How much energy to use AI?
Author
Débora
Published
The price of a click or the weight of a watt?
Foreword: this article will discuss the energy consumption of AI as of 2026.
We often picture it as something hazy, floating in an immaterial “cloud.” Yet Artificial Intelligence has a silicon body, a ravenous appetite for energy, and a very real bill. Behind every ChatGPT answer or every automatic sorting of your photos, an electrical machine spins up.
Still, performance doesn’t have to mean excess. The whole question is where to set the slider between wasteful AI and frugal AI. A journey into the heart of the machine, from the gigawatt-hour to the microsecond.
AI energy consumption: understanding orders of magnitude
From the second hand of a watch to thousands of electric cars, the energy impact of AI mostly depends on the scale at which you look at it. Here, the numbers tell radically different stories depending on the technology chosen and what you do with it—revealing a real energy paradox.
Training AI: understanding the energy cost of creation
Training is when we build the AI’s brain. It’s a massive, one-off energy investment.
GWh
Training a Generative AI (GenAI)
Here we’re talking in Gigawatt-hours (GWh). Picture 10,000 electric sedans charging simultaneously: that’s the price of a single training run for a large model.
KWh
Training Traditional AI (Non-generative)
It depends on the model, but at Neovision our training runs generally consume between 1 and 10 kWh. It’s the modest heat of an electric heater left on for anywhere from an hour to a single night. The difference is staggering.
Inference in AI: the invisible consumption of everyday use
Inference is what we call using an AI once it has been trained.
Wh
Generative AI inference
It consumes Watt-hours (Wh). One LLM answer is far more costly than a standard Google search. This is also where the trap lies: if the individual effort seems tiny, repeating it worldwide creates a colossal mass effect.
mWh
Traditional AI inference (non-generative)
About 3 mWh (milliwatt-hours).
You’d have to click a mouse 6,000 times to reach that consumption.
µWh
On-device AI inference
The queen of frugality with only 3 µWh (microwatt-hours): the energy needed to move the second hand of a watch forward by one second.
The energy paradox
If training is an isolated storm, inference is a fine but endless drizzle. Globally, daily usage weighs far more than the initial creation phase.
The Open Source lever: real power for transparency and frugality
Open Source isn’t an eco-label; it’s the best way to obtain a rigorous energy audit. In AI, Open Source isn’t just a licensing question—it’s a matter of radical transparency.
Auditability and control with Open Source code
Before you optimize, you need to see. Opening the code is like turning on the lights in the engine room. Where closed models (ChatGPT, Gemini, Claude, etc.) are untouchable monolithic boxes, Open Source gives us a precise observatory.
Rather than suffering an “average” consumption, Open Source gives us the levers to ensure the efforts behind optimizing our models’ energy use are real.
Knowing exactly what we consume and having the power to change the code to reduce that footprint is taking back control.
The engineer’s scalpel: forging frugality through expertise
Usage optimization (Inference):
- Sizing: Match the architecture size to the real need.
- Business optimization: Focus the algorithm on relevant data to avoid unnecessary computations.
- Hybrid models: Orchestrate several specialized architectures.
Training optimization:
- Minimizing cycles: Reduce the number of runs thanks to full visibility.
- Data selection: Prioritize quality to shorten compute time.
- Bias control: Correct errors early to avoid costly retraining.
This demand for clarity is now embodied in work carried out at Neovision to measure (and no longer simply estimate) the energy consumption of each training run for the models developed for our clients.
Estimating the invisible: the riddle of “Black Boxes”
Closed models (ChatGPT, Gemini, Claude) don’t disclose their energy “recipe,” but we can still track their footprint.
The race for efficiency
If closed models don’t share their energy “recipe,” the race for efficiency is on. Google, for example, divided the energy consumption of its inferences by 40 between 2024 and 2025.
All these efforts account, among other things, for optimizing AI models, but also the infrastructure required for them to run (we could write a whole article about the footprint of datacenters, HPC, etc.).
Practical guide: choosing between web search and AI
- If you can complete your task with one web search, choose your favorite search engine.
- If your task involves hours of manual research, analysis, and synthesis, an inference with your favorite LLM might be a better trade-off.
Making the right choice takes practice—don’t be afraid: train!
Our suggestion
Tools we recommend
AI’s positive impact, a commitment etched into code
Energy efficiency can no longer be an afterthought. Moving from estimation to real measurement with tools designed for AI, and using Open Source to size models more accurately, are concrete levers to move away from electrical gigantism.
AI performance is now measured by its ability to deliver the most accurate result for every watt consumed. Frugality isn’t a brake on intelligence—and it shows up strongly in how each of us uses the technology.
A projet with similar needs
For techies and science folks
Want to dig into related sources?
Bibliography
- Elsworth C, Huang K, Patterson D, Schneider I, Sedivy R, Goodman S, et al. Measuring the environmental impact of delivering AI at Google Scale. Disponible sur : https://services.google.com
- Touvron H, Lavril T, Izacard G, Martinet X, Lachaux MA, Lacroix T, et al. LLaMA: Open and Efficient Foundation Language Models [Internet]. arXiv; 2023 [cité 17 févr 2026]. Disponible sur: http://arxiv.org/abs/2302.13971
- Fernandez J, Na C, Tiwari V, Bisk Y, Luccioni S, Strubell E. Energy Considerations of Large Language Model Inference and Efficiency Optimizations. Disponible sur: https://aclanthology.org/2025.acl-long.1563.pdf