People searching for the best GPU rental sites usually want one hourly number. The SERP for gpu rental prices is a commercial query: OpenSEO has it at about 210 searches a month, difficulty 26, and a high CPC. That is a buyer query, so this note stays on what a renter can check.
I compared ten sites that show up whenever renters argue about cloud GPUs: NovaPod, Vast.ai, RunPod, Lambda, CoreWeave, Paperspace, TensorDock, Genesis Cloud, Crusoe, and Spheron. I did not invent their live asks. A price I did not read on that site, on that day, is not in the table.
What separates the best GPU rental sites
Three things change the bill before the GPU name does. Who you pay. Whether the clock is a second, a minute, or a reserved month. Whether you get a marketplace offer, a template, or a contract cluster.
A site with a cheap card and a one-hour minimum can cost more than a site with a higher ask and per-second billing. A site that only takes a card is a different product from one that takes USDT on TRON. Those are product differences, not slogans.
| Site | How you pay | Clock | How you launch |
|---|---|---|---|
| NovaPod | USDT, USDC, SOL, ETH, BNB | Per second | Live marketplace and templates |
| Vast.ai | Card, with crypto options | Per second, host asks | Host marketplace |
| RunPod | Card | Per second | Pods and serverless |
| Lambda | Card or invoice | Hourly and reserved | Instances |
| CoreWeave | Contract | Reserved and on-demand | Kubernetes clusters |
| Paperspace | Card | Hourly | Notebooks and machines |
| TensorDock | Card | Marketplace | Host offers |
| Genesis Cloud | Card | Hourly | EU instances |
| Crusoe | Contract or card | Capacity deals | Energy-backed instances |
| Spheron | Crypto | Marketplace | Decentralized offers |
Price advantage is the clock, not a slogan
On November 5, 2026 the NovaPod book had an RTX 4090 from $0.09 an hour, 24 GB, median $0.35, 419 offers. An RTX 5090 from $0.26, 32 GB. An H100 SXM from $1.21, median $1.68, 65 offers, 80 GB. An H200 from $2.76, 141 GB. A B200 was listed from $4.38, 192 GB. I did not rent the B200 that day.
The floor is one host. The median is the middle of that book. I do not average them, and I do not paste another company's homepage into this paragraph. If you need Vast.ai or RunPod's number, open their market and write down the offer you can actually launch.
NovaPod's price advantage, when it exists, is a short job on a live ask, paid in crypto, stopped when the batch ends. A reserved CoreWeave or Lambda cluster is a different purchase. It can be cheaper per hour at a scale I was not buying.
Product advantage is what boots
NovaPod is the crypto path: deposit USDT, USDC, SOL, ETH, or BNB, then launch a template such as vLLM, ComfyUI, or PyTorch, or take a root KVM machine on the same ask. Billing is per second. There is no seat fee in that flow.
Vast.ai and TensorDock are marketplaces of other people's machines. The catalog is wide, including odd cards. You are picking a host, not a curated image. RunPod is the one I would send someone who wants a pod and a serverless endpoint on a card. Lambda and Paperspace are closer to a normal cloud account. CoreWeave and Crusoe are where a team with a contract goes. Genesis Cloud matters if the machine has to sit in Europe. Spheron is another crypto-native book, not a copy of this one.
How I would pick among the ten
Match memory first. A 24 GB card and an 80 GB card are not substitutes. Then pick the payment rail you already have. Then read the offer you can launch today, not a blog's starting price from another week.
If the job is a short LoRA, a ComfyUI batch, or a vLLM test, I start on NovaPod and stop the pod when it is done. If I already have a RunPod serverless script, I do not rewrite it for a two-hour test. If I need a year of H100s on a purchase order, I am not in this list's spot market at all.
Questions renters ask
Which of these is the cheapest GPU rental site? The one whose live offer fits the model and stops when you stop. A floor from another day is not a quote.
Does a comparison page replace the marketplace? No. Use the GPU list to see the card, then the marketplace for the machine you can launch.
Why are only NovaPod numbers dated? Because those are the asks I read on this site on November 5, 2026. I will not invent the others.
- GPU rental prices, side by side
- NovaPod and Vast.ai
- RunPod alternatives
- Live GPU marketplace
- GPU models and asks