Methodology
How the numbers are made.
Where compute data comes from, how it is compared and how we decide what to show. No black boxes.
The pipeline
From source to screen.
Collect
Offerings come from tracked provider records: public pricing and documentation pages, provider submissions and reviewed source packages. Each record keeps its source.
Normalise
The original value and unit are always kept. A normalised value, such as price per GPU-hour, is stored separately so comparisons can be checked.
Grade
Material facts carry a source, a date and a confidence grade from A to D. Missing data stays unknown and is never estimated silently.
Publish and monitor
Only public, current, non-archived records appear as market data. Age is always visible and stale evidence stops counting.
Freshness
Old data looks old.
The live market view counts a record as current only if it was checked in the last 30 days. Supply evidence from providers ages on a shorter, fixed clock.
Provider capacity and availability
- Fresh · 0–3 days
- Aging · 3–7 days
- Stale · 7+ days
Provider price evidence
- Fresh · 0–14 days
- Aging · 14–30 days
- Stale · 30+ days
Confidence grades
How sure are we?
- Grades describe evidence, not price.A grade says how well a fact is supported by its source, not whether the price is good.
- Claims stay claims.A provider’s own submission is shown as an attributable claim, not as verified market truth.
- Unknown is a valid answer.If a value is not supported, it stays unknown rather than being filled in.
The four grades
Compute Index
A benchmark that can’t be gamed by volume.
Raw listings do not vote equally. The Index first reduces each provider to one contribution, then takes the reference level across independent providers, so a provider with many listings cannot outweigh the rest.
Eligibility
Exact GPU, commercial basis, geography, freshness, confidence grade and a comparable normalised price are checked explicitly.
Provider contribution
Each provider’s latest eligible observations become a single contribution.
Cross-provider fixing
The median, quartiles, range and dispersion are calculated across provider contributions.
Published or indicative
Series states are explicit. A candidate is not a reference rate, and corrections are published as new revisions rather than rewriting history.
Availability and capacity are separate evidence layers. A listing is never treated as reserved capacity, and unknown availability is never converted into unavailable.
Read the full rules.
The complete Index methodology, versions and the trust model are public.