NVIDIA H200 Server: Dedicated 1x to 8x H200 GPU Nodes. 8x H200 Server from $26.57/hr
A dedicated H200 server with 1, 2, 4 or 8 GPUs, rented by the minute. The H200 is an H100 with 141GB of HBM3e and 4.8 TB/s of bandwidth, so the same model needs fewer GPUs. The table below shows the cheapest live node at each size.
TL;DR: NVIDIA H200 server
- 8x node: the cheapest 8x H200 server on Spheron is $26.57/hr (H200 SXM5, spot), $3.32/hr per GPU, as of Oct 20, 2026.
- 1x server: a single-GPU H200 server on Spheron is $3.35/hr (H200 SXM5, spot), as of Oct 20, 2026.
- Node pricing: each size is its own offer, not the 1x rate times 8: an 8x H200 SXM5 on-demand node works out to $4.80/hr per GPU against $5.54/hr for 1x, as of Oct 20, 2026.
- Form factors: H200 SXM5 (141GB HBM3e). See the interconnect section for NVLink and PCIe.
- Access: dedicated GPUs, root SSH, per-minute billing. Bare-metal clusters on reserved orders. Single H200 GPU rental
H200 server configurations and node prices
| Node | Node / hr | Per GPU / hr | Host |
|---|---|---|---|
| 1x H200 SXM5On-demand | $5.54/hr | $5.54/hr | 28 vCPU · 300GB RAM · 200GB disk |
| 1x H200 SXM5Spot | $3.35/hr | $3.35/hr | 16 vCPU · 200GB RAM · 256GB disk |
| 2x H200 SXM5On-demand | $11.05/hr | $5.52/hr | 56 vCPU · 600GB RAM · 200GB disk |
| 4x H200 SXM5On-demand | $22.07/hr | $5.52/hr | 112 vCPU · 1,200GB RAM · 200GB disk |
| 8x H200 SXM5On-demand | $38.40/hr | $4.80/hr | 208 vCPU · 1,904GB RAM · 30TB disk |
| 8x H200 SXM5Spot | $26.57/hr | $3.32/hr | 128 vCPU · 1,600GB RAM · 256GB 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 H200 GPU rental or current GPU pricing.
H200 form factors and interconnect
- Memory
- 141GB HBM3e, 4.8 TB/s
- GPU to GPU
- NVLink 4, 900 GB/s per GPU, all GPUs in the node
The HGX H200 board. 1.1TB of GPU memory per 8-GPU node, all on NVLink. 8x node from $26.57/hr.
Which H200 node size to pick
- 1x H200
- A 70B model at FP8 on one GPU, with room for KV cache. LoRA fine-tuning of 13B to 30B models.
- 2x H200
- Llama 3 70B at FP16, or long-context 70B serving at FP8 with high concurrency.
- 4x H200
- 564GB of GPU memory: 200B-class MoE models at FP8, or LoRA fine-tuning of 70B at BF16.
- 8x H200
- 1.1TB of GPU memory in one node. DeepSeek-class 600B+ MoE models at FP8 on a single server, and full fine-tuning of models up to about 50B with FSDP (70B needs LoRA or optimizer offload).
Dedicated H200 server or bare metal
Every H200 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 H200 clusters, where you control the host OS and multi-node networking, are sold on reserved terms. Reserved GPU commitments cover capacity, term and pricing.
H200 server FAQ
What an H200 server costs, which form factor to pick, and what you get on Spheron.
On Spheron, an 8x H200 server starts at $26.57/hr for the whole node (H200 SXM5, spot), $3.32/hr per GPU. By form factor: H200 SXM5 from $26.57/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 H200 server is a machine with one to eight NVIDIA H200 GPUs. The 8-GPU build is the HGX H200 board: eight SXM5 GPUs with 141GB of HBM3e each, 1,128GB in total, connected by NVLink 4 at 900 GB/s per GPU.
H200 when memory is the limit: long contexts, big KV caches, or a model that would need 8x H100 but fits in 4x H200. H100 when the model already fits and you want the lower hourly rate. Compute per GPU is about the same; the H200 adds memory and bandwidth.
Dedicated. Every GPU in the node is passed through to your VM and no one else runs on it. Bare-metal H200 clusters are available on reserved orders.