Open the calculator →
independent resource · zero stock, zero bias

AI that fits
inside your office closet.

Local-AI tools & comparators 100% independent, to pick the right GPU and run Llama 4, Qwen 3, DeepSeek, Mistral and 240 open-weight models in-house — no cloud, no quota, no surprise bill.

See reference builds →$ open calculator
100% independent● no stock, no bias◆ free tools & comparators
Reference AI build example with multiple GPUs in a black case
Pro example
2× RTX 5090
64 GB total VRAM
1500 W PSU
Models
242
+12 this week
GPUs
204
NVIDIA · AMD · Apple · Intel
Sync
24h
via Hugging Face
Cloud needed
0
never. by design.
Break-even
9mo
vs GPT-4o-class API
02 · cloud vs local

The math fits on one page.

simulation 1 agent · 8h/day · 30d

CloudGPT-4o-class API

€
Monthly cost
≈ 480 € — token billing
↗
Data leaving
to US servers — Cloud Act exposure
§
GDPR / AI Act
DPA, register, processor controls
~
First-token latency
120 — 800 ms depending on route
×
Quota / rate limit
RPM, TPM, queue during peak hours
∅
Models
≈ 8 in mainstream API catalogs
∞
Paid off
never · perpetual rent

LocalReference STARTER build · 1× 5090

€
Monthly cost
≈ 32 € — electricity
↘
Data leaving
none. stays inside your walls.
✓
GDPR / AI Act
You host and control the system.
⚡
First-token latency
8 — 40 ms on gigabit LAN
∞
Quota / rate limit
limited by your VRAM, not by a vendor
★
Models
242 · Llama, Qwen, Mistral, DeepSeek
9
Paid off
9 months on STARTER tier
verdict computed by /lib/calc.ts · build #2178net savings · 5 376 € / year / agent · break-even month 9
03 · reference builds

Three chassis. One clear stance.

View reference builds →
// 01 · Solo dev
Starter
4 990 €indicative build cost
  • G1× RTX 509032 GB GDDR7 · 575 W
  • R32 GB DDR5-6000Corsair Vengeance · CL36
  • CRyzen 7 7700X8c · Zen 4
  • S1 TB NVMeSamsung 990 Pro · Gen 4
  • ★180 / 242 models~110 t/s · Llama 8B Q4
See details →
popular
// 02 · Agency · Small team
Pro
11 990 €indicative build cost
  • G2× RTX 509064 GB total VRAM
  • R64 GB ECC DDR5UDIMM · 4800 MT/s
  • CRyzen 9 9950X16c · 5.7 GHz boost
  • S2 TB NVMe Gen 4Samsung 990 Pro
  • ★215 / 242 models~38 t/s · Llama 70B Q4
See details →
// 03 · SME · Mid-market
Enterprise
25 990 €from · build cost
  • G2× RTX A6000NVLink · 96 GB · 112 GB/s
  • R256 GB RDIMM ECC8 channels · Corsair WS
  • CTR Pro 7975WX32c Pro · 8-channel RAM
  • S8 TB NVMe Gen 5RAID-10 · hot-swap
  • ★Architecture RAG souveraineSSO · RBAC · journal d'audit (à monter)
Ask us for advice →
04 · independence

Facts, not marketing.
Compare. Decide.

▲

Sovereignty

GDPR and AI Act friendly. Your data stays in your room, measured and auditable.

◆

Bounded cost

Capex, not rent. Break-even in 9 months for typical agentic usage. No TPM ceiling.

●

Open stack

llama.cpp · vLLM · Ollama · Open WebUI · LlamaIndex · Qdrant. No lock-in.

■

Transparent math

VRAM estimates before you buy. Real tokens/s will be added once the showroom rig is benchmarked.