Custom GPU Inference Rig, Built in Quebec

Credit NVIDIA or AMD for the card you pick. Credit llama.cpp, Ollama, and vLLM for the software. We assemble, burn-in, and hand over a runbook.

When a tower beats a 128 GB brick

• The model fits in 24–48 GB of GDDR and you care about memory bandwidth.
• You want to swap the GPU later.
• You already have a card, a 240 V circuit, or a room that can take 500–1,200 W and fan noise.
• You want x86 + CUDA without Arm/DGX OS.

When it loses: you need a single 128 GB pool at ~240 W on a desk. That is Spark or Strix Halo. Architecture: Spark vs custom GPU, VRAM vs unified.

How a quote works

1. You name the model class (8–14B, 30–32B, 70B quantized), users, and whether the box may reach the internet.
2. We pick a GPU from current street (used 3090-class is still the usual 24 GB conversation — 3090 guide).
3. Chassis, PSU (transient response matters; mining taught that), CPU/RAM, exhaust.
4. You pay before we buy the GPU. Made-to-order is non-refundable once parts are ordered, except DOA/defect.
5. Open stack: Ollama, llama.cpp, or vLLM, plus a private UI if you want it.

We will not quote 70B full precision on a 24 GB card. We will not invent tokens/s. Dual-GPU is a separate quote (split VRAM, not a 48 GB unified pool).

Heat and the rest of the shop

A GPU tower makes heat. If you already duct miners, the same 120 mm language applies — see heating with inference and GPU mining rig to inference. Heater SKUs stay on those pages, not on a Law 25 article.

Request a GPU rig quote · Which local AI computer