Vast.ai
Per-second billing GPU cloud platform, aggregating 40+ data center compute, deploying inference and training in seconds
About Vast.ai
Vast.ai is a GPU cloud platform for AI agent and engineering teams, opening compute from 40+ data centers and 20,000+ GPUs in an API-native way, with real-time quotes and per-second billing that lets you deploy in seconds via CLI, SDK, or API. The same platform offers three deployment forms: GPU Cloud for teams that control their own instances, Serverless where the SDK takes over worker scaling for zero-ops inference, and Clusters for large-scale training. Prices form in real time from data center supply and demand, with on-demand, interruptible, and reserved billing; Serverless starts at $5 minimum spend, and you can get an API key yourself to start, with no contract or sales process.
Best for
Best use cases
Rent GPU instances on demand for training and fine-tuning
Serverless inference endpoints billed per second with automatic scaling
Deploy GPU jobs in seconds via CLI / SDK / API
Assemble large-scale training clusters and schedule multi-node compute
Features
GPU Cloud full control
On-demand instances span 40+ data centers and 20,000+ GPUs, with teams in control of environment and deployment pace
Serverless zero-ops inference
The SDK takes over worker scaling, inference services are billed per second, with no tiers, no usage caps, no hidden markups
Clusters for large-scale training
A cluster deployment form for large-scale multi-node training jobs
Real-time prices and three billing modes
Prices refresh in real time from supply and demand, offering on-demand, interruptible, and reserved billing
Pros
- Resource pool spans 40+ data centers and 20,000+ GPUs
- Per-second billing with real-time price refresh and pay-as-you-go settlement
- On-demand, interruptible, and reserved billing let you trade cost against stability
- Starts as low as $5 with no contract and no sales follow-up, self-service in the dashboard
Cons
- Unit prices are set in real time by supply and demand at each data center, so the price of the same GPU model changes over time
- Model selection means checking the specs and past quotes on each GPU card one by one, and bulk price comparison is manual