SaladCloud review

SaladCloud is a hosted GPU infrastructure platform with two main integration layers: an OpenAI-compatible AI Gateway for model inference and Salad Container Engine for Docker workloads. It schedules containers across a distributed network of consumer GPUs, workstations and data-center servers. Community Cloud supplies on-demand capacity; Salad Dedicated offers reserved bare metal and virtual machines on contract.

In an application stack, the AI Gateway can sit behind an existing client that calls an OpenAI-compatible endpoint. Container Engine sits below the workload, managing placement and reallocation across GPU pools. The homepage shows a REST request to the Salad API that creates a container group and sets image, CPU, memory, GPU class, replicas and restart policy. Teams can also run Salad orchestration on their own devices through Private Cloud, which the vendor says keeps company data on those devices. This is infrastructure, not a model-training library or a local GPU runtime.

The homepage demonstrates a REST API at api.salad.com, using a Salad API key and JSON requests. Its example creates a container group, selects a GPU class, sets resource requirements and requests 40 replicas. Container Engine supports Docker workloads, and the vendor describes the containers as OCI containers. The AI Gateway is OpenAI-compatible and serves open-weight models, so it can fit into clients built for that API shape.

Salad also lists Virtual Kubelet for scheduling Kubernetes pods onto SaladCloud, plus Recipes & Models for one-click deployments. The developer hub links to documentation, getting-started material, recipes and GitHub. The homepage does not name SDK languages or a CLI; confirm SDK availability and supported languages in the vendor’s docs before choosing an integration. It also says the API is shared across Community Cloud, Dedicated and Private Cloud. That offers a consistent deployment surface, but teams should validate which API operations and features apply to each environment.

SaladCloud has no listed free plan or free trial. Community Cloud advertises GPU access from $0.015 per GPU-hour, while the listed GPU table gives model- and priority-specific hourly rates. For example, an RTX 5090 ranges from $0.250 to $0.600 per GPU-hour; an RTX 3080 ranges from $0.060 to $0.155. The vendor says usage is charged for running seconds, and the monthly figures assume 730 hours at the Lowest rate: $182.50 for an RTX 5090 and $43.80 for an RTX 3080.

Your compute bill grows with the number of running instances and their runtime. The AI Gateway is billed per token or through a flat monthly plan, but those prices are not published in the information available to us. Salad Dedicated is contract-based, and no rate is listed. No add-ons are listed. Before launch, ask Salad which priority and GPU availability fit your workload, and get the total cost for API usage, reserved capacity and any required support in writing.

The vendor links to GitHub, but does not publish an open-source license for Salad Container Engine or say that the full platform source is available. Treat the orchestration layer as vendor-managed software unless Salad confirms the license and self-hosting terms. Private Cloud is a deployment option: Salad says organizations can run its orchestration stack on their own devices, with company data staying on those devices.

The AI Gateway’s OpenAI-compatible endpoint can reduce client-side migration work, while Docker and OCI containers give workloads a portable packaging format. Those do not guarantee that model configuration, scheduling behavior, secrets, logs or operational tooling transfer directly to another provider. The vendor does not publish a data-export process in the homepage information. Before adopting the platform, ask how to export logs and configuration, what data the gateway retains, and whether Private Cloud requires a contract or ongoing hosted services.

Salad’s architecture is designed to reallocate work when a node disappears. The vendor says Container Engine verifies nodes, isolates containers behind a hypervisor and streams logs. That is relevant for distributed hardware, but does not remove the need to test retry behavior, checkpointing and workload recovery. GPU capacity can vary across a network of consumer devices; confirm the availability and hardware mix for the regions and GPU classes your service needs.

The homepage cites 16,000+ daily active GPUs and nodes in 190+ countries. It does not publish API rate limits, a service-level agreement, detailed uptime targets or per-model capacity guarantees. Ask about these before making the gateway a critical dependency. The vendor says AI Gateway data is never used for training, and Private Cloud keeps company data on company devices. Review the security and Trust Center materials for your organization’s requirements, and confirm data retention, incident response and compliance coverage directly. Enterprise Support offers security review and a named team for large deployments.