Validation report · July 15, 2026

H100 price persistence

Is today's cheapest published rate stable enough for a budget assumption? The first H100 review cannot support that conclusion yet.

Result: insufficient evidence

No 30-, 90-, or 180-day H100 range is published. A July 15 read-only preflight found zero rights-cleared observations, 41 eligible live H100 days before the rights gate, and no day with more than two providers. The method requires at least three providers on 80% of observed days.

Decision

The latest cheapest listing and a historically supported range answer different questions. A current listing can be inspected on the H100 SXM price page; it is not promoted here as a budget input. With no rights-cleared historical observations, the report cannot calculate a median, percentile band, floor persistence, or supported budget range.

Window review

WindowMinimum continuityResultSupported range
30 days24 observed days; maximum 3-day gapInsufficientWithheld
90 days72 observed days; maximum 7-day gapInsufficientWithheld
180 days144 observed days; maximum 14-day gapInsufficientWithheld

Missing calendar intervals cannot be derived from a public-eligible series because that series currently contains no observations. They are not filled from archives or interpolated. The 41-day preflight count describes rows before the rights gate; it is not presented as report coverage.

Billing-mode review

Each billing mode is a separate series: On-demand, Spot, Interruptible, Reservation / capacity block, Serverless, Cluster. None has passed the evidence gate. Unknown billing and ambiguous H100 variants are excluded rather than blended with H100 SXM, PCIe, or NVL.

Provider changes and revisions

No provider entry or exit effect is reported: without an eligible public series, there is no defensible before-and-after median to decompose. No verified corrections were present in the preflight. Future corrections will append a new version linked to the prior record; source observations will not be overwritten.

Provider-verification checklist

What would change this result

A window can publish only after field-level collection and redistribution rights are recorded, its continuity threshold is met, and at least three independent providers appear on 80% of observed days. Passing reports will show p25, median, p75, calendar coverage, explicit gaps, provider entry and exit effects, floor persistence, revisions, sources, and the latest listing separately.

See the versioned method and JSON contract. Published rates still would not establish availability, realized transaction cost, workload performance, reliability, ancillary costs, or future prices.

Artifact acceptance record