deploybase

Cheapest H100 in US West: Provider Pricing Ranked

Deploybase · August 27, 2025 · GPU Pricing

Contents

Cheapest H100 Us West: Pricing Overview

H100 in US West: constrained supply. RunPod: $2.69/hr on-demand. West cheaper than East by 10-15% due to lower East Coast demand.

Spot instances: $1.60-1.80/hr (35-40% discount). Good for batch work that can survive interruptions.

Regional Pricing Comparison

Lambda Labs does not offer H100 in US West as of March 2026, instead providing H100 variants only in US East. This limits Lambda's competitiveness for teams prioritizing West Coast inference latency or cost optimization.

CoreWeave bundles H100s exclusively in 8-GPU clusters at $49.24 per hour for US-based deployments. Regional pricing variations remain minimal across CoreWeave regions, with US West matching US East rates. Per-GPU cost breaks down to $6.16 per hour.

AWS EC2 p3.16xlarge instances with 8xV100 GPUs provide alternative options in US West but lack current H100 inventory. H100 availability through AWS appears in select regions with different pricing structures. Direct H100 rental pricing through AWS EC2 remains unavailable for standard purchase.

Google Cloud Platform offers H100 GPU access through Compute Engine in US regions. Pricing starts at approximately $2.84 per hour on-demand for single H100 instances. Preemptible (spot) instances cost roughly $0.85 per hour, representing significant savings for interruptible workloads.

Provider Rankings by Cost

  1. RunPod H100 SXM: $2.69 per hour (on-demand)
  2. Google Cloud H100: $2.84 per hour (on-demand)
  3. CoreWeave 8xH100: $6.16 per GPU (on-demand cluster)
  4. RunPod H100 SXM Spot: $1.60-1.80 per hour (interruptible)
  5. Google Cloud H100 Preemptible: $0.85 per hour (interruptible)

RunPod maintains lowest on-demand pricing for single H100 rental in US West. Google Cloud spot instances offer lowest absolute pricing but require fault tolerance. CoreWeave cluster pricing includes infrastructure and networking costs offsetting per-GPU calculations.

Price competitiveness shifts based on workload duration and reliability requirements. Teams planning long-term (30+ days) training benefit from reserved capacity. CoreWeave's cluster model suits multi-week distributed training without frequent job submission overhead.

Total Cost Factors

Egress charges significantly impact total H100 rental costs. AWS and Google Cloud charge $0.12 per GB for data transfer out of regions. Keeping data within provider networks eliminates these fees. A 1TB dataset transferred externally adds $120 to project costs.

Persistent storage costs vary by provider. RunPod charges $0.10 per GB monthly, CoreWeave $0.30 per GB, and Google Cloud $0.025 per GB. Storage selection becomes material for 500GB+ datasets.

Networking overhead in multi-GPU setups influences training efficiency. RunPod's network architecture introduces approximately 5-10% overhead for 8-GPU training compared to on-premises setups. CoreWeave's optimized networking reduces overhead to 2-3% through dedicated infrastructure.

Management overhead differs significantly. CoreWeave clusters require advance provisioning (2-4 hours) but then handle scaling automatically. RunPod instances launch in seconds but require manual multi-instance coordination for distributed training.

Selection Criteria

Workload type determines optimal provider selection. Short-term inference (hours to days) favors RunPod's simplicity and fast provisioning. Long-term training (weeks to months) favors CoreWeave's cluster infrastructure and predictable pricing.

Budget constraints shift the calculation toward spot instances. Teams tolerating 30-50% interruption rates should prioritize Google Cloud preemptible pricing at $0.85 per hour. Budget-conscious teams accept 5-10% job restart overhead to achieve significant cost savings.

Data residency requirements eliminate certain providers. Teams requiring West Coast processing should avoid Lambda Labs (no West option) and prefer RunPod or Google Cloud with confirmed US West presence.

Multi-region deployments optimize costs through load distribution. Distributing workloads across cheaper regions reduces average costs by 10-20%. However, increased operational complexity must be weighed against cost savings.

FAQ

Is H100 worth the cost for US West inference? Yes, for models exceeding 100B parameters requiring sub-100ms latency. A100 GPU alternatives cost 20-30% less while delivering 70% of H100 throughput. Teams should benchmark specific models before committing to H100 infrastructure.

How much cheaper is H100 in US West versus US East? Pricing differences typically range 5-15% depending on provider and instance type. US West generally offers 10% discounts due to lower regional demand. However, availability constraints may offset cost advantages.

Can I use spot instances for H100 training? Yes, with proper checkpointing. Save model weights every 10-30 minutes. Upon interruption, resume from latest checkpoint without complete retraining. This approach reduces effective training cost by 35-40%.

What is the minimum viable budget for H100 training in US West? A one-week training job on single H100 SXM instance costs approximately $450 (168 hours * $2.69). Including storage, egress, and overhead, budget $500-600. Larger workloads spanning months realize better unit economics.

Should I prefer RunPod or CoreWeave for H100 in US West? RunPod suits smaller jobs (single GPU, days to weeks). CoreWeave suits larger operations (8+ GPUs, month-long training). CoreWeave's per-GPU cost decreases at scale, favoring distributed training.

Compare H100 specifications with other GPUs. Explore RunPod GPU pricing details. Review CoreWeave GPU pricing for cluster options. Check AWS GPU pricing for alternative providers. Compare cheapest H100 in US East for regional differences.

Sources