Training a Model vs Using It: Which Costs More?

Every AI model has two lives. First it is trained: it learns from huge amounts of text, images or code. Then it is used: it answers prompts from people, every day, for months or years.

Both use energy and water, but in very different ways. This lesson shows how big each one is, and why the answer to "which costs more?" is changing.

What You'll Learn

• What training and using (also called inference) mean for energy use
• How much energy one well-studied training run used
• Why billions of small prompts can add up to more than one big training run
• Why companies rarely publish training figures for their newest models

Two different jobs

How big is a training run?

The best-studied example is an older model, GPT-3, released in 2020. A 2021 study by Google and UC Berkeley researchers estimated that training it used about 1,287 megawatt-hours (MWh) of electricity and produced about 552 tonnes of CO2-equivalent emissions.

One megawatt-hour is one million watt-hours. So 1,287 MWh is about 1.3 billion Wh. That is a lot of energy for one job, enough to boil roughly 13 million kettles of water.

A separate study, "Making AI Less Thirsty", estimated that training this model also consumed about 700,000 liters of water on site for cooling, and several million liters once the water used to make the electricity was added.

Newer models are much larger than GPT-3 and are trained on far more data. Most companies have not published how much energy their newest training runs used. Outside researchers make estimates, but they vary, so be careful with any single number you see.

Why using AI can add up to more

A single prompt is tiny next to a training run. But prompts never stop.

Here is a simple calculation. Take the GPT-3 training figure of about 1.3 billion Wh and divide it by about 0.3 Wh per text prompt (the rough figure from lesson 1):

• 1,300,000,000 Wh ÷ 0.3 Wh ≈ 4.3 billion prompts

So after about four billion short prompts, the energy used for answering would equal the energy used for that training run. Popular chatbots handle a very large number of prompts every day, so a widely used model can pass that point fairly quickly.

This is a rough comparison. It mixes an older model's training figure with a newer per-prompt figure, and real numbers depend on the model. But it shows the key idea: the more people use a model, the more its total footprint comes from use rather than training.

The hidden costs on both sides

Both training and use have costs that are easy to miss.

Before and around training:

• Experiments. Teams run many smaller test runs before the final one. These are rarely counted in public figures.
• Failed runs. Big jobs sometimes crash and must restart from a saved point.
• Hardware. Making chips and servers uses energy and water before any AI runs on them. Mistral's 2025 life cycle study is one of the few to count this.

Around use:

• Idle capacity. Services keep extra machines ready for busy moments, as you saw in lesson 2.
• Longer, heavier tasks. Reasoning models, long documents, images and video push the cost per use up.
• More features. AI is being added to search, email, phones and office tools, so the number of prompts keeps growing.

Why the balance is shifting

A few years ago, most attention went to the energy cost of training. Today, more attention goes to use, for three reasons:

1. Many more people use AI every day than a few years ago.
2. Each answer can do more work. Reasoning modes and AI agents can run many steps for one request.
3. AI is built into tools you use without opening a chatbot at all.

At the same time, the energy per answer is falling for many services, thanks to better chips and software. Lesson 6 looks at that race between growing demand and growing efficiency.

Key Takeaways

• Training is one huge job per model. Using a model is billions of small jobs that never stop.
• Training an older model, GPT-3, used about 1,287 MWh of electricity, by one well-known estimate.
• Most companies do not publish training figures for their newest models, so estimates vary.
• With heavy use, the total energy of answering prompts can pass the energy of training.
• Both sides have hidden costs: test runs and hardware for training, idle machines and heavy tasks for use.

Sources

• Patterson et al., Carbon Emissions and Large Neural Network Training (2021)
• Li, Yang, Islam and Ren, Making AI Less "Thirsty"
• Mistral AI, Our contribution to a global environmental standard for AI (2025)
• Epoch AI, How much energy does ChatGPT use? (2025)