Does AI running on your own computer use less?

Free software makes it possible to install an open AI model on your computer, PC or Mac, and recent phones ship with one. The question comes back with these local AIs: does an answer computed at home, without a data centre, use less?

We followed the same answer along the 2 paths, in a single unit: the electricity of a half-page answer, 400 tokens (the word fragments the model writes), the quick question of the AI Impact Database.

A computer serves only one person. A server processes several requests at the same time on the same chip, and shares its consumption between them. Researchers at Stanford and Together AI↗ measured this on a data centre card, Nvidia's B200, with the same model of 8 billion parameters: the energy of a question falls from 6,449 to 548 joules when the card serves 64 at once instead of one, a division by 12. Their questions, conversations and reasoning have a length that is not published.

The same study also compares the chip of a Mac Studio with the B200, one request at a time. It attributes 240 to 305 W to the Mac's graphics part alone, more than the 145 W that Apple publishes↗ at most for the whole machine in its base configuration; we do not retain its ratio between the 2 chips. With 145 W and the study's response times, the Mac would use about 6,200 joules per question with the same model of 8 billion parameters, as much as the B200 serving a single request, according to our calculation. One request at a time, no gap between the 2 chips is established.

Served to many users, the same model would therefore probably use less in a data centre. The gap with a recent machine is not measured, and local hardware has progressed markedly since 2019, according to the same Stanford study.

In exchange, the data centre adds steps that the computer does not have. Google published↗ the breakdown of the median question put to Gemini in May 2025, to our knowledge the only measurement of this kind published by a provider on its own service.

Machines in reserve and the building make up 17% of the total. At a large operator, cooling adds about a tenth to server electricity; the Database retains this coefficient at 1.10, and notes that an unknown operator is counted at 1.56. The trip across the network weighs even less. The text of a question and its answer is about 2 kilobytes; even 100 kilobytes exchanged, counted with the factor established for 2015 by Aslan and his co-authors↗ (0.06 kWh per gigabyte), make 0.006 Wh, 2% of a quick question, according to our calculation. This factor was then halving about every 2 years.

Installing an AI on your computer almost always amounts to changing model. The machine's memory sets the size of what it can run: the largest local model in the Stanford study, gpt-oss-120b, fits, compressed to 4 bits, in the 128 GB of memory of the study's Mac, and a phone with 8 GB runs compressed models of a few billion parameters. Most providers do not publish the size of their assistants' models; the Database puts the model served by default in a medium class, and frontier models at several hundred billion parameters.

With a model of 2 to 8 billion parameters, the laptop uses 20 to 80% of the electricity of an assistant's average model, and 5 to 20% of that of a large model, according to our calculation. In their simulation of a service that sends each request to a small local model when it suffices, and to a large online model otherwise, the Stanford researchers obtain 60 to 80% less energy, and attribute most of the gain to model size rather than hardware efficiency. Their starting point is a service that would send every request to a model of 235 billion parameters. Our analysis of AI decarbonisation retains 27.8% for the smallest sufficient model, based on open models weighted by their downloads.

A small model does not always provide the same service. In the same study, a model of 8 billion parameters answers 57.5% of questions correctly, and gpt-oss-120b, the best local model taken alone, 71.4%, while Gemini 2.5 Pro gets 66.5 to 89.5% right depending on the benchmark. An insufficient answer is asked for again, and its electricity adds up.

These figures cannot be compared to the last decimal. The laptop measurement counts only the processor and the graphics card: it is a lower bound of what the socket draws, without the memory, the motherboard or charger losses. The data centre figures count the whole building. Its graphics card dates from 2019, and a recent machine uses markedly less, according to the same Stanford study. The Database values are estimates, which providers do not allow to be checked model by model.

Apple equips its devices with a model of about 3 billion parameters↗, stored on 2 bits each, and runs a larger model on its own servers. On an iPhone 16 Pro, the Stanford study runs a model of 4 billion parameters, compressed to 4 bits, at 9.5 W: 23 times less power than the Nvidia-card workstation tested with the same model, but 15 times more time, 37 seconds per question. That makes about 0.1 Wh per question of the study, according to our calculation. The phone answers 38% of questions correctly, against 44.5% for the workstation; per unit of energy, it nevertheless gives more good answers than the workstation.

When calculated at home, an answer uses household electricity. In 2025, electricity produced in France emitted 41 gCO2e per kWh over its life cycle, 11 times less than the world average (458 g) and 9 times less than that of the United States (384 g). Providers do not publish the country where a question is processed; the Database counts world average electricity by default. The hub's map shows every country.

On equal electricity, the 2 answers keep the ratio of their watt-hours. An answer calculated in France benefits from French electricity, as would a data centre located in France.

Calculating at home also removes the water that the data centre consumes for cooling: 0.26 mL for the median Gemini question according to Google, about 5 drops. The water used by the power plants that produce the electricity remains in the 2 cases. Our article on the water footprint of AI↗ details these scopes.

A computer already in service has already been manufactured, and installing a model on it does not change its manufacturing. The count changes when the machine is bought for this purpose, with more memory or a graphics card.

According to ADEME↗, an average laptop weighs 182 kgCO2e for its manufacture and distribution, over a life cycle of 193 kg. Bought to run models and kept 5 years, the period used by ADEME, at 20 answers per day, it would carry about 5 gCO2e of manufacturing for each answer. That is 36 times the CO2e of the electricity of one answer from the average assistant model, counted with the world average, according to our calculation.

In a data centre, the manufacturing of chips is shared between millions of requests: the Database counts an AI server used 69% of the time for 5 years. Google counts 0.01 gCO2e per Gemini question for its emissions other than electricity, manufacturing included.

For an organisation, the switch to local computing also moves the lines of the Bilan Carbone®: the purchased service leaves purchases, the electricity enters the Energy item and the machine enters fixed assets. Our analysis of AI in a Bilan Carbone® sets this out, including for a server installed on its own premises.