autonomous-ai/autonomous-computer

autonomous‑ai / autonomous‑computer

What it is – An open‑source hardware project that provides detailed designs, parts lists, CAD files, BIOS settings and assembly instructions for building a “Personal AI Computer”. The repo ships four reference builds (2 × RTX 5090, 2 × RTX PRO 6000, 4 × RTX 5090, 4 × RTX PRO 6000, plus larger 8‑GPU configurations) together with step‑by‑step photos and a short software‑setup guide.

Why it matters – The authors argue that by 2026 most AI inference will be run on cloud APIs owned by a few providers. Running the models locally on your own machine gives you:

1. Control – you keep the model weights on‑premises, so no external party can shut you down.
2. Privacy – your prompts never leave your network, protecting proprietary data.
3. Cost – once the hardware is bought, inference has no per‑token API bill.

Key hardware specs (per build):

• GPU: NVIDIA RTX 5090 or RTX PRO 6000 (Blackwell) – up to 8 GPUs, 256 GB VRAM total.
• CPU: Intel Xeon W5 or AMD Threadripper Pro / EPYC 9004, 96‑192 GB RAM, 1‑2 TB NVMe.
• PCIe: Gen 5 ×16 slots for each GPU, 10 GbE networking, optional BMC.
• Power: 1.5‑5 kW draws, with appropriately sized PSUs.
• Form factor: Desk‑sized open‑frame for 2‑4 GPUs, 5U rack for 8‑GPU build.

Software side – The hardware is agnostic; the repo recommends the open‑source Grid orchestrator (autonomous‑grid) for pooling machines into a local AI network, but any local inference engine (vLLM, Ollama, llama.cpp, etc.) can be used.

How to get started

1. Choose a build (e.g., 2x-5090/README.md).
2. Follow the parts list and CAD files to source components and fabricate brackets.
3. Assemble the frame, mount GPUs, connect power and networking.
4. Run the provided BIOS‑tuning script (setup.md) to enable multi‑GPU PCIe lanes and install NVIDIA drivers.
5. Install Grid (or your preferred engine) and load your quantized/open models.

Community & contribution – Users can submit improvements, alternate parts, or showcase their own builds via the issue template. The best community builds are featured in the repo.

License – MIT, allowing anyone to fork, modify, build and even sell derived machines.

---

This summary is based solely on the repository’s README.