For much of the 20th century, AI struggled not from a lack of ambition, but because available hardware wasn't powerful enough. Early systems hit limits on processing speed and memory, contributing to "AI winters" as progress stalled and funding dried up. Today, this problem is largely resolved. AI models are now trained on specialized chips in vast data centers. Compute, which used to be the main bottleneck, can now simply be purchased. Companies like Nvidia mass-produce powerful graphics processing units (GPUs)—originally designed for gaming but perfectly suited to AI calculations. What holds AI back now? The physical limit of electricity.
Modern AI models don't just train once; they operate continuously, powering chatbots, search engines, and autonomous agents. This shift has made AI a constant, large-scale electricity consumer. According to Sampsa Samila of IESE Business School, "the core issue is not a shortage of energy in absolute terms, but rather the availability of reliable, firm capacity at the right place and the right time".
Predictions for AI energy consumption show this strain. The International Energy Agency (IEA) projects data centers will consume more than twice as much electricity by the decade's end. In parts of the U.S., data center power usage already rivals heavy industry.
How AI is used matters as much as how it is trained. Training large models consumes immense power but occurs infrequently. What is growing faster is the everyday work of models responding to users. Samila notes that newer "reasoning" AI systems, which deliberate longer, push energy demands into everyday operations rather than occasional large training runs.
(Adapted from: https://www.livescience.com)