Personal-AI-Computer

Time to build your own Personal AI Computer

https://github.com/user-attachments/assets/36f7aa02-42ac-4b35-b757-a5dd3b43ef1a

Prefer a finished machine?

Building from this guide is the full DIY path. If you'd rather skip sourcing, CNC, and assembly, we have it covered — order a Personal AI Computer.

The future of AI is local

AI in 2026 looks like computing in 1975: the intelligence lives in a few mainframes, owned by a few companies, and you rent it by the token. Everyone knows how that movie ends. The machine moves onto the desk — and the people who own their machines build everything that comes next.

The Personal AI Computer is that machine. The models are good, the GPUs are affordable, and homes and businesses are reaching the same answer: bring it in-house. Three reasons:

1. No one can switch you off. In June 2026 the US government took Fable 5 offline for most of a month and gated GPT-5.6 behind an approved-partner list. Weights on your own disk can't be revoked.
2. Your business is their training data. Every prompt you send OpenAI or Anthropic carries your product, your process, your edge. They will turn your secret sauce into their product.
3. Inference becomes free. Open weights on a cloud API already cut the bill 20–30×. On your own machine there is no bill at all. No API, no meter.

The end-to-end open-source local AI stack

Open software lets you run and shape your AI; open hardware lets you build, repair, and improve the machine it runs on. With both in your hands — the code, the CAD, the BOM, the BIOS settings — you can build the whole thing end to end. So we open-source both:

• Open Source Hardware — this repo. Your Personal AI Computer in five builds, below: every part, every bracket, every BIOS setting, every assembly photo.
• Open Source Software — your choice. The easiest way is Grid, the open orchestrator for local AI — but any local AI engine works: vLLM, Ollama, llama.cpp.

Building your own? Get the case kit with PCIe risers: 2-GPU, 4-GPU or 8-GPU. Not sure your parts fit? Ask for a free fit check or email [email protected], and we answer within a day.

2× NVIDIA RTX 5090

The entry-level build, for personal use. Enough for Llama, Qwen, and DeepSeek with quantization — run OpenClaw, Hermes Agent, or your own LangChain stack locally.

• 2× NVIDIA RTX 5090 — 64 GB VRAM · 3,584 GB/s · 419 FP32 TFLOPS
• Intel Xeon W5 (ASUS W790 ACE) · 96 GB RAM · 1 TB NVMe
• PCIe Gen 5 ×16 per GPU · 10 GbE + 2.5 GbE
• 1,550 W draw · 1,600 W PSU
• 12.5″ × 12.5″ × 16″ · 33 lb

Build the 2× 5090 →

2× NVIDIA RTX PRO 6000

The same desk-sized box, three times the VRAM. 192 GB of ECC memory on two cards — run the big open models as released instead of quantizing them down to fit.

• 2× NVIDIA RTX PRO 6000 Blackwell — 192 GB GDDR7 ECC · 3,584 GB/s (96 GB per card)
• Intel Xeon W5 (ASUS W790 ACE) · 96 GB RAM · 1 TB NVMe
• PCIe Gen 5 ×16 per GPU · 10 GbE + 2.5 GbE
• 1,600 W draw · 1,600 W PSU
• 12.5″ × 12.5″ × 16″ · 33 lb

Build the 2× 6000 →

4× NVIDIA RTX 5090

The team build. Larger open models like Kimi, MiniMax, and GLM. No API bills. Low latency. Private data.

• 4× NVIDIA RTX 5090 — 128 GB VRAM · 7,168 GB/s · 838 FP32 TFLOPS
• AMD Ryzen Threadripper Pro · 96 GB RAM · 1 TB NVMe
• PCIe Gen 5 ×16 per GPU · 2× 10 GbE · BMC
• 2,750 W draw · 4,000 W PSU
• 20″ × 20″ × 24″ · 66 lb

Build the 4× 5090 →

4× NVIDIA RTX PRO 6000

The team build, in ECC. 384 GB of VRAM: fine-tune and serve the biggest open models as released, for a whole team.

• 4× NVIDIA RTX PRO 6000 Blackwell — 384 GB GDDR7 ECC · 7,168 GB/s (96 GB per card)
• Intel Xeon w5-3423 (ASUS Pro WS W790E-SAGE SE) · 192 GB DDR5 ECC · 2 TB NVMe
• PCIe Gen 5 ×16 per GPU — 112 lanes across seven slots
• 4,000 W — 2× 2,000 W, native 12V-2×6 · 240 V circuit
• 19.8″ × 19.8″ × 24.2″ · 66 lb

Build the 4× 6000 →

8× NVIDIA RTX 5090

The on-prem build, for business. Develop, serve, and fine-tune with open models — the work that should never leave your floor.

• 8× NVIDIA RTX 5090 — 256 GB VRAM · 14,336 GB/s · 1,676 FP32 TFLOPS
• 2× AMD EPYC 9004 (ASRock Rack GENOA2D24G-2L+) · 192 GB RAM · 1 TB NVMe
• PCIe Gen 5 ×16 per GPU, over MCIO · 2× 1 GbE · BMC
• 5,100 W draw · 8,000 W PSU
• 5U rack chassis

Build the 8× 5090 →

Software

1. BIOS tuning and GPU testing — after you build the rig, run this to make sure it actually works: BIOS settings for multi-GPU, NVIDIA drivers, then confirm every GPU is detected, linked at full PCIe width, and stable under load.

2. Serve your models — the rig runs, now put it to work. The easiest way is Grid, the open orchestrator for local AI: it pools your machines into one local AI network. Or run any local AI engine — vLLM, Ollama, llama.cpp.

   curl -fsSL https://grid.autonomous.ai/install.sh | bash
   

Contributing

Built one? Improved a part? Found a better component? See CONTRIBUTING.md — and share your build. The best community builds get featured.

License

Open source under the MIT License. Fork it, change it, build your own and sell it — we just want it built.

---