NVIDIA H100 Pricing
Eli Mernit
Renting an NVIDIA H100 PCIe costs $1.83 to $3.29 per GPU-hour on demand at specialist clouds. Hyperscaler VMs run $6.98 to $11.06 per GPU, and the top figure is Google's 8-GPU SXM VM. The median across 40 providers tracked by getdeploying is about $3.40, counting every H100 variant. Beam's on-demand H100 PCIe machine costs $1.83 per hour.
Key takeaways
- Beam's $1.83 machine is the cheapest on-demand H100 PCIe here. Vast.ai ($1.93) and RunPod ($1.99) are within $0.16.
- Google's on-demand H100 is an 8-GPU SXM VM at $88.49 per hour, so its per-GPU rate assumes you use all eight.
- Interruptible capacity is far cheaper: Vast.ai's interruptible H100 PCIe offers run around $0.35 per GPU-hour.
H100 PCIe cloud prices per hour, by provider
| Provider | SKU | On-demand $/GPU-hr | Notes |
|---|---|---|---|
| Beam | H100 PCIe | $1.83 | On-demand machine; vCPU, RAM, and NVMe included |
| Vast.ai | H100 PCIe (marketplace) | $1.93 | Host-set price; median offer $2.57; low availability |
| RunPod | H100 PCIe (Community Cloud) | $1.99 | Per-second billing; Secure Cloud $2.89 |
| Hyperstack | H100 PCIe | $2.50 | Per-minute billing; spot $2.00; out of stock (getdeploying) |
| Spheron | H100 PCIe | $2.64 | Live marketplace rate; per-minute billing, 20-min minimum |
| Lambda | H100 PCIe | $3.29 | Fixed public rate; 1x instance only |
| Azure | NC40ads H100 v5 (H100 NVL) | $6.98 | 1-GPU VM; NVL is a 94 GB PCIe card; East US |
| Google Cloud | a3-highgpu-8g (8x H100 SXM) | $11.06 | 8-GPU VM only ($88.49/hr), us-central1; SXM, not PCIe; spot $52.96/hr |
| Market median | — | $3.40 | Across 40 providers with an on-demand price (getdeploying); all H100 variants |
How these prices were collected
Each figure is an on-demand list price per GPU-hour from the provider's own pricing page, with spot excluded. We used single-GPU shapes where offered and divided 8-GPU VM prices by eight.
- Beam pricing page
- Marketplace data from Vast.ai: Vast.ai H100 PCIe pricing
- RunPod GPU pricing page
- Hyperstack GPU pricing page, including spot VM rates
- Spheron H100 rental page
- Lambda instance pricing page
- Azure virtual machine pricing page (NC40ads H100 v5)
- Google Cloud accelerator-optimized VM pricing
- getdeploying H100 price tracker for the market median
H100 PCIe vs SXM5: what you trade away
The standard PCIe card is the slower H100: fewer SMs, HBM2e instead of HBM3, a lower power limit, and an NVLink bridge to at most one other card, where SXM5 boards join eight GPUs over NVSwitch. It keeps the Transformer Engine and FP8, so single-GPU inference and LoRA fine-tuning run well on it. The gap shows in multi-GPU distributed training, where SXM5 nodes win. Azure's H100 NVL sits between the two, a PCIe card with 94 GB.
Where to rent an H100 PCIe for less
On demand, nobody here undercuts Beam's $1.83, which includes vCPU, RAM, and NVMe. Vast.ai and RunPod come close, but Vast prices are set by individual hosts and change hourly, and RunPod's $1.99 is its Community Cloud; the Secure Cloud tier costs $2.89. Lambda charges $3.29 and bundles 26 vCPUs and 225 GiB of RAM. Paying less than $1.83 means interruptible capacity: Vast.ai's interruptible offers run around $0.35 per GPU-hour and can stop at any time, so checkpoint often.
What this means for your workload
An H100 PCIe suits single-GPU jobs that fit in 80 GB, like serving a mid-size LLM, LoRA or QLoRA fine-tuning, and FP8 batch inference. If weights plus KV cache outgrow 80 GB, an H200 with 141 GB beats splitting the model across PCIe cards; see our H200 pricing comparison.
FAQ
How much does an H100 PCIe cost per hour?
Specialist clouds charge $1.83 to $3.29 per GPU-hour on demand, with Beam at the low end and Lambda at the top. Azure's single-GPU H100 NVL VM costs $6.98 per hour.
What's the cheapest way to rent an H100 PCIe?
Beam's $1.83 on-demand machine is the lowest price we found, ahead of Vast.ai ($1.93) and RunPod Community Cloud ($1.99). Jobs that tolerate interruptions can use Vast.ai's interruptible offers at around $0.35 per GPU-hour.
H100 PCIe vs H100 SXM: which should I rent?
Rent the PCIe card for single-GPU inference and fine-tuning, where it costs less and keeps FP8 support. Rent SXM for multi-GPU training, because SXM boards link all eight GPUs through NVSwitch.
