The cheapest GPU cloud, one table.

Spare GPU capacity a cloud provider sells cheaply because it isn't currently reserved by another customer. The trade-off is that the instance can be reclaimed with little notice if the provider needs the capacity back. For batch training, fine-tuning, and fault-tolerant inference, spot pricing typically runs 50 to 80 percent below the same card's on-demand rate.

Spot pricing works like an auction: it floats with supply and demand for that specific GPU, in that specific region, at that specific moment. When more capacity sits idle, prices drop. When a provider's supply tightens, or when a burst of training jobs starts up, prices climb. That is also why the same GPU model can show wildly different prices across providers at the same time — each provider's pool of idle capacity is separate.

From a live feed of spot offers across RunPod, Vast.ai, AWS Spot, Azure Spot and Crusoe Cloud — the same feed behind AIVory's GPU marketplace and every individual GPU page on this site. For each GPU model we show the single cheapest live offer among its variants. The table is server-rendered from a snapshot at build time, then refreshed in your browser from the live feed where possible. See the methodology section above for the full explanation, including what this list doesn't cover.

It changes by the hour, which is the entire point of this page — check the table above instead of trusting a fixed answer. As a rule of thumb, consumer cards on peer-to-peer marketplaces like Vast.ai tend to undercut datacenter-only providers on older Ampere and Ada cards, while RunPod is frequently the cheapest source for newer Hopper and Blackwell datacenter GPUs.

The provider ends your session, usually with a short warning window. If you're renting directly, you restart on the next cheapest available offer yourself. If you're using Smart Inference, the router can fail over automatically to another provider so a reclaimed instance doesn't take your API down with it.

Match VRAM to model size first, price second. A quantized model under 13B fits an RTX 4090, L4, or T4. A 70B model wants at least one H100 or A100 80GB, more if you're running it unquantized. Anything above 200B parameters needs a multi-GPU cluster regardless of which card you choose. See the "reading the table" section above for the full breakdown by architecture.