Own your intelligence.
Never revoked.
Your models don't get deprecated, rate-limited, or pulled. The weights sit on your disk.
Private by physics.
Prompts never leave the building. Nobody trains on your work. Not by policy, by physics.
No per-token bill.
Own the machine and agents can run around the clock. You pay for electricity, not tokens.
Yours to keep.
Fine-tune on what only you know. The model becomes an asset you own, and it sells with the company.
Eight cards, tuned and burn-tested. We install RTX 5090 or RTX PRO 6000 Blackwell — installed, driver-loaded, and benchmarked as one system before it ships. Plug into a 240V circuit and run.
A server chassis, not a showpiece. At eight cards you're building something that lives in a rack and runs for weeks, so it ships in a 7U GPU server chassis built for exactly this: hot-swap fans, hot-swap drive bays, pull a drive without opening the machine.
Datacenter cabling, not desktop ribbon. Each GPU connects over a PCIe 5.0 ×16 riser fed by two MCIO cables — the shielded connector datacenters use to carry Gen5 without dropping signal. Straight from the CPU, no PCIe switch deciding which card waits.
.webp)
.webp)
ASRock Rack TURIN2D24G-2L+. A dual-socket AMD EPYC server board with enough Gen5 lanes to feed all eight cards direct from the processors — plus server-grade remote management, so you can watch it and restart it without standing in front of the rack.
Two AMD EPYC 9124s. Dual server processors carry every card on its own dedicated Gen5 ×16 — no sharing, no waiting. This is the same silicon that runs the datacenters you're replacing.
Built to stay up. 2000W CRPS server supplies, hot-swap and redundant — pull a failed unit while the box keeps serving. One thing to plan for: this needs a 240V circuit, not a wall outlet. Sort it before the machine arrives.
Air moved on purpose. A wall of twelve hot-swap fans, PWM controlled, wired to one controller and matched to the airflow path through the case. Swap a fan without stopping a job.
ECC everywhere, 8TB of NVMe. Sixteen 32GB DDR5-4800 ECC RDIMMs catch silent bit-flips on jobs that run untouched for weeks. Two 4TB NVMe drives in hot-swap bays, six bays free to grow.
GPU
Compute
Memory & storage
Power & cooling
Dimensions
Software
Warranty
Other
The lineup
Three builds.
Open-source hardware
The whole machine
is open.
Every CAD file, bill of materials, and BIOS setting — free on GitHub. Fork it, change it, build your own, even sell it. This is how the personal computer began: in the open. We’re doing it again, for AI hardware.
GitHub 1.2k2x-5090/ 4x-5090/ 8x-5090/ 4x-6000/
bom/ step_models/ stl-models/ photos/
README.md setup.md
MIT · 147 forksClone it. Build it. Sell it. We just want it built.