What a local model actually costs to run
Short answer
The cash cost is electricity and disk. A 12 GB card drawing around 170 W under load for an hour a day is a few dollars a year in most places — far less than a subscription, and paid in battery rather than currency on a laptop.
Electricity
A mid-range desktop GPU draws on the order of 150–200 W while generating and almost nothing between questions, because inference is bursty: a few seconds per answer, idle the rest of the time.
At an hour of actual generation a day — which is a lot of questions — that is roughly 60 kWh a year, in the region of ten to twenty dollars depending where you live. Realistic use is a fraction of that.
Disk
0.75 GB to 30.4 GB per model. Keeping two or three of the everyday ones is 10 GB; a large one is 16 to 30 GB on its own. Disk is cheap and this is a one-off.
Battery, which is the real cost on a laptop
This is where local inference is genuinely expensive. Sustained GPU load drains a laptop battery quickly, and a long document session is noticeable in a way that browsing is not.
The practical pattern is a small model on battery and a large one on mains. Qwen 3.5 2B is 1.2 GB, scored full marks on our evaluation, and costs far less power than the 12B.
Against a subscription
A hosted assistant is typically $20 a month, which is $240 a year, with a usage cap and an ongoing transfer of your documents. Local is a few dollars of electricity and a weaker model.
Which is the better deal depends entirely on whether your questions need frontier capability. For reading documents in front of you, mostly they do not.
Other ways people ask this
- If I run the model myself, am I paying for electricity instead of a subscription?
Related
- Up a levelGuides: local AI, browser privacy and the hardware to run it
- Closely relatedRunning a local model on integrated graphics
- Closely relatedWhat a local model does to your battery
- Closely relatedRunning a local LLM without a graphics card
- Closely relatedCan you run a local LLM on an NPU?
- Elsewhere on the siteCan your PC run a local AI model?
- Elsewhere on the siteHardware guide for local AI