Purpose
SaladCloud is a distributed GPU cloud platform that provides compute power for AI/ML workloads. Its core idea is to coordinate and utilize idle consumer-grade GPUs scattered around the world, rather than building new centralized data center capacity, thereby creating a structural cost advantage. The company says its service covers more than 190 countries, with more than 16,000 active GPUs participating every day.
Main Capabilities
Three Infrastructure Forms
- Community Cloud: an on-demand distributed compute pool made up of idle consumer-grade GPUs, with official pricing examples listed as RTX 5090 and RTX 4090 starting as low as $0.015/hour.
- Salad Dedicated: contract-based reserved bare metal and virtual machines, which the company says start with RTX PRO 6000 Blackwell, with thousands of cards coming online progressively.
- Private Cloud: deploy the same orchestration stack inside an organization's own infrastructure (such as existing workstations and servers), pooling idle GPUs into a private, OpenAI-compatible token service, with data never leaving the enterprise's devices.
Salad Container Engine
This is the software layer that schedules, monitors, and redistributes Docker workloads across distributed GPU pools, and can be used for Community Cloud, Dedicated, or clusters of owned devices. The company describes its capabilities as including: scheduling OCI containers to nodes never seen before, validating nodes, isolating nodes behind a hypervisor, streaming logs, and immediately redistributing work when a node disappears. The company mentions support for 10,000+ replicas and failover.
OpenAI-Compatible Token API
Provides an OpenAI-compatible endpoint for open-weight models, served by the network. You can pay per token or apply for a monthly plan, and the company states that data will not be used for training. The page mentions support for Qwen, with a context length of 262k (subject to the page's wording).
Performance Benchmarks (Company-Published)
- Image generation: 5,243 Flux.1-Schnell images per dollar; Civitai serves 10 million images per day on 600+ Salad GPUs.
- Speech: 95.1% accuracy on CommonVoice, audio cost From $0.16/hour.
- Text: On RTX-class GPUs, use OpenVoice for up to 6 million words per dollar.
- Agents and RAG: GPUs at flat hourly rates, or token-based pricing through AI Gateway, with input tokens as low as $0.06 per million.
- Computer vision: Label 309,000 images per dollar with RAM++; YOLOv8 object detection costs 73% less than Azure.
- Batch processing and simulation: Transcribe 1,000,000 hours of audio for $1,260; GROMACS is priced more than 90% below data center pricing.
Official sources emphasize that the above benchmarks are published results, and every number links to its methodology.
Operations and Support
Production customers can communicate directly with Salad engineers about architecture, container tuning, and cost; large-volume deployments and committed contracts can receive customized support and pricing.
How to Use
The officially provided paths include:
- Use the managed token API.
- Deploy containers through Salad Container Engine (Docker workloads).
- Choose suitable infrastructure on Community Cloud, Salad Dedicated, or your own devices.
- In private data center scenarios, deploy the orchestration stack on your own infrastructure to form an OpenAI-compatible endpoint.
The same API, portal, and orchestration layer run through the above options. The page also provides tutorials, benchmarks, and product update content, such as "SaladCloud Recipes: the fastest way to deploy on distributed GPUs."
Applicable Scenarios
Based on official customer testimonials and benchmarks, suitable scenarios include: large-scale image generation inference, LoRA training (officially stated to be more than 15,000 LoRAs per month), speech transcription, computer vision annotation, batch processing and scientific simulation (such as GROMACS), and agent and RAG workloads. Official customers mention significantly lower costs and increased scaling on consumer-grade GPUs compared with A100s and mainstream hyperscale cloud providers.
Limitations and Considerations
- Community Cloud nodes are interruptible (similar to spot instances), priced by priority tier (High, Medium, Low, Lowest). Unlike ordinary spot, Salad Container Engine automatically reassigns workloads to nodes of the same tier, charges only for runtime, and does not charge for cold starts or image downloads.
- Not suitable for scenarios requiring sub-10 ms latency, multi-GPU NVLink training, or data residency guarantees that consumer-grade networks cannot provide. For such needs, the official recommendation is to use Salad Dedicated reserved hardware or the private infrastructure of Private Cloud.
- On security, containers run inside Linux virtual machines partitioned by a hypervisor and host machine, and images, configurations, and traffic are encrypted in transit and at rest; host intrusion detection will destroy the environment and blacklist the machine when an access attempt is detected. The official source states compliance with SOC 2 Type I.
- The GPU range includes NVIDIA RTX-class consumer GPUs (from RTX 3060 to RTX 5090 32 GB) and some AMD models; each GPU is benchmarked before being brought online and continuously scored by a proprietary trust rating. Data center and workstation GPUs (starting with RTX PRO 6000 Blackwell) are provided contractually through Salad Dedicated.
- Since 2018, the network has had 450,000+ GPUs contributing compute; node suppliers (Salad Chefs) earn Salad Balance by running the desktop application.
Information Not Yet Provided
Official materials do not explain specific registration process details, the complete price list for each tier, specific SLA terms, availability details by region, or information such as free credits.
