The whole subject reduces to one number, and once you have it you can price any piece of equipment in your head in about two seconds.
A watt left running for a year uses 8.77 kWh. There are 8,766 hours in an average year, a watt is a thousandth of a kilowatt, and that is the entire derivation. So at 30 cents per kWh, every continuous watt on your shelf costs $2.63 a year.
That is the trick. Not efficiency curves, not benchmark scores. A number in watts, multiplied by a number in dollars.
| Your tariff | One continuous watt, per year |
|---|---|
| $0.10/kWh | $0.88 |
| $0.15/kWh | $1.31 |
| $0.20/kWh | $1.75 |
| $0.30/kWh | $2.63 |
| $0.45/kWh | $3.94 |
So a device that idles at 20 W costs somewhere between $18 and $79 a year to leave switched on, and which end you land on has nothing to do with the device.
Standing draw is the bill
Here is the arithmetic that changes how people buy equipment. Take a machine drawing 40 W at rest and 300 W flat out, at 30 cents per kWh, and vary only how many hours a day it actually works:
| Hours under load per day | Annual cost | Share that is idle |
|---|---|---|
| 0 | $105.19 | 100% |
| 0.5 | $119.44 | 86% |
| 1 | $133.68 | 75% |
| 2 | $162.17 | 59% |
| 4 | $219.15 | 40% |
| 8 | $333.11 | 21% |
| 24 | $788.94 | 0% |
Most home machines live in the top three rows. At an hour a day of real work, three quarters of the annual cost is the machine sitting there doing nothing, and that portion is untouched by anything you do to make the workload faster.
The same thing seen from the other side. Hold the workload at two hours a day and change only the idle figure:
| Idle draw | Annual cost |
|---|---|
| 80 W | $258.60 |
| 60 W | $210.38 |
| 40 W | $162.17 |
| 20 W | $113.96 |
| 10 W | $89.85 |
Dropping 80 W of idle to 20 W saves $145 a year at that tariff. There is no optimisation of the two working hours that comes close, because those two hours are only $65 of the total to begin with.
Buy for idle draw. Optimise for idle draw. Idle is what you are paying for.
Where the watts hide
They are rarely in the machine you were thinking about. Everything below is drawing power right now, whether or not anything is being asked of it:
- Spinning drives. Each one has a standing cost, and an array multiplies it by the drive count. The NAS capacity calculator shows what a given array costs to keep standing, alongside what it stores. Going from four drives to eight is not just more money at purchase, and which RAID level you pick decides how many of those drives you are running for redundancy rather than storage.
- Networking. A switch, a router, an access point, a fibre or cable modem. Individually small, permanently on, and there are usually four of them.
- The UPS itself. The conversion is not free and the unit has its own overhead, which is a real cost of the protection rather than an argument against it. Size it with the UPS calculator and read the answer in volt-amps, because that is the unit the label uses. Working out which size you need turns on that unit rather than on runtime, which is not where most advice starts.
- Anything with a clock or a standby light. Small, and there are a lot of them.
Add a modest set of always-on equipment together and 97 W of standing draw is easy to reach. That is 850 kWh and $255 a year at 30 cents, or $85 at 10 cents. It is not a large bill, but it arrives every year and nobody ever decided to spend it.
Nameplate ratings are not draw
Do not build any of this from the numbers printed on the equipment. A power supply rated at 750 W is telling you what it can deliver, not what it takes. A processor’s TDP is a thermal design figure. Both will be wildly above what your machine actually pulls at rest, and using them turns this arithmetic into fiction.
Measure at the wall with a plug meter. It is the cheapest piece of test equipment in the lab and it is the only input here that matters. Two readings are enough: the machine at rest with nothing running, and the machine doing the work you actually do.
Two details worth knowing when you read one. Your meter shows watts, real power, and that is what the utility bills you for; the volt-amp figure on a UPS label is apparent power and it is a larger number for the same load, related by the power factor. And a machine’s draw at rest depends on what is plugged into it and how the firmware is configured, so a figure from someone else’s identical build is a hint rather than an answer.
Energy per token, which is the useful efficiency number
If you run models, the honest efficiency measure is not watts. It is energy per unit of work. At a 300 W load, holding everything else constant:
| Generation speed | Energy per million tokens | Cost at $0.30/kWh |
|---|---|---|
| 10 tok/s | 8.33 kWh | $2.50 |
| 20 tok/s | 4.17 kWh | $1.25 |
| 40 tok/s | 2.08 kWh | $0.63 |
| 60 tok/s | 1.39 kWh | $0.42 |
| 120 tok/s | 0.69 kWh | $0.21 |
A machine twice as fast at the same draw uses half the energy for the same output, because it finishes and goes back to idle. This is why raw wattage is a poor way to compare machines that do work, and why the idle argument above and this one do not conflict: idle dominates the bill, and among the hours that are not idle, speed is efficiency.
I have not measured tokens per second on any machine yet, so that column is arithmetic waiting for inputs. When the first runs are published to the benchmark database, real generation speeds go into this table and it stops being hypothetical.
The levers, in order of how much they move
- Turn things off. An unused machine at 0 W costs nothing, and this beats every other item on this list combined. Wake-on-LAN exists.
- Consolidate. Three machines at 25 W idle cost more than one at 45 W doing all three jobs. Fewer power supplies, fewer standing loads.
- Lower the idle figure. Fewer drives, spin-down where the workload tolerates it, unplugging cards nothing uses, letting the firmware actually reach its low-power states.
- Shift to a cheaper window. If you are on a time-of-use tariff, moving heavy jobs into the cheap hours cuts the load portion of the bill. It does nothing for the idle portion, which is spread across all hours.
- Make the work faster. Real, and last, because it only touches the hours the machine is busy.
Put your own two readings into the power cost calculator and it does all of the above at once: annual kWh, annual cost, the idle share, cost per hour of work, and energy per million tokens.
What this does not count
- Cooling. Every watt that goes in comes out as heat. In a cooled room in summer you pay again to remove it, at a ratio depending on the air conditioner. In winter, in a heated house, the same heat displaces heating and part of the cost comes back. Neither effect is in any figure above.
- Tariff structure. I have used a flat unit rate. Standing charges do not change when you add a server, so leave them out; tiered rates mean the lab may be billed at your highest tier rather than your average, which makes the true cost higher than this arithmetic suggests.
- Taxes. Whether your unit rate includes them varies by country. Use the rate that appears on the bill you actually pay.
- The hardware itself. Purchase price is a separate calculation, and if you are weighing a machine against paying an API for the same work, the electricity here is only one input to that decision.
FAQ
How much does a home server cost to run per month?
Multiply its standing draw in watts by 0.73 to get the monthly kWh, then by your unit rate. A 20 W device is about 14.6 kWh a month, so roughly $4.38 at 30 cents per kWh or $1.46 at 10 cents. For a machine that also does heavy work, split it: idle watts for the hours it rests, load watts for the hours it works.
Is it cheaper to leave a server on or turn it off between uses?
Off, always, and usually by more than people expect. A machine idling at 40 W costs $105 a year to leave switched on before it does a single unit of work. If you use it an hour a day, that is $105 spent on the twenty-three hours you did not use it. The counter-argument is convenience rather than cost, and wake-on-LAN answers most of it.
Do hard drives use much power when idle?
Enough that the drive count is a real line in the annual bill, which is why the NAS calculator reports standing watts next to usable capacity. Spin-down helps if your workload tolerates it, and it is not free: drives take time to spin back up, so anything touching the array frequently will keep it awake anyway and add start cycles for nothing.
What is a reasonable idle draw to aim for?
That depends entirely on what the machine has to do, so I would frame it as a budget rather than a target. Decide what you are willing to pay per year, at your own tariff, and divide by the per-watt figure in the first table. At 30 cents, a $50 annual budget buys about 19 W of permanent draw. Then spend that budget on the equipment that earns it.
Does a UPS increase my electricity bill?
Yes, by its own conversion losses and standby overhead, continuously. Treat it as the running cost of the protection: on a lab drawing 97 W the addition is tens of dollars a year, against the cost of a corrupted array. If it is oversized for the load it is protecting, some of that is waste, which is the argument for sizing it properly rather than for skipping it.
Why is my measured usage higher than this calculation?
Usually one of three things. The idle figure came from a specification rather than a meter. The machine is not reaching its low-power states, which firmware settings and an attached device can both prevent. Or cooling is paying twice for the same watts, which none of the arithmetic here includes.
Sources
- The kWh conversion is definitional: 8,766 hours in an average year, being 365.25 days, times 24. Every annual figure on this page is watts times 8.766 divided by 1,000, times your unit rate.
- Wattage figures are illustrative inputs, not measurements of any product. Nothing on this page was measured on hardware. Take yours from a plug meter, because a power supply rating is a maximum it can deliver rather than a draw.
- Tariffs vary by country, supplier and time of day. Use the unit rate on your own bill rather than a national average.
- The arithmetic is the same code that runs the power cost calculator and the NAS capacity calculator. When benchmark runs are published, the energy-per-token table gains measured generation speeds and this article’s updated date moves.