GRID & SILICONLearn
Get the newsletter

AI's power, chips and data centers, explained in plain English

Learn home › Chips

Training AI and Using AI

What changes when an AI program finishes learning?

How it works

AI (computer programs that learn patterns from examples) needs computers. During training (learning from examples), computers adjust a model (the AI program that makes answers).

Computers check results against goals, change settings and try again. Chips do math. Memory stores the needed numbers.

Inference (using a trained model) happens when people use the model. People send messages or pictures. Computers use models to answer. Longer or harder jobs can take more work.

Both stages need computers, power and cooling. Training may take one long job. Inference repeats as requests arrive. Energy comparisons need power draw, running time and request counts.

Picture the scale

Google measured a middle, or median, Gemini Apps text request at 0.24 watt-hours in May 2025. Its report compares that with less than nine seconds of TV viewing. That is one measured service, not a rule for every AI request. More requests still add work.

Training AI and Using AI processExamples→Training→Trained model→Inference

Why should I care?

An AI tool helping with a game or homework still runs on physical chips. Many repeated requests add work after the model finishes training.

Words to know

Training
Adjusting a model from examples; like practicing a skill with feedback.
Inference
Using a trained model on input; like applying a practiced skill to a new task.

Check yourself

1. A model adjusts its settings from examples. Which stage is this?

  1. Training
  2. Inference
  3. Electricity transmission
Show answer and why

A. Training. Training changes the model using examples.

2. A trained model answers a new request. Which stage is this?

  1. Training from the same examples again
  2. Inference
  3. Changing the model’s settings from feedback
Show answer and why

B. Inference. Inference applies the model to new input.

3. Why count repeated requests when comparing energy?

  1. Training removes later computing needs
  2. Every request uses zero energy
  3. Many uses can add up
Show answer and why

C. Many uses can add up. Total work includes how often a task runs.

Sources and number checks