Deploy any GPU. Any framework. Under 30 seconds.

GPU Pods are dedicated GPU instances you can spin up on Runpod. Unlike abstracted serverless GPUs, Pods give you full control over the underlying VM, drivers, and environment. You get a persistent instance (or ephemeral, if you prefer) with direct access to powerful GPUs, letting you run training, inference, or other workloads exactly how you want.

We offer 30+ GPU models, from entry-level inference cards to top-tier training accelerators. Examples include A100, H100, RTX 6000 Ada, L4/L40 series, and many more—over 30 options in total. You can pick any supported GPU when you launch a Pod, and new models roll out as soon as they’re live on the platform. For the latest availability, check the dashboard or query the API.

Pricing is shown as an hourly rate but billed by the millisecond. You only pay for the exact time your Pod runs—if you start and stop a Pod in one minute, you’re charged just that minute. Storage volumes may incur minimal fees when attached, but compute costs are metered by the millisecond.

Yes. GPU Pods support custom Docker images. You can build an image with your preferred libraries and push it to a registry (Docker Hub, ECR, etc.), then reference it when you launch the Pod. That way you control the OS, drivers, and dependencies.

Any framework that runs on Linux and supports GPUs: PyTorch, TensorFlow, JAX, ONNX, CUDA toolkits, etc. Since you control the container, you can install whatever versions or additional tools you need (e.g., NCCL, Horovod). We provide base images with common ML stacks to speed up setup.

We offer spot instances where GPU capacity is available at a discount, but with the risk of eviction when demand spikes. You can use them for fault-tolerant or batch workloads. The UI/API will indicate current spot availability and pricing.