6,221 Self-Hosters Were Asked About AI. Only 849 Keep It on Their Own Hardware.
The 2026 Self-Host User Survey asked about AI for the first time. Of the 3,761 self-hosters who use it, only 849 keep the model on their own metal.
You run your own photos. Probably your own media, your own passwords, your own Git, and if you are one of the two-thirds of respondents who told the 2026 Self-Host User Survey they keep a NAS, you have spent real money on the principle that your data belongs on your own shelf. Then there is the thing you type your hardest questions into at eleven at night, when the reverse proxy has stopped answering and you cannot work out why. That one, the numbers say, you rent.
Until 12 November there were no self-hosting AI usage statistics worth citing. Everything in circulation traced back to a vendor with something to sell, or to GitHub stars and Docker pull counts standing in for people. Then selfh.st published its annual survey with 6,221 completed responses and, for the first time, a section about AI. Of those respondents, 3,761 use AI in some capacity to help them self-host. Of that group, 849 use self-hosted models only.
That second number is the one worth sitting with, and it is not the number the headlines have been repeating.
The self-hosting AI usage statistics that did not exist a week ago
The survey is run by selfh.st, a self-hosting newsletter, and published on 12 November 2026. It drew 6,221 completed responses against 4,081 the year before — a 52% rise, which is a lot of new people answering a long questionnaire about their home infrastructure.
The 2025 edition is worth loading once for the comparison, because it makes the point by omission: it has no AI section, no AI question, and no mention of AI anywhere on the page. So nothing here is a trend. There is no “up from” available for any AI figure in this survey, and anybody who writes one is inventing it. This is a first measurement of a population that had never been measured on this.
Who that population is matters, and the survey is unusually honest about it. Most respondents self-host for personal rather than commercial purposes. Hobby and privacy remain the two prime motivators. Only 16% avoid putting sensitive data on their own hardware. About two-thirds self-host home automation. And 4,761 of them work in technology against 1,271 who do not, which is roughly 79% — this is a self-selected newsletter audience of technically employed hobbyists, not a random sample of anyone. Read every figure below with that in mind. It is a precise measurement of a specific crowd, which is worth far more than a vague measurement of everybody, as long as nobody pretends it is the latter.
The 23% is smaller than it sounds
The survey’s own one-line summary reads that 23% use self-hosted models only, and that is the sentence the aggregators have copied. It is accurate. It is also missing its denominator, which is where the story is.
The 849 people who answered “self-hosted only” are 22.6% of the 3,761 who use AI at all. They are 13.6% of the 6,221 people who filled in the survey. So among self-hosters who have brought AI into their workflow, roughly three in four send it to somebody else’s API. Across the whole sample, about one in seven keeps AI entirely on their own metal.
One caveat belongs here rather than in a footnote. That question allows multiple answers — the options total 6,687 selections across 3,761 people — so “self-hosted only” is an odd fit for a multi-select, and it should be read as self-reported exclusivity rather than a clean partition of the group. It is still the only direct count anyone has published of self-hosters who keep their models at home.
What self-hosters actually use AI for
The use-case question shows what they are asking it to do, and the shape is consistent: these are the jobs of somebody maintaining infrastructure, not building a product.
Out of 3,761 AI users, troubleshooting takes 3,249 — 86% — followed by configuration at 2,919 and learning and research at 2,863. Scripts or automations takes 2,670. Documentation sits at 1,941. Application development comes last at 1,838, just under half.
Look at the top four as tasks rather than categories. Read this log and tell me which line matters. Explain what this block in a compose file is doing. Write me a cron job that prunes these snapshots. Explain how this protocol works. They are short-input, bounded, well-documented problems, and they are the ones a model running on your own hardware is genuinely competent at. The single job where a frontier model is still clearly ahead — writing an application — is the one fewest people are using it for.
Which makes the 849 figure stranger, not more sensible. The crowd that most wants its data at home is mostly using AI for exactly the work that does not need to leave home.
Why Claude, in a crowd that distrusts clouds
Among the 3,761, Anthropic’s Claude takes 2,360 selections, ahead of OpenAI’s Codex and ChatGPT at 1,436 and Google’s Gemini at 1,035. The open-weight names show up far down the list — DeepSeek at 169, Mistral at 67, Z.ai’s GLM at 65 — and the survey counts them as services, on the same question and alongside the hosted names, rather than as local installs. Running one of those models on your own hardware is what “self-hosted only” is for.
The honest reading is not hypocrisy. It is a population that has thought carefully about where its data lives and has decided, case by case, that this particular category is worth the exception. These are people who already run a firewall, a reverse proxy and their own DNS. They know what they are sending out. They are sending it anyway because the quality gap at the top end is real and because a debugging session at eleven at night is not the moment to discover your 8B model has lost the thread.
That is a defensible trade. It is just worth noticing that it has never been tested against what local models can do now, because most of the people making it made it when a local model was noticeably worse at all six of those jobs.
Moving the four cheap jobs onto your own box
If you are in the three-in-four and want to move some of it home, the useful move is not to replace your cloud subscription. It is to take the top four use cases off it and leave the application development where it is.
Start with the model sizes, because this is where most guides get vague. Ollama’s library page for qwen3 lists sizes from 0.6B up to 235B, with the download size and context window printed against each one. The one worth starting with for general troubleshooting is qwen3:8b at a 5.2GB download with a 40K context window. If you have VRAM to spare, qwen3:14b is 9.3GB. At the small end, qwen3:1.7b is 1.4GB and qwen3:0.6b is 523MB, which will run on almost anything with a fan.
The interesting tag is qwen3:4b. It is only 2.5GB — smaller than the 8B — but Ollama lists it with a 256K context window against the 8B’s 40K. For the configuration use case, that trade is usually the right one. When the input is a 400-line compose file or a stack trace with twenty frames of context, the window is the constraint, not the parameter count. A larger model that cannot see the whole file is worse than a smaller one that can.
One warning, because it is the most common mistake in this corner: a download size is not a VRAM requirement. A 9.3GB download does not mean 9.3GB of video memory, and the relationship between the two depends on quantisation, context length and how much of the model you offload. If you want the actual arithmetic, this site has a piece on what a 35B model really needs from an 8GB card that works it through properly.
Practically, the path is short. Get Ollama answering on port 11434 — the complete guide to running local LLMs with Ollama covers the install and model selection and there is no reason to repeat it here. Pull qwen3:8b and qwen3:4b so you have both the general and the long-context option. Put a self-hosted chat interface that speaks to Ollama in front of it, because pasting logs into a terminal gets old within a day. Then, for a fortnight, send the troubleshooting and configuration questions there first and only escalate to the cloud when the local answer is not good enough. Keep a tally of how often you escalate. Whatever that number turns out to be, it is yours, measured on your hardware and your actual questions — which is a better basis for the decision than anybody’s benchmark, this survey included.
And if you are among the two-thirds who self-host home automation and are tempted to wire the model into it, read what to withhold from a local model pointed at Home Assistant first. The defaults there give a model more reach than most people expect.
What breaks when you go local-only
Application development is the obvious casualty, and the survey already shows people know it — it is the use case with the fewest takers even among those paying for frontier models. A local 8B or 14B will write you a working script. It will not hold a multi-file refactor in its head, and pretending otherwise is how people end up disappointed and back on a subscription within a week.
Context is the second thing that bites, and it bites in a way that reads as stupidity rather than as a limit. The window Ollama prints against each tag is the ceiling on what the model can be shown at once; feed it a log longer than that and the answer you get back is an answer about the part that fit. Nothing in the reply tells you which part that was. If your work is mostly reading long outputs, pick for the window first and the parameter count second — which is the whole argument for the 2.5GB qwen3:4b and its 256K window over the 5.2GB qwen3:8b and its 40K.
The third limit is not technical. Among all 6,221 respondents, 3,922 — 63% — said they are fine with AI assistance in software development but draw the line at vibe coding, and 422 said they avoid AI-assisted projects altogether. That second group is small, about 7%, but it exists and it is not confused. Running somebody’s AI-generated application on hardware in your own house is a provenance question, not a performance one, and the survey shows it is the majority position that there is a line, even where people disagree about where it falls.
There is also a case for not bothering. If your threat model is “I do not want my logs in a training set” and your provider contractually says it does not train on your input, a local model may be solving a problem you do not have. The counterweight is worth reading honestly: when a local model is actually the right answer, and what it costs goes through the cases where the answer is no.
What the survey still does not tell you
The gaps are as informative as the figures. The survey does not ask which models the 849 run, on what hardware, at what quantisation, or whether they are satisfied with the results. It does not ask what any of them tried and abandoned. “Self-hosted only” covers both somebody running a 70B on two 3090s and somebody running a 4B on a Mac mini, and those are not the same answer.
It is worth appreciating how little existed before this. Findstack published a self-hosting statistics piece in October 2026, measured on data from August, which tried to answer roughly this question from Docker Hub pull counts and Ollama’s GitHub stars. Its authors state plainly that they “did not survey any users”, that hardware and cost breakdowns “require survey data we don’t hold”, and that their numbers are “consistent with” AI tooling being the fastest-growing self-hosted category but “cannot prove it”. That is a careful piece of work admitting it cannot see the thing it wants to measure. Four weeks later, 6,221 people simply answered the question.
The number to watch is the 849. If the local models keep closing the gap on the four jobs this crowd actually uses AI for, it should move next November — and because this year is the baseline, that will be the first self-hosting AI figure anybody can honestly say has changed.
Sources
Every figure in this article traces to one of these. Dates are when the page was read.
- 2026 Self-Host User Survey Resultsselfh.stAccessed Nov 16, 2026Independent
- 2025 Self-Host User Survey Resultsselfh.stAccessed Nov 16, 2026Independent
- qwen3 model libraryOllamaAccessed Nov 16, 2026Official docs
- Self-Hosting Statistics: We Measured It Two Ways and Got Opposite AnswersFindstackAccessed Nov 16, 2026Independent
Written by
Full-stack developer who builds and runs AI automation systems in production. Runs local LLMs on personal hardware, builds N8N pipelines that actually ship, and deploys on Cloudflare Pages. Every guide on Pipeline Monk is tested on real consumer hardware — a Ryzen 7 5800HS with 16GB RAM and a GTX 1650. If it works on that, it works.