Max

I've spent more than 20 years in IT and operations, and I've been building custom PCs since 2000 — my first build ran an Intel i486DX, assembled from parts when most people were buying prebuilts. Choosing every component myself never stopped; it just moved from gaming towers to the GPU rigs and AI boxes this site is about.

I started on the datacenter floor as a technician — racking servers, tracing cables, and learning what hardware reliability actually means for machines that run 24/7. Capacity planning, power budgets, thermals, and virtualization aren't abstractions to me; they were the job. That's the lens I bring to a GPU buying decision: not just the spec sheet, but what it costs to run, how hot it gets, and whether it'll still make sense in a year.

Since the original ChatGPT release I've been deep in local AI. I run open models — Gemma, Qwen, Mistral and others — on my own hardware: discrete GPUs plus an Apple Mac Studio M2 with 32GB of unified memory and a Mac Mini M4 for daily duty. When this site says what a given amount of VRAM means for local LLM inference, or how a model behaves at a given quantization, that comes from machines on my own desk.

How I work

Local AI Rigs is research-driven. Every product recommendation is verified against the live retailer listing and manufacturer documentation; specs that matter for local AI — VRAM, memory type, bandwidth, power draw, software support — are cross-checked against multiple sources; and every pick names at least one buyer who should skip it. Where I have first-hand experience — Apple Silicon, local model serving, infrastructure planning — I say so directly. Where a number comes from someone else's bench or from transparent planning math, I label it as such and never publish invented benchmark figures. The full process is on the about page.

Find me on GitHub. Spotted a price, spec, or compatibility note that looks wrong? Tell me — accuracy is the whole point.