A starting point for a
more transparent GPU market.
A rental price only becomes useful when you know what hardware, terms, and available supply it describes.
The first observation
At 04:35 UTC on Sep 13, 2026, our first H100 SXM 80 GB snapshot contained 9 eligible single-GPU offer configurations from 6 providers. The provider-floor median was $3.08 per GPU-hour, with qualifying provider floors ranging from $2.79 to $4.29.
| Provider | Lowest eligible quote | Region |
|---|---|---|
| Lyceum | $2.7900 / GPU-hour | Europe |
| Massed Compute | $2.8900 / GPU-hour | us-central-12 |
| Ori | $2.9000 / GPU-hour | dallas-2 |
| VERDA | $3.2500 / GPU-hour | FIN-HEL |
| RunPod | $3.4900 / GPU-hour | global |
| Lambda Labs | $4.2900 / GPU-hour | us-south-3 |
Why the basket matters
An H100 PCIe offer is not an H100 SXM offer. A price per GPU inside an eight-GPU instance does not mean that a single GPU can be rented at that price. The initial series fixes the model, memory, instance size and pricing lane before calculating a reference.
A reference, not a transaction
The lowest provider quote is an observed floor. The median across providers describes the middle of those floors, with equal provider weight. It does not estimate the price of a typical completed transaction, the amount of supply, or every user’s total bill.
From an observation to a time series
This first point establishes a record. Future changes require future observations. We have not backfilled this series with older catalog averages that mix unavailable stock, spot offers and different configurations.
The complete methodology describes those decisions and the data API provides access to the daily record. This note remains tied to the first saved observation even as the live dashboard changes.