| Qté | Price |
|---|---|
| 3-4 | $76,630 |
| 5-9 | $75,050 |
| 10-29 | $74,260 |
| 30+ | $73,470 |
Your models don't get deprecated, rate-limited, or pulled. The weights sit on your disk.
Prompts never leave the building. Nobody trains on your work. Not by policy, by physics.
Own the machine and agents can run around the clock. You pay for electricity, not tokens.
Fine-tune on what only you know. The model becomes an asset you own, and it sells with the company.
We build the whole machine. You skip the build. We source every component, mill and assemble the chassis, install and tune a set of 8x RTX Pro 6000 BlackWell or RTX 5090, then burn-test the full system as one build - loaded with drivers, benchmarked. Ready to run. You just plug it in and go.
A server chassis, not a showpiece. The Autonomous Computer and Autonomous Computer Studio ship in a case we mill ourselves. The Autonomous Computer Rack does not, and that is on purpose. At eight cards you are building something that lives in a rack and runs for weeks, so it goes in a 7U GPU server chassis built for exactly this. It holds eight full-size cards, pushes air through eight hot-swap 8056 fans plus four more, and puts eight hot-swap bays on the front so you can pull a drive without opening the machine. Nothing here is styled. It is built to be cooled and serviced, and to keep running while it does.
Datacenter cabling, not desktop ribbon. Each GPU connects over a PCIe 5.0 x16 F-adapter riser fed by two MCIO cables, the shielded connector datacenters use to carry Gen5 across a cable without dropping the signal. Every card gets a full Gen5 x16 straight from the CPU, no PCIe switch in the middle, so no card shares lanes or waits on another. That is the same bandwidth a card gets in a top desktop slot, times eight, all at once. The risers and cabling come matched to the chassis, routed for airflow and built to stay put in a machine that runs for weeks.
.webp)
.webp)
The board that never makes a card wait. On an eight-GPU machine the board is the whole game. It has to hand every card its own lanes, hold steady under load for weeks, and let you manage it from somewhere else. The Autonomous Computer Rack runs a dual-socket AMD EPYC server board, the ASRock TURIN2D24G-2L. Two EPYC 9005 CPUs carry enough PCIe Gen5 lanes that all eight cards connect straight to the processor, each on its own dedicated Gen5 x16, with no PCIe switch in between deciding which card waits. Memory runs across 24 channels of ECC DDR5, so the cards are never starved for data. And it ships with server-grade remote management, so you can watch it and restart it without standing in front of the rack. Chosen for lanes, stability, and the fact that it is meant to be run in a rack, not looked at.
Power built to stay up. Eight cards at full draw is a serious electrical load, and this is where a deployment machine either holds or falls over. The Autonomous Computer Rack runs four 2000W CRPS supplies, hot-swap and redundant, feeding a distribution board that splits clean power to every card. One thing to plan for: this needs a 240V circuit, not a standard wall outlet. That is normal for hardware at this scale, but it is the kind of thing you sort out before the machine arrives, not after.
Air moved on purpose. The fans are mapped to the chassis before putting components in, not bolted on after. The rack runs a wall of twelve hot-swap fans, PWM controlled, wired to one controller and matched to the airflow path through the case.
Memory built to run for weeks. The AC8 is meant to serve models around the clock, so ECC is not optional here, it is the point. Error-correcting memory catches silent bit-flips before they corrupt a long-running job, which is exactly the failure mode that bites hardest when a machine runs untouched for weeks. It's sized to your total VRAM across eight cards, so memory never becomes the limit.
Open-source hardware
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 9402x-5090/ 4x-5090/ 8x-5090/ 4x-6000/
bom/ step_models/ stl-models/ photos/
README.md setup.md
MIT · 125 forksClone it. Build it. Sell it. We just want it built.
The lineup
