Data centers are buildings full of servers that store data and run programs on the internet. They all need electricity to operate and to stay cool. With the advance of AI, the consumption of these facilities has drawn attention from companies, governments and journalists.
How it works
Consumption comes from three main parts. The first is the servers and their chips, such as GPUs, which do the calculations for the models.
The second is cooling. Chips get very hot when working hard, and the heat needs to be removed with industrial air conditioning or liquids.
The third is support infrastructure: network, storage, lighting and backup power systems.
To compare efficiency, a metric called PUE is used, which relates all the building's energy to the energy that reaches only the servers. The closer to 1, the less energy is lost to support.
Practical example
Suppose you have an app that calls a large model for simple tasks, like classifying emails into categories. Each call uses energy in the cloud, and you pay for it indirectly.
If you switch to a smaller model that solves the same task, you reduce cost, response time and energy. Another option is to cache repeated responses and not call the model twice for the same question.
cache = {}
def responder(pergunta):
if pergunta not in cache:
cache[pergunta] = chamar_modelo(pergunta)
return cache[pergunta]Why understanding this matters
Who plans an AI product needs to know that energy is part of operating cost. This influences price, choice of cloud provider and even the region where servers are located.
For companies, the topic also involves environmental goals and local energy availability. For devs, it becomes a practical incentive: write concise prompts, use the smallest model that gets the job done and avoid unnecessary calls.
Exact consumption numbers vary a lot between sources and years. To cite values, use reports from energy agencies or from the companies themselves.
Difference between training and inference
Training a model is the phase where it learns, and it requires a lot of computing power for a limited time. Inference is daily use, when millions of people ask questions. With heavy use, inference also weighs heavily on the energy bill.