Custom local AI PC assembled in France
“Custom PC,” “assembled in France,” “local AI”: three overused terms that many vendors use without them really meaning much. Here is what they actually entail, misconception by misconception, followed by a concrete account of a build as we do it at Radiance Systems.
Five misconceptions about custom AI PCs
The priorities are reversed. A gaming PC targets frames per second; an AI PC targets VRAM, memory bandwidth, and stability over long computations. A card with a lot of video memory but average graphics performance will be excellent for AI and mediocre for gaming. Designing for AI means making different trade-offs for every component.
True customization is about VRAM, RAM, storage, and cooling, tailored to your specific use. A 14B LLM user, an AI video creator, and an engineer fine-tuning models do not need the same machine, even with the same budget. Color is the last question, not the first.
It is the best way to pay for power that will never be used. If you are running 14-billion-parameter models, a 16 GB card is enough, and a 32 GB card will sit half-idle. Customization means not overspecifying as much as it means sizing correctly. The goal is suitability, not overkill.
The difference is concrete when something goes wrong. With a local build, it is the same team that built the machine, tested it, and answers when you call, in French, knowing your configuration. Not a hotline disconnected from production. And the machine remains upgradeable: we can take it back for an upgrade.
That was true two years ago. Today, a machine can be delivered with everything pre-installed: Ollama, Open WebUI, or LM Studio, and the models already downloaded. The user opens their browser and chats with their AI like a consumer tool, without typing a single command.
What does a build actually look like
Customization is not an abstract promise. Here is how the manufacturing of a machine unfolds, real step after real step, taking the example of a creator who wants to do AI image and video generation.
Example: a station for AI visual creation
The need. The client generates images with Flux and wants to get into AI video. They therefore need a lot of VRAM and bandwidth, but not necessarily an extreme computing processor. We start from there, not from a catalog. This approach applies to every local AI station or server we assemble, regardless of your profession.
The heart. We choose an RTX 5090 32 GB: its memory and bandwidth are what AI video demands most. It is the decision that structures everything else.
The balance. Around it, we size 64 GB of RAM for the encoders and the pipeline, a fast NVMe for large files, a power supply with headroom, and cooling calibrated for long renders without throttling.
The preparation. Upon request, we install ComfyUI, the desired models, and the CUDA environment, so that the first generation comes out in a few minutes, not after a day of configuration.
The workshop. The machine is assembled by hand in Auriol, neatly cabled, then tested under real load on AI renders to check temperatures and stability before shipping.
The delivery. It leaves packed and insured, anywhere in the EU, ready to plug in. And if the client later wants a second card or more storage, the machine is designed to be upgraded.
Change the initial need, and the entire chain adapts: a doctor will want a silent server that facilitates GDPR compliance, a team a shared local AI server, an ML engineer an ECC card for long training, a gamer a versatile gaming and AI base. The common point: we always start from the use, never from an imposed product.
Sovereignty, taken seriously
Running AI locally already keeps your data at home, outside of any foreign cloud. Assembling it in France adds end-to-end consistency: design, assembly, and support are local.
For an individual, it is the assurance of a reachable contact and a durable machine. For a professional subject to confidentiality (doctor, lawyer, accountant) it is a direct answer to the requirements of the GDPR: no data leaves the office, on an infrastructure controlled locally.
Some basics to get started
These configurations are starting points. The towers are adjusted component by component.
CoreIA RTX 5060 TiTo get started with local AI, 14B LLM.
€2,068
CoreIA RTX 5090The basis of the example above: AI video, models up to 30B in 8 bits.
€10,291
CoreIA GB10 (ASUS Ascent GX10)Team AI server, 128 GB unified, compact.
€6,999
What to remember
A custom local AI PC is neither a rebranded gaming PC, nor a static catalog. It is a machine sized for what you really do, assembled and tested by hand in France, delivered ready to use, and designed to last and evolve. We start from your use, we adjust every component, and your data never leaves your hardware.
The rest (the color of the case, the lighting) comes at the end. It is the right order of priorities.
