SaladCloud

• Deploy your existing container. Package an inference service or worker as an OCI-compatible image, then configure a container group through the portal or API. Resource choices include CPU, memory, GPU class and replica count, allowing the same application image to be evaluated on different configurations.
• Use distributed consumer GPUs. Community Cloud schedules containers onto available GPU machines rather than reserving a conventional dedicated server. It suits jobs such as image generation, transcription and batch inference when the application can handle variable availability and resume or retry interrupted work.
• Manage replicas and recovery. Container groups expose operational controls such as priority, automatic startup and restart behavior. Scheduling and replacement mechanisms help restore capacity after a machine disappears, but applications still need their own durable state, idempotency and sensible request or job handling.
• Integrate hosted AI endpoints. The AI Gateway offers access to open-weight model services through an OpenAI-compatible interface. This can reduce the infrastructure work needed to evaluate supported models, while model availability, token billing or contracted pricing should be checked for the intended workload.
• Monitor the deployment lifecycle. Inspect container status and logs when troubleshooting image downloads, initialization or running workloads. Community Cloud billing is described in terms of running containers rather than image download time; compare the full cost of useful output, including retries and startup delays.
• Evaluate other deployment arrangements. Salad also describes Private Cloud orchestration for an organization's own fleet and dedicated compute options. These have a different procurement and availability path from Community Cloud, so confirm the actual hardware, reservation, data handling and service agreement before designing around them.

Check before you buyCommunity Cloud can be interrupted and is a poor match for strict low-latency, multi-GPU interconnect or residency requirements without further validation. Dedicated offerings need availability confirmation. No startup credit entitlement was verified.

Official sources · checked 2026-10-02Official product and deployment options ↗