H100 GPU Pricing Guide 2026: Every Provider Compared
The NVIDIA H100 remains the workhorse of AI training and inference. But prices have dropped 70%+ from 2023 peaks. Here's every provider compared, with hidden costs exposed.
H100 SXM 80GB — Full Price Comparison
Prices as of March 2026, sorted from cheapest to most expensive on-demand pricing:
| Provider | Type | On-Demand | Spot/Preemptible | Monthly (24/7) |
|---|---|---|---|---|
| Vast.ai | Marketplace | $1.49/hr | $1.19/hr | $1,073 |
| Lambda | Specialized | $1.85/hr | N/A | $1,332 |
| FluidStack | Marketplace | $2.15/hr | $1.72/hr | $1,548 |
| CoreWeave | Specialized | $2.23/hr | N/A | $1,606 |
| DigitalOcean | Cloud | $2.50/hr | N/A | $1,800 |
| TensorDock | Marketplace | $2.59/hr | $1.81/hr | $1,865 |
| RunPod | Marketplace | $2.69/hr | $1.89/hr | $1,937 |
| Vultr | Cloud | $2.85/hr | N/A | $2,052 |
| RunPod Secure | Managed | $3.29/hr | N/A | $2,369 |
| GCP | Hyperscaler | $3.67/hr | $1.47/hr | $2,642 |
| AWS (p5) | Hyperscaler | $3.93/hr | $1.57/hr | $2,830 |
| CoreWeave | Specialized | $4.76/hr | N/A | $3,427 |
| Azure | Hyperscaler | $6.98/hr | $2.79/hr | $5,026 |
Price Drop Context
In mid-2023, H100s were $7.50-$11.00/hr on-demand. Today's cheapest at $1.49/hr represents a 80% drop. AWS's 44% price cut in June 2025 triggered an industry-wide correction that's still playing out.
Best Deals Right Now
Based on current pricing and reliability, here are our top picks:
Perffeco may earn a commission from provider links. This does not affect our pricing data.
Hidden Costs to Watch Out For
The hourly rate is just the beginning. Here's what catches most teams off guard:
| Cost Type | Typical Range | Impact |
|---|---|---|
| Data Egress | $0.08-0.12/GB | Adds 5-15% for data-heavy workloads |
| Storage | $0.02-0.08/GB/mo | Model checkpoints add up fast (100GB+ per save) |
| InfiniBand Premium | 15-30% extra | Required for multi-GPU training |
| Spot Interruptions | Save 60-90% | But you lose progress on preemption |
| Enterprise Support | $5-15K/mo | Required for production SLAs |
| Reserved Commitment | 1-3 year terms | 30-60% savings but lock-in risk |
When to Use Spot vs On-Demand
- Spot/preemptible: Fine-tuning with checkpointing, batch inference, experimentation. Save 40-60%.
- On-demand: Production inference, time-sensitive training, demos. Pay the premium for reliability.
- Reserved: Only if you need 24/7 GPUs for 6+ months. The savings are real but the commitment is too.
H100 vs Alternatives: Should You Even Use H100s?
Depending on your workload, cheaper GPUs might be a better fit:
| GPU | Best For | From | vs H100 |
|---|---|---|---|
| A100 80GB | Inference, fine-tuning | $1.10/hr | 26% cheaper, 60% perf |
| L40S 48GB | Inference, small models | $0.79/hr | 47% cheaper, 40% perf |
| RTX 4090 | Dev/testing, small inference | $0.34/hr | 77% cheaper, 25% perf |
| H200 141GB | Large models, 1.4x VRAM | $3.49/hr | More VRAM, 20% faster |
| B200 192GB | Next-gen, 2.4x VRAM | $2.99/hr | Newer arch, 2x faster |
Bottom Line
The H100 market is a buyer's market in 2026. Prices have cratered 70%+ from peaks, and competition between marketplace providers (Vast.ai, RunPod) and hyperscalers is driving further reductions.
For most teams, Vast.ai or Lambda offer the best value for on-demand H100s. Use spot instances on GCP/AWS for interruptible workloads, and consider A100s or L40S for inference-only deployments where you don't need full H100 power.
Compare All GPU Prices in Real Time
Track live pricing across 12 providers, 32 GPU configurations, with hidden cost analysis and a GPU calculator.
Open GPU Economics Dashboard