NVIDIA H100 Price in 2026: Cloud Pricing, Cost Analysis & Where to Rent
The NVIDIA H100 is the gold standard GPU for AI and machine learning in 2026. But how much does it actually cost — and where should you rent one? This complete guide covers H100 pricing, cloud provider comparisons, buy vs rent analysis, and total cost of ownership calculations.
Quick Answer: H100 cloud rental prices range from $2.00–$4.50/hr depending on the provider. The cheapest on-demand H100 is on Voltage Park and Vast.ai at ~$2.00–$2.10/hr. Purchasing an H100 outright costs $25,000–$40,000 per GPU.
What Is the NVIDIA H100 and Why Does It Dominate AI/ML?
Built on NVIDIA's Hopper architecture, the H100 delivers up to 3–6x the performance of the A100 for transformer-based models. With 80GB of HBM3 memory, 3.35 TB/s bandwidth, and native FP8 support, it has become the default choice for LLM training, fine-tuning, and high-throughput inference. Every major AI lab — from OpenAI to Anthropic to Meta — relies heavily on H100 clusters.
How Much Does an H100 Cost? Buying vs Renting
Purchasing an H100 GPU
- H100 SXM (80GB HBM3): $25,000–$35,000 per GPU
- H100 PCIe (80GB HBM3): $25,000–$30,000 per GPU
- DGX H100 system (8x H100): $250,000–$350,000
Purchase prices have dropped 20–30% from the 2024 scarcity-driven peak. Buying only makes sense for organizations running GPUs at high utilization 24/7 for multiple years.
Renting an H100 on the Cloud
- On-demand cloud rental: $2.00–$4.50/hr per GPU
- Spot/preemptible pricing: $1.50–$2.50/hr (with interruption risk)
- Monthly reserved: 10–25% discount over on-demand
At $2.50/hr, running an H100 for 1,000 hours costs $2,500 — a fraction of the $30,000+ purchase price. For most teams, renting is the clear winner.
H100 Cloud Pricing Comparison (March 2026)
| Provider | H100 Price/hr | Variant | Billing |
|---|---|---|---|
| Voltage Park | $2.00–$2.50 | SXM 80GB | Per-hour |
| Vast.ai | $2.10 | SXM / PCIe | Per-second |
| TensorDock | $2.20–$2.80 | SXM / PCIe | Per-second |
| RunPod | $2.49 | SXM 80GB | Per-second |
| Lambda Labs | $2.89 | SXM 80GB | Per-hour |
| CoreWeave | $2.95 | SXM 80GB | Per-minute |
| Hyperstack | $2.95 | SXM 80GB | Per-hour |
| Google Cloud (A3) | ~$4.10 | SXM 80GB | Per-second |
| AWS (P5) | ~$4.15 | SXM 80GB | Per-second |
| Microsoft Azure | $5.40 | SXM 80GB | Per-hour |
Cheapest on-demand H100: Voltage Park and Vast.ai at $2.00–$2.10/hr. Best value with reliability: Lambda Labs at $2.89/hr with strong uptime and ML-focused support.
H100 80GB SXM vs PCIe: Price and Performance Differences
| Feature | H100 SXM | H100 PCIe |
|---|---|---|
| Memory | 80GB HBM3 | 80GB HBM3 |
| Memory Bandwidth | 3.35 TB/s | 2.0 TB/s |
| FP16 Performance | 989 TFLOPS | 756 TFLOPS |
| FP8 Performance | 1,979 TFLOPS | 1,513 TFLOPS |
| NVLink | Yes (900 GB/s) | No |
| TDP | 700W | 350W |
| Cloud Price | $2.50–$4.15/hr | $1.89–$3.00/hr |
Key takeaway: The SXM variant is 30–40% faster for training workloads due to higher memory bandwidth and NVLink support for multi-GPU communication. PCIe is cheaper and sufficient for single-GPU inference. For multi-GPU training, always choose SXM.
When Is the H100 Worth It vs Cheaper Alternatives?
H100 vs A100
- H100 advantage: 3–6x faster for transformer training, FP8 support, 67% more memory bandwidth
- A100 advantage: 30–50% cheaper ($1.50–$2.50/hr vs $2.00–$4.15/hr)
- Choose H100: Training 13B+ parameter models, production inference for 70B+ models, time-sensitive runs
- Choose A100: Budget-constrained teams, smaller models (<10B), non-urgent training
H100 vs L40S
- H100 advantage: 2x memory bandwidth (HBM3 vs GDDR6), NVLink for multi-GPU, superior training performance
- L40S advantage: 30–40% cheaper ($1.50–$2.00/hr), 48GB VRAM, strong FP8 inference
- Choose H100: Multi-GPU training, memory-bandwidth-bound workloads, models >30B parameters
- Choose L40S: Single-GPU FP8 inference, 7B–30B models, tighter budgets
Total Cost of Ownership: H100 Cloud Scenarios
How much will an H100 actually cost for your project? Here are three common usage scenarios:
100 Hours (Experimentation / Fine-tuning)
| Provider | Total Cost |
|---|---|
| Vast.ai ($2.10/hr) | $210 |
| Lambda Labs ($2.89/hr) | $289 |
| CoreWeave ($2.95/hr) | $295 |
| AWS ($4.15/hr) | $415 |
500 Hours (Serious Training Project)
| Provider | Total Cost |
|---|---|
| Vast.ai ($2.10/hr) | $1,050 |
| Lambda Labs ($2.89/hr) | $1,445 |
| CoreWeave ($2.95/hr) | $1,475 |
| AWS ($4.15/hr) | $2,075 |
1,000 Hours (Production / Extended Training)
| Provider | Total Cost |
|---|---|
| Vast.ai ($2.10/hr) | $2,100 |
| Lambda Labs ($2.89/hr) | $2,890 |
| CoreWeave ($2.95/hr) | $2,950 |
| AWS ($4.15/hr) | $4,150 |
At 1,000 hours, choosing Vast.ai over AWS saves over $2,000 — enough to fund another 1,000 hours of H100 compute on Vast.ai.
H100 Price Trends: 2024–2026
H100 cloud pricing has dropped significantly as supply caught up with demand:
- Late 2023: $4.00–$8.00/hr (extreme scarcity, waitlists everywhere)
- Mid 2024: $3.00–$5.00/hr (supply improving, new providers launching)
- Early 2025: $2.50–$4.00/hr (market maturing, competition increasing)
- March 2026: $2.00–$4.15/hr (stable market, wide availability)
Prices are expected to continue a gradual decline through 2026 as NVIDIA ships more H100/H200 units and the B100/B200 generation enters cloud data centers. The H100 will remain the workhorse GPU for AI workloads for at least another 12–18 months.
Frequently Asked Questions
How much is an H100 per hour?
H100 cloud rental costs range from $2.00/hr (Voltage Park, spot) to $5.40/hr (Azure, on-demand). The market average for on-demand H100 SXM is approximately $2.50–$3.50/hr on dedicated GPU clouds, and $4.00–$5.40/hr on hyperscalers (AWS, GCP, Azure).
Is the H100 worth the premium over A100?
Yes, if you are training models with 13B+ parameters or need fast turnaround on large training runs. The H100 is 3–6x faster for transformer workloads, meaning a 3-day A100 training job can finish in 8–16 hours on H100. For smaller models and inference, the A100 often provides better cost-efficiency.
Should I buy or rent an H100?
Rent unless you will run the GPU at high utilization for 2+ years. At $2.50/hr, the break-even point for buying ($30,000) is approximately 12,000 hours — about 16 months of 24/7 operation. Factor in cooling, power ($0.50–$1.00/hr in electricity for SXM), maintenance, and depreciation, and renting is usually the better financial decision.
What is the cheapest way to get H100 access?
Spot instances on Vast.ai ($1.50–$2.10/hr) or Voltage Park ($2.00–$2.50/hr) offer the lowest H100 pricing. The trade-off is potential interruption, so always checkpoint your training jobs. For reliable on-demand access, Lambda Labs at $2.89/hr is the best value.
When will H100 prices drop further?
H100 prices will likely continue to decrease gradually through 2026 as NVIDIA's next-generation B100/B200 GPUs become available in cloud data centers. Expect H100 pricing to settle around $1.50–$2.50/hr by late 2026 as supply fully normalizes.
Compare H100 Prices Across 17+ Providers
Stop overpaying for H100 compute. Find the cheapest H100 instance in seconds with real-time price comparison.
Compare H100 Prices Now →