# H100 cloud Europe guide: renting H100 GPUs in EU countries

URL: https://cloudj.net/blog/h100-cloud-europe
Published: 2026-11-12 · Updated: 2026-11-12

> An H100 cloud Europe checklist: SXM, PCIe and NVL variants, searching EU country codes, the account level datacenter GPUs need and the Datacenter only switch.

Renting H100 cloud Europe capacity on Cloudj is a search, not a reservation. GPUs come from a live inventory of independent host machines in many countries, and you find European H100s by filtering that inventory by model and country code. Because "H100" covers three different products on the GPU list, and because datacenter GPUs need a higher account level, it pays to know what you are searching for before you open the console. This article works as a checklist for that.

For the wider European picture, start with the [GPU cloud Europe guide](https://cloudj.net/blog/gpu-cloud-europe-guide).

## Checklist item 1: know which H100 you want

The GPU list has three H100 variants, each with its own page and its own live stock and price:

| Variant | VRAM | Form factor | Typical reason to choose it |
|---|---|---|---|
| [H100 SXM](https://cloudj.net/gpu-cloud/h100-sxm) | 80 GB | SXM module on a multi-GPU board | Multi-GPU training where GPUs exchange data constantly; SXM boards usually connect GPUs with NVLink |
| [H100 PCIe](https://cloudj.net/gpu-cloud/h100-pcie) | 80 GB | PCIe card | Single-GPU fine-tuning and inference; lower power than SXM |
| [H100 NVL](https://cloudj.net/gpu-cloud/h100-nvl) | 94 GB | PCIe card, designed for pairing | Large-model inference where the extra memory per GPU helps |

The form-factor notes are general NVIDIA product background; what a particular machine offers is what its listing shows. The practical point is that stock differs by variant. A search for H100 SXM in Germany can be empty while an H100 PCIe is listed, or the other way round. If your code does not depend on NVLink, search all three.

## Checklist item 2: check your account level first

Workstation and datacenter GPUs, including the H100 family, need a higher account level than a new account has. New accounts start with consumer GPUs, a small number of instances and an hourly spending limit. If your level does not allow H100s, the console tells you: "This GPU class is not available at your account level".

Sort this out before a deadline. Contact Cloudj support from the console to ask about a higher level, and plan your first H100 run for after that is settled.

## Checklist item 3: search H100 cloud Europe listings by country code

Open the deploy page with a variant preselected, for example [deploy an H100 SXM](https://app.cloudj.net/deploy/gpu?model=h100-sxm), and enter the European codes you accept in the **Country** field:

```text
DE, NL, FR, SE, FI
```

You can separate codes with commas or spaces. Results are sorted by total price, lowest first, up to 50 machines. Location is shown at country level only, so a `DE` listing tells you the machine is in Germany, not which city or data centre.

If a machine is listed, you can deploy it. If not:

- switch to another H100 variant, or to a different GPU count (1x, 2x, 4x or 8x);
- add more EU codes your policy allows, such as `BE`, `PL`, `ES` or `IT`;
- loosen reliability or price filters;
- search again later, because stock changes during the day;
- for training that has no location requirement, clear the Country field.

If no H100 suits the job, an A100 may. The [A100 vs H100 comparison](https://cloudj.net/blog/a100-vs-h100-cloud) explains when the older card is the better value.

## Checklist item 4: decide on "Datacenter only"

Machines hosted in a data centre carry a **Datacenter** badge, and the **Datacenter only** switch hides everything else. Turn it on if your policy or your client requires data-centre hosting; leave it off if you would rather see every option.

The badge describes where the host is. It is not an SLA. Every listing is an independent machine without a hardware SLA, and the console reminds you to keep backups of important data.

## Checklist item 5: read duration, reliability and bandwidth

For H100 jobs, which are often long and expensive, these fields matter as much as the price:

- **Max duration and "available until".** When the date arrives, the instance and its disk are removed. Your run, plus time to copy checkpoints off, must fit before it.
- **Reliability.** Prefer higher values for multi-day runs.
- **Bandwidth price per TB.** Large checkpoints and datasets add up in both directions.
- **Disk size.** It is fixed at deploy, so size it for the model, data, checkpoints and headroom.

If the host goes offline, billing pauses while it is down, but there is no automatic move to another machine. Save checkpoints to storage you control on a regular schedule.

## Checklist item 6: data location, without assumptions

Filtering to EU codes keeps the host machine inside the countries you choose. That is useful for teams with a policy about where data is processed, but the filter alone does not make a workload compliant with any regulation, and Cloudj makes no compliance claim for itself or for hosts. Check your own legal obligations. The guide on [EU data location for GPU cloud](https://cloudj.net/blog/eu-data-location-gpu-cloud) covers how to turn a policy into a search.

## Checklist item 7: plan the first hour

GPU billing is per minute from a prepaid balance, the price shown at deploy is locked, and a deposit (about an hour of GPU, some storage hours and a bandwidth reserve) is held when you deploy. Use the first hour to confirm the environment:

```bash
nvidia-smi
python -c "import torch; print(torch.cuda.get_device_name(0), torch.cuda.device_count())"
```

The "PyTorch + JupyterLab" and "Ubuntu 22.04 + CUDA (SSH)" templates are good starting points; custom public Docker images also need a higher account level. For a full fine-tuning workflow in Europe, read [LLM fine-tuning on GPUs in Europe](https://cloudj.net/blog/llm-fine-tuning-gpu-europe).

## FAQ

### Which H100 should I search for?

If your training uses several GPUs that communicate heavily, start with SXM. For single-GPU work, PCIe is fine. For large-model inference, NVL's 94 GB helps. If unsure, search all three.

### Why can't I select an H100?

Datacenter GPUs require a higher account level. Contact Cloudj support from the console.

### Is a German H100 always available?

No. Search with `DE` and the variant you need; if nothing is listed, add other EU codes or try another variant.

### Does the Datacenter badge mean an SLA?

No. It shows the host is in a data centre. There is no hardware SLA on any listing.

## Search for an H100 in Europe

Open the [H100 SXM deploy page](https://app.cloudj.net/deploy/gpu?model=h100-sxm), enter `DE, NL, FR, SE, FI`, and try the PCIe and NVL variants if the SXM search is empty. Live "from" prices are on [GPU cloud](https://cloudj.net/gpu-cloud).