What does it cost to run an LLM at home? Electricity calculator with EIA prices

How the cost is calculated

power while generating = device watts x devices + rest of the system
power while idle       = idle watts x devices + rest of the system
energy per day (kWh)   = [busy hours x power while generating + idle hours x power while idle] / 1000
                         busy hours = hours on x utilization
cost per month         = energy per day x 365/12 days x price per kWh
per million tokens     = cost per month / generated tokens per month x 1,000,000
hardware per month     = what you paid / months you spread it over
API per month          = input tokens x input price + output tokens x output price
payback                = what you paid / (API per month - electricity per month)

A month here is 365/12 days, the same 730-hour month the break-even calculator uses.

Rated power is a ceiling, not your bill

Each device in the list carries the vendor's own power figure, under the vendor's own name for it. For NVIDIA GeForce cards that is Total Graphics Power (Graphics Card Power on the RTX 30 series), for AMD and Intel cards Total Board Power, and for workstation cards Max Power Consumption. Each is the most the card is designed to draw, not what it draws while generating text, which is mostly limited by memory bandwidth rather than compute.

Apple publishes a different kind of figure: the maximum power of the whole Mac, measured at the wall while running a compute-intensive test, for one configuration of each desktop Mac. That already includes everything, so the rest-of-system default for Macs is 0 W. AMD's Ryzen AI Max figure is a chip TDP that each computer maker sets between 45 and 120 W; the list uses the top of that range. NVIDIA lists 240 W for the whole DGX Spark.

Rest of the system is an assumption: 100 W for a desktop around a graphics card (CPU, board, memory, drives, fans and power-supply losses), 30 W around an APU whose figure covers only the chip, and 0 W when the vendor's figure is already the whole computer (Apple's desktop Macs, DGX Spark). A plug-in power meter at the wall replaces all of this with a measurement.

To measure, put the meter between the wall and your computer. Note the reading while a long answer is generating and while the machine sits idle, choose "Measured", and set the rest of the system to 0.

Electricity prices

The default is the U.S. average price of electricity to residential customers for July 2026, 18.31 cents per kWh, from the EIA's Electric Power Monthly, data for July 2026, released September 24, 2026 (next release October 23, 2026). EIA marks these values as preliminary estimates based on a cutoff model sample (Form EIA-861M). Pick your state to use its average instead. In July 2026 they ranged from 12.72 cents in Louisiana to 48.00 cents in Hawaii, so the same machine costs 3.8 times as much to run in one state as in the other.

These are averages: total residential revenue divided by kWh sold, so fixed customer charges are spread over every kWh. Your marginal rate, or a time-of-use rate, can be lower or higher. The rate on your bill is the better number, and the calculator takes it. The full state table is under the calculator.

A worked example

These are the calculator's defaults: an RTX 4090 at its 450 W rating plus 100 W for the rest of the system, on 8 hours a day and generating 25% of that time, 200,000 generated and 600,000 prompt tokens a day, at the U.S. average of 18.31 cents.

• Energy: 51.7 kWh a month, $9.47, or $1.56 per million generated tokens.
• The same tokens from an API: Groq GPT OSS 20B (openai/gpt-oss-20b) lists $0.075 per million input and $0.30 per million output tokens, $3.19 a month. gpt-oss-20b fits on a 24 GB card, so this compares the same model on both sides.
• Implied speed: those tokens over 2.0 busy hours a day are 27.8 tokens per second, under the RTX 4090's bandwidth ceiling of 392 tokens per second for this model at an 8k context in the Can I run it? calculator.

At these volumes the electricity alone costs more than the API, before counting the card. The same defaults on other hardware:

Rated power is an upper bound for each device, so these are upper bounds too, apart from the rest-of-system assumption. They are arithmetic, not measurements.

When running it at home is cheaper, and when it is not

What usually decides the comparison is the hardware: what it cost, spread over its useful life, against how many tokens you actually generate. An API bills only for tokens; the computer costs the same whether it generates a million tokens a day or none. Enter what you paid and the calculator shows the months to pay it back from the API bill you no longer pay, or says it never pays back.

There are good reasons to run models locally that this calculator does not price: your data stays on your machine, it works offline, nobody changes or retires the model under you, and you can run fine-tuned variants no API offers. Compare like with like: pick an API serving the same open model you run at home (the list puts those first), because a larger closed model is a different product.

What is not included

• The computer's idle draw outside the hours you enter, and power-supply efficiency beyond what the rest-of-system figure covers.
• Cooling: in summer, an air conditioner removes the heat the computer adds; in winter, that heat offsets some heating.
• A separate figure for prompt processing: it is folded into the busy hours you enter.
• Taxes, fees and tiered or time-of-use rates beyond the average price.
• The resale value of the hardware at the end of the period you spread it over.

Also available as Markdown.