Guide

What does local AI actually cost?

"Free" is the usual answer and it is not quite true. Running a model on your own machine costs electricity, disk and an hour of your first evening. Those are all real. They are also, put together, smaller than most people assume and in a different order than they expect — the power is noise, the disk is the bill.

Electricity: cents, not euros

The arithmetic is short enough to do in public. A machine draws some number of watts while it generates, and it generates some number of tokens a second. Divide one by the other and you have kilowatt hours per million tokens.

Priced at 37 cents per kilowatt hour, the German household average for 2026 reported by the BDEW. That is deliberately the high number: new contracts in July 2026 were nearer 24 cents, and every figure here would shrink by a third. Power draw and throughput are from published measurements of sustained inference on Apple silicon, not from your machine.

A million generated tokens is roughly 750,000 words. Heavy personal use — an agent running most days, long conversations, code — lands somewhere around two million tokens a month. So the electricity bill for running AI on your own hardware, hard, for a month, is between 16 cents and about 1.50 euros. It is genuinely not the cost worth thinking about.

The interesting row is the first one. A fanless MacBook Air is the cheapest machine per token and the slowest per hour, because efficiency and throughput are different questions. If you are optimising for the meter rather than the clock, a small model on a small machine wins outright.

Disk: the actual bill

Models are large files and you will end up with several, because different work wants different models. A realistic working set — a fast general model, a coding model, something tuned for the language you write in — comes to about 40 GB.

What that costs depends entirely on where you put it. Apple's first storage step buys 256 GB for 250 euros in the eurozone, which is close to a euro a gigabyte, so 40 GB of models is about 40 euros of built-in SSD. The same 40 GB on an external USB-C drive, at current prices for a terabyte, is about 3 euros.

So keep the library on an external SSD if your internal disk is tight — with one condition. Weights are memory-mapped, so the drive is read while the model loads and again whenever memory gets tight. A fast NVMe enclosure is fine. A USB-A stick or a spinning disk will make loading painful and paging catastrophic.

Time: one evening, once

The first real cost is the first download, and it is the one that decides whether people stay. A 10 GB model takes about 28 minutes on a 50 Mbit line, seven minutes on 200 Mbit, and a minute and a half on gigabit. After that it is a file on your disk that loads in seconds and never needs the network again.

The second time cost is choosing what to download, and it is worse than the waiting because it has no obvious end. This is why Slate reads your Mac's memory and recommends one model on first run instead of opening onto a list — the memory guide explains the arithmetic it uses, and the 8 GB page covers the machines where the answer is tightest.

Battery, which nobody mentions

Sustained generation pulls roughly 20 to 40 watts on a laptop. A MacBook Air battery holds about 53 watt hours. That is one and a half to two and a half hours of continuous generation from full, and it is why an agent left running on battery in a café is a different proposition from asking a few questions.

It is also an argument for the small model again. A 4B answering in four seconds costs a fraction of what a 14B answering in twenty seconds costs, and for a great deal of everyday text work you cannot tell them apart.

What you stop paying

No per-token metering, so a long document costs the same as a short one: nothing. No rate limit, so the expensive question and the twentieth question in a row are the same question. No seat, no account, no monthly renewal to forget about.

And one that only shows up later: the model cannot be retired out from under you. A hosted model you have built a habit or a workflow around gets deprecated on someone else's schedule. A file on your disk does what it did last year.

The comparison, done honestly

The going rate for a consumer AI subscription is 20 US dollars a month — both ChatGPT Plus and Claude Pro listed at that in July 2026 — which is roughly 240 dollars a year before VAT. Against that, local AI costs a couple of euros a year in electricity, a one-off 40 GB of disk, and a first evening.

That comparison is only fair if the local model does your work, and often it will not. A model that fits on a Mac is a fraction of the size of a frontier model in a data centre and it answers like it. The right way to read the numbers is not "local is 100 times cheaper", it is: the running cost is small enough that the only question left is whether the quality is enough for the task. That question you answer by trying it, which is free.

Where a subscription genuinely wins: occasional use. If you ask a handful of questions a week, a free cloud tier costs nothing and takes no disk, and 40 GB of weights on your SSD is a worse deal than an account you never pay for.

Slate itself costs nothing to run: the core is free and MIT-licensed, and Slate Pro is a one-time 39.99 euros rather than a subscription. It will also use a cloud model with your own key when that is the better tool, and it tells you plainly when a turn is leaving your machine. The comparison page includes the cases where another tool fits you better.