A homelab on a budget: what you actually need (and what is a waste)

Most people build their first homelab on a budget — and most of them overspend on the wrong things. Not because the gear is expensive, but because they buy for a workload they imagine instead of the one they will actually run. So let me start with measurements, not a shopping list.

I run this blog from a VPS with 981 MB of RAM and one vCPU. Not a lab machine, not a rack — the cheapest thing that serves static files. Let me ask it what it is actually doing right now.

$ ssh -p622 root@88.218.79.95
$ free -m | awk '/Mem:/{print $2" MB total, "$3" MB used, "$7" MB available"}'
981 MB total, 84 MB used, 753 MB available

$ for p in nginx easytier-core vaultwarden mariadbd sms.smtp.server syncserver; do \
    printf '%-16s %s MB\n' "$p" \
      "$(pgrep -f "$p" | xargs -r ps -o rss= -p | awk '{s+=$1} END{printf "%.1f", s/1024}')"; \
  done
nginx            6.3 MB
easytier-core    6.1 MB
vaultwarden      9.2 MB
mariadbd         7.4 MB
sms.smtp.server  6.4 MB
syncserver       3.2 MB

That is a web server, a VPN daemon, a password manager, a database, an SMTP service and a sync service — about 39 MB of RAM for six services. The box reports 84 MB used in total (the rest is the OS and a build that happened to be running). Disk: 14 GB total, 8.5 GB used, most of which is the blog repo at 613 MB.

That is the honest requirement for a large class of home stacks: tens of megabytes, one core, a few gigabytes of disk. Not 64 GB and a Xeon.

What a home stack actually needs (measured)

The dev box I am typing this on is a laptop CPU from 2017 — an i7-8550U, 4 cores / 8 threads, 11.8 GB of RAM, a 234 GB root disk and two more disks hanging off it. It runs five Docker containers, a VPN daemon and a local model server, and it is not struggling.

$ nproc; lscpu | grep -E 'Model name|^CPU\(s\)'
8
Model name:          Intel(R) Core(TM) i7-8550U CPU @ 1.80GHz
CPU(s):              8
$ free -m | awk '/Mem:/{print $2" MB total, "$3" MB used"}'
11833 MB total, 8832 MB used
$ df -h /
Filesystem              Size  Used Avail Use% Mounted on
/dev/mapper/cryptroot   234G  152G   70G  69% /

So the real requirements of a home stack are modest and easy to state:

• Compute: one modern core is enough for a static site, a few containers and a VPN. Two to four cores give you headroom for a CI job or a build.
• Memory: 4 GB is generous, 8 GB is comfortable. The blog host does its job in under 100 MB.
• Storage: a few hundred GB of SSD. Media libraries are the only thing that actually needs terabytes, and that is a different hobby.
• Backups: off-box, tested, and boring. This is the requirement people skip and then regret.
• Network: the router you already own. A VPN daemon is a few megabytes, not a separate box.

What is genuinely cheap and useful

A used mini PC. An off-lease 6th–8th generation Intel mini PC (think the small office boxes that flood the second-hand market) is the sweet spot. Realistic bands are ¥800–1,500 for the machine and ¥250–450 for a 500 GB SSD — prices vary a lot by region and timing, so check locally rather than trusting any single number, mine included. It idles around 8–12 W, is silent, and is small enough to sit on a shelf.

One SSD. Buy the storage you will use this year. A single 500 GB drive holds more containers, images and notes than most people will ever create.

A UPS is optional. If your power is stable, a mini PC survives brief outages fine — a static site does not care, and most services recover on their own. A small line-interactive UPS (roughly ¥200–400) is worth it if you have frequent brownouts or a machine that does something write-sensitive. It is insurance, not a requirement.

A Pi for one job. A single-board computer is excellent at exactly one thing: being always-on for almost no power (5 W class). Ad-blocking DNS, a VPN endpoint, a small uptime monitor. It is not a general-purpose server and you should not pretend it is.

The old laptop you already own. This is the cheapest homelab on the list: ¥0 in hardware, because the sunk cost is already sunk. A laptop even has a built-in UPS (its battery) and a screen for setup. If you are unsure whether you need a homelab at all, this is the experiment to run first.

What is a waste for most people

• A rack. You have one or two machines. A shelf is a rack. Racks exist to hold dozens of units and manage their airflow and cabling; at home they hold dust.
• 10GbE. Your internet connection is not 10 Gb, your disks are the bottleneck, and a single stream from your phone will not saturate 1 Gb either. Fast Ethernet is a lab-shaped solution to a problem you do not have.
• Enterprise gear. Old 1U servers are cheap to buy and expensive to run — see the arithmetic below. They are also loud enough that you will turn them off, which defeats the purpose.
• A hypervisor with one VM. Virtualisation earns its complexity when you run several isolated guests. For one workload you want a package manager and systemd, not a hypervisor. I already argued this in Proxmox vs plain Linux, and the measured stack above is the proof: everything fits in a normal OS.
• “Just in case” capacity. The blog host uses 84 MB of 981 MB and 8.5 GB of 14 GB. Capacity you buy “just in case” is capacity you paid for, powered, cooled and backed up for nothing. Buy for the workload you have; the used market will still be there next year.
• Untested redundancy. A second disk that has never been restored from is not a backup. See below.

The real homelab total cost over 3 years

Here is where a budget gets real. I wrote a small script to compute the homelab total cost over three years — electricity plus hardware amortisation — and compared it to renting a VPS. Two honest caveats first: I have no power meter on my box, so the watt figures are typical published idle/load figures for each hardware class, and the tariff is an assumption (China residential tier-1, roughly ¥0.56/kWh; real rates run about ¥0.5–0.6 and vary by region and tier).

// 3-year TCO: owning a homelab box vs renting a VPS.
const HOURS_3YR = 3 * 365 * 24;      // 26,280 h
const TARIFF = 0.56;                 // ¥/kWh (assumption)
const kwh  = (w) => w * HOURS_3YR / 1000;
const elec = (w) => kwh(w) * TARIFF;

const scenarios = [                  // name, upfront ¥, avg watts
  ['Used mini PC + SSD (+UPS)',   1200 + 350 + 300, 18],
  ['Used mini PC + SSD (no UPS)', 1200 + 350,       18],
  ['Used 1U enterprise server',   900,              90],
  ['Old laptop (already owned)',  0,                12],
  ['Pi-class board + PSU + SD',   400,               5],
];

Run it and you get a real table, not a vibe:

Assumptions: 26,280 h over 3 yrs, tariff ¥0.56/kWh

| Setup                        | Upfront | Electricity (3 yr) | TCO (3 yr) | ¥/mo equiv | Elec % |
|------------------------------|---------|--------------------|------------|------------|--------|
| Used mini PC + SSD (+UPS)    | ¥1,850  | ¥265 (473 kWh)     | ¥2,115     | ¥59        | 13%    |
| Used mini PC + SSD (no UPS)  | ¥1,550  | ¥265 (473 kWh)     | ¥1,815     | ¥50        | 15%    |
| Used 1U enterprise server    | ¥900    | ¥1,325 (2365 kWh)  | ¥2,225     | ¥62        | 60%    |
| Old laptop (already owned)   | ¥0      | ¥177 (315 kWh)     | ¥177       | ¥5         | 100%   |
| Pi-class board + PSU + SD    | ¥400    | ¥74 (131 kWh)      | ¥474       | ¥13        | 16%    |
| VPS at ¥20/mo                | ¥0      | — (included)       | ¥720       | ¥20        | —      |
| VPS at ¥40/mo                | ¥0      | — (included)       | ¥1,440     | ¥40        | —      |
| VPS at ¥60/mo                | ¥0      | — (included)       | ¥2,160     | ¥60        | —      |

Payback vs "mini PC, no UPS" (¥1,550 upfront, ~¥7.4/mo in power):
  vs VPS ¥20/mo: 123 months (10.2 yrs) — beyond the 3-yr window
  vs VPS ¥40/mo:  47 months (4.0 yrs)  — beyond the 3-yr window
  vs VPS ¥60/mo:  29 months (2.5 yrs)

The electricity cost arithmetic

The formula is trivial: watts × hours × ¥/kWh. Over three years that is watts × 26,280 ÷ 1000 × 0.56. Two things fall out of it that a shopping list will never tell you:

1. At Chinese residential rates, electricity is not the biggest line for a mini PC. 18 W × 26,280 h = 473 kWh = ¥265 over three years — about 13% of the total. Hardware dominates.
2. Electricity becomes the biggest line the moment you buy the wrong hardware or live where power is dear. A used 1U server idling at 90 W burns 2,365 kWh = ¥1,325 over three years — 60% of its total cost, and more than the machine cost to buy. At a higher tariff like ¥1.00/kWh, the same mini PC’s power bill nearly doubles to ¥473, and the server’s jumps to ¥2,365.

So the honest version of “electricity is often the biggest cost” is: it is the cost that never stops, and it is the line that punishes cheap-looking deals. A ¥900 server that costs ¥1,325 to run is not a ¥900 server. When people say a homelab is expensive, they usually mean the watts.

A mini PC homelab vs a VPS

This is the trade-off people get wrong in both directions. Over three years:

• A budget VPS (¥20–40/month) is often cheaper than buying a new mini PC. At ¥40/month it costs ¥1,440 over three years — less than the ¥2,115 mini PC — and you never touch the hardware, never hear a fan, never lose power. The crossover, from the script above, is around four years at ¥40/month and about 2.5 years at ¥60/month.
• A mini PC homelab wins when you already own the hardware (the old laptop row: ¥177 total), when the VPS you need is expensive (more RAM or CPU than a ¥40 plan gives), when you need local storage or a GPU, or when you simply want to own it.

The pragmatic answer most people should hear: rent a cheap VPS for the public-facing bits, and use the machine you already own as the lab. That is close to what I do — the blog lives on the ¥-cheap box, the experiments live on the laptop. I keep a deliberately short list of what I self-host in 2026; the list stays short because the cheapest stack is the one you have to maintain least.

Honest advice

• Start with the smallest thing that solves your problem. Often that is a single VPS or the laptop you already have. Upgrade when something hurts, not before.
• Back up before you buy more. A second machine that mirrors an unbacked-up disk is two copies of the same mistake. Get one tested, off-box backup working first.
• Buy the second disk before the second machine. Redundancy for your data beats capacity for your compute. A mirror of the disk you care about is worth more than another node to run containers on.
• Measure before you upgrade. Run free -m, df -h, ps aux --sort=-%mem. If you are using 84 MB of 981 MB, your problem is not RAM. The numbers above are why I have not bought a bigger box: I cannot justify it against a real measurement.

What this blog’s own stack actually uses

So the whole point, measured once more, plainly:

The blog serves 254 articles in two languages, an OG-image pipeline and a VPN endpoint from under 100 MB of RAM. The load average you might see on it (3.09 when I looked) was an Astro build running at that exact moment, not steady state — which is itself the lesson: measure at the right time, or you will buy hardware to fix a spike that lasts ninety seconds.

A homelab on a budget is not a smaller version of a datacenter. It is the realisation that most home workloads are tiny, that the used market sells perfectly good small machines, and that the cost that actually compounds is the electricity — so buy small, buy used, and spend the savings on a backup disk.