Did the Rebuild Actually Save Power?

In the first post I listed power as one of four reasons for retiring two 1U Supermicro servers in favour of two small Proxmox mini PCs. It was honestly one of the main ones. The old pair ran around the clock to do work that fits comfortably on a single modern node, and Maine has some of the highest residential electricity prices in the country.

So: did it work? Did the rebuild actually save power?

I want to give you a clean before-and-after chart. I can’t — and the reason is the honest heart of this post.

I can’t measure it

The only power metering I had through the rebuild was at the whole-house level — no meter on the rack, no per-circuit monitoring, and, while the swap was actually happening, no smart plug on any of the servers. And a house is a noisy place to look for a 300-watt signal: we charge electric vehicles, run an electric dryer, and cook on electric ovens. Any one of those swamps the servers by an order of magnitude. Switching off two old machines should drop the baseline by something like a quarter-kilowatt — but in a whole-home trace next to an EV pulling 7,000 watts, that step is invisible.

I went looking for it in the Home Assistant energy data anyway, and I couldn’t pick it out with any confidence. That is a real limitation, not a rhetorical one. The right tool would have been a metered PDU or a couple of smart plugs logging each box before I shut it down, and I didn’t have them in place while it mattered. I eventually fixed that — more on that below — but too late for the clean, continuous before-and-after at the panel I really wanted.

So the body of this post is estimates — bounded as tightly as I can, and labelled as estimates throughout — followed, near the end, by a direct spot measurement that finally puts a real meter behind them.

The one number I do know: the rate

The electricity rate is real, straight off my bill. My utility is Central Maine Power (CMP), and in Maine the bill splits into two parts that you pay on every kilowatt-hour:

• Supply (the standard offer): $0.127210 / kWh¹
• Delivery: $0.136474 / kWh²

Add them and the all-in cost of a kilowatt-hour is $0.263684 — call it 26.4 cents. That is well above the US residential average of roughly 16–17 cents.³ It is the number that turns watts into dollars, and it is the main reason power was worth caring about here at all. For what it’s worth, the standard offer alone rose from $0.106128 in 2025 to $0.127210 for 2026 — about a 20% jump in the supply half of the bill in a single year,¹ which doesn’t make idle servers any easier to justify.

What the old servers drew

Two 1U Supermicro servers, each with two Xeon E5-2620 v2 CPUs and 128 GB of DDR3, idle most of the day but never fully asleep. Without a meter I can’t give you the real figure, but for this generation of dual-socket hardware an idle-to-light-load draw on the order of 150 watts per server — about 300 watts for the pair is a reasonable estimate, and the pair figure is the one I used in the first post. Every wattage in this post is for both boxes of a pair unless I say otherwise.

Running the pair flat out, 300 watts continuous works out to:

300 W × 24 h × 365 days ≈ 2,600 kWh per year

At 26.4 cents that’s about $690 a year to keep the pair powered on. If the true idle was nearer 250 watts — entirely possible — it’s closer to 2,200 kWh and $580. Either way it’s real money, spent largely to keep idle capacity warm.

What the mini PCs draw

The replacement is two Minisforum mini PCs running Proxmox. These are mobile-class parts with a fraction of the power envelope of a dual-Xeon server. I don’t have a meter on these either, but a draw on the order of under 20 watts per mini PC — about 35 watts for the pair under the lab’s actual workload is a conservative estimate for this class of machine.

35 W × 24 h × 365 days ≈ 310 kWh per year

That’s about $80 a year for the pair.

The before-and-after

Putting the two estimates side by side:

Roughly 2,300 kWh and $610 a year, give or take. Even if every estimate here is optimistic, the shape doesn’t change: replacing two dual-Xeon servers with two mini PCs is close to a tenfold cut in continuous draw. That’s the kind of margin that survives a lot of error in the inputs. I’m confident in the direction and the rough magnitude; I’m explicitly not confident to three significant figures.

What they actually draw — measured

I almost shipped this post as pure estimate. Then, before pulling the plug on the old servers for good, I did the thing the whole post had been apologising for not doing: I put a smart plug on every one of the four boxes — chimaera and basilisk on the old side, both Proxmox mini PCs on the new — and logged each one through Home Assistant.

This is a spot measurement, not a year of data: roughly five hours of overnight, near-idle running. It won’t capture every busy daytime peak, and one night is not a full duty cycle. But idle is the state these machines sit in most of the time, and it’s the state the whole estimate rested on. Time-weighted over the window, the plugs recorded:

Two things stand out. The old pair came in at ~297 W — within about one percent of the 300 W I’d guessed, which is a closer hit than a back-of-envelope number has any right to be.

Some of that tight fit is aggregation, with an honest wrinkle behind it. chimaera was up under its usual load and pulled ~161 W. basilisk I’d already powered down for good as part of the downsizing; I only spun it up briefly — idle and all but empty — to take this reading, where it drew ~136 W. Carrying its old share of VMs it would have drawn more, so if anything ~297 W understates what the pair pulled when both ran around the clock. The pair was the number I actually estimated, and the meter — even with one box under-loaded — lands within a percent of it.

The other standout: the mini PCs drew roughly half my deliberately conservative estimate: about 19 watts for the pair, idle.

So the measured gap is ~278 watts, if anything a touch wider than the ~265 W I estimated — the savings come out slightly better than I claimed, because the new hardware sips even less than I feared. Annualised the same way I annualised the estimates (24/7), that’s about 2,400 kWh and $640 a year, right in the neighbourhood of the ~2,300 kWh / ~$610 estimate. And the ratio is starker than the “tenfold” above: at idle the old pair draws about 16× what the new pair does.

The caveat I keep flagging still applies. This is a short, idle window, not a year of mixed load, so read the annual figures as the estimate confirmed rather than a fresh measurement in their own right. But the direction and the rough magnitude I argued from estimates now have a real meter behind them — and the meter agrees.

The part that nags at me: the AI

Power was the goal, so it’s worth turning the same lens on the rebuild itself.

I ran this whole project as a collaboration with Claude Code, and I’ll be honest: the energy cost of large AI models bothers me. It’s one of my bigger reservations about using them at all. It would be a poor result to save electricity on the hardware and quietly spend it back — and worse — talking to a datacenter. So I tried to bound it.

From my own usage logs for this project, the rebuild took roughly 16,000 conversational turns and about 1.3 billion tokens — though around 94% of those were cached reads rather than freshly computed, with on the order of 18 million tokens of generated output. Those counts are real. Turning them into energy is where the estimating starts, and it gets shaky fast:

• No vendor I use publishes a real per-query energy figure for the model I’m running, so I can’t compute this directly.
• The closest public numbers are for other models. Google has put a median text prompt to its Gemini assistant at about 0.24 Wh.⁴ Epoch AI estimated a typical ChatGPT query at roughly 0.3 Wh.⁵
• The model I used is larger than those and often ran with a long context, so I scaled up generously — call it 1 to 6 Wh of useful work per turn, meant to cover the whole turn (generating output and prefilling the ~6% of tokens that weren’t served from cache), not the output alone.

That puts the inference energy for the entire rebuild somewhere in the range of 15 to 100 kWh — tens of kilowatt-hours, with wide error bars. At my own electricity rate that would be about $4 to $26 of power, though the datacenter pays a very different rate than I do, so the kilowatt-hours are the honest unit of comparison, not the dollars.

The comparison that matters:

• The old servers, left running, burn ~2,300 kWh every year.
• The whole rebuild’s worth of AI is a one-time ~15–100 kWh.

So the energy I spent talking to the AI is paid back by switching off the old hardware in somewhere between a few days and about two weeks of runtime, and it amounts to under 5% of a single year’s electricity saving. On those numbers, the AI’s energy is a rounding error against what the rebuild saves.

Two honest caveats keep me from being smug about that. First, this counts only inference — the energy to run the model — and excludes the very large, shared cost of training it, which I have no way to attribute to my slice of usage. Second, “it pays for itself” is an argument about this project, where the AI helped retire a genuinely wasteful piece of hardware. It is not a general licence; the same tool pointed at make-work would just be more load on the grid. The thing that makes the maths work here is that there was real waste to remove.

What I’d do differently

The fix was metering, and I did get there in the end — just later than I should have. Smart plugs on all four boxes cost a few tens of dollars and took an evening to wire into Home Assistant. Had I put them in place months ago, I’d have a clean, continuous before-and-after at the socket — including a full duty cycle with daytime load — instead of a five-hour idle snapshot bolted on at the last minute. The lesson isn’t “measure”; I knew that. It’s measure before you think you need to, because the cheapest moment to capture the “before” is while the old thing is still running, and that window closes the day you power it down.

The plugs are staying on. By the time chimaera finally powers off for good, I’ll have the continuous, full-duty-cycle trace I wish I’d started with — and the numbers will be there whether or not they’re worth a word more.

So: the rebuild saves power — and now I can show it, not just argue it. The estimate and the meter agree to within a percent on the old pair, the new pair draws even less than I’d hoped, and the AI that helped do the work cost a small fraction of a single year’s saving to run.

AI-assistant disclaimer: This post was drafted by Claude Code (Claude Opus 4.8, claude-opus-4-8) from my homelab session notes, commit history, usage logs, and architecture decision records, then reviewed and edited by me. The server power figures are bounded estimates, spot-checked against smart-plug measurements over a short idle window; the AI-energy figures lean on third-party proxies. Verify specifics before relying on them.