The electricity bill for a home AI box

By Emmanuel Sauvage · Published · Updated

This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.

Leaving Ollama on (the app that runs language models on your machine) changes the electricity bill the way a NAS or a small media server does, except the spikes arrive when the GPU (graphics card) or the CPU (processor) is generating from a model. This guide estimates kilowatt-hours and euros for a household. It stops short of a promise to the cent: real watts depend on the case, the model that is loaded, and how you use it.

LocAIhome’s frame: households that host their own AI, and professional server rooms are outside this page. Always cross these figures with the 3-year comparison of Ollama and ChatGPT, where electricity sits next to the hardware purchase.

Why the watts belong in the budget before you buy

A cloud subscription folds hosting and energy into the monthly price. At home the computer is spread over several years, and the power bill returns every month. Leave average watts out and a project that looked paid back in 24 months can slip to 36 or more, especially if a gaming tower stays on around the clock for a few questions in the evening.

The point is a calm estimate. Many mini PC and Mac mini setups stay in a modest yearly band in October 2026, in the neighborhood of a TV box or an old fridge that seals poorly. The job is to pick a profile that matches noise and power of a mini PC left on and the VRAM your model needs.

From watts to euros, step by step

Step 1: estimate average power

Instant power jumps between idle and inference. For a yearly budget, use a realistic average over a week:

  1. Idle with Ollama installed: model unloaded or asleep, disk busy now and then.
  2. Idle with the model in memory: ready to answer, fans low.
  3. Inference: generating tokens, GPU or CPU busy.
  4. Short spikes: downloading a model, indexing, or another app (Docker, Open WebUI, and so on).

Write down a watt estimate for each state, then weight it by time. A simplified example: 15 W × 20 h + 80 W × 2 h + 120 W × 0.5 h ≈ (300 + 160 + 60) / 22.5 ≈ 23 W average across the day.

Step 2: convert to kilowatt-hours per year

The formula:

kWh/year = (average watts × 24 × 365) / 1,000

The shortcut 8.76 is handy: kWh/year ≈ average watts × 8.76.

Average watts kWh/year A sense of scale
10 W 87.6 kWh A mini PC that is almost always idle
20 W 175 kWh A typical mini PC, 24/7
40 W 350 kWh High idle, or a small NAS plus Ollama
80 W 701 kWh A tower that is often busy
150 W 1,314 kWh Heavy use, or an old desktop

Step 3: apply your price per kilowatt-hour

In France, the blue tariff (the regulated retail price of electricity) is a common public reference for households. Market contracts and the off-peak option change the effective price. In October 2026, many household budgets use a cautious band of about €0.22 to €0.28 per kWh, tax included, for a quick estimate. Read the “price per kWh” line on your bill. The grids move.

Yearly cost ≈ kWh/year × tax-inclusive price per kWh

Example with 20 W average and an indicative €0.25/kWh:

  • kWh/year ≈ 20 × 8.76 = 175 kWh
  • Cost ≈ 175 × 0.25 = €43.75/year (about €44)

That is the same order of magnitude as a streaming subscription: useful in a long comparison, and over ten years it adds up.

If your utility bills in dollars, or at a different euro rate, keep the 8.76 formula and substitute the unit price on your own statement. The worked examples on this page stay in the October 2026 euro band above.

Step 4: tighten it with a real measurement

The tables above are orders of magnitude. After the machine is built:

  • measure an outlet with a display of watts and kilowatt-hours for 7 days;
  • or compare your Linky curve before and after the install, with the rest of the home held steady;
  • if you change cases, cross-check choosing hardware for home AI.

Watt profiles: mini PC, Mac, GPU tower

Vendors rarely print “power draw under Ollama.” Treat these as indicative bands for October 2026, and check them on your own machine:

Hardware profile Idle / Ollama idle Typical inference Comment
N100 mini PC / low-TDP Ryzen 8–25 W 30–80 W A strong candidate for 24/7 in the living room
Mac mini M4 10–30 W 25–70 W Often quiet. See the dedicated page
Mini PC with a modest iGPU 15–35 W 40–100 W GPU, eGPU, and OCuLink
Tower with a 12–16 GB dedicated GPU 40–90 W idle 150–350 W peak Heavier models, a higher bill
Synology NAS plus Ollama 15–40 W plus the disks varies Add the disks that spin all day

As you move up in 32 vs 64 GB of RAM and in model size, idle can stay high if the model remains loaded in memory.

Note: the advertised TDP (thermal design power) is the chip’s indicative heat envelope. The wall reading is a different number. The power supply, fans, disks, and an eGPU (an external graphics card) add to it.

Around the clock, or a planned shutdown

Two habits:

A server that is always ready

What you gain: the family assistant, Home Assistant, or n8n can call Ollama without waiting for a boot. What you pay: the hours when nobody is chatting.

If the average stays 15–25 W, the yearly cost often stays under about €60 at an indicative €0.25/kWh. Compare that with the convenience.

Night shutdown, or an aggressive sleep

If you use it 2 hours a day on weekdays, an “equivalent” average near 5–10 W (a large share of the time at 0 W) changes the picture. The electricity slice of the 3-year cost can fall by half or more.

A compromise: stop Ollama at night with cron or systemd, unload models you are finished with (ollama stop), and let the mini PC sleep lightly if the BIOS allows it.

Scenario Active hours/day Indicative average watts Indicative kWh/year €/year at €0.25/kWh
24/7, low idle 24 18 W ~158 ~€40
24/7, high idle 24 45 W ~394 ~€99
2 h/day of inference 2 h at full load, off the rest of the time ~12 W equivalent ~105 ~€26
Desk, 8 h/day 8 h mixed, off at night ~20 W equivalent ~175 ~€44

Software, Docker, and loads that hide

A Docker stack of Ollama and Open WebUI often adds 5–15 W that never go away. LM Studio, ComfyUI, or Whisper running at the same time create their own spikes. Installing Ollama on a headless Linux box sometimes trims overhead compared with a Windows session that has a full desktop.

For images, see ComfyUI and FLUX locally. The power bill can exceed text chat alone.

A lower bill, with the use you wanted

  1. Smaller or quantized models put less load on the GPU. See choosing a first model.
  2. Unload models after use (ollama ps, then stop).
  3. Give it air. Heat that forces the fans up also pushes the watts up.
  4. Careful undervolting on a CPU or GPU that supports it (see the FAQ).
  5. A dedicated mini PC rather than a do-everything gaming tower. See the 2026 mini PC comparison.
  6. An SSD for the models: disks of that kind draw less than a spinning hard drive. See storage for models.

Privacy and the bill: local AI skips a cloud subscription. The household legal side is in privacy when the AI stays home.

Put electricity into the whole budget

Over 36 months, a machine at €800–1,200 plus about €120–180 of electricity (assuming 20–25 W average) stays in the same neighborhood as about €720 and up of cloud subscriptions alone for a family, and the computer is still yours afterward. Use the hardware selector to line up the purchase price and the watt profile.

To try this before a dedicated always-on box: a first local AI in 15 minutes or Ollama on Windows on a PC you already have. Measure the change on the bill before you buy a box that stays on.

Meters (optional)

To tighten the math, a smart plug that measures consumption separates the AI box from the rest of the home:

Check the current price (October 2026) before you buy. The ranges on this page do not replace your utility bill, whether that is EDF or another supplier.

Keep going

Seasons, and electric heat

In winter, the heat from a mini PC adds a little comfort in a small room. In summer, the same heat can push the air conditioning and raise the whole-home bill beyond the kilowatt-hours counted at the plug. Give it ventilation, keep ComfyUI sessions for cooler hours, and leave the case out of a closed TV cabinet. Overheating makes the hardware draw more power for the same Ollama reply.

In short: estimate an average wattage, multiply by 8.76, apply your tax-inclusive price per kilowatt-hour, then measure a real week. That is the most honest missing piece in the budget for a home AI server.

Common questions

How do watts become a yearly bill?

Multiply average power in watts by 8.76 (24 hours times 365 days, divided by 1,000) to get kilowatt-hours per year, then multiply by your tax-inclusive price per kilowatt-hour. Example: 20 W times 8.76 is about 175 kWh/year. At an indicative €0.25/kWh that is about €44 a year. Check the rate on your own bill.

Should the machine stay on all night, or shut down?

If you only use Ollama for 2 hours in the evening, turning it off the rest of the time cuts electricity use to about a tenth or less, in exchange for a restart and a possible reload of the model into RAM. For a family assistant that should answer anytime, leaving it on makes sense when idle power stays low, as on an efficient mini PC.

What is undervolting, and does a home AI box benefit?

Undervolting lowers the processor or GPU supply voltage a little while the clock speed stays stable, which cuts watts and heat and leaves the language model unchanged. It is a BIOS/UEFI setting or a vendor tool. Move in small steps, test stability under an Ollama load, and skip it on a one-and-only work PC that has no backup.

Does a Linky meter show what Ollama uses?

Linky, the French smart meter, records the whole home, not the mini PC’s outlet by itself. Turn off the big appliances, note the curve before and after you add the box, or use a plug-in meter on the AI machine for a typical week.

Does a tower with a graphics card cost a lot more to run than a mini PC?

During GPU inference a tower can hit 150–350 W at the peak, against an idle that is often 8–25 W for a mini PC. Left on all day with lots of idle time, the mini PC wins on the bill. If you generate text or images continuously, the tower draws more and can run heavier models. See the VRAM guide and the hardware comparisons.

Should I estimate with France’s regulated rate or with off-peak hours?

For a first estimate, the regulated retail rate (often called the blue tariff) at the tax-inclusive price per kilowatt-hour is a public reference. Indexed contracts, and off-peak versus peak hours, change the result. Use the unit price on your latest bill, the October 2026 one if that is what you have, rather than a national average.

Sources

Narrow it down with the hardware selector · French version of this section