AI workstations and GPU servers
We design and build computers for artificial intelligence: quiet desktop AI workstations and GPU servers in 4U rack cases. They run local language models (LLMs), power RAG systems, fine-tune models and generate images and video. Every configuration is matched to your models, number of users and budget.
What an AI workstation is for
Local language models (LLMs)
Running models such as Llama, Qwen, Mistral, Gemma or DeepSeek on your own hardware with Ollama, llama.cpp, LM Studio or vLLM. The model works as a private assistant for text, code and documents.
RAG and work on company documents
Searching and answering questions based on company documents, contracts or technical documentation. Data never leaves the company network, which makes it easier to meet GDPR and confidentiality requirements.
Fine-tuning and training
Fine-tuning with LoRA and QLoRA, training vision models and experiments in PyTorch. These tasks need more VRAM than inference and benefit from several graphics cards.
Image and video generation
Stable Diffusion, Flux and video models in ComfyUI. More graphics memory allows higher resolutions and keeping several models loaded at once.
A model server for the team
A single GPU server makes models available to many users through an OpenAI-compatible API. The team uses the assistant in the browser, in the code editor or in its own applications.
Your own hardware instead of the cloud
Data stays in the company. Prompts, documents and source code are not sent to external services. This matters in law firms, design offices, healthcare, public administration and wherever professional secrecy or NDAs apply.
Predictable cost. Hardware is bought once, and using the models is not billed per token or per GPU hour. The cost of the workstation is spread over years of use, and larger orders can be financed through leasing.
Full control. You choose the models and their versions, there are no request limits or changes in a provider's offer to worry about, and the workstation also works without internet access.
Desktop AI workstation or rack server?
We build both from the same components. They differ in the number of graphics cards, where they run and how they are used.
| Feature | Desktop AI workstation | 4U rack AI server |
|---|---|---|
| Graphics cards | 1-2 GeForce RTX 5090 or NVIDIA RTX PRO cards | 2-4 NVIDIA RTX PRO cards with blower cooling |
| VRAM | 32-96 GB | 96-384 GB |
| Location | at the desk, quiet operation | server room or rack cabinet, loud operation |
| Who it is for | one person or a small team | many concurrent users, continuous operation |
| Access | directly or remotely | over the network, API for applications |
| Power | standard 230 V socket | a dedicated circuit is recommended with four cards |
Desktop AI workstations
A tower AI workstation runs at the desk and does not need a server room. Configurations with a single GeForce RTX 5090 (32 GB of VRAM) are a good fit for local language models of around 30 billion parameters, image generation and developers working with a coding assistant.
The Threadripper platform holds two RTX 5090 cards or an RTX PRO 6000 with 96 GB of memory and leaves room for upgrades in the years ahead. Every workstation is also suitable for GPU rendering and graphics applications.
Multi-GPU configurations
| System | Processor | Graphics cards | Memory / storage | Use |
|---|---|---|---|---|
| AI 2× RTX 5090 | Threadripper 9970X (32 cores) | 2× GeForce RTX 5090 · 64 GB VRAM | 256 GB DDR5 ECC · 4 TB NVMe | models of around 70 billion parameters, LoRA fine-tuning |
| AI 96 GB | Threadripper 9970X (32 cores) | RTX PRO 6000 Blackwell · 96 GB VRAM | 256 GB DDR5 ECC · 4 TB NVMe | models above 100 billion parameters, long context |
Multi-GPU configurations are quoted individually. The price depends mainly on the number and model of graphics cards, and their availability and prices change frequently.
4U rack AI servers
A GPU server in a 4U rack case is built for continuous operation and many users. It holds up to four NVIDIA RTX PRO cards, a Threadripper PRO processor with eight-channel ECC memory and fast NVMe drives for models and datasets.
The server is mounted in a rack cabinet or placed on its own in a technical room and is accessed over the network. A node with a single RTX PRO card is a good start when a model server for a few people is needed.
| Build | Processor | Graphics card | Memory / storage | Price from | details |
|---|---|---|---|---|---|
| Rack 4U Pronode with an RTX PRO card | Ryzen 9 9950X | RTX PRO 4000 · 24 GB | 128 GB DDR5 · 2 TB NVMe | from€5,170 |
Multi-GPU configurations
| System | Processor | Graphics cards | Memory / storage | Use |
|---|---|---|---|---|
| AI server 2× GPU | Threadripper PRO 9975WX (32 cores) | 2× RTX PRO 6000 Blackwell · 192 GB VRAM | 256 GB DDR5 ECC · 2× 4 TB NVMe | model server for a team, RAG |
| AI server 4× GPU | Threadripper PRO 9985WX (64 cores) | 4× RTX PRO 6000 Blackwell · 384 GB VRAM | 512 GB DDR5 ECC · 2× 4 TB NVMe | the largest open models, training, many users |
Multi-GPU configurations are quoted individually. The price depends mainly on the number and model of graphics cards, and their availability and prices change frequently.
All 4U rack servers →How to choose a computer for AI
VRAM decides model size
A model runs fastest when it fits entirely in graphics memory. With popular 4-bit quantisation a model needs roughly 0.6-0.7 GB of memory per billion parameters, plus room for the conversation context.
| VRAM | Example cards | What you can run |
|---|---|---|
| 16 GB | GeForce RTX 5070 Ti, RTX 5080 | models up to around 14 billion parameters, Stable Diffusion, quantised Flux |
| 24 GB | RTX PRO 4000 Blackwell | models of around 20-30 billion parameters with a shorter context |
| 32 GB | GeForce RTX 5090 | models of around 30 billion parameters with a long context, Flux and video models |
| 48 GB | RTX PRO 5000 Blackwell | models of around 70 billion parameters in 4-bit quantisation, shorter context |
| 64 GB | 2× GeForce RTX 5090 | models of around 70 billion parameters with room for context |
| 96 GB | RTX PRO 6000 Blackwell | quantised models above 100 billion parameters, fine-tuning smaller models |
| 192-384 GB | 2-4× RTX PRO 6000 Blackwell | the largest open models, many concurrent users, training |
GeForce RTX 5090 or NVIDIA RTX PRO
The GeForce RTX 5090 offers the most performance for the money and is the best choice for workstations with one or two cards. NVIDIA RTX PRO cards have more memory (up to 96 GB per card), ECC memory and a design made for several cards side by side. The Max-Q versions draw 300 W and are intended for systems with up to four cards, which is why we use them in servers.
Several graphics cards in one computer
Tools such as llama.cpp, Ollama and vLLM can split a model across several cards, so their memory can be used together. This requires a motherboard with enough full PCIe slots, which is why we build multi-GPU systems on the Threadripper or Threadripper PRO platform.
Processor, RAM and storage
There should be at least as much RAM as total VRAM, and more for training and data preparation. Models and datasets take up tens or hundreds of gigabytes, so we use fast NVMe drives of 2-4 TB or more. The processor handles data preparation and drives several cards.
Power, cooling and noise
An RTX 5090 draws up to 575 W, so a workstation with two cards needs a power supply of at least 1600 W and a case with very good airflow. In servers with four cards we size the power supply with headroom and use blower cooling that pushes warm air out of the case. Every system goes through a 24-hour test with all cards under full load.
System and software
AI workstations and servers ship with Ubuntu Linux or Windows 11 Pro, current NVIDIA drivers and CUDA. On request we also prepare Docker with GPU support and selected tools such as Ollama, Open WebUI or ComfyUI, so the hardware is ready to work once connected.
How much does an AI workstation cost?
An AI workstation with a GeForce RTX 5090 (32 GB of VRAM) costs from €7,890. A Threadripper configuration prepared for additional cards costs from €11,950. Multi-GPU systems and rack servers are quoted individually, because the price depends mainly on the number and model of graphics cards.
Desktop AI workstation or rack server?
A desktop workstation suits one person or a small team: it sits at the desk, runs quietly and holds one or two cards. A 4U rack server makes sense when many people will use the models at once, more than two cards are needed, or the hardware has to run around the clock in a server room.
Which graphics card for a local LLM?
The most common choice is the GeForce RTX 5090 with 32 GB of memory, which fully holds models of around 30 billion parameters. Larger models need two RTX 5090 cards or an RTX PRO 6000 with 96 GB. If you are not sure which models you will use, tell us what the workstation is for and we will propose a configuration.
Do two graphics cards add up their VRAM?
They do not form one shared memory, but llama.cpp, Ollama and vLLM split the model layers between cards. Two 32 GB cards can therefore run a model that needs around 60 GB. Some performance is lost to communication between the cards.
Can a local model replace ChatGPT?
Open models running locally handle summarising and analysing documents, translation, writing code and answering questions from company knowledge well. In the most complex tasks the largest commercial models still have an edge, so many companies combine both: confidential data is processed locally and the rest in external services.
Linux or Windows for AI?
For servers and container work we recommend Ubuntu Linux, where most AI tools are simplest to run. Windows 11 Pro is chosen when graphics or CAD applications run on the same workstation. Tools such as Ollama, LM Studio and ComfyUI work on both.
Can an AI workstation be upgraded later?
Yes. We choose the motherboard, power supply and case with upgrades in mind, so a Threadripper workstation can take another graphics card, more RAM and more drives. When ordering, it helps to say how many cards the system should eventually hold.
Can an AI workstation be leased?
Yes. We issue VAT invoices, and for larger orders leasing or long-term rental is available through a leasing company. We ship across Europe; delivery outside Poland is quoted individually.