SaladCloud

What it is for

SaladCloud is a distributed GPU container cloud: push a Docker image, pick a GPU class and priority, and run on idle community machines (or reserved Dedicated boxes). Use it when the job is cheap interruptible inference or batch GPUs, not when you need a reserved data-center SLA on every replica.

What makes it unique: the cover is Salad’s own homepage deploy snippet — POST …/container-groups with gpu_classes, priority, and replicas — not a pricing card.

Key features

• Container Engine (SCE) — deploy a Linux AMD64 image as a container group; Salad places replicas on eligible GPUs (Quickstart; homepage curl). A deploy, not a token API.
• Priority / reallocation — High is never preempted by other workloads; Lowest is cheapest and most interruptible; nodes can still disconnect and SCE reallocates (salad.com/pricing High / Lowest). Fleet behavior, not a second deploy.
• AI Gateway — OpenAI-compatible https://ai.salad.cloud/v1 with no container to run (docs Salad AI Gateway). A token API, not SCE.
• Transcription API — managed speech-to-text billed per audio-minute, not per GPU-hour (docs Transcription API). A SaaS job, not a container.
• Dedicated / Private Cloud — reserved RTX PRO 6000 Blackwell (and on-prem license) quoted per allocation, off the community list (pricing Salad Dedicated / Private Cloud). A contract box, not Lowest-priority community.

Recipes are SCE templates, not a fifth product. S4 (30-day, 100 MB uploads) is storage. The portal/API are install / SDK. GPU-seconds (and audio-minutes on Transcription) are the meter. GPU SKUs (RTX 5090, 4090…) are SKU.

Who it's for

Best fit: inference teams who will prepay credits and accept interruptible community nodes — official FAQ: no free trial.

Pricing

Verified 2026-10-22 (Asia/Shanghai) from salad.com/pricing. Published USD per GPU-hour; billed per second while Running. vCPU and RAM ride with the GPU. Allocation / image pull / cold start are not billed.

• Community Cloud — no seat. No free trial (pricing FAQ). Additional classes from $0.015 / GPU-hour at Lowest.
• RTX 5090 (32 GB) — High $0.500 · Medium $0.417 · Low $0.333 · Lowest $0.250 / hr ($182.50 / mo at Lowest × 730 h).
• RTX 4090 (24 GB) — High $0.330 · Medium $0.273 · Low $0.217 · Lowest $0.160 / hr ($116.80 / mo Lowest).
• CPU-only — $0.005 / vCPU-hour + $0.001 / GB RAM-hour (example: 2 vCPU / 2 GB = $0.012 / hr).
• Dedicated / volume — sales quote (100+ GPUs, reserved Blackwell, Private Cloud).

Higher priority can preempt lower. Confirm the live table; classes and counts move.

Limitations / costs

• No free trial. You add credits before a replica runs.
• Community nodes disconnect. Replicas reallocate; local disk is ephemeral (quickstart).
• Distributed GPU cloud, not a reserved RunPod pod story. For a different GPU desk, see RunPod.

Alternatives

• RunPod — GPU pods / serverless when you want a different marketplace.
• Modal — scale-to-zero functions when you do not want to own a container group.
• Replicate — hosted model API when Salad’s Docker-on-community-GPUs is not the job.

Official tutorial

Salad’s docs point at YouTube, but third-party “cheapest GPU” walkthroughs are not first-party getting-started. We do not embed them. Use written Container Engine Quickstart plus Pricing.

Website

https://salad.com

Source

Catalog row curated for AI Tools Nav. Cover is Salad’s official homepage container-group API snippet (POST gpu_classes + priority). Dollars checked on salad.com/pricing (5090 High $0.500 / Lowest $0.250; from $0.015 Lowest; no free trial). No getting-started embed.

Last verified

2026-10-22 (Asia/Shanghai) editorial check.