GPU compute markets

Methodology

Where the figures come from, how they are made comparable, and their limits.

Sources

Every source is public and free to read. Each is polled on its own schedule, and a failed poll is retried after five minutes.

Loading…

Making prices comparable

Every source names and prices GPUs its own way.

  • One name per GPU. Each source's GPU name is matched to one model and memory size, so "RTX4090", "4090" and "GeForce RTX 4090 24GB" are the same GPU. Names that cannot be matched are left out rather than guessed.
  • Per GPU, per hour. Prices are US dollars for one GPU for one hour. Where a cloud only sells whole machines, the machine's price is divided by its GPUs, and the smallest machine carrying the GPU is used, as the closest match to renting one GPU elsewhere. A month is 730 hours. Price per GB of VRAM divides the hourly price by the GPU's memory.
  • One price per provider and tier. When a provider lists a GPU several times in one tier, such as in different regions, the lowest price is shown and the full range is kept beside it.
  • Tiers are kept apart. On-demand, community, marketplace, serverless and spot prices are different products. Medians and best deals leave out spot capacity, which can be taken back at any time, and price indexes, which cannot be rented.
  • Machines and regions. Shadeform, Spheron, AWS and Azure describe whole machines: GPU count, CPUs, memory, disk and region. Each region is placed in an area, with the EU kept apart from the rest of Europe for data protection. A machine sold in several regions of one area is shown once, at its cheapest.
  • Checks on each price. A listing is left out of lowest prices, best deals and comparisons when it cannot be had: sold out, or from a source that has missed three polls in a row. Prices more than three times below or above a GPU's median are flagged, not hidden: on open marketplaces a very low price is often real.
  • History is averaged. Charts average each listing's prices over buckets, from 30 minutes for the last day to a week for five years, so lines stay readable. A gap in a line means the listing was missing, not that it was free.

Keep in mind

A listed price is not always what a rental ends up costing.

  • Listed prices are asking prices. Ornn's index is the exception: it is built from prices actually paid.
  • Marketplaces such as Akash and Vast.ai are open to anyone with a GPU, so their cheapest offers can be unusual hardware, poorly connected or gone within minutes.
  • Taxes, storage, network transfer and minimum billing periods are not included. Region is not shown, though prices often differ by region.
  • A listing can be priced while no capacity is free. Stock is shown where the source reports it.
  • Always check the final price on the provider's own site.