About and method

We list parts for 1–4 GPU local-AI machines that run headless: boxes built to serve or run models, not gaming PCs. This page says how the list is made, what it rests on and what it does not.

What we have not tested

We have no lab and no rig of our own. Nothing on this site is bench-tested by us. The Catalog and the guides are compiled from two kinds of evidence:

• Manufacturer specs. Every spec value cites the page that states it: the maker’s product page first, then its datasheet, manual or official announcement. Retailer pages, reviews, forum posts and this site never count as a spec source. When no acceptable source states a value, the cell stays empty rather than unsourced.
• Documented Builds. Multi-GPU machines that their owners wrote up in public with a parts list. They back the Community Builds Survey and the build tiers, and they are the only source we accept for how many GPUs a board really takes. Each one links to its write-up.

Where a guide states something from experience, it is someone else’s documented experience, and the guide links to it.

How Parts are chosen

A Part is listed only because it fits a 1–4 GPU headless local-AI build. Each Category has its own bar. A few examples:

• GPU: at least 12 GB of VRAM, and a current or recent consumer or workstation card. How much VRAM a local model needs explains why 12 GB is the floor.
• Motherboard: at least two x16-length slots, each running at x8 electrical or better. PCIe lanes, bifurcation and risers explains why.
• PSU: 1200 W or more, ATX 3.x preferred. PSU sizing and power-limiting shows how we count the GPUs a supply can feed.
• Risers and cables: rated by their maker for the PCIe generation they are listed at. No unrated Gen4 ribbons.
• Storage: NVMe, PCIe 4.0 or newer, 1 TB or more, with a published endurance (TBW) rating.

The Catalog aims for about 60 Parts across 9 Categories. Those are targets, not quotas. A Category runs short rather than include a Part that does not meet its bar.

No paid placements

Nobody pays to be listed, and nobody pays for a better spot. Store links carry affiliate tags, so a purchase through them may earn us a commission. Commission never decides inclusion: not whether a store pays, not how much, and not whether a tag exists for a region. A Part that qualifies is listed even when no store link pays. A Part that pays well but does not qualify is not listed.

We show no prices. Prices change often and differ by region and seller, so the store link is where you see the current one.

Why some Parts are recommended used

A few Parts are marked Used-Recommended. That happens only when both hold:

1. Bought used, the Part is the best value per GB of VRAM in its class.
2. A buying checklist for it exists on this site.

The RTX 3090 is the clearest case. NVIDIA lists it with 24 GB of memory, and it is the GPU most often found in the documented multi-GPU builds in the Survey. The used RTX 3090 checklist covers what to check before and after you pay. The RTX 3060 12 GB is the low-cost case, with its own checklist. Used Parts also carry used and renewed store links, and the value case is put in relative words only, never in prices.

Review every 60 days

Every Explainer, build tier, the Survey, this page and the Privacy page show a Last reviewed date. The site build fails when any of them is more than 60 days past review, so no build can ship a page in that state. The build also fails on any Part spec value that does not cite an https source, or that cites this site.

How Hardware Reads are chosen

Hardware Reads is a short, curated feed of articles about hardware for local-AI builds.

• Approved Sources only. Reads come from a fixed list of publishers. Each one was checked first, including what its terms and feed say about linking.
• Every Read is reviewed by hand. A new article is a candidate until a person reads it, approves it and writes the short note shown with it. Nothing reaches the feed automatically.
• No copied text. A Read keeps the article’s title, link and date, plus our own note. No excerpts, no images and no full text. The article itself stays on the publisher’s site.

Who writes this

This site is written in one voice, as “we”. There is no named author or persona, and every claim on it should trace to a cited source, the curation rules above or a check the build runs.