Where Kubernetes meets thousands of distributed GPUs.
Out of GPUs in your cluster, or overpaying for them in the cloud? SaladCloud's Virtual Kubelet integration registers the network as a node in your Kubernetes cluster. Pods scheduled to it run as Salad container groups. Your manifests, CI, and observability stay exactly as they are.
# Schedule a pod onto SaladCloud like any other node apiVersion: v1 kind: Pod metadata: { name: inference } spec: nodeSelector: { type: virtual-kubelet } tolerations: - key: virtual-kubelet.io/provider operator: Exists containers: - name: app image: ghcr.io/acme/inference:latest resources: { limits: { nvidia.com/gpu: 1 } }
Extend the cluster you have to the GPUs you need.
Deploy without workflow changes
Pods target the Salad virtual node through standard selectors and tolerations. Nothing else in your pipeline changes.
Built for GPU applications at scale
Thousands of RTX-class GPUs worldwide, scheduled and reallocated by Salad Container Engine underneath.
The lowest unit cost in the market
Community Cloud pricing from $0.015 per GPU-hour, billed per second while instances are running.
Good for bursty, parallel, GPU-heavy work.
Batch inference, rendering, data pipelines, and experiment sweeps that already run as Kubernetes jobs move well. Pods that need persistent volumes, inter-pod networking, or sub-second start times should stay on your cluster; we'll help you draw the line.