Space · GPU
GPU rental prices
Public GPU rental prices from seven sources, in USD per GPU-hour. Every number keeps its source, its date and the configuration it was quoted for.
Research compilation, not a quote. List prices, indices and transaction prices are different bases. None of them is a multi-year contract price, and nothing here is investment advice. Cite the source and the date with any number. Commercial reuse, or using an index in a financial product, needs a license from the data owner.
How fresh each source is
The data date is the newest data point a source holds, not the day we fetched it. A source is marked stale when its data is older than about twice its usual update interval: four days for daily sources, two weeks for weekly ones.
Model
Price history
Colour shows the source. Line style shows the statistic: solid for an index or a median, dotted for the lowest price, dashed for spot and reserved medians. Select a legend item to hide or show its line.
Market summary by billing type
GetDeploying, latest weekly file: the lowest price and the median of each provider's lowest price, for every billing type and term, with how many providers and listings sit behind each figure.
Latest prices by provider
Per-provider list prices from American Compute's compilation of provider websites, dstack's open-source catalog of provider price APIs (AWS, Google Cloud, Lambda, Nebius, Runpod, Verda and CloudRift), the Azure Retail Prices API and Oracle's price list. Each row shows the configuration the price was quoted for and the age of the quote. Azure and dstack rows show the cheapest region, with the number of regions and the highest regional price.
Medians by configuration
Medians never mix configurations. Each row is one billing type, GPU count and form factor under the filters above. Each provider's lowest price counts once per row, and flagged rows are left out.
Sources and licenses
How the numbers are built
- A list price is not a transaction price. American Compute, GetDeploying, the dstack catalog, Azure and Oracle publish list prices. Ornn is a transaction-price index. Silicon Data is an index built from normalized rental records.
- GetDeploying and Silicon Data do not split H100 into SXM and PCIe, or A100 into 40GB and 80GB; this page files them under H100 SXM and A100 80GB. American Compute, dstack, Azure and Oracle rows carry their own configuration.
- Configuration comes from the source. American Compute gives GPUs per instance, interconnect and region where the provider states them; form factor and deployment are read from the instance name and stay blank when it says nothing. Azure sizes and Oracle shapes are mapped from their published specifications. The dstack catalog states GPUs, vCPU and RAM per instance (and disk for some providers); SXM is read from instance names and families such as AWS P5 and Google Cloud A3, and fractional GPUs are left out. vCPU, RAM and storage appear only where a source states them.
- Azure reservation prices are published as a total for the term and are divided by the hours in the term. Windows, Low Priority and Dedicated Host prices are left out. Oracle publishes one price per GPU-hour for each GPU type.
- Each observation is stored once per source, statistic, model, billing, term, provider and data date; list prices also keep the instance, region and price. A second run on the same day overwrites the observation instead of adding a row. dstack listings keep one row per on-demand price and one current row per spot listing; the daily dstack medians keep the history.
- The payback model starts its market rate from GetDeploying's on-demand provider median for the latest week. Where GetDeploying has no current figure for a model, it uses Silicon Data's neocloud index, then Ornn's transaction index, then American Compute's on-demand median, then the dstack catalog's on-demand median. Any series on this page can be chosen there instead.
Attribution: GetDeploying (CC BY 4.0); American Compute GPU cloud price comparison; Ornn Compute Price Index; Silicon Data indices; Azure Retail Prices API; Oracle Cloud price list; dstack gpuhunt GPU catalog (MPL-2.0), built from the providers' own price APIs. Full Ornn and Silicon Data history, and commercial use, need their own licenses. This page keeps only values observed on public pages, for research.