Mini PCs vs Old Servers: The Power-Bill Maths

The advert says free. It is a dual-socket rack server with 128GB of registered ECC memory, twelve drive bays, two redundant power supplies and hardware management, and the company that owned it has moved to the cloud and wants it gone by Friday. You would be mad to say no.

I said yes. Twice. And the second time, I did the arithmetic first, which is the only reason I want to tell you about it.

Here is the specific trap: a homelab machine’s cost is dominated by a variable that does not appear on the price tag and accrues silently for years. That machine idles at somewhere between 130W and 180W. Not under load — sitting there, doing nothing, with the drives spun down. A mini PC that will run most of the same workloads idles at 6-10W. Over four years at UK domestic rates, that gap is roughly £1,600. The free server is one of the most expensive things in the flat.

That is the headline and it is not the whole story, because there are workloads where the server wins outright and I want to be fair to both sides.

Measure, do not estimate

Before any of this means anything, get a plug meter. They cost about fifteen pounds and every estimate you make without one will be wrong, usually by a factor of two in whichever direction is most convenient for what you already wanted to do.

What you need is the idle figure, because that is what you pay for. A homelab machine spends the overwhelming majority of its life idle. Peak draw under a synthetic benchmark is close to irrelevant — the machine is at that number for minutes a week and at idle for the other hundred and sixty-seven hours.

The arithmetic:

1
2

Annual kWh = watts × 24 × 365 / 1000
Annual cost = Annual kWh × price per kWh

At roughly £0.25/kWh, a useful shorthand: every idle watt costs about £2.20 a year. Memorise that. It turns “this box draws 150W” into “this box costs £330 a year” instantly, and it reframes every hardware decision you will make.

Some real figures, measured at the wall, at idle:

The dual-socket machine costs more per year in electricity than a new mini PC costs to buy. That is the sentence that ended my rack-server phase.

Where the watts actually go

Understanding why old servers are so thirsty stops you making the same mistake with a different machine.

Idle power management barely existed. Server CPUs of that era have C-states, but the platform around them — the chipset, the PCIe lanes, the vast board — was designed on the assumption that a server is always busy, because in a datacentre with 90% utilisation it is. Deep package C-states often do not engage at all. Mobile-derived silicon in a mini PC will sit in a state drawing under two watts; a 2013 Xeon platform frequently will not go below thirty per socket no matter what you ask of it.

Registered ECC memory has a floor. Every stick draws several watts continuously. That glorious 128GB is sixteen sticks and perhaps 40W of permanent overhead, whether you use it or not.

Redundant PSUs. Two power supplies, each running at maybe 15% of rated load, which is precisely where a PSU’s efficiency curve is at its worst. An 80+ Gold unit rated 92% at 50% load might manage 80% at 15%. You lose a fifth of the draw as heat before the components see any of it. Two of them is worse than one.

Fans. Six to ten high-static-pressure fans, drawing 3-6W each at idle speed. That is 30W of fans keeping cool a machine that is mostly hot because of its fans.

The BMC. Five to ten watts, permanently, whether the machine is on or off. It is a genuinely useful five watts, as I argued in IPMI and the old server nobody wants to reboot, and it still costs about £18 a year and it never stops.

None of this is fixable. It is a property of the platform rather than a setting you have missed.

Measuring properly, without a plug meter per machine

A plug meter gives you the wall figure, which is the honest one. But if you want a continuous record — and you should, because idle draw drifts as you add services — the machine can report on itself. Intel’s RAPL counters expose package power directly, and they cost nothing to read:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16

#!/usr/bin/env bash
# /usr/local/bin/pkg-watts — average package watts over N seconds
set -euo pipefail
SECS="${1:-60}"
D=/sys/class/powercap/intel-rapl:0/energy_uj

[ -r "$D" ] || { echo "RAPL unreadable (try sudo, or AMD needs a different path)" >&2; exit 1; }

start=$(cat "$D"); sleep "$SECS"; end=$(cat "$D")

# energy_uj wraps at max_energy_range_uj
max=$(cat "${D%/*}/max_energy_range_uj")
delta=$(( end >= start ? end - start : end + max - start ))

awk -v d="$delta" -v s="$SECS" \
  'BEGIN { w = d / 1e6 / s; printf "%.1f W package, %.0f kWh/yr, £%.0f/yr\n", w, w*8.76, w*8.76*0.25 }'

RAPL reports the CPU package only, so it undercounts the wall figure by everything else — the board, the RAM, the fans, the PSU losses. On a mini PC that gap is a few watts. On a dual-socket server it can be 60W, because the platform overhead is the whole problem. Use RAPL for tracking change over time and the plug meter for the number you put in a spreadsheet.

The other measurement worth taking is the one nobody does: idle draw after a fortnight of real use. A freshly installed machine idles beautifully. The same machine three weeks later, with a monitoring agent polling every ten seconds, a container runtime, a filesystem doing background scrubs and a mesh client keeping a connection alive, has lost its deep C-states entirely and gained fifteen watts. powertop --auto-tune recovers some of it. Knowing that the number moves is worth more than the number.

The honest case for the old server

I have been one-sided, so here is where the mini PC loses, and these are real.

RAM ceiling. A typical mini PC tops out at 64GB, occasionally 96GB. A ten-year-old dual-socket board takes 384GB of DDR3 that is now nearly free. If your workload is genuinely memory-bound — a large in-memory database, dozens of VMs, big local model inference — no quantity of mini PCs substitutes for one machine with enough RAM. Memory does not shard across boxes for free.

ECC. Most mini PCs do not support it. Some Ryzen models do, unofficially, in ways that are hard to verify. If you are running ZFS with data you actually care about, this is a real argument, and the endless internet fight about how much it matters obscures a simple truth: ECC costs almost nothing on a server and is nearly unobtainable on a mini PC.

Drive bays. Twelve hot-swap bays with a backplane and a proper HBA. A mini PC has one or two NVMe slots and a USB port, and USB-attached storage for a ZFS pool is a decision you will come to regret in a way that involves the word “resilver”. If you need bulk spinning storage, you need a chassis that holds it.

PCIe. Full-height slots for an HBA, a GPU, a 10GbE card. Mini PCs have, at best, one M.2 slot you could adapt and an OCuLink port if you are lucky.

Cost of entry. Free, or near enough. If your budget is genuinely zero right now, £363 a year spread over twelve monthly bills is invisible in a way that £500 today is not. That is a bad financial argument and a completely real human one.

The four-year total

Total cost of ownership over four years, for a machine doing typical homelab work:

The free server costs nearly three times as much. And the mini PC is silent, fits on a shelf, does not need a rack, does not heat the room in July, and has a CPU whose single-thread performance embarrasses the Xeon by a factor of two — which matters enormously, because most homelab work is one thread at a time waiting on something.

The “noise mitigation” line is not padding. A rack server in a flat is genuinely, seriously loud — 55-65dB at idle, which is a vacuum cleaner in the next room, permanently. You will spend money on a sound-dampened cabinet, and the cabinet will restrict airflow, and the fans will spin up to compensate, and you will have spent £150 to increase your power draw. Ask me how I know.

When I would still say yes

Free and you need 256GB of RAM this month. Take it. Nothing else gets you there.

You have somewhere it can live that is not your home. A garage, a basement, a rack at a friend’s place, a business with a machine room and a flat electricity arrangement. The noise and heat arguments evaporate and only the electricity remains.

It is genuinely occasional. A machine that is off 90% of the time costs 10% of the electricity. That combination — big machine, rare use, woken on demand — is the one configuration where old server hardware makes complete financial sense in a home. A 165W box that runs three hours a day costs £45 a year rather than £363.

Learning. If the point is to understand ECC, IPMI, out-of-band management, SAS backplanes and real server firmware, a mini PC teaches you none of it. That is a legitimate reason to own one, and it is worth being honest that you are paying £300 a year in tuition.

The mistakes I made

I counted the purchase and ignored the operating cost. £2.20 per idle watt per year. Every hardware decision, run it through that first.

I bought the “free” one twice. The second time, I had already learned the lesson and reasoned that this one had a newer CPU generation. It idled at 95W instead of 165W. That is still £209 a year, or a new mini PC every two years, forever.

I underestimated the noise. I genuinely believed I would get used to it. I did not. Nobody does. The fans modulate, and the modulation is the part that gets you.

I confused peak capability with useful capability. Twenty-four cores sounds transformative. Almost every homelab workload — a reverse proxy, a media server’s metadata scan, a container build, a database with four users — is bounded by one thread or by IO. Those twenty-four slow cores lose to four fast ones for nearly everything I actually do.

I did not measure for a year. Fifteen pounds, five minutes, and it would have changed a decision that cost me four figures. I go further into that reckoning in the real cost of self-hosting, and the electricity is only the part you can put a meter on.

The verdict

For most homelabs, in most homes, in 2026: buy the mini PC. Two mini PCs, if you want redundancy, and they will together draw less than a fifth of what one rack server does while being quieter, faster at the things you actually run, and cheaper over any horizon longer than eighteen months.

Buy the old server if you need RAM or drive bays that mini PCs cannot reach, if you have somewhere for it to live that is not a room you sleep near, or if the point is to learn the enterprise stack. Those are all good reasons. “It was free” is not one of them.

And whatever you buy, put a meter on it. The gap between what people believe their homelab draws and what it actually draws is the largest single error in this hobby, and it is fifteen pounds and five minutes to eliminate.