Every number has
a methodology.

A reference price should be understandable, reproducible, and clear about what it leaves out. Here is how ours works.

01 / Define a comparable basket

Each series covers one exact GPU model and memory size, in a single-GPU instance, with on-demand billing. H100 SXM 80 GB and H200 SXM 141 GB are separate series. We never combine PCIe, SXM, different memory capacities, or multi-GPU instance sizes into the same reference.

B300 SXM uses NVIDIA’s 288 GB nominal memory specification. The existing source catalog labels the full b300-sxm model as 262 GB; we accept that legacy label and 288 GB for this exact SKU only. GB300 and MIG partitions remain separate and are excluded. Quote exports preserve the original catalog memory value alongside the nominal value. See the NVIDIA B300 specification. RTX A6000 (Ampere, 48 GB) and RTX PRO 6000 Blackwell (96 GB) are separate series.

02 / Give each provider equal weight

For each provider, select the lowest qualifying price across its regions and offers. Then take the median of those provider floors. A provider with a hundred listings receives the same weight as a provider with one.

This is a provider-floor median, not an average transaction price or an estimate of what every customer pays. Provider entries are distinct names in the source catalog; they are not a verified count of economically independent operators.

03 / Publish only with enough coverage

A reference needs at least three qualifying providers. With one or two providers, we publish the observed quotes and label the series “Limited.” With no qualifying supply, the reference is unavailable. We never fill a missing quote with an old price, a zero, or an estimate.

Quote and provider counts describe observed configurations. They do not measure physical GPU inventory, guaranteed capacity, or market share.

04 / Keep an honest historical record

The live reference updates from recent source observations. The daily archive records one observation per UTC day and preserves its capture timestamp. A seven-day change requires a published observation on the corresponding date seven days earlier. Missing dates remain missing; lines and returns are not extrapolated across them.

History starts with the first observation captured under this methodology. Earlier GPUPerHour catalog history mixes billing and availability states and is not included in these benchmark series. Methodology versions are recorded with every snapshot; incompatible versions are not silently joined.

A newly added series has no historical rows before it was recorded. If the live source is unavailable and an older snapshot predates that series, it is labelled “Not recorded”; absence from that snapshot is not reported as zero availability.

What the price does not include

CPU, RAM, storage, network transfer, taxes, credits, billing increments and provider-specific terms can differ. “Secure” reflects the source listing’s classification, not an independent security audit. Availability is an observation, not a reservation; verify the final bill and capacity with the provider.

We use the earlier upstream stock timestamp when it is supplied. Otherwise, freshness reflects the last successful observation recorded by GPUPerHour. Source collection and display caches can add delay.

Ownership and corrections

GPUIndexes is operated by Quanta Cloud LLC, which also operates GPUPerHour and DeployGPU and offers GPU infrastructure. QuantaCloud listings receive the same eligibility rules and provider weight as other listings. We do not describe this benchmark as independent of those commercial interests.

Source prices come from GPUPerHour’s provider observations. Each published quote includes its source offer identifier, region and observation time. To report an issue, use GPUPerHour’s contact page and include the index, timestamp and source offer.