We collect published cloud GPU rates and normalize every listing to USD per GPU per hour. The headline figure for each listing is total_price_usd_per_hour — the instance total divided by GPU count. This is always explainable.
The core index median is built from comparable on-demand listings. Reservation products such as AWS Capacity Blocks, spot or interruptible listings, and other non-comparable billing modes can still appear in live tables, but they are labeled by billing and excluded from the median.
EST.When a bundled instance (e.g. a hyperscaler node) can't be separated into a GPU component, we show not separable rather than a fabricated number.
Each listing is tagged Reliable, Community, or Experimental. Only index-eligible listings enter the median. A listing is excluded with an explicit reason — community/experimental tier, unknown form factor, aggregator source, reservation or spot billing, suspicious validation, non-comparable serverless, or staleness over 12 hours.
For SKUs that mostly trade on marketplaces, we may also show a separately labeled Community marketplace median. It is calculated from one cheapest current community-tier row per non-aggregator provider. Aggregators such as Shadeform are excluded from this median to avoid double counting underlying providers. This metric is not the core reliable index median.
< 2h — fresh, in the median.2–12h — stale warning, still in the median.12–24h — excluded from the median, still shown.> 24h — hidden by default.Historical curves are built from immutable, billing-mode-separated daily snapshots. Each version carries its methodology version and content hash. Missing dates remain gaps; we do not interpolate them or publish Wayback-derived observations.
A 30-day range requires 24 observed days with no gap over 3 days; 90 days requires 72 observed days with no gap over 7 days; 180 days requires 144 observed days with no gap over 14 days. In every window, at least three providers must be present on 80% of observed days. Archive observations are excluded from public calculations. A failed coverage, gap, provider, rights, provenance, or classification gate returns insufficient evidence and no supported range.
Historical collection and redistribution are denied by default. A provider can enter the analysis only after its rights record documents the source, permitted fields, collection method, attribution, retention, redistribution scope, and approval evidence.
This is an informational index of published list prices, not a settlement-grade transaction benchmark. It does not establish negotiated prices, deployment availability, or ancillary costs. Verify the current rate and terms with the provider before provisioning.
How are cloud GPU prices normalized?
Every listing is normalized to USD per GPU per hour — the instance total divided by GPU count. Currency is converted to USD. This figure is always present and always explainable.
What does "index-eligible" mean?
Only reliable-tier on-demand listings with a clear form factor, from non-aggregator sources, that pass validation and freshness checks enter the index median. Each excluded listing carries an explicit reason.
How often are prices updated?
Prices are scraped hourly. The current index median reflects each provider’s cheapest eligible listing from its most recent scrape.
Is this a transaction price index?
No. These are published rates for developers comparing rental options — not settlement-grade transaction benchmarks. List price differs from negotiated price.
Where does the price history come from?
History accrues from permitted live observations after launch. Missing dates remain explicit gaps; we do not interpolate or publish Wayback-derived observations.