One watt running around the clock is 8.76 kWh a year — about $1.30 at the US average of ~$0.15/kWh, or roughly €2.60 at European rates near €0.30/kWh. That single conversion prices every homelab decision: a 10W mini PC costs ~$13 a year, a 40W NAS ~$53, and a rescued dual-Xeon server idling at 120W costs ~$160 a year in the US and over €300 in Germany — more than the machine did. And because a homelab sits idle 95%+ of the time, idle watts are effectively the whole bill.

The formula, and the table to sanity-check against

watts × 8.76 = kWh/year
kWh/year × your rate = annual cost

Example: 40W NAS → 350 kWh → $53/yr at $0.15, €105/yr at €0.30

Typical measured idle draws (estimates; your BIOS settings and disk count move these):

DeviceIdle watts$/yr @ $0.15€/yr @ €0.30
Raspberry Pi 53–4W~$5~€10
N100 mini PC6–9W~$10~€21
Used ThinkCentre (8th-gen i5)8–12W~$13~€26
4-bay NAS, drives spinning30–45W~$50~€100
Tower with idle discrete GPU45–70W~$75~€150
Dual-Xeon rack server90–140W~$150~€300

Find your actual rate on your bill, not from a national average — US states span roughly $0.11 to $0.40/kWh, which is the difference between electricity being a rounding error and being the dominant cost of the hobby.

Measure, don't guess

Spec sheets list maximums; you care about idle, and idle varies wildly with configuration. Two tools: a plug-in power meter (a Kill A Watt or equivalent, ~$25) for a spot check, or — better — a smart plug with energy metering (~$10–15) feeding Home Assistant's energy dashboard, which turns one week of real usage into a kWh number that includes every drive spin-up, backup window, and nightly transcode. Measure a week, not a minute; my "10W" server averages 14W once the 2am restic run and the weekend Jellyfin sessions are amortised in. A UPS with a load readout gives the same answer for the whole rack at once.

Load barely matters; duty cycle is everything

The counterintuitive result: heavy use is cheap because it's brief. A transcode pulling an extra 30W for an hour a day is 11 kWh a year — under $2. An LLM query spiking a GPU to 200W for thirty seconds costs effectively nothing. What costs money is anything continuous, and the list of continuous draws is short: base platform power, spinning disks (5–8W each, so four drives ≈ $30–40/yr), and — the one that surprises people — a discrete GPU existing at all. An RTX 3060 idles at 12–17W doing nothing, call it $16–22 a year as a standing charge for the slot being occupied. Fine if it earns its keep; worth questioning if it ran one experiment in March.

Per-app accounting follows directly: your twenty containers are collectively a rounding error on an idle CPU. Attribute costs to hardware states, not software — disks that can't sleep because one app scans continuously, a GPU held resident, a second box that exists for one service that could co-locate.

Upgrades that pay back, ranked

Retire enterprise gear. Replacing a 120W rack server with a $150 used mini PC saves ~$145/yr at US rates — payback around 13 months, and closer to 6 in Europe. Nothing else on this list is in the same league; the off-lease hardware guide covers what the big iron is still legitimately for.

Consolidate boxes. Two half-idle machines burn nearly double the platform power of one. Folding a Pi-and-old-desktop sprawl into a single mini PC typically saves 15–30W — $20–40/yr — and simplifies everything else too.

Let the CPU sleep. powertop --auto-tune plus enabling ASPM and package C-states in the BIOS gets many machines 3–8W lower for free. Verify with powertop that you're reaching C6 or deeper; a single USB device or NIC setting can pin the package awake, and finding it is a classic satisfying evening.

Spin down disks — selectively. Sleeping drives save 5–8W each, but frequent spin-ups add wear and latency, and any always-scanning app defeats the timer anyway. Worth it for a cold archive drive; not for the array your media server watches.

What doesn't pay: replacing a working 10W machine with a newer 7W one ($4/yr, never recoups hardware cost), and PSU efficiency chasing on machines this small — though if you're repurposing a desktop, know that a 750W PSU at 15W load can run under 70% efficient, one more point for the right-sized mini PC.

What I'd do

Put a metering smart plug on the whole stack for one week and get the real number — most people's guess is off by 2× in one direction or the other. Under 25W total (the healthy range for a mini-PC-plus-drives lab): you're done, spend the evening elsewhere, since even at German rates you're near €65/yr for the whole hobby — compare that against the SaaS subscriptions it replaces and it's carried by a single service. Over 60W: something on the list above is mispriced, it's usually the old server or the resident GPU, and the fix funds itself inside a year. The goal isn't minimalism for its own sake — it's knowing that every watt on the meter is one you decided to spend.