The NVIDIA H100 is the most sought-after GPU in the cloud. It's also the most price-volatile: GPU Tracker currently tracks 70 single-H100 instances across 13 providers, with prices ranging from $0.80/hr to $11.56/hr. That's a 14x gap for the same chip.
Below is a complete, current ranking of every provider offering single H100 GPUs — no multi-GPU instances, no inflated list prices, just the cheapest available entry point per provider as of March 2026.
All 13 H100 Providers Ranked by Price
| Rank | Provider | Cheapest H100 | Type | Notes |
|---|---|---|---|---|
| #1 | Verda | $0.80/hr | Spot | Cheapest H100 in market |
| #2 | AWS | $1.16/hr | Spot | p5 spot instances |
| #3 | RunPod | $1.25/hr | Spot | Community cloud |
| #4 | Nebius | $1.25/hr | On-demand | EU-based, GDPR |
| #5 | Vast.ai | $1.47/hr | Spot | Marketplace, variable |
| #6 | GCP | $1.54/hr | Spot | A3 spot instances |
| #7 | Crusoe | $1.60/hr | On-demand | Carbon-neutral compute |
| #8 | Cudo Compute | $1.87/hr | On-demand | Distributed compute |
| #9 | Hyperstack | $1.90/hr | On-demand | UK-based |
| #10 | Lambda Labs | $2.49/hr | On-demand | Popular for ML teams |
| #11 | OVHcloud | $2.99/hr | On-demand | European provider |
| #12 | Scaleway | $3.22/hr | On-demand | EU, H100 SXM |
| #13 | Latitude.sh | $7.97/hr | On-demand | Bare metal |
Data sourced from GPU Tracker's live feed, updated every 6 hours. Spot prices fluctuate. Check real-time H100 prices before renting.
Spot vs On-Demand H100 Prices
The H100 spot market has genuine deals. Spot H100 instances start at $0.80/hr with a median of $1.75/hr. On-demand H100 starts at $1.58/hr with a median of $2.99/hr. The spot-to-on-demand discount for H100s averages around 40%.
Important: H100 spot instances on decentralized platforms (Vast.ai, RunPod Community) can be interrupted with 30-second to 5-minute notice. Always use checkpoint saving if running training jobs on spot H100s. For inference APIs, use on-demand or reserved instances only.
H100 vs H200: Should You Upgrade?
The H200 is now available from 7 providers at prices starting at $0.33/hr spot and $1.99/hr on-demand. Given that the H200 offers 141GB HBM3e (vs H100's 80GB) and 4.8 TB/s bandwidth (vs 3.35 TB/s), the H200 spot price is a compelling alternative to on-demand H100 for large-model inference.
| Spec | H100 SXM5 | H200 SXM5 |
|---|---|---|
| VRAM | 80GB HBM3 | 141GB HBM3e |
| Memory Bandwidth | 3.35 TB/s | 4.8 TB/s |
| FP16 TFLOPS | 989 | 989 (same die) |
| Spot Price (from) | $0.80/hr | $0.33/hr |
| On-Demand Price (from) | $1.58/hr | $1.99/hr |
| Median Market Price | $2.59/hr | $2.29/hr |
The H200's median market price ($2.29/hr) is actually lower than the H100's median ($2.59/hr), while offering 76% more VRAM. This is the result of H100 demand still being higher — more providers stock it, and buyers have more negotiating power on H200 because supply is less constrained. Check the H200 price comparison before defaulting to H100.
When to Choose Each Provider
- Verda ($0.80/hr): Best pure spot price if available. Check availability first — supply is limited.
- RunPod ($1.25/hr spot): Good mix of price and UX. Community Cloud is spot; Secure Cloud is on-demand at $1.99+/hr. Well-suited for ML teams.
- Nebius ($1.25/hr on-demand): Strong choice for EU-based teams needing GDPR compliance without paying Lambda or AWS prices.
- Lambda Labs ($2.49/hr on-demand): Reliable, developer-friendly, no spot interruptions. Worth the premium for teams that need consistency.
- AWS ($1.16/hr spot): Best if you're already in the AWS ecosystem and can tolerate spot interruptions on p5 instances.
- Latitude.sh ($7.97/hr): Bare metal H100 — justified for workloads that need dedicated hardware and zero noisy-neighbor interference.
All H100 prices update every 6 hours. See current availability and filter by region, commitment type, and VRAM at the H100 price comparison page.
H100 Prices by Region
H100 availability and price vary significantly by region. EU data centers typically cost 15–35% more than US East. APAC has the highest premiums due to supply constraints. Here is the typical cheapest H100 by region across our tracked providers:
| Region | Cheapest Spot | Cheapest On-Demand | Median On-Demand | Best Provider |
|---|---|---|---|---|
| US East (Virginia) | $0.80/hr | $1.58/hr | $2.49/hr | Verda / AWS spot |
| US West (Oregon/CA) | $0.99/hr | $1.99/hr | $2.79/hr | RunPod / Lambda |
| EU West (Frankfurt/Dublin) | $1.29/hr | $2.49/hr | $3.29/hr | Nebius / OVHcloud |
| EU North (Stockholm) | $1.39/hr | $2.79/hr | $3.49/hr | Crusoe / Nebius |
| APAC (Tokyo/Singapore) | $1.79/hr | $3.49/hr | $4.99/hr | AWS / GCP |
| UK (London) | $1.49/hr | $2.99/hr | $3.79/hr | Hyperstack |
H100 SXM vs H100 PCIe: Which Should You Rent?
The H100 comes in two physical variants: SXM5 (used in 4/8-GPU server pods with NVLink) and PCIe (used in single-GPU instances over PCIe Gen 5). They share the same die, but performance and pricing diverge:
| Spec | H100 SXM5 | H100 PCIe |
|---|---|---|
| TDP | 700W | 350W |
| Memory Bandwidth | 3.35 TB/s | 2.0 TB/s |
| NVLink | 900 GB/s (full) | 600 GB/s (bridged pair) |
| FP16 TFLOPS | 989 | 756 |
| Typical cloud price | $2.49–$11.56/hr | $1.49–$3.99/hr |
| Best for | Multi-GPU training, MoE inference | Single-GPU inference, dev workstations |
Rule of thumb: If you're using a single GPU or doing inference on a 70B-or-smaller model, H100 PCIe is the cheaper, more available option. If you're training across 4+ GPUs, the NVLink bandwidth on SXM5 is worth the premium — the GPUs effectively share memory.
Methodology
Data source: Every 6 hours we scrape publicly listed cloud GPU pricing from each provider's API or pricing page. The dataset covers single-H100 instances only — multi-GPU instances are normalized to per-GPU pricing for comparability.
Spot vs on-demand: We label any instance with a stated risk of interruption (Vast.ai marketplace, AWS spot, GCP spot, RunPod Community) as spot. Reserved capacity contracts are not included — only publicly available rates.
Regional pricing: Regional medians reflect the cheapest available H100 in each region during the most recent snapshot window. Hyperscaler regions (AWS p5, GCP A3, Azure ND H100) are tracked separately from specialty clouds.
What is excluded: Private enterprise contracts (CoreWeave reserved, Lambda 1-year, AWS Savings Plans) — these can be 30–50% lower but are not publicly listed. Bare-metal-only providers requiring sales calls.
H100 Pricing FAQ
What is the cheapest H100 cloud GPU in 2026?▾
The cheapest publicly listed H100 is $0.80/hr spot on Verda. RunPod Community Cloud and AWS p5 spot also start around $1.16–$1.25/hr when available. On-demand floor is $1.58/hr. All numbers update every 6 hours in our live feed.
How much does an H100 cost per hour on AWS?▾
AWS p5.48xlarge (8× H100 SXM5) lists at $98.32/hr on-demand — that's $12.29/hr per GPU. Spot pricing drops to $1.16–$3/hr per GPU when capacity is available. Reserved Instances and Savings Plans cut another 30–60% but require multi-year commitment.
Are spot H100 instances reliable enough for training?▾
Yes, with checkpointing. Spot H100s on AWS, GCP, and specialty clouds get 1–5 minute interruption notices. Use frameworks like Hugging Face Accelerate or torch.distributed.checkpoint to save state every 5–30 minutes. For inference APIs that must not drop requests, use on-demand or reserved capacity instead.
What is the difference between H100 80GB and H100 NVL?▾
H100 NVL is a paired-card variant — two H100 PCIe cards bridged via NVLink to expose 188 GB combined VRAM, optimized for LLM inference. It is priced higher per GPU but cheaper than two separate H100s plus the bridge. Most providers list it as a single "H100 NVL" SKU.
Why is the H100 price range so wide ($0.80 to $11.56/hr)?▾
Three factors: (1) Hyperscalers like AWS, Azure, and Latitude.sh charge premium rates because they bundle networking, support, and SLAs. (2) Specialty clouds and marketplaces (Vast.ai, RunPod, Verda) compete on raw price. (3) Spot vs on-demand: spot capacity is ~40–60% cheaper but interruptible. The right price for you depends on workload sensitivity and uptime requirements.
Should I rent an H100 or H200?▾
The H200 has 76% more VRAM (141 GB vs 80 GB) and 43% more memory bandwidth, while its median market price ($2.29/hr) is currently lower than the H100's median ($2.59/hr). For new inference workloads, H200 is the better default. The exception: if you need an H100-specific tested setup or your model is already optimized for 80 GB, the H100 still works fine.
Do H100 prices vary by region?▾
Yes, significantly. US East has the cheapest H100s (spot from $0.80/hr). EU regions run 30–50% higher (spot from $1.29/hr). APAC is the most expensive (spot from $1.79/hr). The driver is supply: NVIDIA prioritizes US deliveries to the largest cloud buyers, leaving EU and APAC capacity tighter.
Can I get an H100 without a long-term commitment?▾
Yes. All providers in our list except Latitude.sh allow hourly on-demand H100 rental with no contract. Lambda Labs requires a 1-hour minimum. Vast.ai bills by the minute. RunPod, Verda, AWS, GCP, Crusoe, Cudo Compute, Hyperstack, and Nebius all support hourly on-demand or per-second billing.
How fast can I get an H100 instance running?▾
On RunPod, Vast.ai, Verda, and Lambda Labs, instance start time is typically under 60 seconds. AWS p5 and GCP A3 take 2–5 minutes due to longer provisioning. CoreWeave and Nebius reserved clusters take longer to provision (hours) but offer better networking once running.
Is renting an H100 still cheaper than buying in 2026?▾
For most teams, yes. A single H100 SXM5 costs $30,000–$40,000. At a 50% utilization rate, rental breakeven is roughly 18–24 months at current $2/hr cheapest rates. Below 30% utilization (most teams), renting is cheaper indefinitely. Above 60% utilization for 24+ months, buying or a 3-year reserved contract starts to win.
