We Have AI at Home

Chapter 1: A Box of Scraps

Date: August 13, 2026

Being a software developer right now is a wild experience. In my view, the emergence of the transformer large language model represents the first truly groundbreaking shift in our field in ages.

The rise of AI coding agents has fundamentally altered the workflow. It feels less like a gradual change and more like a leap in tooling:

It's the equivalent of switching from traditional hand-carved woodworking to using industrial power tools—efficient, but with a different set of risks.

The Philosophy of Local Control

I have a personal rule: I can't truly trust a piece of technology unless I can dismantle it and rebuild it in my own garage. The current trend of connecting to a proprietary service to use a coding agent—where the "brain" is controlled by a distant corporation—makes me uneasy.

The risks are too high:

• Dependency: You integrate it into your life, then the service changes or vanishes.
• Control: You are at the mercy of someone else's API and guardrails.

While the tech is thrilling, I don't want to be a client of a massive, remote data center. I want a surplus shortage-era data center in every single home!

---

The Hardware Hunt: Mining E-Waste

To make this happen, I need GPUs. The fundamental requirement for AI is fast memory paired with a processor capable of massive parallel matrix multiplication, which can be represented as:

$\text{Output} = \sigma(W \cdot X + b)$

Unfortunately, any hardware explicitly marketed for AI is currently priced for enterprises, not hobbyists. However, I found a loophole in the graveyard of "cloud gaming."

The AMD "Ghost" Cards

Around 2022, companies tried to push cloud gaming (running games in a data center to avoid buying a console). AMD created a specific workstation GPU for this:

1. Modified Specs: Extra RAM added.
2. Stripped Down: All video outputs removed.
3. Market Failure: They were overproduced, and the venture failed due to the inevitable problem of latency (lag).

Because these aren't "AI cards" and AMD's software support is famously great terrible, they are dirt cheap on eBay from e-waste resellers. Mine looked practically brand new.

The Build Specifications

Since we are in a parts shortage, I built the rest of the rig from salvaged components.

A note on the CPU: The 10900X is arguably one of Intel's worst offerings from the last two decades—a desperate attempt to compete with Ryzen's lithography. But for me, that just means it was cheap.

A note on Memory: Since the heavy lifting happens in the GPU's VRAM, the system RAM and SSD are mostly just for loading the model.

---

Engineering the Cooling Solution

Server GPUs are designed for chassis with screaming blower fans; they have no onboard cooling. I looked at existing 3D-printable shrouds, but they were all flawed:

• Option A: Tiny, deafening 40mm fans.
• Option B: Too wide to allow multiple cards to sit side-by-side.

The Custom Design

I used OnShape to design a custom shroud. My goal was to fit two 80mm server fans across four dual-slot cards (approx. 160mm total width).

Technical Details:

• Material: Carbon fiber ASA (though PLA would have sufficed).
• Attachment: Used the existing screw-holes on the metal cable guide fins.
• Airflow: 10,000 RPM fans to ensure maximum thermal headroom.

---

The Final Configuration

The initial setup was a nightmare. I spent an hour in my garage wearing ear protection because the fans were screaming while I fought with the BIOS.

To get the motherboard to recognize four GPUs with massive amounts of VRAM, I had to dive deep into the advanced settings to enable:

1. Resizable BAR
2. MMIO High Size

Build Checklist:

• Source e-waste GPUs
• Salvage X299 Motherboard
• 3D Print Cooling Shrouds
• Install 1600W PSU
• Configure BIOS settings

As I finally got the system to post, I realized I didn't have Et...