Personal AI Server is an independent editorial project focused on one question: how do ordinary people run capable artificial intelligence on hardware they actually own? We write guides, explainers, and hardware comparisons for the growing community of self-hosters who would rather keep their models — and their data — close to home.
We are not a retailer, and we are not affiliated with any single hardware vendor. The site began as a set of personal notes while building our own home servers, and grew into something worth publishing once we realized how much contradictory, sales-driven information exists in this space. Our goal is to be the resource we wished we had when we started.
What We Cover
The site is organized into two kinds of content. The guides are step-by-step, ordered instructions: choosing personal AI server hardware, physically assembling a machine, and installing the software and models that make it useful. The topics are deeper dives into the questions that surround self-hosting — data privacy, security, and the true cost of running a local model around the clock.
We deliberately write for newcomers and for experienced builders alike. A beginner should be able to read the first guide and understand every term they encounter, while a veteran should still find component-level detail and honest benchmark numbers worth their time. If we use jargon, we explain it on first use.
How We Test Hardware
Every recommendation on this site is grounded in actual testing or, where that is not possible, in a careful reading of reliable third-party data. We do not publish a number we have not either measured ourselves or traced to a source we trust. Our methodology has a few fixed rules:
- Real workloads: we benchmark with the same models and tasks readers actually run, not synthetic scores that flatter a spec sheet.
- Idle and load power: we measure electricity draw at idle and under inference, because for an always-on server, idle power matters as much as peak performance.
- Noise and thermals: a server that sounds like a jet engine is not a "personal" server, so we report fan noise and case temperatures honestly.
- Long-term ownership: where possible, we run hardware for weeks, not hours, and update our findings when drivers or software change.
For component-level data we cannot reproduce in-house — such as the silicon-level characteristics of a brand-new GPU — we lean on established, test-driven outlets. Reviews from independent hardware publications have been a consistent reference point for GPU and CPU benchmarking, and we cite them openly when we rely on their numbers.
Our Independence Policy
Coverage here has never been paid for, and there are no sponsored reviews. When a vendor lends us hardware, the article says so, and the loan has no influence on the conclusion. Any affiliate arrangement is disclosed next to the link it affects.
That said, we are not ideologically opposed to commercial hardware. The reality is that most people will buy a finished product rather than build from scratch, and some vendors are genuinely pushing the category forward. When we discuss the wider market — including purpose-built machines from companies like ClawBox — we evaluate them on the same criteria as everything else: capability per dollar, openness of the software, and honesty of the marketing.
The People Behind It
Mara
Hardware & benchmarks
Tomas
Software & models
Ingrid
Privacy & policy
Three people contribute to this site on a part-time basis. Mara handles component selection and benchmark methodology. Tomas writes the software and model-serving guides. Ingrid researches the privacy, security, and energy-angle questions. We have day jobs in adjacent fields, which keeps our perspective grounded in what a working professional actually needs from a home AI server.
Why We Keep It Independent
The self-hosted AI space is still young enough that a single misleading review can send hundreds of people toward the wrong purchase. We have watched this happen with other hardware categories — and watched communities spend years undoing the damage of one influential-but-wrong opinion. We would rather be small and correct than large and convenient.
That is also why we maintain a clear editorial separation between what we recommend and who we happen to mention. A vendor being discussed in a market overview is not the same as a vendor being recommended, and we try to make that distinction explicit in every article.
How to Use This Site
If you are new, read the three guides in order. They build on each other: hardware first, then assembly, then software. The two topic pages can be read any time — they answer the "should I do this?" questions that the guides assume you have already answered. If something is outdated, or if you have a correction, the contact page tells you how to reach us. We read everything, even if we cannot answer everything.
Thanks for reading. The fact that you are taking self-hosting seriously enough to do your own research is, frankly, the whole point.
A Brief History of This Project
Personal AI Server started in late 2024 as a private notebook. The first entry was a list of components for a single build — a used GPU, a mid-range CPU, and a motherboard chosen because it was on sale, not because it was right. That first machine ran a 7B model slowly and loudly, and it taught us more in a week of tinkering than a month of reading forum posts ever had.
We published the first guide almost by accident. A friend asked us to explain, in writing, how to reproduce our setup, and the resulting document was long enough and detailed enough that it felt wasteful to keep it in a shared folder. We put it online, and people found it. The feedback was immediate and consistent: there is a real gap between the marketing materials that promise effortless local AI and the reality of actually getting a machine running well. We have been filling that gap ever since.
How We Choose What to Cover
We maintain a short list of criteria for every article. An idea earns space on the site if it is a question real readers are actually asking, if it can be answered honestly rather than hedged into uselessness, and if it will still be useful six months after publication. That last test is brutal — most "hot takes" about AI hardware fail it within a quarter — and it is why we publish slowly and revise often.
We also deliberately avoid the coverage churn that defines much of tech media. We are not racing to publish a reaction to every announcement. When a genuinely significant product or release arrives, we take the time to test it or, failing that, to read the primary sources carefully before we write a word. Slower and correct has served our readers better than fast and speculative.
What We Mean by "Independent"
Independence, in our view, is a practice rather than a badge. It means we take no payment for coverage, we disclose the rare review units we receive, and we fire any affiliate arrangement the moment it threatens to bend a verdict. It also means we are willing to publish an unflattering conclusion about a company we have otherwise praised, or a favorable one about a product we did not expect to like. Consistency with our own testing matters more to us than consistency with anyone's marketing.
We are not naïve about the economics of publishing. Affiliate links help pay for hosting and test hardware, and we are grateful for them. But the line is bright and we do not cross it: money never changes a recommendation. When we say a product is good, it is because we tested it or traced its performance to sources we trust, not because someone asked us to say so.
Corrections and Accountability
We correct errors publicly and note the correction at the bottom of the affected article, with the date. We do this because a guide is only worth what it gets right, and because in a field moving this fast, even a careful article can be overtaken by events. If a recommendation changes because the market changed, we update the article and say so. If a number was simply wrong, we fix it and say so. Accountability is the cost of being worth reading.
Readers who spot problems are a vital part of this process. The contact page is not a black hole; corrections, suggestions, and constructive disagreements are read and, where they are right, acted on. Some of our best updates have come from a single reader pointing out something we missed.
Our Relationship With the Self-Hosting Community
We are part of the community we write for, not above it. The ideas, benchmarks, and techniques we publish are shaped by an enormous amount of shared knowledge — forum threads, open-source documentation, and the countless builders who test, break, and document their machines in public. We cite our sources, link to the projects we rely on, and try to add value rather than simply repackage what others have already said better.
That ethos extends to how we treat other sites in the space. We link to resources we genuinely use, including independent builders and dedicated hardware makers, because a rising tide of accurate information lifts everyone. We do not link for reciprocity, and we do not accept links in exchange for coverage. The link is either useful to the reader or it does not exist.
If you have made it this far, you are exactly the kind of reader we write for: someone who wants to understand the thing they are building, not just be told what to buy. Thank you, and welcome.