RunPod is a GPU cloud for AI workloads, covering training, fine-tuning, batch processing, and inference on one account. Three compute forms sit side by side: Pods, Serverless, and Clusters, so work moves from experiment to production without replatforming.
A Pod is a GPU-enabled container environment that starts in under a minute and runs your own containers, framework, and code. More than 30 GPU SKUs are offered, from RTX 4090s to B200s, across 31 regions.
Serverless takes a handler you write and turns it into an inference endpoint that scales from zero to thousands of concurrent workers and bills only for active time. RunPod reports sub-200ms cold starts through a feature called FlashBoot, and the platform queues and distributes tasks so that an endpoint does not need its own orchestration layer.
Persistent network storage holds a full pipeline with no egress fees. Global Volumes, in beta, provide region-independent storage that mounts into a Pod in any RunPod data center.
Features
- Pods: GPU-enabled environments that start in under a minute and run your own containers, framework, and code
- Serverless workers: a handler becomes an autoscaling inference endpoint, from zero to thousands of workers
- Zero idle cost: an endpoint costs nothing while it is not running
- Cold starts: FlashBoot is reported at sub-200ms on serverless workers
- Hardware range: over 30 GPU SKUs, from RTX 4090s to B200s
- Global regions: workloads run across 31 regions
- Network storage: persistent storage for full pipelines, with no egress fees
- Global Volumes: a beta feature mounting region-independent volumes into a Pod in any data center
- Observability: real-time logs, monitoring, and metrics without a custom framework
