GPU servers in Germany: compute for AI, on our own hardware

We provide AI and ML teams with GPU servers in Germany for training and inference, run on Kubernetes with the NVIDIA GPU Operator and, if you want, attached to your own cluster. We run our own hardware in a data centre in Munich. There is no price list: you get a quote for each configuration and term.

Request an intro call

Typical first stepTypical entry point: an intro call about your model, your load and the term you need. After that you get a quote for that configuration and term.

Where teams usually start

  • The GPU bill grows with every experiment. Your load is predictable, but the cloud still charges it by the hour.
  • The cards you need are not available. Your cloud provider has a waiting list, or capacity only with a long commitment.
  • The data has to stay in Germany. Training data or customer documents cannot go just anywhere, and your customer wants to know where they are processed.
  • The data must not leave the building. Then the inference stack belongs in your on-site data centre, hardware included.

How we work

  1. Requirements.

    Which model, training or inference, how much GPU memory, how much data, what term. You get a configuration that fits your load.

  2. Quote.

    Per configuration and term, in writing.

  3. Provisioning.

    The GPU servers run on our hardware, as nodes in Kubernetes, as virtual machines with a GPU, or attached to your cluster, whichever you need. If the right card is not in stock, we buy it for you and build it in. You get the access your workloads need.

  4. Operations.

    We keep hardware, drivers and Kubernetes up to date, within the agreed scope. You run your models and applications.

What we run ourselves

We run our own hardware in a data centre in Munich. Proxmox VE and our Kubernetes clusters run on it. We attach GPUs to Kubernetes with the NVIDIA GPU Operator, which installs the drivers and runtime and makes the cards usable for workloads in the cluster.

We have delivered and commissioned the same setup in a customer's on-site data centre: servers and GPUs sourced, Kubernetes with GPU nodes set up, vLLM for inference and LiteLLM as the company-wide entry point for every application. How that project went: Case study: on-premises LLM inference.

How GPUs fit into a platform you run yourselves: GPU nodes in your Kubernetes platform. Location, encryption and backups: Where your data lives.

Stack

  • Proxmox VE and our Kubernetes clusters on our own hardware.
  • GPUs in Kubernetes through the NVIDIA GPU Operator: drivers, runtime and the cards made available to workloads in the cluster.
  • Inference with vLLM, with LiteLLM as the shared entry point for applications, as in the inference stack we delivered to a customer.

Case studies

Company with an on-site data centre

Case study: on-premises LLM inference, delivered with the hardware

For a company we sourced, built and commissioned an inference stack for its on-site data centre: three 2U servers with NVIDIA A100 GPUs, Kubernetes with GPU nodes, vLLM for inference and LiteLLM as the shared entry point for the company's applications. Plus high availability for the network underneath, with tested failovers for firewalls and switches.

Our own product

Case study: running a SaaS on our own infrastructure

Fortlauf is our product. We built it, and we run it on our own hardware. In production since July 2026: tenant isolation in the database, immutable document storage, delivery by GitOps, measured recovery.

Who you will talk to

One fixed contact from the intro call until your servers are running.

Send a message

What we do not do

  • No price list and no one-click ordering. Every configuration gets its own quote.
  • No resale of other providers' GPU capacity. The GPU servers run on our own hardware.
  • No on-call cover. Operations run in agreed hours.

Frequently asked questions

Which GPUs are available?

That depends on what you need. Whatever is not in stock, we buy for you and build in. The model, GPU memory and term are set out in your quote.

Does the data stay in Germany?

On the GPU servers we provide, yes. We run our own hardware in a data centre in Munich. Where your applications send data beyond that is your decision.

How is it billed?

You get a quote for each configuration and term. There is no public price list at the moment.

Can we add the GPUs to our own Kubernetes?

Yes, as additional nodes in your cluster, or you use them in a cluster we build for you. How we build Kubernetes platforms: Kubernetes platforms, built and run.

Is this suitable for running our own language model?

Yes, that is a typical use: an open model for inference on a GPU server in Germany, connected to your applications. Which model size fits which card is something we settle when we go through your requirements.

Can you build the inference stack on our premises instead?

Yes. If the data must not leave your building, we source the servers and GPUs, set up Kubernetes with GPU nodes, vLLM and LiteLLM, and commission the stack in your data centre. That is what we did for one customer: Case study: on-premises LLM inference.

Insights

Want to talk about GPU capacity?

Tell us which model you want to run, what load you expect and for how long.