The NVIDIA H100 price ranges from $2.99 to $12.29 per GPU-hour to rent, and about $36,000 per card to buy. It's still a widely rented data-center GPU, and where you rent it moves the price more than anything else.
This guide covers NVIDIA H100 pricing: on-demand cloud rates across 16 providers, what an H100 costs to buy, and when buying pays off. It also breaks down the H100's SXM, PCIe and NVL variants, and compares it against the A100 and H200.
Takeaways
- Renting an H100 costs $2.99–$12.29 per GPU-hour, with a median of $4.40 across 16 providers.
- Neoclouds charge less than half what hyperscalers do ($4.12 vs $10.04 per GPU-hour on average).
- Vast.ai's H100 supply is thin and its prices swing daily, going from 9 verified H100 hosts a day in August to 0–2 in November.
- A new H100 PCIe costs about $36,000, and an 8-GPU HGX server costs around $300,000.
- Thunder Compute charges $3.20/hr for on-demand access with per-minute billing.
On-demand GPUs
Starting at $0.35/hr, billed per minute.
NVIDIA H100 Cloud Pricing
| Provider | GPU / Instance | On-Demand $/GPU-hr* | Notes |
|---|---|---|---|
| Vultr | H100 80GB | $2.99 | |
| Vast.ai | H100 80GB | $3.00 | Median of 2 verified US/CA hosts on Nov 8; ranges $2.40-$3.59/GPU-hr. Priced as compute plus 100GB storage, with bandwidth billed separately. |
| Hyperbolic | H100 80GB | $3.19 | |
| Thunder Compute | H100 80GB | $3.20 | Per-minute billing. |
| Lambda | H100 80GB | $3.79 | |
| Crusoe Cloud | H100 80GB | $3.90 | |
| Modal | H100 80GB | $3.95 | |
| Runpod | H100 80GB | $4.39 | |
| DigitalOcean | H100 80GB | $4.41 | |
| Nebius | H100 80GB | $4.50 | |
| Paperspace | H100 80GB | $5.95 | |
| CoreWeave | 8× H100 80GB | $6.16 | Normalized from an 8-GPU node. |
| AWS | p5.48xlarge | $6.88 | Normalized from an 8-GPU node. Cheapest US region (us-east-1); ranges $6.88-$8.60/GPU-hr across 7 US options. |
| Oracle Cloud | BM.GPU.H100.8 | $10.00 | Normalized from an 8-GPU node. |
| Google Cloud | a3-highgpu-8g | $10.98 | Normalized from an 8-GPU node. Cheapest US region (us-central1); ranges $10.98-$13.20/GPU-hr across 10 US options. |
| Azure | Standard_ND96isr_H100_v5 | $12.29 | Normalized from an 8-GPU node ($98.32/hr). Cheapest US region (eastus); ranges $12.29-$15.98/GPU-hr across 9 US options. |
| *Published U.S. on-demand list rates in USD. Last reviewed on November 13, 2026. | |||
Vast.ai H100 Prices and Availability
Vast.ai's H100 rate depends on which hosts are available on a given day. We have checked Vast.ai's verified H100 80GB listings daily since August 5, 2026, and both supply and price have been unstable.
| Month | Days with H100 availability | Avg. daily hosts | Daily median $/GPU-hr |
|---|---|---|---|
| August | 96% | 8.6 | $2.38–$6.59 |
| September | 90% | 3.6 | $2.51–$6.80 |
| October | 97% | 1.4 | $2.23–$10.95 |
| November 1–9 | 44% | 1.3 | $2.23–$3.00 |
| Verified US/CA hosts, checked daily since Aug 5, 2026. | |||
- Supply has dried up. Vast.ai averaged 9 verified H100 hosts a day in August. Since October, most days have had a single host, and no H100 was listed on 5 days in November.
- The price moves daily. The daily median changed by 29% on average from one day to the next. It hit $10.95 on October 16 and fell to $2.23 on October 31.
H100 Cost Benchmark
| Provider | On-Demand $/GPU-hr | 10-Hour Cost |
|---|---|---|
| Vultr | $2.99 | $29.90 |
| Vast.ai | $3.00 | $29.98 |
| Hyperbolic | $3.19 | $31.90 |
| Thunder Compute | $3.20 | $32.00 |
| Lambda | $3.79 | $37.90 |
| Crusoe Cloud | $3.90 | $39.00 |
| Modal | $3.95 | $39.49 |
| Runpod | $4.39 | $43.90 |
| DigitalOcean | $4.41 | $44.10 |
| Nebius | $4.50 | $45.00 |
| Paperspace | $5.95 | $59.50 |
| CoreWeave | $6.16 | $61.55 |
| AWS | $6.88 | $68.80 |
| Oracle Cloud | $10.00 | $100.00 |
| Google Cloud | $10.98 | $109.79 |
| Azure | $12.29 | $122.90 |
NoteH100 capacity is often tight. If your job fits on an A100, Thunder Compute's A100 80GB is a cost-effective alternative at $1.09/hr.
H100 Hardware Price: Buy vs Rent
NVIDIA doesn't publish list prices for data-center GPUs, so the prices below come from listings at Newegg and SabrePC.
The H100 PCIe can be purchased as a single card, while the H100 SXM is often sold in 8-GPU systems.
This breakeven calculation is based on Thunder Compute's H100 rate of $3.20/hr. It includes electricity at about $0.13/kWh, but it doesn't factor in depreciation or other costs.
| Option | Price | Rent | Pays off (24/7) | Pays off (8h/day) |
|---|---|---|---|---|
| 1× H100 PCIe 80GB1 | $35,989 | $3.20/hr | 1.3 yr | 4.0 yr |
| 8× H100 SXM HGX server2 | $308,935 | $25.60/hr | 1.5 yr | 4.4 yr |
|
1 H100 80GB PCIe at Newegg, November 2026. 2 SabreEDGE 6U HGX H100 server, base configuration, November 2026. Thinkmate's 8-GPU HGX H100 systems start at $296,378. |
||||
At 8 hours a day, a purchased H100 takes over four years to pay off. Buying only makes sense for steady 24/7 use.
H100 Specifications: SXM vs PCIe vs NVL
All three H100 variants use the same Hopper GH100 die. They only differ in memory, power and interconnect.
- The H100 SXM runs at up to 700W and comes in 4- and 8-GPU HGX systems.
- The H100 PCIe is a standard 350W card with slower HBM2e memory. It fits ordinary servers and is sold as a single card.
- The H100 NVL is a PCIe card with 94GB of HBM3, built for LLM inference. NVLink bridges connect pairs of cards.
| Spec | H100 SXM | H100 PCIe | H100 NVL |
|---|---|---|---|
| Architecture | Hopper (GH100) | ||
| VRAM | 80 GB HBM3 | 80 GB HBM2e | 94 GB HBM3 |
| Memory bandwidth | 3.35 TB/s | 2 TB/s | 3.9 TB/s |
| FP81 | 3,958 TFLOPS | 3,026 TFLOPS | 3,341 TFLOPS |
| FP16 / BF161 | 1,979 TFLOPS | 1,513 TFLOPS | 1,671 TFLOPS |
| TF321 | 989 TFLOPS | 756 TFLOPS | 835 TFLOPS |
| FP64 | 34 TFLOPS | 26 TFLOPS | 30 TFLOPS |
| NVLink | 900 GB/s, 8 GPUs | 600 GB/s bridge | 600 GB/s, 2 GPUs |
| TDP | Up to 700W | 300–350W | 350–400W |
| MIG | Up to 7 instances | ||
| 1 With sparsity. Dense is half. Full breakdown in our NVIDIA H100 specs guide. | |||
H100 vs A100
The A100 80GB has the same memory capacity as the H100, but the H100 adds FP8 support through its Transformer Engine, about 64% more memory bandwidth, and roughly three times the FP16 tensor throughput.
| Spec | A100 80GB SXM | H100 SXM |
|---|---|---|
| Architecture | Ampere (GA100) | Hopper (GH100) |
| CUDA Cores | 6,912 | 16,896 |
| VRAM | 80 GB HBM2e | 80 GB HBM3 |
| Memory Bandwidth | 2.0 TB/s | 3.35 TB/s |
| FP8 Throughput1 | Not supported | 3,958 TFLOPS |
| FP16 / BF16 Throughput1 | 624 TFLOPS | 1,979 TFLOPS |
| NVLink Bandwidth | 600 GB/s | 900 GB/s |
| TDP | 400W | Up to 700W |
| Thunder Compute price | $1.09/hr | $3.20/hr |
| 1 With sparsity. Dense throughput is half. | ||
At Thunder Compute's rates, the H100 costs almost 3 times as much; it needs to finish a job 3 times faster to cost the same. That happens for FP8 inference and large compute-bound training runs. For memory-bound inference, fine-tuning and development, the A100 is usually the cost-effective choice.
Read nextA100 pricing comparison for rates across providers.
H100 vs H200
The H200 uses the same GH100 die as the H100, so compute throughput is identical. The difference is memory: the H200 replaces the H100's 80 GB of HBM3 with 141 GB of HBM3e, increasing capacity by 76% and bandwidth by 43%.
| Spec | H100 SXM | H200 SXM |
|---|---|---|
| Architecture | Hopper (GH100 die) | |
| VRAM | 80 GB HBM3 | 141 GB HBM3e |
| Memory Bandwidth | 3.35 TB/s | 4.8 TB/s |
| FP8 Throughput1 | 3,958 TFLOPS | |
| FP16 / BF16 Throughput1 | 1,979 TFLOPS | |
| TDP | Up to 700W | |
| Cloud price (on-demand) | $2.99–$12.29/GPU-hr | $3.99–$10.85/GPU-hr |
| 1 With sparsity. Dense throughput is half. | ||
Read nextFull comparison of AWS P5 instances powered by NVIDIA H100 and H200 GPUs.
When to Choose the H100
For development, fine-tuning, and any inference workload that fits inside 80GB, the H100 is the cost-effective Hopper choice. The compute matches the H200, and Thunder Compute's H100 80GB is available at $3.20/hr with per-minute billing.
What LLMs Fit on H100 GPUs?
With 80GB of VRAM per card, an H100 runs small and mid-sized models on a single GPU, while 70B models at FP16 and the largest MoE models need multi-GPU nodes. The estimates below are published weight sizes only; add 10–30% for KV cache, activations, and framework buffers at a moderate context window.
| Model | Parameters | Precision | VRAM (weights only)1 | H100 setup |
|---|---|---|---|---|
| Llama 3 70B | 70B | FP8 | ~70 GB | 2× H100 node |
| Llama 3 70B | 70B | FP16 | ~140 GB | 2× H100 node |
| Qwen 2.5 72B | 72B | FP8 | ~72 GB | 2× H100 node |
| Llama 4 Scout | 109B MoE | 4-bit | ~65 GB | 1× H100, limited KV headroom |
| Llama 4 Maverick | 400B MoE | 4-bit | ~245 GB | 4× H100 node |
| GLM-5.2 | 744B MoE | INT4 | ~372 GB | 8× H100 node |
| Kimi K2.7 | 1T MoE | INT4 | ~594 GB | 8× H100 node, limited context headroom |
| 1 VRAM estimates include weights only. | ||||
Last Thoughts on NVIDIA H100 Pricing
The H100 rents for $2.99–$12.29 per GPU-hour, and the provider matters more than the GPU: hyperscalers charge more than twice as much as neoclouds for the same card. Buying only pays off for steady 24/7 use.
If your model fits in 80GB and you don't need FP8, an A100 often gets the same job done for about a third of the hourly price.
FAQ
How much does an NVIDIA H100 cost?
A new H100 PCIe 80GB lists at about $36,000 on Newegg. The SXM version is sold inside 8-GPU HGX servers, which start at about $296,000–$309,000 at Thinkmate and SabrePC. Renting costs $2.99–$12.29 per GPU-hour on-demand across 16 cloud providers, with a median of $4.40 (November 2026).
What is the cheapest NVIDIA H100 cloud provider?
Vultr has the lowest fixed H100 rate at $2.99/GPU-hr as of November 2026. Vast.ai's median was $3.00/GPU-hr on November 8, but that's from 2 hosts listing an H100 and the price changes daily. Hyperbolic ($3.19), Thunder Compute ($3.20) and Lambda ($3.79) follow. Among hyperscalers, AWS is cheapest at $6.88/GPU-hr.
How much does an H100 cost per hour on AWS, Azure and Google Cloud?
AWS charges $6.88/GPU-hr on the p5.48xlarge (normalized from the 8-GPU node, us-east-1). Azure charges $12.29/GPU-hr on the ND96isr H100 v5 (8-GPU node). Google Cloud charges $10.98/GPU-hr on the a3-highgpu-8g in us-central1 (8-GPU node).
How much VRAM does the NVIDIA H100 have?
The H100 SXM and H100 PCIe have 80GB of memory: HBM3 at 3.35 TB/s on SXM and HBM2e at 2 TB/s on PCIe. The H100 NVL has 94GB of HBM3 at 3.9 TB/s.
Is the H100 worth it over the A100?
Only if your workload uses FP8 or is compute-bound. The H100 has the same 80GB of memory as the A100 80GB, about 64% more memory bandwidth and roughly 3x the FP16 tensor throughput. On Thunder Compute, an H100 costs $3.20/hr and an A100 80GB costs $1.09/hr, so the H100 must finish a job about 2.9x faster to cost less.
What is the difference between the H100 and H200?
Both use the same Hopper GH100 die, so compute throughput is identical. But the H200 has 141GB of HBM3e at 4.8 TB/s versus the H100's 80GB of HBM3 at 3.35 TB/s. The extra memory matters only for models or context windows that don't fit in 80GB.