Salad

What makes Salad different

Salad inverts the traditional cloud model by tapping latent consumer GPU resources rather than building hyperscale datacenters. The platform aggregates 450,000+ geographically distributed nodes—consumer GPUs shared part-time by individual operators—and orchestrates them via the proprietary Salad Container Engine. This enables GPU pricing starting at $0.02/hour, with customers reporting up to 90% cost savings compared to AWS, GCP, and Azure.

The trade-off is transparency: nodes are subject to interruption and have longer cold starts than datacenter GPUs, and maximum vRAM is capped at 24 GB. Salad mitigates availability risk through a proprietary trust-rating system and automatic workload reallocation when hosts go offline. Security relies on container isolation, encryption in transit and at rest, and runtime intrusion detection.

This model appeals specifically to workloads tolerant of latency variance and interruption—inference, batch jobs, and training tasks that don’t require guaranteed uptime—while offering radical cost efficiency for price-sensitive AI teams.

Pricing model

Salad charges $0.02–$0.04 per GPU hour for entry-level RTX/GTX instances, with pricing varying by GPU type and availability. Published examples include:

• Text-to-image (Flux.1-Schnell): ~$0.10 per image
• LLM inference: $0.04–$0.12 per hour to deploy; $0.12 per million tokens average for TGI
• Speech-to-text: From $0.10/hour
• Computer vision: Pricing at 73% less than Azure for comparable tasks

For high-volume use (>10 GPUs, long-running jobs, committed contracts), Salad offers custom discounts through sales engagement. There are no pre-paid commitments, no minimum contracts, and usage is transparent and hourly.

This pricing is 3–10× lower than hyperscaler equivalents for equivalent GPU tiers, with customers like Civitai and Blend reporting 60–85% cost reductions.

When it fits

• AI inference at scale: Serving millions of inferences per day across distributed consumer GPUs (Civitai’s 600+ GPU inference pipeline)
• Model training & fine-tuning: LoRA training, Dreambooth, and other workloads where interruption is manageable
• Batch processing & HPC: Rendering queues, data processing jobs, and scientific computing insensitive to latency
• Cost-driven startups: Teams prioritizing GPU access over uptime guarantees and willing to architect for interruption
• Hybrid/multi-cloud: Offloading bursty GPU workloads to reduce spend while maintaining primary infrastructure on traditional clouds

When it doesn’t

• Low-latency, real-time inference: Salad nodes have longer cold starts and are subject to interruption; workloads requiring less than 100ms response times or guaranteed availability should use hyperscalers.
• High-memory workloads: Maximum vRAM is 24 GB; workloads needing 48+ GB on a single GPU are not supported.

Inclusion criteria

Salad meets all three alt-cloud.org inclusion criteria:

1. Transparent pricing: Hourly rates published on homepage ($0.02–$0.04/hr); per-task pricing examples available throughout site
2. Self-service signup: Direct access to SaladCloud via web portal (https://salad.com/)
3. Public SLA & status: Status and reliability information provided in FAQ; proprietary trust-rating system publicly described