GPU rental prices · Hopper
NVIDIA H100 rental starts at $2.89 per GPU-hour (Runpod, PCIe) across the 7 offers we track from 4 providers. The typical price is $3.99, and one GPU running all month costs about $2,110 at the lowest rate. Rent the SXM version for multi-GPU training, where its NVLink and higher power limit matter. For one-GPU inference, the PCIe version has the same 80 GB of memory and costs about $1 less per hour at the same provider.
On-demand prices
Every H100 offer in our catalog, cheapest first. Multi-GPU machines are divided by their GPU count, so you can compare one GPU with one GPU. Monthly cost assumes the GPU runs all 730 hours.
| Provider and offer | Version | GPU memory | GPUs | Per GPU-hour | Per month, 1 GPU | Checked |
|---|---|---|---|---|---|---|
| RunpodRunpod Pod · NVIDIA H100 PCIe | PCIe | 80 GB | 1 | $2.89 | $2,110 | 3 Nov 2026 |
| RunpodRunpod Pod · NVIDIA H100 NVL | NVL | 94 GB | 1 | $3.19 | $2,329 | 3 Nov 2026 |
| LambdaLambda NVIDIA H100 PCIe · 1 GPU | PCIe | 80 GB | 1 | $3.29 | $2,402 | 3 Nov 2026 |
| RunpodRunpod Pod · NVIDIA H100 SXM | SXM | 80 GB | 1 | $3.99 | $2,913 | 3 Nov 2026 |
| LambdaLambda NVIDIA H100 SXM · 1 GPU | SXM | 80 GB | 1 | $4.29 | $3,132 | 3 Nov 2026 |
| DigitalOceanNVIDIA HGX H100 GPU Droplet · 1 GPU | SXM (HGX) | 80 GB | 1 | $4.41 | $3,219 | 3 Nov 2026 |
| PaperspacePaperspace H100 · 1 GPU | SXM | 80 GB | 1 | $5.95 | $4,344 | 3 Nov 2026 |
On-demand list prices before tax, from each provider's pricing page. Community and reserved prices, where a provider offers them, are lower; see the notes below. Full offer details, storage and billing terms.
Before you rent
SXM boards sit in 8-GPU HGX servers with 900 GB/s NVLink between GPUs and run at up to 700 W, which is what multi-GPU training needs. PCIe and NVL cards fit ordinary servers. The NVL has 94 GB and 3.9 TB/s of memory bandwidth, which suits large-model inference on one or two cards.
Lambda charges $3.99 per GPU-hour on its 8-GPU HGX H100 machines, against $4.29 for a single H100 SXM. DigitalOcean charges the same per-GPU rate for 1 or 8 GPUs.
Runpod's Community Cloud, which runs on third-party hosts' machines, lists the H100 SXM at $2.69 an hour against $3.99 on its Secure Cloud. DigitalOcean offers the H100 at $3.26 an hour on a 12-month commitment.
AI models
Memory to load each model size on 80 GB GPUs, with a 20% allowance for runtime buffers, and what that many GPUs cost per hour at the lowest rate above. Long contexts and many parallel users need more memory. More than one GPU only works if your inference server splits the model across them.
| Model size | 4-bit | 8-bit | 16-bit |
|---|---|---|---|
| 8BLlama 3.1 8B, Qwen3 8B | 4.5 GB1 GPU · $2.89/hr | 9.0 GB1 GPU · $2.89/hr | 17.9 GB1 GPU · $2.89/hr |
| 32BQwen3 32B, Gemma 4 31B | 17.9 GB1 GPU · $2.89/hr | 35.8 GB1 GPU · $2.89/hr | 71.6 GB1 GPU · $2.89/hr |
| 70BLlama 3.3 70B, Qwen 72B | 39.2 GB1 GPU · $2.89/hr | 78.3 GB1 GPU · $2.89/hr | 156.5 GB2 GPUs · $5.78/hr |
| 120Bgpt-oss-120b class | 67.1 GB1 GPU · $2.89/hr | 134.2 GB2 GPUs · $5.78/hr | 268.3 GB4 GPUs · $11.56/hr |
| 685BDeepSeek-V3.2 | 382.8 GB5 GPUs · $14.45/hr | 765.6 GB10 GPUs · more than one 8-GPU server | 1,531.1 GB20 GPUs · more than one 8-GPU server |
Weights only, from the parameter count. Check a specific model, context length and user count in the model hardware calculator.
Hardware
| Specification | H100 SXM | H100 NVL |
|---|---|---|
| GPU memory | 80 GB | 94 GB |
| Memory bandwidth | 3.35 TB/s | 3.9 TB/s |
| FP8 Tensor (sparse) | 3,958 TFLOPS | 3,341 TFLOPS |
| FP16 Tensor (sparse) | 1,979 TFLOPS | 1,671 TFLOPS |
| Max power | Up to 700 W | 350–400 W |
| Form factor | SXM | PCIe, dual-slot |
| NVLink | 900 GB/s | 600 GB/s |
From NVIDIA's product pages. Tensor figures marked “sparse” assume structured sparsity; dense throughput is half.
Questions
$2.89 per GPU-hour is the lowest on-demand price we track, from Runpod. The typical price is $3.99, and one GPU running all month costs about $2,110 at the lowest rate. Community and reserved prices go lower.
At 4-bit, yes: the weights need about 39.2 GB with a 20% allowance. At 8-bit they take 78.3 GB, which fits but leaves under 2 GB of the 80 GB for the conversation cache, so use two GPUs for longer chats or several users. At 16-bit a 70B model needs 156.5 GB, or two H100s.
The H200 has the same Hopper architecture and the same 3,958 TFLOPS of sparse FP8 as the H100 SXM, with more and faster memory: 141 GB at 4.8 TB/s against 80 GB at 3.35 TB/s. Rent the H200 when a model and its cache don't fit in 80 GB, or when you serve long contexts. At DigitalOcean the two cost almost the same ($4.47 against $4.41 per GPU-hour).
Compare
Prices and availability change, and we don't check stock. Confirm the current price and region with the provider before you rent. Sources: www.nvidia.com, www.runpod.io, lambda.ai, www.digitalocean.com, docs.digitalocean.com . See our affiliate disclosure.
