Meter readings: an always-on RTX 3090 local AI stack costs €2 a month in electricity

Metering the boring question

Debates about running AI models at home usually revolve around tokens per second and VRAM budgets. A post on dev.to, part of a series the author calls The Local LLM Lab, took a different angle: what does the electricity bill actually say? The author instrumented an NVIDIA RTX 3090 running an always-on AI stack and recorded 30 days of power draw, priced against a dual-rate Bulgarian tariff of 0.30 BGN/kWh by day and 0.18 BGN/kWh at night (22:00–06:00). The total was €2.00 for the month, measured at the card rather than estimated.

What was on the GPU

Three services lived permanently on the 24 GB card:

• WhisperX, a speech-to-text endpoint that sits resident waiting for voice memos, meetings and podcast clips — 7.9 GB of VRAM on average, 10.3 GB at peak
• nomic-embed-text under Ollama, feeding a retrieval pipeline — 308 MB of VRAM
• Immich's machine-learning component — negligible

A separate machine with two more 3090s runs a 27-billion-parameter chat model through vLLM on demand; that box is outside the €2 figure.

The numbers

According to the dev.to post, the Whisper service averaged 22 W and consumed 13.24 kWh (€1.76), the embedding model averaged 3 W and consumed 1.75 kWh (€0.23), and Immich effectively nothing — about 25 W average and 15.03 kWh overall. Although the card is power-limited at 260 W, it rarely came close. The 30-day trace shows a roughly 35 W baseline with the models held in VRAM, brief spikes near 120 W when audio arrives, stretches of 50–54 W when a transcription queue builds, and a single two-hour block averaging 121 W from a batch of long files. In other words, a typical home inference stack is an appliance that is almost always idle and briefly busy — not a card under sustained load.

How it compares

The author frames the alternatives: a 3090 pegged at full utilisation around the clock — a training rig, not an inference box — would cost roughly €23 a month at the same tariff. Equivalent cloud spend is estimated at €14–29 a month for about 10 hours of OpenAI Whisper API transcription, or around €14 for a million GPT-4o tokens at list price. Local comes in under a tenth of those figures, and its cost is flat rather than scaling with usage. The genuinely dominant recurring number is the hardware itself: a used 3090 at €800–1,200 amortises to about €25 a month over three years, more than ten times the electricity bill.

Limits of the measurement

The boundaries matter. The meter covers the GPU only, not the CPU, RAM or the rest of the system; the tariff is specific to Bulgaria; and the workload is light — a personal transcription and embedding pipeline rather than continuous text generation. The €23 figure for a saturated card shows how different the arithmetic becomes under sustained load. The author describes the €2 as a standby cost, with usage spikes too short to move the monthly total.

Why it matters

The cloud-versus-local argument often stalls on impressions rather than data. This is one of the few accounts with actual meter readings behind it, and it suggests electricity is a rounding error for a mostly idle home inference stack — a few euros a month — while hardware amortisation and maintenance time are the real costs. For developers who already own a GPU for gaming or rendering, the marginal cost of keeping resident local models is close to zero. The author has open-sourced the monitoring setup, a project called homelab-monitor that exposes its readings through a read-only MCP server, so the experiment can be reproduced on other machines. The answer to anyone objecting to local inference on power grounds is, on this evidence, that the meter disagrees.