GPU Cloud Reliability: Which providers report H100 stock? (2026)

The cheapest H100 isn't useful if you can't actually get one.

GPU Finder asks two separate questions: what does a provider report right now, and how often did it report an H100 as available during a defined historical window? Provider availability integrations are polled hourly and the history is stored for reliability research.

Below is the live picture for a one-GPU H100 configuration. The current-signal column is point-in-time; the reliability column is a rolling, duration-weighted 30-day measure. Because this table changes as new checks arrive, it is not evidence for a dated conclusion—cite a frozen Availability Index edition for that.

Four signals that should not be conflated

• Current availability is the latest provider-reported status for a configuration. available and limited are observations, not reservations or launch guarantees.
• Seven-day availability is the provider-by-day heatmap on GPU pages. It is useful for recent patterns but is not the 30-day score in the table above.
• Thirty-day reliability is the duration-weighted share of known covered hours reported available. limited and unavailable are known signals but do not count as available.
• Unknown availability means there is no usable provider signal. It is missing data, not zero reliability. Stale availability means the last signal is more than six hours old; the live UI treats it as unknown until a fresh check arrives.

How the rolling 30-day score works

The collector records available, limited, unavailable or unknown with a timestamp. History is change-deduplicated, with a heartbeat record for unchanged status, so record counts are not a valid denominator. Each row marks the start of an interval that holds until the next row (or the calculation time).

Reliability is the duration-weighted ratio: total time within the 30-day window where status was available, divided by total time covered. So a provider that's available for 6 days then drops for 1 day reads as ~86%, not 50%.

Important boundaries:

• Thirty-day window. This is a rolling live window, unlike the immutable reporting period in an Availability Index edition.
• Per-region aggregation. Configuration-hours are aggregated across tracked regions. A provider can have a current in-stock badge because one region reports stock while its historical aggregate is lower.
• 48-hour coverage floor. Below that, the live table shows — rather than manufacturing a percentage.
• Unknown is excluded. An unknown interval is missing coverage and does not count as downtime.
• Six-hour current-stale rule. A current signal older than the shared six-hour freshness threshold is labelled stale and treated as unknown. Historical heartbeat coverage uses a separate interval methodology described in the Index.

How to read a difference between providers

The score describes the signals GPU Finder observed, not fleet size, operational quality or a promise that a launch will succeed. A difference can reflect real capacity, the regions or node sizes exposed by an API, marketplace host churn, or gaps in provider reporting. Those factors should be checked against coverage hours and sample counts before drawing a conclusion.

A provider showing — is not assigned a zero-percent score. It either lacks a usable availability API or does not meet the coverage floor. The published price can still be valid while launch capacity remains opaque.

Where to use the data

Use the H100 live page when deciding where to look now: it combines current price, current availability, a 7-day heatmap and rolling 30-day reliability. Use the Availability Index for research, citation or reuse, because its reporting period, generation timestamp and downloads do not move after publication.

When recommending a provider or sending a user to deploy, GPU Finder uses stable /go/<provider> action links. Those are distinct from research citations: cite canonical report, dataset, methodology, GPU or API URLs, never a referral redirect.

Limitations and disclosure

Provider APIs are not standardized. Some report exact stock, some expose a catalogue state, and some provide no public capacity signal. Hourly polling can miss changes between checks. available or limited never guarantees that account limits, regional quotas or a race with another customer will permit a launch.

GPU Finder founder Alexander Salikov is employed by Nebius, one of the providers measured here. The project is independent, the same calculation applies to every provider, and the raw coverage fields are published for scrutiny. Some provider action links generate referral revenue at no extra cost to users; referral relationships never influence rankings, prices or availability results. See About and methodology.

Sources

• Availability signals come from provider APIs or provider-operated feeds and are normalized into the four statuses above.
• Rolling reliability uses stored AvailabilityRecord intervals; the monthly research calculation and data dictionary are published in the Availability Index.
• Live pricing is sourced separately from the SkyPilot catalog, provider APIs and public pricing pages. Price data is not included in the Availability Index download.
• Methodology, limitations, conflict and referral disclosures: About GPU Finder.

Got feedback or want a specific GPU added to the table? Email me or DM on LinkedIn.