AI workstation
AI workstation for local models
Training or running AI models locally keeps your data in-house and costs nothing per request. That takes hardware designed for video memory, system memory and sustained load. NoMadPC builds such workstations by hand in the workshop near Frankfurt.
Video memory is the key figure
Whether a language model or an image generator runs on the graphics card is mainly decided by video memory (VRAM). Larger models or longer contexts need more of it; quantisation lowers the requirement. The GeForce RTX 5090 has 32 GB of VRAM; for larger models, multiple cards or professional cards are an option.
- VRAM determines which models fit entirely on the graphics card
- System memory for datasets and for models that are partly offloaded
- Fast NVMe storage for models and training data
- Cooling and power supply for hours of training under full load
Training or inference
For inference, meaning running finished models, what matters most is that the model fits into video memory. Training and fine-tuning need considerably more memory and often run for hours or days. We design the workstation for your workflow and say openly where a cloud solution would be cheaper.
Your data stays with you
A local system processes customer data, source code or internal documents without sending them to an external service. For many companies that is the main reason for their own AI workstation.
How a project works
You describe models, data volumes and goal. We put the components together, build the system by hand, tune BIOS, memory and fan curves for sustained load and hand it over after a thorough stress test, with an invoice showing VAT.
Matching systems
Frequently asked questions
That depends on model size, quantisation and context length. Rule of thumb: the model plus its context has to fit into video memory. The RTX 5090 offers 32 GB; for larger models we plan multiple cards or professional cards.
Yes, fine-tuning and smaller training runs work well locally, but need more memory and cooling for sustained load. For very large training runs we tell you openly when cloud compute makes more sense.
Your data doesn't leave the building, there are no costs per request and the system is available at any time. For very large, infrequent training runs the cloud can be cheaper.
Yes. You receive a proper invoice with VAT shown; we record your VAT ID during the consultation.
Let's talk about your AI project
Tell us your models, data volumes and goal. We tell you which hardware handles it and where you can save.