Your own graphics cards, not shared and not billed per token. Ubuntu 24.04 with NVIDIA driver, CUDA and Docker is already installed. Ready in under a minute after payment: log in via SSH, start your model. Hosted in Frankfurt.
Prefer a whole server with a card of your own? See the GPU-Dedicated servers
The professional card from NVIDIA's Blackwell generation, built for round-the-clock operation in the data center. Every card in your GPU-VPS is yours alone, at full performance. The numbers show you where everything sits on the card.
Plus 752 fifth-generation Tensor cores and 188 fourth-generation RT cores. 4 PFLOPS in FP4, 120 TFLOPS in FP32.
512 bit wide, 1,597 GB/s. Spread across 32 chips of 3 GB each, half of them on the back.
Right in the chip: H.264, H.265 and AV1 in hardware, including 4:2:2 from professional cameras.
Full power draw of the Server Edition, built for sustained load in the data center.
About 64 GB/s in each direction between card and server.
Schematic top view without the heatsink. Performance figures according to NVIDIA, the split of cores and memory chips derived from those figures.
A language model has to fit entirely into GPU memory, or it gets slow. With 96 GB per card, a single RTX PRO 6000 Blackwell runs what would otherwise take several cards.
Pick a size and see which open models fit entirely into GPU memory. Whether gpt-oss-120b on one card or DeepSeek on eight: weights, context and data stay on your server.
Rough figures for the model weights. Rule of thumb: parameters × bytes per parameter (BF16 = 2, FP8 = 1, 4-bit = 0.5). On top comes room for the context (KV cache), the longer the context, the more. With several cards you spread one model, for example with vLLM via --tensor-parallel-size.
You choose how many cards your GPU-VPS gets. vCPUs, memory and storage grow with it. Moving up is possible at any time, the cards are added physically.
All prices excl. VAT. You pay in advance, monthly or for longer periods. Once the included traffic is used up, we throttle to 10 Mbit/s at no extra cost. You can add extra traffic at any time.
Need something special that isn't listed here? We are happy to put together an individual offer. Get in touch
Your GPU-VPS starts with our AI-Ready image. NVIDIA driver, CUDA 13.4 and Docker with GPU access are set up, so you go straight to the interesting part.
Preinstalled and tuned for the RTX PRO 6000 Blackwell. You can also choose any other image and set everything up yourself.
--gpus all works right away$ nvidia-smi $ docker run --rm --gpus all \ ubuntu nvidia-smi
One command starts Ollama and the Open WebUI web interface together. Then you load a model like gpt-oss-120b with one click, and your chats stay on your server in Frankfurt.
docker run -d -p 3000:8080 --gpus=all \ -v ollama:/root/.ollama \ -v open-webui:/app/backend/data \ -e WEBUI_SECRET_KEY=$(openssl rand -hex 32) \ --name open-webui --restart always \ ghcr.io/open-webui/open-webui:ollama
http://your-ip:3000 and create your admin account right away: the first account becomes admin. Then load, for example, gpt-oss:120b or llama3.3:70b under "Models".vLLM serves your model through the same interface as the OpenAI API. Point your existing apps, n8n workflows or agents at your own server.
docker run -d --gpus all --ipc=host -p 8000:8000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --name vllm vllm/vllm-openai:latest \ --model openai/gpt-oss-120b \ --api-key your-secret-key
curl http://localhost:8000/v1/chat/completions \
-H "Authorization: Bearer your-secret-key" \
-H "Content-Type: application/json" \
-d '{"model":"openai/gpt-oss-120b","messages":[{"role":"user","content":"Hello!"}]}'
ComfyUI is the node-based interface for image and video models like FLUX.1, Stable Diffusion 3.5 or Wan 2.2. With 96 GB you load large models at full precision.
apt install -y python3-venv git git clone https://github.com/Comfy-Org/ComfyUI && cd ComfyUI python3 -m venv .venv && . .venv/bin/activate pip install torch torchvision torchaudio \ --extra-index-url https://download.pytorch.org/whl/cu130 pip install -r requirements.txt python main.py
ssh -L 8188:localhost:8188 root@your-server-ip
http://localhost:8188 in your browser. This way ComfyUI is not reachable from the internet, only by you.Four encoders and four decoders per card process H.264, H.265 and AV1 in hardware. Ideal for large video libraries, streams or footage from professional cameras.
apt install -y ffmpeg ffmpeg -hwaccel cuda -hwaccel_output_format cuda -i input.mov \ -c:v av1_nvenc -preset p5 -cq 28 -c:a copy output.mp4
av1_nvenc for hevc_nvenc.NVIDIA's official container ships PyTorch, CUDA and cuDNN in matching versions. Inside it you train LoRA adapters on your own data, for example with Unsloth or Hugging Face TRL.
docker run --gpus all -it --rm --ipc=host \ -v $PWD:/workspace \ nvcr.io/nvidia/pytorch:26.10-py3
python -c "import torch; print(torch.cuda.get_device_name())"
# NVIDIA RTX PRO 6000 Blackwell Server Edition
From the first idea to continuous operation: everything that needs a lot of graphics memory and compute.
An internal assistant or the AI building block of your app. Your data never goes to an outside AI company.
Let a model read your manuals, tickets or contracts and answer questions about them.
Product images, concept art or short clips, without credits and without waiting in a queue.
188 fourth-generation RT cores and 96 GB for large scenes and photorealistic animation.
Transcribe podcasts, meetings or videos, generate subtitles and synthesize voices.
More than 24,000 CUDA cores per card compute flows, molecules or models in parallel.
Both come with an NVIDIA card just for you. The difference is how fast you start and how conveniently you manage your server.
Whether one card or eight.
Start, reinstall, rescue, browser console, firewall and SSH keys, all in one place.
Save a state, try a new model or driver, jump back if needed. Nightly backups with 7 days can be added.
Your system, your software, your models. We do not dictate anything.
Basic Protection is always on. For applications that are targeted, you can add zero-lag Protection. More about DDoS protection
Your data and computations stay in Germany. More about our data centers
After that we throttle to 10 Mbit/s at no extra cost. You can add more traffic at any time.
One IPv4 address and IPv6 at no extra charge. Up to 4 more IPv4 addresses at € 2.52 each per month excl. VAT.
You pay in advance and cancel whenever you like. All payment methods
We answer your questions before you buy. Your question is not listed? Just write to us.
Contact usdocker run --gpus all works right away. You can also choose another image and install everything yourself.--tensor-parallel-size 4. Alternatively, run a separate instance on each card, which multiplies throughput for many simultaneous users.The professional card from NVIDIA's Blackwell generation, physically just for you. Ready with driver, CUDA and Docker in under a minute after payment.