The NVIDIA H100 is the gold standard for LLM training and high-throughput inference. But H100 pricing varies dramatically across providers — from $2.29/hr to $5.00/hr for comparable specs. Picking the wrong provider for a week-long training run can cost you $500–$1,500 extra. Here's the current price landscape, updated regularly.
H100 Pricing Across Providers (2025)
| Provider | H100 Variant | Price/hr (1 GPU) | Price/hr (8 GPU) | Notes |
|---|---|---|---|---|
| Hyperstack | H100 NVL 94GB | $2.29/hr | $18.32/hr | EU-based, strong NVLink |
| Lambda Labs | H100 SXM5 80GB | $2.49/hr | $19.92/hr | Best reliability, 99.9% SLA |
| CoreWeave | H100 SXM5 80GB | $2.79/hr | $22.32/hr | Enterprise focus, InfiniBand clusters |
| RunPod On-Demand | H100 SXM5 80GB | $2.79/hr | $22.32/hr | Good developer experience |
| RunPod Spot | H100 SXM5 80GB | $1.20–2.10/hr | $9.60–16.80/hr | Interruptible, 40–60% off |
| Vast.ai | H100 SXM5 80GB | $1.80–2.80/hr | $14.40–22.40/hr | Marketplace, varies by host |
| FluidStack | H100 NVL 94GB | $2.39/hr | $19.12/hr | European provider |
| DataCrunch | H100 SXM5 80GB | $2.49/hr | $19.92/hr | European, ISO 27001 certified |
Total Cost for Common H100 Workloads
| Workload | GPU Config | Runtime | Cost at $2.49/hr | Cost at $1.50/hr (spot) |
|---|---|---|---|---|
| Fine-tune LLaMA 3 70B (QLoRA) | 1× H100 | 6–8 hrs | $15–$20 | $9–$12 |
| Pre-train 7B model to 100B tokens | 8× H100 | ~4 days | $1,900 | $1,150 |
| Fine-tune LLaMA 3 70B (full) | 8× H100 | 2–3 days | $960–$1,440 | $575–$860 |
| Production inference API (24/7) | 1× H100 | 1 month | $1,793/month | Not recommended (spot) |
| Benchmark / experiment (2hrs) | 1× H100 | 2 hrs | $5.00 | $3.00 |
When to Use Spot vs On-Demand H100
H100 spot instances on RunPod and Vast.ai offer 40–60% discounts over on-demand, at the cost of potential interruption. The rule of thumb: use spot for any workload with automatic checkpointing (Axolotl, HuggingFace Trainer, DeepSpeed all support this). Use on-demand for production inference APIs, time-sensitive experiments, and any job where interruption would mean re-running from scratch.
What About Reserved / Committed H100 Pricing?
If you're running H100s continuously or near-continuously, committed contracts cut costs by 30–50%. Lambda Labs offers 1-month reserved H100 at roughly $1,800/month (vs $1,793/month on-demand — essentially the same). CoreWeave and Hyperstack offer 3-month and 12-month contracts with meaningful discounts. At 6+ months of continuous use, reserved pricing on CoreWeave or Hyperstack drops to ~$1.50–$1.80/hr equivalent.
Which Provider Should You Use?
- →Best price (spot): RunPod or Vast.ai — H100 from $1.20/hr with checkpointing
- →Best price (on-demand): Hyperstack at $2.29/hr or Lambda Labs at $2.49/hr
- →Best reliability: Lambda Labs — 99.9% SLA, purpose-built AI infrastructure
- →Best for EU data residency: Hyperstack (Iceland/Netherlands) or DataCrunch (Finland)
- →Best for large clusters (32–256 GPUs): CoreWeave — InfiniBand fabric, enterprise SLAs
- →Best for one-off experiments: RunPod — easiest signup, fastest instance provisioning
Will H100 Prices Drop?
H100 prices have declined roughly 15–25% over the past 12 months as supply from NVIDIA increased and providers expanded capacity. The H200 and Blackwell B200 are entering the market in 2025–2026, which will put further downward pressure on H100 pricing. If your training run is flexible, waiting 3–6 months for lower H100 prices or better B200 availability could save 20–30%.