The cheapest way to run your own AI assistant
What a self-hosted assistant actually costs to run, comparing a Pi, a used mini-PC and a VPS — including the electricity.
The honest answer depends on whether you already own hardware and whether you want local models. Here are real numbers rather than a range.
Free: the machine you already have
On Windows, WSL2 imports a tarball and runs. On a Mac, UTM or OrbStack runs a VM. Cost: nothing.
The catch is that it is not always-on. It stops when you close the lid. Good for evaluating, wrong for an assistant that should answer at 3am.
Cheapest always-on: a VPS
Hetzner CX22 — 2 vCPU, 4 GB — runs the standard tier at about €4/month, or roughly €48 a year. Nothing to buy, nothing to maintain physically, and it is reachable from anywhere by definition.
You are renting someone else’s computer, so the data-residency argument weakens. And local inference on shared vCPU is slow enough not to bother with.
Cheapest owned hardware: a Raspberry Pi
A Pi 5 with 8 GB is around $80, plus roughly $25 for power supply and storage. Call it $105 one-off.
Power draw is 3–7 W, which at UK rates is about £10–15 a year. It pays back against a VPS in roughly two years, and against a single SaaS seat in about five months.
No usable local inference — the Pi 5 has no GPU offload, so expect a few tokens per second on small models. Most Pi deployments use a cloud provider for the model.
Best value overall: a used mini-PC
A refurbished business mini-PC — OptiPlex Micro, ThinkCentre Tiny, EliteDesk Mini — with 16 GB is often under $200. That gets you KVM acceleration and enough memory to run 7–8B models locally at 4-bit.
Idle draw is 6–15 W, so £15–35 a year.
This is the recommendation if you want local models at all. See deputyOS on a mini-PC.
The part people underestimate
Model API usage. deputyOS does not proxy or meter it — you bring your own key and pay the provider directly.
Light personal use runs a few dollars a month. An agent polling and acting on a schedule can be considerably more, which is why cost guardrails and quiet hours exist and are worth configuring on day one. See cost guardrails.
Running a local model makes this near-zero, which is the real financial argument for the mini-PC over the Pi.
Summary
| Path | Up front | Running | Local models |
|---|---|---|---|
| Existing laptop | £0 | £0 | Depends on the machine |
| Hetzner VPS | £0 | ~€48/yr | No |
| Raspberry Pi 5 | ~$105 | ~£12/yr | Not usefully |
| Used mini-PC | ~$200 | ~£25/yr | Yes, 7–8B |
costhardwareuse-case