Energy for AI: easier said than done [Spanish]

Author: Not Specified | Source: El Periódico de la Energía | Read the full article in Spanish

The world of artificial intelligence (AI) is experiencing a massive energy challenge that's far more complex than most people realize. Training sophisticated AI models requires an enormous amount of electrical power, with some models like GPT-4 consuming around 30 megawatts during their development process. This incredible energy demand is pushing technology companies to rethink how they build and power their data centers.

Major tech giants like Amazon Web Services, Google, Meta, and Microsoft are at the forefront of this infrastructure transformation. Currently, these four companies control nearly half of the data center capacity in the United States. Their ambitious expansion plans include significant increases in power capacity, with AWS alone planning to grow from 3 to 12 gigawatts. However, building these massive data centers isn't a quick process – it typically takes around seven years from initial planning to full operational status.

Interestingly, the tech industry isn't just focusing on raw power, but also on efficiency. Innovative approaches like DeepseekV3's "Expert Mixture" architecture are emerging, which use networks of smaller, specialized models working together. These strategies aim to reduce the overall energy consumption while maintaining high-performance computing capabilities. The goal is to balance the incredible computational needs of AI with sustainable and manageable energy infrastructure.

[Read More (translated)]