One of the arguments for local AI is that once you own the hardware, it's free to run. It isn't quite free: your graphics card draws power while it generates text or images. The good news is that the cost is easy to work out. Forge PC hasn't measured any system for this guide, so the examples below use manufacturers' rated power figures and Ofgem's published price cap rate, and they're our own arithmetic, not measurements.
Electricity is billed in kilowatt-hours (kWh). One kWh is 1,000 watts for one hour.
Ofgem says that from 1 October to 31 December 2026, someone on a standard variable tariff paying by Direct Debit pays on average 26.32p per kWh for electricity, averaged across England, Scotland and Wales. There's no VAT on electricity from 1 October 2026 to 31 March 2027. Your own rate may differ, so check your bill. Ofgem will announce the cap for 1 January to 31 March 2027 on 25 November 2026.
The daily standing charge (54.83p a day on average for electricity, says Ofgem) is a fixed charge per day, so leave it out of this calculation.
Worked examples: an hour at full rated GPU power
These use each card's official maximum board power, so they're a worst case for the graphics card alone. A card running a chatbot may sit idle between prompts rather than at its limit the whole time.
Graphics card | Rated power (manufacturer) | Energy for 1 hour | Cost at 26.32p/kWh |
|---|
GeForce RTX 5090 | 575W total graphics power | 0.575kWh | about 15.1p |
GeForce RTX 5070 Ti | 300W total graphics power | 0.3kWh | about 7.9p |
GeForce RTX 5070 | 250W total graphics power | 0.25kWh | about 6.6p |
Intel Arc B580 | 190W total board power | 0.19kWh | about 5.0p |
GeForce RTX 3060 | 170W graphics card power | 0.17kWh | about 4.5p |
Two longer examples:
Two hours a day for a month on a card drawing 250W: 250 x 60 hours / 1,000 = 15kWh, or about £3.95 at 26.32p.
An overnight image-generation batch of eight hours with the whole PC drawing 300W: 2.4kWh, or about 63p.
Your total will be higher than the GPU figure alone, because the CPU, motherboard, drives, fans and power supply losses all add to it.
Step 1: Measure what your GPU actually draws
On an NVIDIA card, open a terminal while your model is running and use nvidia-smi's power display:
nvidia-smi -q -d POWER -l 1
-q -d POWER shows the power readings and -l 1 repeats them every second until you press Ctrl+C. NVIDIA's documentation says the "Average Power Draw" reading is the average for the whole board over the last second, supported on Ampere or newer cards; "Instantaneous Power Draw" is the last measured reading. Note the figure while text or images are generating, and while the card is idle with the model loaded.
For other graphics cards, check the power readings in a monitoring tool. See our guide to monitoring PC temperatures and performance with HWiNFO and Afterburner.
Step 2: Measure the whole PC
For the most accurate figure, use a plug-in energy monitor between the wall socket and your PC. It shows the whole system's draw, including power supply losses. Note the reading while your AI task runs.
Step 3: Estimate your hours and do the sum
Be honest about how long the GPU is actually busy. A chat session might keep the card working for a few minutes in an hour; a batch of images or a long video job can keep it busy for hours. Multiply your measured watts by those busy hours, divide by 1,000 and multiply by your unit rate.
How to cut the cost
If you're weighing this against a subscription, our cloud AI vs local LLM PC comparison covers the other costs and trade-offs.
Our verdict
Multiply watts by hours, divide by 1,000 and multiply by your unit rate: at Ofgem's average of 26.32p per kWh, even an RTX 5090 at its full 575W rating costs about 15p an hour. Measure your own PC with nvidia-smi or a plug-in energy monitor for a real figure, and remember the cap changes on 1 January 2027.