Cloud GPU Pricing Comparison 2025
Compare GPU pricing across AWS, GCP, Azure, Lambda Labs, RunPod, and Vast.ai. Find the best value for your AI workloads.
Table of Contents
1. Cloud GPU Market Overview
Cloud GPU pricing varies dramatically between providers and instance types. Understanding the landscape helps optimize costs for training and inference workloads.
Key factors: On-demand vs spot pricing, commitment discounts, egress costs, and availability.
2. Major Cloud Providers
AWS (p4d.24xlarge - 8x A100 40GB): ~$32/hr on-demand, ~$19/hr spot.
GCP (a2-highgpu-8g - 8x A100 40GB): ~$29/hr on-demand, ~$9/hr spot.
Azure (ND96asr_v4 - 8x A100 40GB): ~$27/hr on-demand, ~$8/hr spot.
H100 instances: $8-12/hr per GPU on major clouds.
Note: Spot/preemptible instances offer 60-70% savings but can be interrupted.
3. Specialized ML Clouds
Lambda Labs: A100 80GB at $1.29/hr, H100 at $2.49/hr. Simple pricing, good availability.
RunPod: A100 from $1.19/hr, community cloud options even cheaper. Flexible.
Vast.ai: Marketplace model, A100 from $0.80/hr. Variable quality/reliability.
CoreWeave: H100 at $2.06/hr. Kubernetes-native, good for production.
4. Cost Optimization Strategies
Use spot instances for fault-tolerant training with checkpointing.
Reserved capacity for predictable workloads (30-60% savings).
Right-size instances: Don't pay for 8 GPUs if 4 suffice.
Consider egress: Training data upload and model download costs add up.
Hybrid approach: Develop locally, burst to cloud for large runs.
5. Recommendations by Use Case
Experimentation: Vast.ai or RunPod community cloud. Cheapest option.
Training runs: Lambda Labs or GCP spot. Good balance of price/reliability.
Production inference: CoreWeave or major cloud with reserved capacity.
Enterprise: AWS/GCP/Azure with committed use discounts and support.
◈ Related Guides
Need Help Choosing Hardware?
Compare specs and pricing for all AI hardware in our catalog.
Open Compare Tool →