Electricity cost of running an LLM at home in Europe: per month and per million tokens, 2026

At the EU household average of EUR 0.2832 per kWh for the first half of 2026 (Eurostat, released October 2026), generating one million tokens with gpt-oss-20b for one user costs about EUR 0.11 in electricity on a DGX Spark, EUR 0.14 on an RTX 4090 PC and EUR 0.15 on a Mac Studio M3 Ultra. Four hours of generation a day adds EUR 4 to EUR 18 a month, depending on the machine. In Germany and Ireland those figures are 28% and 37% higher; in Hungary they are 60% lower.

• EU average price, H1 2026: EUR 0.2832/kWh; highest Ireland EUR 0.3871, lowest Hungary EUR 0.1139.
• Energy per million tokens (gpt-oss-20b, one user): Spark 0.39 kWh, RTX 4090 PC 0.51 kWh, M3 Ultra 0.55 kWh.
• Power during generation: Spark 122.8 W, M3 Ultra 227 W, RTX 4090 PC 509.1 W (peak, whole system).
• Idle: Mac Studio 6 to 9 W, Spark 40 to 45 W at launch software, RTX 5090 card 31 W, RTX 4090 card 21 W.

What does electricity cost in each EU country in 2026?

Between EUR 0.11 and EUR 0.39 per kWh for households, a 3.4x spread. Eurostat's latest bi-annual release covers January to June 2026 for medium households using 2,500 to 5,000 kWh a year, with all taxes and levies included; its Statistics Explained article gives the EU weighted average as EUR 0.2832. I pulled the country figures from the nrg_pc_204 dataset (updated 19 October 2026) and multiplied them by the energy per million tokens worked out in the next sections.

Eurostat nrg_pc_204, band 2,500 to 4,999 kWh, all taxes included, read 7 November 2026. Per-token columns are my calculation: price times 0.3878, 0.5057 and 0.5469 kWh per million tokens, derived below from Jeff Geerling's measurements of gpt-oss-20b. Your own tariff, not the national average, is what to multiply.

The ranking moves more than people expect from one half-year to the next. Denmark fell from EUR 0.3312 to EUR 0.2375, Czechia from EUR 0.3217 to EUR 0.2884, while Finland and Sweden rose by four and three cents.

How much power does a local LLM machine draw?

From under 10 W idle on a Mac Studio to over 500 W at the wall for a desktop with an RTX 4090 while generating. The most useful single source is Jeff Geerling's ai-benchmarks repository, because it runs the same model file (gpt-oss-20b Q4_K_M) through llama-bench on each machine and records power; its README calls the figure "Power (Peak)" and the issues do not say how it was metered.

"Peak" figures are the highest reading during a llama-bench run that includes prompt processing, so they overstate steady single-user generation somewhat. GPU idle figures are card-only, measured by ComputerBase on a Windows desktop at 3,840 x 2,160 and 60 Hz; the rest of the PC adds its own idle draw, which none of these sources measured. Spark idle fell after NVIDIA's January 2026 update; the DGX Spark power guide has those later figures and is not repeated here.

What does a million tokens cost in electricity?

About 11 to 16 euro cents at the EU average for one user, by this formula: kWh per million tokens = watts x (1,000,000 / tok/s) / 3,600,000. The steps for each machine with a measured pair of watts and tokens per second:

• DGX Spark: 1,000,000 / 87.97 = 11,368 s. 122.8 W x 11,368 s = 1.396 MJ = 0.388 kWh. At EUR 0.2832: EUR 0.110.
• RTX 4090 PC: 1,000,000 / 279.65 = 3,576 s. 509.1 W x 3,576 s = 0.506 kWh: EUR 0.143.
• Mac Studio M3 Ultra: 1,000,000 / 115.29 = 8,674 s. 227 W x 8,674 s = 0.547 kWh: EUR 0.155.

For the RTX 5090 and 3090 I only have a ceiling. If the card alone drew its full rating while generating at the llama.cpp guide's speed, that is 575 W x 3,540 s = 0.565 kWh (EUR 0.160) for the 5090 and 350 W x 6,182 s = 0.601 kWh (EUR 0.170) for the 3090, before the rest of the PC. Single-user decode is limited mostly by memory bandwidth, so I would expect the card to sit below its rating while generating; the RTX 5090 power limit article covers the measured trade-off between watts and tokens.

The fast GPUs and the slow, frugal boxes end up close per token, because the RTX 4090 draws about four times the Spark's power and generates about three times faster. The Mac's low idle draw does not help here; this is energy while working.

What does it cost per month?

Single-digit euros for most home use, and up to about EUR 150 a month for a 5090 at its rating around the clock in Germany. The table splits idle from generation, since most local machines spend most hours waiting. Add the two columns that match your habits.

My calculation. Idle month = watts x 24 h x 30.42 days; 4 h a day = watts x 4 h x 30.42 days; 24/7 = watts x 730 h. Prices EUR 0.2832 (EU), 0.3618 (Germany) and 0.1139 (Hungary) per kWh, Eurostat H1 2026. Spark idle uses ServeTheHome's launch-software figure; later firmware idles lower.

A Spark left on all month and used four hours a day comes to EUR 11.12 at the EU average: 20 idle hours at 40 W and 4 working hours at 122.8 W, every day. For a desktop PC the idle figure is the card's 21 to 31 W plus the CPU, board, drives and fans, which none of my sources measured; my inference is that it sits above the Spark's 40 W, so the PC costs more while doing nothing.

Does serving several users at once change the cost per token?

Yes, by a large factor, because power rises much less than throughput when you batch requests. On one Spark, GPUwerk measured Qwen3-Coder 30B-A3B at 80.9 tok/s for one request and 1,078 generated tok/s across 256 concurrent requests (vLLM 0.27.1, 2 September 2026, on the benchmarks page). GPUwerk did not meter power in that run. If the Spark drew 200 W while doing it, ServeTheHome's "just under 200W" for combined CPU and GPU load, then a million tokens took 928 s and 0.052 kWh: EUR 0.015 at the EU average, about a seventh of the single-user figure above. That is an estimate under a stated assumption, not a measurement.

The practical version: electricity per token is a small cost for one person and a negligible one for a busy shared server. Capacity, not power, is what limits a home machine.

Is running at home cheaper than renting or an API?

On electricity alone, home wins against renting and roughly ties with the cheapest APIs. Generating 24/7 on a Spark costs about EUR 25 a month in power at the EU average and EUR 32 in Germany, against $576.70 for a 730-hour month of a rented Spark at $0.79 an hour, which includes power and cooling. The gap is the hardware: a 128 GB Spark lists at $6,950, so the rent or buy break-even is about utilization, not the power bill.

Against an API the comparison is closer than people assume. EUR 0.11 to 0.16 of electricity per million tokens at single-user speed is in the range of the whole output price for gpt-oss-20b at budget providers ($0.14 at DeepInfra, $0.30 at Groq on 7 November 2026), before you have paid anything for the hardware. The open model API price comparison has the full list.

My recommendation: if your tariff is above EUR 0.30 per kWh, prefer a machine that idles low, because idle hours outnumber working hours for almost everyone. A Mac Studio idles at 9 W, a Spark at 40 W or less, and a big-GPU PC at more than its card's 31 W once the CPU, board and fans are counted. The DGX Spark vs Mac Studio comparison covers the other differences between those two. If you run batch jobs all day, the fast GPU's higher draw is repaid in tokens. And if the machine sits in a bedroom, the noise figures may matter more than the euros.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• Electricity price statistics, Eurostat Statistics Explained, 2026-10
• Electricity prices for household consumers, bi-annual data (nrg_pc_204), Eurostat, 2026-10-19
• AI/LLM Benchmarks (llama.cpp and Ollama), Jeff Geerling, GitHub
• Benchmark Intel 265K / Pi CM5 with Nvidia RTX 4090 (issue #43), Jeff Geerling, GitHub, 2025-12-11
• Benchmark Dell Pro Max with GB10 (issue #34), Jeff Geerling, GitHub, 2025-11-22
• Benchmark M3 Ultra Mac Studio 512GB (issue #42), Jeff Geerling, GitHub, 2025-12-11
• guide: running gpt-oss with llama.cpp, ggml-org/llama.cpp discussion #15396, 2025-08-18
• DGX Spark benchmark results, ggml-org/llama.cpp, 2026-02-05
• NVIDIA DGX Spark review, page 4 (power), ServeTheHome, 2025-10-14
• Mac Studio power consumption and thermal output, Apple Support
• Nvidia GeForce RTX 5090 im Test: Leistungsaufnahme und Effizienz, ComputerBase, 2025-01-23
• Nvidia GeForce RTX 5090 Founders Edition review, Tom's Hardware
• GeForce RTX 5090, NVIDIA