Vast.ai
A bidding marketplace for GPU rental, where independent operators list spare hardware and prices land well below the major clouds.
Pricing Plans
| Plan | Price/mo | CPU | RAM | Storage | Notes |
|---|---|---|---|---|---|
| RTX 3090 (interruptible) | — | 8 vCPU | 32GB | 100GB | Often under $0.20/hr |
| RTX 4090 (on-demand) | — | 12 vCPU | 48GB | 150GB | Around $0.30/hr |
| H100 SXM (on-demand) | — | 24 vCPU | 200GB | 500GB | Around $2.20/hr |
· Updated
About Vast.ai
Vast.ai is a marketplace rather than a host. Anyone with GPUs, from a data center to someone with a rack of 3090s, can list capacity, and you rent it. The result is the cheapest GPU compute available anywhere, and the most variable.
Machines carry a reliability score, a measured internet speed, and a datacenter or hosted flag. Reading those before you rent is the whole skill. A cheap listing on a residential connection with a 0.9 reliability score will cost you more in failed runs than a dearer one in a real facility.
Interruptible instances work like spot pricing. You set a bid, and you keep the machine until someone outbids you. For a training job that checkpoints every few minutes this is close to free money. For anything stateful it is a way to lose work.
Pricing
There is no monthly plan. RTX 3090s frequently go for under $0.20/hr interruptible, RTX 4090s around $0.30/hr, and H100s around $2.20/hr, which undercuts the mainstream clouds substantially. Storage and bandwidth are billed by the individual host and vary, so read the listing.
Prices here move by the day, because they are set by supply. Treat any figure as a snapshot.
Service Details
Instances run as Docker containers with SSH and Jupyter access. You supply an image or use one of the PyTorch and TensorFlow templates. The CLI is the practical interface: you can search offers by GPU type, price ceiling, reliability, and bandwidth, then create an instance from the result in one command.
The honest caveat is trust. You are running your workload on hardware belonging to a stranger, on a machine you do not control at the hypervisor level. Do not put sensitive data or proprietary weights on Vast without thinking hard about it. For hobby training, fine-tuning open models, and batch inference, the economics are hard to argue with.
Available in: Global, distributed across independent operators
Learn More
For more information about Vast.ai, visit their website at https://vast.ai/