52 providers priced today, 2026-09-18. Each one has its own page: every GPU it sells against the market median, its price history, and where it is the cheapest place to rent. Sorted by how many GPUs each prices — the ones at the top are the ones you can compare across the most hardware.
vs market is the mean, over the GPUs a provider sells, of how far its on-demand median sits from the index median for that GPU. It is a rough summary, not a ranking: a provider that is 30% below on the A100 and 30% above on the H100 averages to zero. The per-provider page shows each GPU on its own line.
Hyperscaler prices bundle CPU, RAM and network into the instance; neoclouds sell the GPU more directly; marketplaces are independent hosts, so their price is a spread. Read the methodology for how each number is produced.
There is no single answer, which is why every provider has a page. On 2026-09-18, among providers pricing three or more GPUs, the lowest average against the market is GMI Cloud (-41.2%), then SaladCloud (-36.1%), Packet.ai (-33.9%), RunPod Community Cloud (-30.8%). io.net is the single cheapest place for the most GPUs (7). Marketplaces win on price and lose on guarantees; check the provider's kind before comparing.
RunPod Community Cloud prices 25 of the GPUs we track today, followed by RunPod Secure Cloud (23), Vast.ai (23), GPU.ai (21). Coverage is what makes a provider comparable: the more GPUs it sells, the more of the market it can be measured against.
A hyperscaler (AWS, Azure, Oracle, OVHcloud, DigitalOcean, Vultr, Scaleway) sells an instance — CPU, RAM, storage, network — with the GPU inside it, priced for the ecosystem. A neocloud (Lambda, Nebius, CoreWeave, RunPod Secure, Crusoe, and most of the list) sells GPUs as the product, priced per GPU-hour. A marketplace (Vast.ai, RunPod Community, io.net, Akash, SaladCloud, TensorDock, GPU.ai, Spheron) is many independent hosts, so its price is a spread and its cheapest listing is one machine.
For each GPU a provider prices, its own on-demand median is compared with the index median for that GPU; the percentages are then averaged over its GPUs. It is a summary, not a ranking — a provider 30% under on one GPU and 30% over on another averages to zero. The per-provider page shows every GPU separately.
Providers are added when their public pricing can be read reliably and checked by hand; a provider whose page changes shape drops to zero rather than showing garbage, and reappears when the reader is fixed. If you run a GPU cloud with published prices and are not listed, or a price here does not match your page, write to us with the URL of the pricing page.