Cheapest GPU cloud by GPU class: lowest published price per GPU-hour

“Cheapest” depends on what you will accept in return. This page separates the lowest price you can get with no strings attached from the lower prices that come with conditions.

How the comparison works

1. First table. The lowest on-demand list price per GPU class across providers in the GPU Cloud Index: no interruption, no term, no third-party host. This is the fair like-for-like comparison.
2. Second table. The lowest price per GPU class once you accept a condition: spot capacity that can be reclaimed, a community host, or a reservation where the provider says “starting from”.
3. Marketplace snapshot. Vast.ai prices are set by hosts and move all the time, so they are shown separately as the lowest and median listing at the time we checked.

Lambda is missing from the on-demand table on purpose. Its current on-demand prices are visible only after login, and the public list takes effect on 16 November 2026. Announced prices are never shown as current.

Reading the result

• The cheapest row is not always the cheapest bill. A provider may require a whole 8-GPU node (CoreWeave), bill CPU, memory or idle storage separately, or have a minimum billing period (DigitalOcean: 5 minutes). Check the row’s provider in the index before deciding.
• SXM and PCIe are different classes. A cheap H100 PCIe row is not equivalent to H100 SXM for multi-GPU training. See renting an H100.
• Spot prices are per region and can change. CoreWeave’s figures are its North America table.
• A low price does not mean capacity. We record whether you can launch yourself, not live stock.

When the cheapest option is wrong

If a run takes days and cannot checkpoint, a reclaimed spot instance or a failed community host costs more than the hourly saving. For inference with idle periods, a serverless platform can cost less overall despite a higher hourly rate.