Cheapest H100 GPU cloud in 2026: provider-by-provider comparison
Where to rent NVIDIA H100 GPUs cheapest in 2026 — per-hour pricing across CoreWeave, AWS, Azure, Lambda Labs, and RunPod with reserved vs on-demand notes.
H100 pricing in 2026 has compressed roughly 35% from peak 2024 levels as B200 capacity comes online and providers free up Hopper inventory. Here is what an 8x H100 SXM5 node actually costs across major providers, on-demand.
Per-hour pricing (8x H100 SXM5, on-demand)
- RunPod (EU-West-1, H100 PCIe): $18.40/hr — cheapest credible on-demand.
- CoreWeave (EU-Central-1, H100 NVL 4-GPU): $12.90/hr ($25.80/hr at 8x).
- Lambda (US-East-2): $24.50/hr — strong SXM5 availability.
- AWS P5 (US-East-1): ~$32/hr H100 SXM equivalent.
- Azure ND H100 v5: ~$31/hr.
Reserved vs on-demand
Three-year reserved commitments cut another 30–45% across all providers. Lambda and CoreWeave are the most aggressive on multi-year H100 contracts in 2026; if the workload is steady-state, signing 1–3 year terms is the single biggest cost lever.
Hidden costs to budget for
- Egress: AWS and Azure charge for cross-region data movement; neoclouds typically don't.
- Storage: persistent NVMe for checkpoints adds $0.10–$0.20/GB/month.
- Idle: per-second billing (Lambda, RunPod) matters when jobs restart frequently.
Live pricing across all providers and SKUs is indexed at /marketplace and exposed via the structured Server Index at /intelligence.