What Does One AI Token Cost?
The electricity used to generate an AI token can be measured independently of its selling price. The first task is to say exactly what is being measured.
An AI service can tell you what a million tokens will cost before it generates any of them. Ask how much electricity those tokens require and the answer becomes less convenient. The model matters, the hardware matters, and the amount of text already in the conversation matters. Even the other requests being processed alongside yours can change the result.
None of that makes the electrical cost unknowable. It makes the conditions important. I want to work through those conditions, from the activity inside a processor to a number on an electricity bill, without borrowing the answer from an AI company’s price list.
Scope note: this series calculates the electricity used to run a model. Hardware manufacture, training, and the full cost of operating an AI business are separate questions.
Start with what a token counts
A language model handles text in pieces called tokens. Depending on its tokenizer, a piece might be a word, part of a word, punctuation, or another sequence of characters. A token count describes the text representation. It does not tell you how many electrical operations produced it.
That difference is easy to miss because tokens appear as neat units on a bill. A provider can charge by the million, distinguish input from output, and offer a discount for reused input. Those are useful commercial categories. They do not make a token a standard quantity of energy in the way that a joule is.
In a conventional text-generating transformer, the computer calculates possible next tokens from the preceding sequence, selects one, and continues. The calculation uses learned numerical values, called weights, and information retained from earlier text. Those numbers have to be stored, moved, multiplied, added, and made available for the next step. The original transformer paper describes the attention and learned transformations underlying that work.
The visible result can be very small. Producing a comma still involves the model’s computation. Reading the comma does not reveal how much computation occurred, any more than counting the tokens in the finished answer reveals the exact hardware that ran it.
This is why I would start an energy comparison by asking which model produced the tokens and under what conditions. Two systems producing the same number of tokens have completed the same output count. That is all the count establishes by itself.
Joules before dollars
A joule measures energy. A watt measures the rate at which energy is being used: one joule per second. A machine drawing an average of 100 watts for ten seconds uses 1,000 joules during that interval.
Suppose, purely as a worked example, a measurement records 1,000 joules while a model produces 1,000 tokens. Divide the energy by the output count and the result is one joule per token. That number belongs to the measured run. It is neither an industry average nor a property of language itself.
Electricity bills normally use kilowatt-hours. One kilowatt-hour equals 3.6 million joules, so one million tokens at one joule each require about 0.278 kilowatt-hours. At a declared electricity price of $0.0862 per kilowatt-hour, their electrical cost is approximately $0.0239. About 2.4 cents.
The reference price comes from the US Energy Information Administration’s preliminary 2025 industrial average. It is a public comparison point, not a claim about any provider’s electricity contract. A different rate changes the dollar result; it does not change the joules already measured.
Here is the whole conversion, with the output expressed per million tokens:
Electrical cost = joules per token × dollars per kilowatt-hour ÷ 3.6.
The small divisor is simply the result of putting both quantities into the same units. There is no assumption about an API tariff hidden in it.
A small number needs a precise description
Published experiments have measured less than one joule per generated token in particular configurations. Jae-Won Chung and colleagues report 0.151 joules for Qwen 3 32B running a conversation workload on one NVIDIA B200 with a maximum batch size of 512. A batch groups requests for processing together. Their metric divides the energy used by the graphics processing unit, or GPU, during steady operation by the tokens generated during that period. GPUs are processors widely used for model calculations; this result does not measure every electrical load in the building. (Where Do the Joules Go?)
At the declared reference price, that result converts to roughly $0.00362 per million generated tokens: around a third of a cent in measured GPU electricity. The conversion is straightforward; choosing what to infer from it requires more care.
It establishes that a fraction-of-a-cent GPU electricity result is physically observed for that workload and configuration. It does not establish the energy of your latest chatbot conversation. A small local batch, a longer context, another model, or a system generating many internal reasoning tokens can produce a different result. Nor does the figure include the server’s other components or cooling simply because they also need electricity.
Those distinctions preserve the usefulness of the number. Without them, the same measurement can be made to support almost any argument about AI, from claims of negligible impact to claims of implausible profit margins.
The question the price list cannot answer
An already-purchased processor is not purchased again when it produces another token. Within a calculation of additional electrical expenditure, the original purchase is fixed and cancels when one run is compared with another. The computer still uses energy while it works. That energy is the physical quantity this series follows.
A business has expenses beyond that quantity, and a retail price can reflect much more than electricity. Subtracting the electrical cost from a selling price does not calculate profit. But those other expenses do not prevent an independent measurement of the electricity, either. Both statements can be kept intact.
The practical question is now narrower and answerable. Which circuits worked, which numbers moved, how long did the equipment remain powered, and how many tokens did the measurement include? Once those questions have answers, the cost conversion can be repeated at any declared electricity price.
The next article begins inside the hardware, where a numerical operation and a memory transfer can consume very different amounts of energy.
