Home / Learn / Data & AI / Offline AI / Lesson 3.9
All lessons

Module 3: Tokens: the unit of everything · Lesson 3.9 · about 20 min

Why your own machine makes tokens "unlimited"

Subscriptions cap use per few hours or per week, and APIs charge per token. On your own machine the only limits are speed, memory and electricity. Work out your own numbers.

Key points, meanings and formulas on one page: choose "Save as PDF" in the print box.

Module 3 · about 20 minutes · read first: 3.4 and 3.5. You’ll be able to compare your monthly token use with plan limits, API prices and the cost of running it yourself.

On your own machine nobody meters your tokens; the only limits are how fast it writes, how much memory it has and the electricity you pay for.

1. Why this matters

“Unlimited” is one of the strongest reasons to run AI yourself, but it is not “free” and not “infinite”. Putting real numbers on it shows whether it matters for you.

2. Three ways to pay for tokens

  • Subscriptions (Claude, ChatGPT, Google plans): a flat monthly price with usage limits per few hours or per week. For example, Codex use on ChatGPT plans is metered per 5-hour window, and the allowance varies by model.
  • APIs: pay per million input and output tokens, no flat fee.
  • Your machine: pay once for hardware, then electricity. No meter, no window, no throttling. The ceiling is tokens per second × seconds in the day.

It is the difference between a gym that lets you in three times a week, a pay-per-visit pool, and a bicycle you own: ride it as much as your legs (the machine’s speed) allow.

3. Your numbers

Enter your own use and prices. The example numbers are made up; real cloud prices are on each provider’s pricing page, and your electricity price is on your bill.

Calculator: your token budget
Prices are in whatever currency you type; the example numbers are made up. Find real ones on the provider's pricing page and your electricity bill. Words become tokens at the English rule of thumb (words ÷ 0.75). Local time counts only writing time, the bigger part (Lesson 3.4); the computer itself is not counted here (Module 16 adds it).

4. What “unlimited” really means

  • Volume: a machine writing 20 tokens per second can produce about 1.7 million output tokens a day running flat out.
  • Freedom: run a 6-hour batch job, retry 50 times, let an agent loop all night (Module 13), without watching a quota.
  • Still bounded: by speed (one big job at a time on one machine), memory (model size) and electricity (Module 6).
  • Not counted here: the price of the hardware. Module 16 adds it and works out break-even.

“Unlimited means free.” Unlimited means unmetered: nobody counts or caps your tokens. You still pay for the hardware (once) and the electricity (every hour it works). For a laptop the electricity is tiny; for a big desktop running all day it shows on the bill (Module 6).

Maths and principle

Principle: local cost per token is energy per token times electricity price.

Equation: cost per million output tokens = (watts ÷ tokens per second) × (10⁶ ÷ 3.6 × 10⁶) × price per kWh = watts ÷ (3.6 × tokens per second) × price per kWh.

Analogy: a home bread maker: after you buy it, each loaf costs only flour and power, and you can bake as many as the hours allow.

Try it: 150 W at 20 tokens per second and 0.20 per kWh? (150 ÷ 72 ≈ 2.08 kWh per million tokens, about 0.42 per million.)

Here is that sum, one step at a time.

How long for a million tokens? At 20 tokens per second: 1,000,000 ÷ 20 = 50,000 seconds, about 13.9 hours.

Energy used. The machine draws 150 watts, which is 0.15 kW. 0.15 kW × 13.9 hours = 2.08 kWh (the unit on your bill).

Cost. At 0.20 per kWh: 2.08 × 0.20 = 0.42 per million output tokens.

A whole day. Flat out for 24 hours (86,400 seconds) it writes 20 × 86,400 = 1.73 million tokens, for about 0.15 kW × 24 h × 0.20 = 0.72 of electricity.

watts
how much power the machine draws while it writes
tokens per second
how fast it writes
3.6
joins the units: a million tokens and 3.6 million joules per kWh (1 kWh = 1,000 W for 3,600 s)
price per kWh
what you pay for one kilowatt-hour, from your electricity bill
Physical meaning
Each token costs a tiny sip of electricity, and that is all. A faster machine at the same power makes each token cheaper, because it finishes the work sooner.

5. Worked example

Made-up numbers. A researcher tags 40,000 abstracts a month, about 8 million output tokens. Her desktop writes 40 tokens per second at 300 W: about 56 hours and 17 kWh a month, so about 3.40 of electricity at 0.20 per kWh. The same job on a plan would hit limits within days; on an API it would cost whatever 8 million output tokens cost there.

6. Practice

  1. Put your own daily use into the calculator.
  2. At 10 tokens per second and 60 W, what is the electricity per million tokens at your price?
  3. Name one job you would run locally only because there is no limit.
Answers
  1. Most light users find the running cost tiny; the question becomes hardware cost (Module 16).
  2. 60 ÷ 36 ≈ 1.67 kWh per million tokens, times your price per kWh.
  3. Examples: re-tagging an archive, trying 20 prompt versions on 500 items, an agent that retries until tests pass.

New words: quota or usage limit, per-token pricing, kWh (kilowatt-hour, the unit on your electricity bill).

  • Subscriptions cap use per few hours or per week; APIs charge per token; your own machine has no meter.
  • Your daily ceiling is tokens per second × 86,400: about 1.7 million output tokens at 20 per second.
  • Electricity per million tokens = watts ÷ (3.6 × tokens per second) kWh, times your price per kWh.
  • Unlimited means unmetered, not free: hardware and electricity still cost money.
  • It matters most for batch jobs, retries and long agent runs that would hit plan limits.

Check yourself

1What limits how many tokens you can use on your own machine?
2Your machine writes 20 tokens per second. Running all day, how many output tokens is that?
3When is unlimited local use most valuable?

Sources

Your ticks stay in this browser. Save them as a file on the course home.

Updated 2026-11-22. Prices and model names change fast: check the official source before you buy.