GPU Servers: Dedicated H100, H200 and A100 Nodes, 1x to 8x, by the Minute
Dedicated GPU servers from 5+ providers in one marketplace. Pick a GPU and a node size, deploy in minutes, and pay per minute for whole GPUs that nobody else shares.
TL;DR: GPU servers on Spheron
- Cheapest server: 1x A100 PCIe 80GB at $1.20/hr (spot) on Spheron, as of Oct 20, 2026.
- 8x H200: from $26.57/hr per node (H200 SXM5, spot), as of Oct 20, 2026.
- 8x A100: from $12.00/hr per node (A100 PCIe 80GB, on-demand), as of Oct 20, 2026.
- Access: whole GPUs passed through to a dedicated VM with root SSH, billed per minute after a 20-minute minimum. Bare metal on reserved orders. Per-GPU rates for every model
GPU server prices by model
| GPU server | Form factors | 1x node / hr | 8x node / hr |
|---|---|---|---|
| NVIDIA H100 server | H100 SXM5, H100 NVL, H100 PCIe | $2.21/hrH100 PCIe, spot | No offer right now |
| NVIDIA H200 server | H200 SXM5 | $3.35/hrH200 SXM5, spot | $26.57/hrH200 SXM5, spot |
| NVIDIA A100 server | A100 SXM4 80GB, A100 PCIe 80GB | $1.20/hrA100 PCIe 80GB, spot | $12.00/hrA100 PCIe 80GB, on-demand |
Each figure is the cheapest node on offer at that size when this page was built, or a dated last-known estimate when none was live. Spot nodes can be reclaimed without notice. For B200, B300 and other GPUs, see the GPU rental catalog.
Dedicated GPU server vs GPU VPS vs bare metal
| Feature | Dedicated server | GPU VPS | Bare metal |
|---|---|---|---|
| GPU access | Whole GPUs passed through to your VM | A slice or time-share of one GPU | Whole GPUs, and the host OS too |
| Noisy neighbours | None on the GPUs | Yes, other tenants share the card | None |
| Sizes | 1x, 2x, 4x, 8x | Fractions of 1 GPU | 8x nodes and multi-node clusters |
| Billing | Per minute, no contract | Often monthly plans | Reserved term |
The servers on this page are dedicated. Bare-metal clusters are sold through reserved GPU commitments.
GPU server FAQ
What a GPU server is, what it costs, and dedicated vs VPS vs bare metal.
A GPU server is a machine with one or more datacenter GPUs attached, plus its own CPUs, RAM and local disk. Renting one gets you the same hardware an on-premise cluster would, without buying it. The servers on this page have 1, 2, 4 or 8 NVIDIA H100, H200 or A100 GPUs, billed per minute on Spheron.
The cheapest on-demand GPU server on Spheron is $1.43/hr for 1x A100 PCIe 80GB, as of Oct 20, 2026. Spot nodes, which can be reclaimed without notice, start at $1.20/hr (1x A100 PCIe 80GB). Rates come from 5+ providers and move with supply; the table on this page shows the cheapest node for each GPU.
A GPU VPS shares one physical GPU between tenants, by slicing it or by time-sharing it, so your throughput depends on who else is running. A dedicated GPU server gives you whole GPUs that nobody else is scheduled on. For training and for latency-sensitive inference, dedicated is what you want.
Yes, on reserved orders. The on-demand and spot servers on this page are single-tenant VMs with the GPUs passed through whole. Bare-metal H100, H200 and A100 clusters, where you control the host OS and the multi-node network, are sold on reserved terms.