How much electricity does a local LLM use? The real monthly cost

Before buying a GPU for local AI, the fair question is what it does to the power bill. We worked it out from the manufacturers' own power ratings and the US government's average electricity price, then checked it against one user's measured month. Short version: for how most people use a local model, electricity is a small line item. The hardware is the big one.

How much power does a GPU draw running a local LLM?

At most, its rated board power plus the rest of the PC. NVIDIA rates the RTX 3090 at 350 W, the RTX 4090 at 450 W and the RTX 5090 at 575 W; AMD rates the RX 7900 XTX at 355 W typical board power. Apple lists a Mac mini M4 at 4 W idle and 65 W max, and an M4 Pro at 5 W idle and 140 W max.

Two things pull real use below those figures. First, a GPU only draws heavily while it is generating; between prompts it drops to idle. Second, generating text one token at a time is usually limited by memory speed, not raw compute, so a card often runs below its rating. The rated figure is a ceiling, which makes the sums below a cautious estimate.

For a whole PC we add 150 W for the CPU, board, drives and fans. That's our assumption, not a measurement; your system may be lower or higher.

What it costs per month, card by card

All figures at 18.16 cents/kWh, the US average residential price for August 2026 in the EIA's Electric Power Monthly (preliminary, released October 23, 2026). Hours are hours of full-load generation, not hours the PC is switched on.

The sum is simple enough to redo with your own numbers: watts ÷ 1,000 × hours × your price per kWh. Your bill shows your rate; it can be well above or below the national average. If yours is double, double the table.

What one user measured over 30 days

The best real-world check this month came from r/LocalLLM. A user measured 30 days of local inference on an RX 7900 XTX at the wall: 30.20 kWh, which cost them 12.38 CHF (about $14.05) at Swiss rates, for 964 million input and 13 million output tokens. They priced the same token counts on an equivalent API, qwen3.8-flash at $0.15 per million input and $0.47 per million output tokens, at $150.71.

Two things stand out. 30.20 kWh over 30 days is an average of about 42 W, far below the card's 355 W rating, because most of the month was idle time and prompt reading, not flat-out generation. And at the US average rate the same 30.20 kWh would cost about $5.48.

The replies raised the right objections. The comparison leaves out the price of the card (and whether it loses or gains value), and the time spent setting it up. Several pointed out that a fair test needs the same quantization the API provider runs, since a smaller local quant is cheaper to run but not the same model. Treat it as one person's month, not a benchmark, but it matches what the ratings predict.

When does electricity matter?

Mostly in three cases:

• You run it around the clock. A batch job or an always-busy API server at full load is the $65 to $95 a month row, more than many cloud subscriptions.
• Your rate is high. At 40 or 50 cents per kWh, every figure in the table more than doubles.
• You're choosing between hardware. Over three years of daily use, an RTX 5090 PC and a Mac mini M4 Pro differ by about $230 in electricity at two hours a day, and about $920 at eight hours; that can narrow the price gap between them.

For on-demand use (ask, read, ask again), the GPU sits idle most of the time and the monthly cost is a few dollars. Capping the card's power limit lowers the full-load figure further; see how to power-limit your GPU for local AI.

Is electricity the cost that decides it?

Rarely. The card, and the RAM to go with it, cost far more than a year of power for most people. Whether local AI pays off depends on how much you use it and how much privacy matters to you, which we cover in is running local AI worth it?. If you only need a big GPU now and then, compare renting against buying, and use cost to run to see what a specific model needs.