NVIDIA H100 Server: Dedicated 1x to 8x H100 GPU Nodes. 8x H100 Server from $25.61/hr
A dedicated H100 server with 1, 2, 4 or 8 GPUs, rented by the minute. Every GPU in the node is yours: no slicing, no time-sharing, root SSH into the node. The table below shows the cheapest live node at each size, for SXM5, NVL and PCIe boards.
TL;DR: NVIDIA H100 server
- 8x node: the cheapest 8x H100 server on Spheron is $25.61/hr (H100 SXM5, spot), $3.20/hr per GPU, as of Oct 20, 2026.
- 1x server: a single-GPU H100 server on Spheron is $2.21/hr (H100 PCIe, spot), as of Oct 20, 2026.
- Node pricing: each size is its own offer, not the 1x rate times 8: an 8x H100 SXM5 spot node works out to $3.20/hr per GPU against $2.79/hr for 1x, as of Oct 20, 2026.
- Form factors: H100 SXM5 (80GB HBM3), H100 PCIe (80GB HBM2e). See the interconnect section for NVLink and PCIe.
- Access: dedicated GPUs, root SSH, per-minute billing. Bare-metal clusters on reserved orders. Single H100 GPU rental
H100 server configurations and node prices
| Node | Node / hr | Per GPU / hr | Host |
|---|---|---|---|
| 1x H100 SXM5On-demand | $4.22/hr | $4.22/hr | 24 vCPU · 200GB RAM · 200GB disk |
| 1x H100 SXM5Spot | $2.79/hr | $2.79/hr | 16 vCPU · 200GB RAM · 256GB disk |
| 2x H100 SXM5On-demand | $8.41/hr | $4.20/hr | 48 vCPU · 400GB RAM · 200GB disk |
| 4x H100 SXM5On-demand | $16.79/hr | $4.20/hr | 96 vCPU · 800GB RAM · 200GB disk |
| 8x H100 SXM5On-demand | $33.54/hr | $4.19/hr | 192 vCPU · 1,600GB RAM · 200GB disk |
| 8x H100 SXM5Spot | $25.61/hr | $3.20/hr | 192 vCPU · 960GB RAM · 20.1TB disk |
| 1x H100 PCIeOn-demand | $3.96/hr | $3.96/hr | 24 vCPU · 240GB RAM · 1TB disk |
| 1x H100 PCIeSpot | $2.21/hr | $2.21/hr | 28 vCPU · 180GB RAM · 850GB disk |
| 2x H100 PCIeOn-demand | $7.92/hr | $3.96/hr | 48 vCPU · 480GB RAM · 2TB disk |
| 2x H100 PCIeSpot | $4.41/hr | $2.20/hr | 60 vCPU · 360GB RAM · 1.6TB disk |
| 4x H100 PCIeSpot | $8.81/hr | $2.20/hr | 124 vCPU · 720GB RAM · 3.3TB disk |
Each row is the cheapest live offer at that size, form factor and tier when this page was built. A size missing from the table had no offer at that moment. Spot nodes can be reclaimed without notice. Need one GPU, not a node? See H100 GPU rental or current GPU pricing.
H100 form factors and interconnect
- Memory
- 80GB HBM3, 3.35 TB/s
- GPU to GPU
- NVLink 4, 900 GB/s per GPU, all GPUs in the node
The HGX board. Pick it for multi-GPU training, where tensor and pipeline parallel traffic crosses NVLink every step. 8x node from $25.61/hr.
- Memory
- 94GB HBM3, 3.9 TB/s
- GPU to GPU
- PCIe Gen5 to the host, NVLink bridges pair GPUs at 600 GB/s
14GB more memory per GPU than SXM5 and higher bandwidth. Good for inference fleets where each model fits on one GPU or a bridged pair.
- Memory
- 80GB HBM2e, 2.0 TB/s
- GPU to GPU
- PCIe Gen5, 128 GB/s per GPU. These offers list no NVLink bridge
Hopper on PCIe cards in standard servers. Fine for single-GPU jobs and data-parallel work that rarely talks across GPUs. 4x node from $8.81/hr.
Which H100 node size to pick
- 1x H100
- LoRA fine-tuning of 7B to 13B models, or QLoRA on a 70B model. Serving one model up to about 70B at 4-bit.
- 2x H100
- Serving Llama 3 70B at FP8 with tensor parallel 2, with room left for KV cache at moderate concurrency.
- 4x H100
- Full fine-tuning of 13B-class models with FSDP, or serving 70B at FP16 with long contexts.
- 8x H100
- 640GB of GPU memory in one box. Pretraining runs, full fine-tuning of 30B-class models, LoRA on 70B, and serving 400B-class MoE models at FP8. Pick SXM5 when the job is tensor parallel.
Dedicated H100 server or bare metal
Every H100 node above, on-demand or spot, is a single-tenant VM: the GPUs are passed through whole, nobody else is scheduled on them, and you get root SSH into the node. That is different from a GPU VPS, where a host slices one card between tenants or caps your share of it.
Bare-metal H100 clusters, where you control the host OS and multi-node networking, are sold on reserved terms. Reserved GPU commitments cover capacity, term and pricing.
H100 server FAQ
What an H100 server costs, which form factor to pick, and what you get on Spheron.
On Spheron, an 8x H100 server starts at $25.61/hr for the whole node (H100 SXM5, spot), $3.20/hr per GPU. By form factor: H100 SXM5 from $25.61/hr (spot). Rates are live as of Oct 20, 2026 and move with marketplace supply. Spot nodes can be reclaimed without notice; on-demand nodes run until you stop them.
An H100 server is one machine with one to eight NVIDIA H100 GPUs attached, plus its own CPUs, RAM and local disk. The standard 8-GPU build uses the HGX H100 board, where all eight SXM5 GPUs talk over NVLink at 900 GB/s each. PCIe and NVL servers put the GPUs on PCIe cards instead: NVL cards pair up over NVLink bridges at 600 GB/s, and all other GPU-to-GPU traffic crosses PCIe Gen5.
Dedicated. Each node is a VM with the whole GPUs passed through to it, so no other tenant runs on your GPUs. Bare-metal H100 clusters, with the host OS under your control, are available on reserved orders.
SXM5 for multi-GPU training and any tensor-parallel serving, since NVLink carries that traffic at 900 GB/s per GPU across the whole node. NVL for inference where you want 94GB per GPU. PCIe when each job fits on one GPU. Compare the live node rates in the table; they move with supply.