Your data stays on your server
Prompts, uploaded documents and answers are processed and stored on the server you administer. Nothing has to be sent to a model provider's API.
GPU servers · Private AI
Price list updated
A private AI server is a server you control that runs an open-weight AI model, so prompts, documents and answers are processed on that server instead of being sent to a model provider's API. RyzenRDP rents GPU servers for this in Frankfurt, Germany: 1, 2, 4 or 8 NVIDIA RTX PRO 6000 Blackwell SE GPUs with 96 GB of VRAM each, root access over SSH, from $1,951.09 a month.
Prompts, uploaded documents and answers are processed and stored on the server you administer. Nothing has to be sent to a model provider's API.
Run the open-weight model and version you have evaluated. It changes when you change it, not when a provider retires or updates it.
The server has one monthly price whatever the number of tokens. Heavy, steady use does not raise the bill.
You decide who reaches the model: an SSH tunnel for one person, a VPN for a team, or a reverse proxy with sign-in.
The same idea goes by several names. A private LLM is a language model that runs for one organization only. An on-premise LLM runs on hardware in your own building. Sovereign AI means keeping the model, the data and the infrastructure under your own control or within your own jurisdiction. A rented GPU server sits between the two ends: you do not own or house the hardware, but you run the system on it and decide what goes in and out.
There are three ways to give an application or a team a language model. They differ in where the prompts are processed, who runs the hardware and how the cost behaves.
| What matters | Hosted model API | Private AI server on a rented GPU | On-premise hardware |
|---|---|---|---|
| Where prompts are processed | On the model provider's servers, under its terms | On your server in Frankfurt, Germany | On your own hardware, in your own building |
| Who runs the hardware | The model provider | We run the server; you run everything on it | You do |
| Cost | Per token, rising with use | Fixed per month, from $1,951.09 | A large purchase up front, then power, cooling and space |
| Models | The provider's models, often the most capable | Any open-weight model that fits in the VRAM | Any open-weight model that fits in the VRAM |
| Time to start | Minutes | After your order is built, then your own setup | Weeks to months of procurement |
| What you must do yourself | Write the application | Install, secure, update and back up the system | All of that, plus the hardware and the room it is in |
Many teams end up with a mix: a hosted API for work that needs the most capable models and carries no sensitive data, and a private server for material that must not leave their control.
Not automatically. A hosted API charges per token, so light use costs little. A GPU server costs the same every month whether it is busy or idle, so its cost per token depends on how much work you give it.
A private server makes sense in two situations:
If your use is light and the data is not sensitive, a hosted API with an AI VPS on CPU for the application around it is the simpler setup.
Be clear about the split before you put sensitive material on any rented server.
RyzenRDP provides the GPU server in Frankfurt, Germany, its network connection and the console you manage it from. We do not install, run or monitor your model, and your application's prompts are not sent to us: they go to the software you run on the server.
You are responsible for the operating system updates, the model and its license, who can reach the endpoint, encryption of connections and of anything you want encrypted at rest, backups, and the lawfulness of the data you process.
A server located in Germany does not by itself make a system compliant with the GDPR or any other regulation. Compliance depends on what you process, on what basis and how you secure it. The location is one fact you can rely on; our privacy policy and terms set out the rest of what we do.
Pick an open-weight model whose license fits your use, then the plan whose VRAM holds it: the sizing table shows how.
Install Ollama or vLLM and load the model. Both give you an OpenAI-compatible API on the server itself.
Open WebUI gives your team a browser chat with accounts and conversation history, connected to that API.
For answers from your own files, add retrieval: an embedding model and a vector database on the same server, so the documents stay on it.
Keep the model API on 127.0.0.1, publish only the chat interface through a reverse proxy with TLS and sign-in or over a VPN, and open no other port.
Every plan is a Linux server in Frankfurt, Germany with NVIDIA RTX PRO 6000 Blackwell SE GPUs and 96 GB of VRAM per GPU, running Ubuntu 24.04 LTS with NVIDIA CUDA.
| GPUs | VRAM per GPU | Total VRAM | vCPU | RAM | Disk | Monthly | Stock and order |
|---|---|---|---|---|---|---|---|
| 1 × NVIDIA RTX PRO 6000 Blackwell SE | 96 GB | 96 GB | 16 | 128 GB | 500 GB | $1,951.09 | |
| 2 × NVIDIA RTX PRO 6000 Blackwell SE | 96 GB | 192 GB | 32 | 256 GB | 1,000 GB | $3,900.60 | |
| 4 × NVIDIA RTX PRO 6000 Blackwell SE | 96 GB | 384 GB | 64 | 512 GB | 2,000 GB | $7,799.63 | |
| 8 × NVIDIA RTX PRO 6000 Blackwell SE | 96 GB | 768 GB | 128 | 1,024 GB | 4,000 GB | $15,597.68 |
Stock follows free GPU capacity and is shown live: a row marked Out of stock cannot be ordered until capacity returns. If the plan you need is out of stock, ask the sales team and tell us which plan and how many servers you need. For a physical server built to your specification, or one in another data center, ask for an enterprise quote.
A private AI server is a server you control that runs an AI model for you alone. Prompts and documents are processed on that server, not on a model provider's shared API. It can be hardware you own or a GPU server you rent, such as a RyzenRDP GPU server in Frankfurt, Germany.
A private LLM is a large language model deployed for one organization, on infrastructure that organization controls. In practice it is an open-weight model served with a runtime such as Ollama or vLLM on your own GPU server.
No. On-premise means the hardware stands in your own building. A rented GPU server stands in a data center, in this case in Frankfurt, Germany, and you administer it remotely. You get the same control over the model and the software without owning, housing or cooling the hardware.
Your prompts go to the software you run on your server, not to us, and we do not install, run or monitor your model. As with any hosting provider, we operate the infrastructure the server runs on; our privacy policy describes what we process. Encrypt what you need encrypted.
A location does not make a system compliant. The server is in Frankfurt, Germany, which settles where the data is stored and processed. Whether your processing complies with the GDPR depends on your data, your legal basis and your security measures, which remain your responsibility.
Mostly VRAM: enough for the model's weights and the context of its users. A RyzenRDP GPU server has 1, 2, 4 or 8 GPUs with 96 GB each, 96 GB to 768 GB in total, which covers models from a few billion to several hundred billion parameters depending on precision.
Yes. Install a chat interface such as Open WebUI next to the model. Your team signs in through the browser, and the conversations stay on your server.
From $1,951.09 a month for the single-GPU server with 16 vCPU, 128 GB RAM, 500 GB disk and a 10 Gbps port, up to $15,597.68 a month for 8 GPUs. The price is fixed per month, with no per-token charge.
GPU servers in Frankfurt, Germany with 96 GB of VRAM per GPU. From $1,951.09 a month.