Tutorial 01
How we built a 6,000€ LLM PC (and what we learned the hard way)
In October 2025, we built a local‑AI desktop that runs large language models on our own hardware without cloud GPU bills or uploading private data.
We set a strict budget of about 6,000€ for the whole setup, including a serious monitor and the small comforts that make a workstation pleasant to use every day.
Think of an “LLM PC” like a fast kitchen for a huge recipe book. The recipe book is the model. The GPU does most of the cooking. Everything else exists to feed it, keep it stable, keep it cool, and keep your day‑to‑day work smooth.
This story is written in plain language so you can understand why each part was chosen, what trade‑offs we accepted, and what to watch out for if you build your own.
The one decision that drives everything: the GPU
The graphics card is the engine of an LLM PC. The key number is GPU memory (VRAM). If VRAM is big enough, the model stays fast. If it is too small, the model spills into system RAM or SSD storage and becomes dramatically slower.
We chose an NVIDIA GeForce RTX 5090 with 32GB of VRAM. It is expensive (about 2,700€) but it is the main instrument for local LLM work. Cut the GPU tier too far and you do not get a “slightly slower” machine—you get a different class of machine with smaller models and constant compromises.
The CPU: keeps the system feeling “snappy”
The CPU is the sous‑chef. It keeps everything responsive while the GPU does the heavy lifting. We selected an AMD Ryzen 9 7950X, a 16‑core, 32‑thread desktop processor, so the workstation stays calm under stress.
The motherboard: the road system
We chose the GIGABYTE X670E AORUS MASTER for its strong power delivery, PCIe 5.0 readiness, and multiple M.2 slots. A good board prevents unstable power, port shortages, and noisy thermals.
System RAM: the cabinet next to the desk
RAM is not the GPU desk—it is the cabinet right next to it. We chose 128GB DDR5 (4×32GB Kingston FURY Beast DDR5‑5200) so the system does not choke when several AI tools, browsers, and files are open at once.
Storage: where models live
We installed two Samsung 990 PRO 4TB NVMe drives (8TB total). One holds the OS and apps, the other holds model files and datasets. Fast SSDs make local AI feel convenient, not fragile.
Power supply: the part beginners underestimate most
Modern GPUs spike in power. We chose a Corsair HX1200i (1200W) with ATX 3.1 support and the PCIe 5.1 12V‑2x6 cable. Stability is worth the money.
Cooling: keeping the kitchen calm
High‑performance parts create heat. We used an ARCTIC Liquid Freezer III 360 AIO cooler so the CPU stays under control without the fans screaming constantly.
Case and airflow: performance you can feel
We chose the Fractal Design Torrent (Black Solid) and added a Fractal Dynamic X2 GP‑14 fan. For a flagship GPU, airflow is performance.
Monitor and peripherals: your human interface
We used the Dell UltraSharp U3223QE 32‑inch 4K USB‑C hub monitor, a Logitech MX Keys S keyboard, and a Logitech M330 Silent Plus mouse. Powerful hardware is wasted if the daily experience is cramped or annoying.
Putting it together: what the build actually feels like
Build the motherboard outside the case first. It is the least stressful way to see every step clearly.
Seat the CPU carefully—if it is aligned, it drops in without force. RAM takes more pressure than most beginners expect.
Install SSDs early; it is fast and satisfying. Route cables slowly for airflow and safety.
The GPU is the moment of truth. Use any support bracket available, and double‑check the power connector—fully seated, no sharp bends.
First boot can involve restarts for RAM training. That is normal. Update BIOS early for stability improvements.
Pitfalls we hit (so you do not have to)
- First boot often feels wrong before it feels right—RAM training can cause multiple restarts.
- High‑end parts do not guarantee stability. The “boring” parts (PSU, cooling, airflow) carry the build.
- GPU power connectors must be fully seated. “Almost in” is not in.
- Flagship GPUs are physically large. Plan for airflow and cable clearance.
The honest drawbacks of our system
- 32GB VRAM is huge, but not infinite—some models will still not fit comfortably.
- High‑power hardware means heat and noise under sustained load.
- The system is physically large and heavy.
- Flagship GPUs dominate the budget and the value curve moves fast.
Rough budget breakdown (rounded on purpose)
Core hardware
GPU ~2,700€ · CPU + motherboard ~800€ · RAM ~300€ · Storage ~600€
Power + comfort
PSU + case ~500€ · Cooling + fans ~100€ · Monitor ~800€ · Keyboard + mouse ~150€
If you build your own, copy the logic, not the exact parts
Decide your GPU first, then build a stable support system around it. If your budget is lower, choose a more modest GPU tier but keep the foundation solid.
The next tutorial will cover operating system installation and the first AI software setup, step by step.