RunPod Overview
RunPod is a serverless GPU cloud with approximately 15,000+ GPUs of compute capacity across US, Europe, Asia: 15 regions. The platform offers H100, A100, L40S, 4090, 6000 Ada GPUs, positioning itself in the GPU infrastructure market for AI workloads. Key differentiators include $1.60-$2.50/hr H100 pricing, Instant on-demand lead times, and support for popular AI frameworks including PyTorch, TensorFlow, and JAX.
Pricing and Cost Structure
RunPod offers on-demand, reserved (1-12 month), and spot/preemptible pricing. On-demand $1.60-$2.50/hr H100. Reserved contracts provide 30-50% discounts. Spot/preemptible instances are 60-80% below on-demand. Additional costs include inter-node networking ($0.50-$2.00/GB), persistent storage ($0.08-$0.25/GB-month), and support tiers ($500-$5,000/month). For 8x H100 24/7, monthly on-demand is approximately $9,216-$14,400.
Performance Benchmarks
Performance benchmarks on RunPod show competitive results. H100 training of Llama 4 Scout (17B) with 8 GPUs at BF16 achieves 28,000-32,000 tok/s with ZeRO-3. Inference for Mixtral 8x7B at FP8 on 2x H100 reaches 3,200-3,800 tok/s. NCCL all-reduce benchmarks show 85-92% of theoretical peak for 8-GPU nodes. Inter-node InfiniBand or RoCEv2 networking provides 200-400 Gbps depending on configuration.
Availability and Lead Times
GPU availability at RunPod: Instant on-demand. High-capacity configs (8+ GPU NVLink) and liquid-cooled deployments require longer timelines. Reserved contracts receive priority provisioning. Teams needing guaranteed capacity should contract 4-8 weeks ahead. On-demand is available within standard lead times, though peak demand may cause intermittent shortages for specific GPU models.
Contract Terms and SLAs
RunPod offers month-to-month on-demand billing. Reserved contracts require 1, 6, or 12-month commitments. SLA: 99.9% uptime for reserved, 99.5% for on-demand. Service credits: 5-30% for violations. Support tiers range from community (free) to premium 24/7 with 15-min SLAs ($1,000+/month). Data egress typically $0.05-$0.12/GB, negotiable for high-volume customers.
When to Choose RunPod
RunPod is best for workloads of 8-64 GPUs where competitive pricing and responsive support matter more than hyperscaler ecosystem. Consider alternatives for 256+ GPU clusters where hyperscaler networking tooling excels, when specific GPU models are unavailable, or for highly regulated workloads requiring specific compliance certifications not offered by RunPod.