AI's Growing Thirst: Power Demands Skyrocket

The computational power required for training artificial intelligence models is increasing exponentially, doubling each year. This trend, known as a "scaling law," suggests that larger models trained on more data yield better performance.

Researchers predict that by 2030, the electricity demand for training the most advanced AI models could be comparable to New York City's current usage. Energy experts project that data centers may consume up to 12 percent of America's electricity by the same year.

This immense power requirement stems from the trillions of calculations and vast datasets used to adjust the parameters within AI models. Training these models involves extensive arithmetic, leading to substantial energy consumption and operating costs.

Consequently, the industry's rapid expansion of data centers is making ambitious net-zero electricity goals increasingly difficult to achieve. Many centers rely on natural gas, and even nuclear power is being considered to meet the escalating energy needs.