Train, fine-tune and deploy AI models on bare metal with
RTX 3090 or RTX 5090, 100% dedicated resources and a ready-to-use stack.
All servers include Ubuntu 24.04 LTS, NVIDIA Drivers and CUDA pre-installed. Optional AI stack installable with a single command (Docker, Ollama and Open WebUI).
Dedicated and symmetric 1000 Mbit/s connectivity
One public IPv4 included.
IPv6 ranges available.
Full control: "root" access
Ready to use: Ubuntu 24 LTS + NVIDIA Drivers + CUDA
Setting up a GPU server from scratch can take between 3 and 6 hours: drivers, CUDA, containers, tools. Our servers arrive with that base already set up. The first command you run is for your model.
Base included
The server arrives with Ubuntu 24.04 LTS, NVIDIA 550+ drivers and CUDA installed and verified. Connect via SSH and run nvidia-smi - your two GPUs are active from the first login.
AI stack with a single command
With the script we send you along with access, you install Docker, NVIDIA Container Toolkit, Ollama and Open WebUI in a single line. In minutes you'll have a ChatGPT-like interface connected to your GPUs, running on your own server.
Private LLM inference
Run Llama 3, Mistral, DeepSeek and more. No rate limits, no per-token costs. OpenAI-compatible API.
Fine-tuning and training
Dedicated hardware to train your own models. cuDNN and NCCL included for multi-GPU.
Embedding generation and RAG
Local embeddings with nomic-embed-text. Combine with your vector database. Without sending data to third parties.
Computer vision
Ideal for training and inference of object detection, image classification and segmentation models.
AI automations and agents
Deploy frameworks like n8n, LangChain or AutoGen with direct access to local models.
Environments for R&D teams
Run JupyterHub with GPU access from the browser. Ideal for academic and data science teams.
Production with data sovereignty
Your data never leaves your server. Regulatory compliance without compromises.
You'll have a dedicated server (Bare Metal) 100% for your project. You won't share hardware or bandwidth with anyone else, so you'll get performance and absolute privacy for your projects.
Our World Class datacenters have n+1 redundant services guaranteeing that your business will stay online even when something fails.
Plus, minimal regional latency and carrier-level DDoS protection.
Pro-Active Monitoring
Critical services (networks, power, cooling, etc.) are monitored 24x365 from our NOC.
Expert support
A team with more than 15 years of experience ready to help you quickly and effectively.
High-speed networks
Reliable and ultra-fast networks, on the most important and recognized carriers in the region.
Certified provider
We are the only company in the industry with triple certification: ISO 9001, ISO 14001 and OHSAS 18001
Sustainable infrastructure
Datacenters with efficient and innovative technologies to provide you with the best service without sacrificing the future.
Grow without limits
Your business grows and you need more...perfect! We'll help you reconfigure your server or migrate to different hardware.
Dedicated GPU for AI, with a ready-to-go base and simplified deployment.
Frequently asked questions
We offer you multiple payment methods for your convenience. Learn about them here.
The servers are hosted in one of our four own datacenters in Rosario, Santa Fe, Argentina. This means lower latency for LATAM region clients and that processed data doesn't leave the continent, which can simplify regulatory compliance in certain cases.
At DonWeb we have several layers of protection to mitigate denial-of-service attacks such as: SYN, ICMP and UDP flood, Ping of Death, Smurf, etc. Also those attacks that try to saturate our network's bandwidth.
In case of severe, large-scale attacks, the mitigation systems located upstream, at the level of our connectivity providers (carriers), will be activated. These are activated automatically and within the first 5 minutes of the attack starting.
If the attacks are smaller in scale and directed at specific servers, the mitigation systems within our datacenters will be activated. These filtering systems have rules that are updated periodically and work automatically.
Additionally, our staff assigned to the Network Operation Center monitor the networks 24x7x365 and will act promptly against any anomaly detected in traffic.
Also, beyond DDoS, there are many different types of cyber attacks that can affect your site or application. That's why we recommend:
Dedicated server activation will begin immediately after our administrative team confirms your payment. This process includes implementing all the necessary configurations to get your server up and running.
The estimated time for activation varies depending on when payment is confirmed:
It's worth noting that if you require special configurations, such as adding disks or a balanced infrastructure, the activation process could extend up to 72 hours. However, we'll do our best to minimize this time and ensure your server is available as soon as possible.
As soon as your server is active, you'll receive an email with all the necessary information to access it.
All servers are delivered with Ubuntu 24.04 LTS, NVIDIA drivers, CUDA, cuDNN and NCCL pre-installed and verified. You can connect via SSH and have your GPUs operational from the first login, without additional configuration.
6020 How do I activate the optional AI stack?
With a single command that runs a script we provide on the server. There are two levels:
You'll find instructions for using the script in the server console welcome message.
The base server includes the operating system and drivers. Docker, NVIDIA Container Toolkit and Ollama are activated with a single command that we provide. If you also want Open WebUI or GPU monitoring with nvitop, there's an optional second level that also installs with a single command.
No. They are fully dedicated bare metal servers. The hardware, bandwidth and compute resources are 100% yours. There's no "noisy neighbor" or throttling.
In shared cloud GPU services, the hardware is distributed among multiple clients. This can mean variable performance, unpredictable latency, and your workload's data passing through third-party infrastructure.
With a dedicated GPU server from DonWeb, the hardware is 100% yours: GPUs, RAM, bandwidth and storage are not shared with anyone. Plus, you pay a fixed monthly price, with no surprises for compute hours or data transfer.
Any workload that benefits from GPU, for example:
You have full root access, so you can install and configure whatever you need. The only exception is that game servers and crypto mining are not allowed.
You're not limited to Docker. You can run any model compatible with the tools or frameworks you install on the server. This includes models used with Ollama, vLLM, containers Docker or directly from frameworks like PyTorch or TensorFlow, since you have full root access.
This includes families like Llama, Mistral, DeepSeek, Qwen, CodeLlama and Hugging Face models in GGUF format, among others compatible with your stack.
With 48 GB of combined VRAM (RTX 3090) or 64 GB (RTX 5090), you can run models up to 70B parameters in Q4 quantization.
Yes, without restrictions. You can upload your own models in formats like GGUF or safetensors, run your own Docker images, mount volumes with your data and configure the environment exactly as you need.
Yes. Ollama exposes an API compatible with the OpenAI format. Any application using the official OpenAI SDK can point to your server by only changing the base_url and the api_key, without code changes in your application.
Yes. You have full root access. CUDA is compatible with PyTorch 2.x, TensorFlow 2.x and JAX. You can also use official Docker images from Hugging Face, vLLM and other frameworks without additional configuration.
It's useful for both. The configuration includes cuDNN and NCCL, necessary for distributed training and fine-tuning on multiple GPUs. You can use frameworks like Hugging Face Transformers, Axolotl or Unsloth directly on the hardware, with full access to both GPUs via PyTorch or TensorFlow.
Yes, as long as the total VRAM allows it. With 2 GPUs you can distribute workloads manually or use frameworks like vLLM with --tensor-parallel-size 2 to use both GPUs for a single large model.
vLLM is an inference engine optimized for high throughput and low latency with multiple simultaneous requests. Ollama is simpler and recommended for individual development. If you're going to serve a production API with many concurrent users, vLLM is the option.
The RTX 5090 is NVIDIA's most recent generation, with more VRAM per GPU (32 GB vs 24 GB) and better performance in inference and training workloads.
The RTX 3090 is still very capable for production, especially for inference of models up to 70B parameters in Q4 quantization using both GPUs. If your workload includes intensive fine-tuning or next-generation models, the RTX 5090 is the recommended option.
Includes dedicated and symmetric 1 Gbit/s connectivity. This allows you to download large models or transfer datasets without issues or delays.
Yes. We can set up LAN connections between two or more dedicated servers. This allows you to build distributed inference clusters or separate roles, for example between a model server and a vector database server. Consult with our advisors to define the configuration since it has special requirements.
Yes. The published configurations are the most common, but we can build custom options based on your needs for RAM, storage or connectivity. If your project requires connecting multiple GPU servers to each other via LAN, more bandwidth or any other variation, contact our infrastructure team to evaluate it.
It includes DonWeb technical support for hardware, network and connectivity. The server is delivered with full root access, so software and application management is the client's responsibility. If you need assistance with specific configurations, you can consult our team.
