Our approach / Data & methodology

Better decisions.
Start with better context.

What our numbers mean, where they come from, and how to use them. Every benchmark needs a little context.

01 /

Three kinds of price, never mixed

Reference prices are weekly aggregates from licensed datasets. They are the only numbers used for trend charts.

Listed prices are what a provider publishes on its own pricing page, shown with the source and the time we observed it. They are not inventory.

Broker-verified capacity is supplier-submitted and checked by our desk. We show only an anonymized summary and a price range.

02 /

Units and billing types

Prices are normalized to USD per GPU-hour when the source makes that conversion reliable. When it does not, we keep the original unit and do not compare it with others.

On-demand, spot and reserved prices are always reported separately. Reserved prices state their commitment term.

03 /

Sample size and staleness

Every weekly figure carries its observation count. Weeks with fewer than five observations are marked low sample; fewer than that, and the trend is hidden.

If a source fails, we keep the last valid data and mark it stale rather than showing a gap or an invented value.

04 /

News and AI assistance

News items from official and licensed feeds are grouped into events. A model drafts short summaries, and every fact must point to a source passage.

Forecasts, price attribution and rankings are always reviewed by an editor before publication.

See the list of sources and licenses.