About Dedicated AI Hardware
This site argues one position: that a machine built to do nothing but run models is a different and better object than a general-purpose computer that also can.
There is no company behind these pages, no affiliate program, and no newsletter collecting addresses. It is a set of static files written to make a specific engineering case to people who are deciding how to run AI at home or in a small business. Everything here loads without scripts, without trackers, and without a cookie banner, because nothing on the page tries to identify you.
Why the site exists
Most writing about local AI hardware is either a product announcement or a benchmark chart. Both are useful and both skip the part that actually decides whether a purchase was a good idea: the shape of the machine, and whether that shape matches the job.
You can buy a spectacular general-purpose workstation and have a mediocre experience running a model all day. You can buy a small, cheap, unglamorous box and have an excellent one. The difference is not brand or budget. It is whether the components were chosen for one workload or for twenty, and whether the person buying understood which trade they were making.
That gap is worth writing about, and it is not being written about in plain enough terms. Hence this site.
What the argument actually claims
Three claims, in order of how much they matter.
- Shape beats specification. A machine tuned for one continuous workload behaves differently, in ways you notice daily, from a machine tuned for the average of many.
- Bandwidth is the limit. Token generation is memory-bound, so the specification that predicts real speed is bytes per second plus the size of the largest model that fits, not peak arithmetic.
- Ownership is a trade, not a triumph. A dedicated machine shifts cost from a recurring meter to a capital purchase and shifts labour from a provider back onto you. Whether that trade is good depends entirely on your volume, your latency needs, and whether your data is sensitive.
The third claim is why this site keeps publishing lists of what does not work. A page that only argues for its position is advertising. The useful version of the argument includes the conditions under which the position is wrong, and there are several.
How the material was assembled
Everything here comes from three sources, in this order of trust.
- Measurements on real machines. Throughput, time to first token, memory in use, and idle draw, taken with the model actually loaded and the machine doing whatever it normally does.
- Vendor and platform documentation. Used for facts about memory paths, capacities, and power states, and read skeptically where it concerns performance.
- Reputable independent testing. Sustained-load reviews, thermal testing, and long-run reporting. Sites that publish methodology and repeat runs are worth far more than sites that publish peak numbers.
Where a figure depends on the environment, this site says so rather than quoting a single number as if it were universal. Where a claim is an opinion, it is written as an opinion. Where the evidence is thin, the writing says the evidence is thin, which is a sentence more sites should be willing to publish.
Two external references are used often enough to be worth naming. Vendor documentation for power states and memory architecture, because it is the only authoritative source for those facts. And work on web performance and measurement discipline, such as the material published at web.dev performance, because the habit of measuring what you intend to improve is exactly as useful for a model server as it is for a website.
Editorial rules
These are the rules the pages try to follow, written down so they can be held to.
- No fabricated numbers. If a figure is an estimate, it is labelled an estimate. If it came from a specific machine, the machine's characteristics are described.
- No recommendation without a limit. Every piece of advice includes the case where it is wrong.
- No hidden position. This site favours dedicated machines. That bias is stated at the top of the homepage and repeated here, so a reader can discount for it.
- No copying. Every page is written for this site. Nothing is scraped, spun, or recycled from elsewhere.
- Plain language. Technical terms appear when they are the right word, and are explained the first time they are used.
- No dark patterns. No interstitials, no autoplaying media, no fake urgency. The page loads and the text is there.
Who this is for
Three kinds of reader, and the site tries to serve all three without dumbing anything down.
The person deciding whether to buy
You want to know whether an owned machine makes sense for your usage, and you want the arithmetic to check rather than a recommendation to trust. Start with the homepage, then the rent-versus-own article.
The person building one
You have decided to own and now need to avoid the classic mistakes: too little memory, a link that becomes the wall, an idle floor high enough to regret, storage that never parks. The bandwidth article is the one to read first.
The person who wants the argument
You are interested in why a single-purpose machine is a better object than a general-purpose one, independent of any purchase. That argument runs across all three articles, and the homepage states its shape in one page.
What this site is not
It is not a store and it sells nothing. It is not affiliated with any hardware vendor and receives nothing for mentioning one. It is not a support channel for a product, and it cannot help with a machine that will not boot. It is not updated on a schedule, because pages are revised when the underlying facts change rather than to satisfy a calendar.
It is also not neutral about its subject, and being upfront about that is more honest than performing objectivity. The bias is disclosed, the limits are stated, and the reasoning is laid out so it can be checked. That is the most a site like this can offer, and it is more than most do.
Corrections
If a number here is wrong, or a specification changed, or a claim has been overtaken by newer hardware, the right response is to fix the page rather than defend it. Details for getting in touch are on the contact page, and specific corrections with a source are the most welcome kind of message this site receives.
Start with the main argument if you arrived here first, or go straight to the bandwidth article if you are already shopping.