Energy and compute

What your machine actually costs to run

A £15 plug meter settles the energy question for your own hardware in an afternoon: idle draw, load draw, watt-hours per task, and what that comes to per year.

2026-07-07 · 4 min read

Almost everything written about the energy cost of computing is an argument with a citation attached. Very little of it contains a figure the writer measured on hardware they own, which is odd, because the equipment to take that figure costs about £15 and the measurement takes an afternoon.

The kit

A plug-in energy monitor. The cheap ones report instantaneous watts and cumulative kWh, which is everything you need. Anything that logs over time is a luxury; you can write numbers down.

You are taking four readings.

Reading 1: idle

Machine on, logged in, nothing running. Wait two minutes for it to settle — the number drops for a while after boot as background jobs finish.

Write it down. On a desktop with a discrete GPU this is commonly 60–90W. On a laptop, 8–15W. On a Raspberry Pi, 3–5W.

This reading is the important one, because idle is what your machine does for most of its existence. A home server drawing 70W at idle costs, at 28p/kWh:

70W × 24h × 365 ÷ 1000 = 613 kWh/year × £0.28 = £172/year

That is a number worth knowing before you decide to leave something on permanently, and it is independent of what the machine is doing. Most of the cost of a home server is the cost of it being switched on.

Reading 2: under load

Run the actual thing — your inference job, your build, your video encode — and watch the meter during it. Take the sustained figure, not the spike.

The delta between this and idle is what the work costs. If idle is 70W and load is 310W, the work draws 240W above baseline.

Reading 3: watt-hours per task

Now convert to something comparable:

Wh per task = (load watts − idle watts) × seconds ÷ 3600

A local generation that takes 40 seconds at 240W above idle:

240 × 40 ÷ 3600 = 2.7 Wh

At 28p/kWh that is £0.00075 — three quarters of a tenth of a penny. Do a thousand of them a day and it is 76p. Do a thousand a day for a year and it is £276, which is the point at which it stops being a rounding error and starts being a line item.

This is also the number that makes local-versus-hosted comparisons honest. The hosted equivalent has a price per token you can look up; the local equivalent has an electricity cost plus the amortisation of the card. Now both sides are in pence and the comparison is arithmetic rather than ideology.

Reading 4: grams of CO₂

Electricity is not equally clean at all hours. In the UK, carbonintensity.org.uk publishes grams of CO₂ per kWh at half-hour resolution, free and without a key:

curl -s https://api.carbonintensity.org.uk/intensity | jq '.data[0].intensity.actual'

That figure swings between roughly 50 and 350 gCO₂/kWh depending on wind and demand — a factor of seven. So:

gCO₂ per task = Wh per task ÷ 1000 × current intensity
2.7 Wh at 180 gCO₂/kWh = 0.49 g

Half a gram. Roughly a metre of driving. That is not nothing across a million requests, and it is also not the thing to feel guilty about while a 70W idle draw runs unmeasured for a year.

The factor-of-seven swing is the genuinely actionable part: a batch job that is not urgent can run at 3am when the grid is windy and cost a fifth of the carbon for identical work. That is a cron schedule, not a sacrifice.

What this measurement does not tell you

Be careful with the conclusions, because this is exactly where energy writing goes wrong:

• It is your machine only. It says nothing about what training a frontier model cost, or what a data centre's PUE is. Those are real questions with different methods.
• The meter reads at the wall. That includes PSU inefficiency, which is correct for billing and wrong if you are trying to compare chip efficiency.
• One task is not a workload. Measure your median task and your worst one; the ratio between them is frequently the surprise.

What it does give you is a set of figures nobody can argue you out of, about the hardware in front of you. In practice the three that change behaviour are: idle costs more per year than you think, the grid swings by 7× across a day, and a single local inference costs a fraction of a penny.

Where the energy actually goes takes each of those further — including what happens to the arithmetic when the machine is a laptop on battery, and how to compare a local run against a hosted one without cheating on either side.