Energy and AI is about the electricity needed to train and run artificial intelligence models. These models run in data centers full of powerful chips that consume energy and need to be cooled.

Como funciona

Energy consumption happens in two main phases.

During training, thousands of chips work together for a long time to teach the model. It is an intense phase, but it happens only occasionally.

During inference, the trained model responds to user requests. Each response uses little energy, but together, millions of responses per day represent an important share of consumption.

Beyond the chips, data centers also use energy for cooling and for electrical infrastructure.

Exemplo prático

A company decides to use a smaller model instead of a giant model to classify customer emails. The smaller model performs well for that task and uses fewer computing resources.

Another case: a team that runs heavy batch jobs chooses to do so at times and in regions where the cloud provider has more clean energy available, when that information is offered.

Texto
Tarefa simples -> modelo pequeno
Tarefa difícil -> modelo maior
Resposta repetida -> cache

This practical rule reduces cost and consumption at the same time.

Para que serve entender o tema

Understanding the relationship between energy and AI helps to:

  • Choose the right model for each task without wasting resources.
  • Understand why companies invest in new data centers and energy sources.
  • Assess the environmental impact of AI projects.
  • Plan costs, since energy weighs on infrastructure pricing.

For those who build flows with multiple agents, like in bridgecode, it pays to choose the model according to the difficulty of each step, instead of always using the largest.

Cuidados ao ler números

Consumption estimates vary a lot depending on the method, the model and the hardware. Be skeptical of numbers without a source, and prefer reports from energy agencies, peer reviewed research and reports from the companies themselves, always citing the date.

Diferença entre energia na treinamento e na inferência

Training is a large, concentrated effort. Inference is a small, continuous effort, repeated with each use. As more people use AI, inference gains weight in the total bill.