Renting 2- and 4-GPU servers instead of a full node
Compare European providers that rent 1, 2 or 4 GPUs instead of a full 8-GPU node, with per-GPU rates from each pricing page and the NVLink trade-off explained.
In this article7

Who rents partial nodes, and at what rate per GPU
A team that needs two or four GPUs rather than a whole eight-GPU node has fewer published options than the market's full-node marketing suggests; the wider field of Europe's AI infrastructure providers is mostly built around the full node. A reader in that position can take a Lyceum on-demand GPU VM in a 1, 4 or 8 GPU shape with NVLink at a published per-GPU rate, billed per second with no minimum (read 2026-09-22). The page lists 1x, 4x and 8x configurations and no 2-GPU shape, so a two-GPU requirement means either two single-GPU VMs or one 4-GPU shape, not a 2-GPU slice.[1]
| Provider | Smallest rentable GPU shape | Per-GPU rate at 1 GPU | Per-GPU rate at 4 GPUs | Per-GPU rate at 8 GPUs | Source page (read 2026-09-22; Exoscale, IONOS and Hetzner re-read 2026-10-02) |
|---|---|---|---|---|---|
| Lyceum | 1-GPU on-demand VM (shapes: 1, 4, 8) | Published per-GPU rate, billed per second | Same published per-GPU rate | Same published per-GPU rate | lyceum.technology/pricing/ |
| Exoscale | 1-GPU A40 instance (A30 and RTX Pro 6000 instances also start at 1 GPU) | A40 €1.05 per GPU-hour (1-GPU Small) | A40 €4.18/hr for 4 GPUs (about €1.05 per GPU, flat) | A40 €8.36/hr for 8 GPUs (about €1.05 per GPU, flat) | exoscale.com/pricing/ |
| Hetzner | GEX dedicated server, one GPU per server, not configurable with more | Monthly server price is rendered by script and could not be read from the page | No 4-GPU shape (one GPU per server) | No 8-GPU shape (one GPU per server) | hetzner.com/dedicated-rootserver/matrix-gpu/ |
| IONOS | Cloud GPU VM H200 - S (1x H200 PCIe) | H200 $3.261 per GPU-hour | H200 $13.043/hr for 4 GPUs (about $3.26 per GPU, flat) | H200 $26.086/hr for 8 GPUs (about $3.26 per GPU, flat) | cloud.ionos.com/cloud-gpu-vm |
| Nebius | HGX H100 instance (16 vCPU, 200 GB RAM) | H100 $3.85 per H100 GPU-hour on demand | $3.85 per H100 GPU-hour list price; the prices page names no 4-GPU shape | $3.85 per H100 GPU-hour list price; the prices page names no 8-GPU shape | nebius.com/prices |
| OVHcloud | h100-380 (1x H100) | H100 $2.99 per H100 GPU-hour, flat across the 1-, 2- and 4-GPU H100 shapes | H100 $2.99 per H100 GPU-hour, flat | No 8-GPU H100 shape on the prices page (largest: h100-1520, 4 GPUs) | ovhcloud.com/en/public-cloud/prices/ |
| Scaleway | L4-1-24G (1 GPU) | L4 €0.79 per L4 GPU-hour, flat (1-8 GPUs) | L4 €0.79 per L4 GPU-hour, flat | L4 €0.79 per L4 GPU-hour, flat | scaleway.com/en/pricing/gpu/ |
| STACKIT | GPU Server-n2.14d.g1-EU01 (1 GPU) | n2 €1.74, n1 €2.66, n3 €8.04 per GPU-hour | n2 €1.74 per GPU-hour (€6.97/hr); n1 €4.08 (€16.34/hr, not flat: above 4x its 1-GPU rate); n3 €8.04 (€32.17/hr) | n3 €8.04 per GPU-hour (€64.33/hr for 8 GPUs) | stackit.com price list (Compute Engine GPU) |
Hetzner's GEX servers each have one GPU and cannot be configured with multiple GPUs, per the FAQ on its GPU server page[3], which renders the server price by script. IONOS prices its H200 PCIe VMs in US dollars[7], and Exoscale's A40 instances run 1, 2, 4 or 8 GPUs at the same per-GPU rate[2].
What a fraction of a node costs in interconnect
The trade-off a partial node makes is interconnect bandwidth, and the numbers are on NVIDIA's H100 specification page: fourth-generation NVLink offers 900 GB/s of GPU-to-GPU interconnect on the SXM part, while PCIe Gen5 carries 128 GB/s.[9] A workload that moves tensors between GPUs therefore runs on a different fabric depending on whether the partial shape sits inside an NVLink domain or is PCIe-attached.
Which workloads genuinely need all eight
- What a multi-GPU VM keeps: a 4-GPU VM on an NVLink node still has its four GPUs on that node's NVLink fabric, which is what Lyceum's VM page states for its 1-8 GPU configurations; a 1-GPU VM has no GPU peer at all.[10]
- What genuinely needs all eight: tensor-parallel training and large-model inference where the weights span more than four GPUs and every GPU talks to every other each step.
- What does not: single-GPU fine-tuning, small-model inference, rendering and embarrassingly parallel batch jobs that shard independently lose nothing by taking a fraction of a node.
Reading a per-GPU rate across shapes
Reading a per-GPU rate across shapes is the second check. OVHcloud, Scaleway's L4 line, Exoscale's A40 line, IONOS's H200 VMs and Lyceum publish flat per-GPU rates, so a 4-GPU shape costs four times a 1-GPU shape. STACKIT's n1 line does not: its 4-GPU shape prices well above four times its 1-GPU rate, so the shape price, not the headline per-GPU rate, is what the buyer checks there.
Choosing a shape for your own job
- Bandwidth first: if the job is tensor-parallel, take the largest multi-GPU shape inside one NVLink domain, and take the full eight only when the model or batch genuinely spans them.
- VRAM arithmetic: four GPUs with 80 GB each hold the same weights as one 320 GB pool only if the framework shards them, so compare total VRAM per shape, not GPU count; GPU memory bandwidth sets the token rate either way.
Applied to the table: for a team that needs four GPUs, Lyceum's 4-GPU on-demand VM, billed per second with no minimum, and OVHcloud's 4x H100 instance (h100-1520) at a flat $2.99 per H100 GPU-hour are the shapes that answer the question directly.[5] For a single GPU, Nebius lists L40S from $1.55 per L40S GPU-hour.[4] The decision rule to carry away: match the shape to the fabric your job needs, then compare shape prices, not headline rates.
Frequently asked questions
Can I rent fewer than 8 GPUs?
Yes. Lyceum offers on-demand GPU VMs in 1, 4 and 8 GPU shapes with NVLink, billed per second with no minimum. OVHcloud rents 1, 2 and 4 GPU instances, Scaleway's L4 line runs from 1 to 8 GPUs, and STACKIT's GPU Server line (n1, n2 and n3 machine types) runs 1-, 2-, 4- and 8-GPU shapes. IONOS's Cloud GPU VMs and Exoscale's A40 instances run 1, 2, 4 or 8 GPUs. Hetzner's GPU servers carry one GPU each and cannot be configured with more. Shape listings read 2026-09-22 (IONOS and Exoscale 2026-10-02).
What does a 4-GPU server cost per hour?
It depends on the GPU and provider. OVHcloud's 4x H100 instance bills at a flat $2.99 per H100 GPU-hour. Lyceum's 4-GPU on-demand VM is billed at a published per-GPU rate, per second, with no minimum. Nebius bills its HGX H100 instance at $3.85 per H100 GPU-hour on demand. IONOS's 4x H200 PCIe VM costs $13.043 per hour and Exoscale's 4x A40 instance €4.18 per hour. Rates read 2026-09-22 (IONOS and Exoscale 2026-10-02).
Do I lose NVLink on a partial node?
Not necessarily. A multi-GPU VM on an NVLink-equipped node keeps the fabric among its own GPUs, and Lyceum states NVLink on its 1-8 GPU VM configurations. But many partial shapes are PCIe-attached, where GPU-to-GPU traffic drops from NVLink's 900 GB/s on an H100 to PCIe Gen5's 128 GB/s, per NVIDIA's specification page. The deciding factor is whether your workload moves data between GPUs at all.
Which providers rent single GPUs?
Lyceum offers a 1-GPU on-demand VM at a published per-GPU rate. Nebius bills GPU instances per GPU-hour. OVHcloud's 1x H100 instance carries one H100 at $2.99 per H100 GPU-hour, and Scaleway's L4-1 runs at EUR 0.79 per L4 GPU-hour. Hetzner sells dedicated GPU servers with a single GPU each. Rates read 2026-09-22.
Sources
Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score
Related reading
All Accelerators →NVIDIA AI chip price
Compare NVIDIA AI chip pricing, from H100 and Blackwell B200 capital costs to the option of renting the same GPUs from a European cloud.

LLM APIs with zero data retention: A provider comparison
Compare 7 LLM inference providers that offer zero data retention. Learn what each commitment covers and how to get API logging restrictions in writing.

Batch OCR on serverless GPUs: cost and throughput
Price a 10,000-page batch OCR run two ways, per token and per GPU-hour, from published European rate cards, with the crossover shown.
