NVIDIA H100 price guide 2026: cost per hour, $/TFLOP, and reservation breakeven

The NVIDIA H100 is the most-deployed AI training GPU in 2026, and its price has compressed roughly 35% since the 2024 peak. This guide pulls live numbers from the Servers.computer index to answer the practical questions: what does an H100 actually cost per hour, what is the $/TFLOP across providers, and when does a 1- or 3-year reservation break even versus on-demand?

On-demand H100 price per hour (2026)

Per-hour pricing for a full 8x H100 SXM5 node, on-demand, indexed across the five major providers Servers.computer tracks:

• RunPod (EU-West-1, 8x H100 PCIe): $18.40/hr — the cheapest credible on-demand 8-GPU H100 node.
• Lambda Labs (US-East-2, 8x H100 SXM5): $24.50/hr — strong SXM5 availability, per-second billing.
• CoreWeave (EU-Central-1, 4x H100 NVL): $12.90/hr (≈$25.80/hr at 8 GPUs) — InfiniBand fabric included.
• AWS P5 (US-East-1, 8x H100 SXM): ≈$32/hr — hyperscaler premium for region depth and compliance.
• Azure ND H100 v5: ≈$31/hr — comparable to AWS P5, integrated with Azure ML.

$/TFLOP efficiency — which H100 SKU gives you the most compute per dollar

Per-PFLOP per-hour is the only honest cross-SKU comparison. Lower is better. Numbers below are dense FP8 throughput from the Servers.computer index:

• RunPod 8x H100 PCIe — 6.4 PFLOPS @ $18.40/hr → $2.88 / PFLOP-hr (cheapest).
• Lambda 8x H100 SXM5 — 7.9 PFLOPS @ $24.50/hr → $3.10 / PFLOP-hr.
• CoreWeave 4x H100 NVL — 3.8 PFLOPS @ $12.90/hr → $3.39 / PFLOP-hr.
• AWS H200 (reference) — 8.4 PFLOPS @ $32.60/hr → $3.88 / PFLOP-hr.
• CoreWeave 8x B200 (reference) — 9.6 PFLOPS @ $38.00/hr → $3.96 / PFLOP-hr.

Headline result: a budget-tier H100 PCIe node from RunPod delivers ~27% more PFLOPS per dollar than a Blackwell B200 node. H100 PCIe wins on $/TFLOP; B200 wins on memory bandwidth and wall-clock time per epoch. Pick by bottleneck, not by spec sheet.

H100 reservation pricing vs on-demand

Reserved commitments are the single biggest cost lever on H100. Typical 2026 discounts:

• 1-year reserved: 25–35% off on-demand. Breakeven at ~70% utilization.
• 3-year reserved: 40–55% off on-demand. Breakeven at ~55% utilization.
• Spot / preemptible (RunPod, Lambda): 50–70% off on-demand. Best for fault-tolerant training with frequent checkpoints.

Worked example: a Lambda 8x H100 SXM5 node at $24.50/hr on-demand becomes ~$14.70/hr on a 3-year reservation (40% off). For a workload that runs 16 hours/day, 7 days/week, the reservation pays back in roughly 3 months versus on-demand. Below ~9 hours/day average, on-demand wins.

H100 vs B200 — when the upgrade is worth it

B200 lists at ~2× H100 per hour but only ~1.2× the FP8 throughput. The B200 upgrade pays off when memory bandwidth (8 TB/s vs 3.35 TB/s) or HBM capacity (192 GB vs 80 GB) is the actual bottleneck — long-context inference, MoE models, or 70B+ dense training where KV cache dominates. For 7B–13B training and routine fine-tuning, H100 remains the better $/result.

Hidden H100 costs to budget for

• Egress: AWS and Azure charge $0.05–$0.09/GB for cross-region transfer; neoclouds typically do not.
• Persistent NVMe storage for checkpoints: $0.10–$0.20/GB/month.
• InfiniBand fabric: bundled at CoreWeave and Lambda, extra (~10–15%) at hyperscalers.
• Idle time: per-second billing (RunPod, Lambda) saves 5–15% on jobs that restart frequently.

The bottom line

If you want the cheapest H100 in 2026: RunPod EU-West-1 at $18.40/hr on-demand, or a 3-year Lambda reservation at ~$14.70/hr if utilization is steady. If you need maximum throughput per cluster, CoreWeave H100 SXM5 with InfiniBand. If you need a hyperscaler for compliance, AWS P5 or Azure ND H100 v5. Live pricing across all providers is indexed at /marketplace and exposed via the structured GPU index at /intelligence.