Bernstein Warns of Potential US Power Shortage Due to AI

The United States may face an electricity shortage in two years if the current growth in demand from AI data centers continues. This is according to a report by Business Insider, citing Bernstein Research. 

Analysts noted that the imbalance between supply and demand in the electricity sector could lead to increased costs for computing power. However, any estimate of AI energy consumption should be approached with skepticism due to rapidly changing variables: the number of inferences over a certain period, their complexity, model size, and the performance of chips and servers. 

Bernstein estimates that by 2030, the number of AI search queries will reach 10 billion per day, with responses averaging 375 words. These figures are comparable to Google, yet queries to ChatGPT require about ten times more computing power, experts note.

Why Is AI So Energy-Intensive? 

Business Insider highlights that the graphics processors used for training and operating AI generate more heat compared to regular CPUs, and a server rack for neural networks requires three times more energy. 

The development of AI hardware and software could lead to increased efficiency—more computations with less power. Bernstein’s estimates are based on a steady improvement in the ratio of computing power to energy consumption. 

Bill Gates Urges Caution Over Energy Concerns

Microsoft co-founder Bill Gates urged environmentalists and governments not to overreact to concerns about the massive energy consumption of generative artificial intelligence. He stated that data centers will increase global energy consumption by 2-6%. 

In May, Microsoft reported that its greenhouse gas emissions had risen by nearly a third since 2020, largely due to the construction of data centers. Gates noted that tech companies will drive the development and adoption of clean energy because they will pay a “green premium”—a higher price for it.

“Tech companies are the ones willing to pay more and help develop green energy capacities,” he said.

According to Bitwise expert Juan Leon, the infrastructure of miners and their access to cheap electricity could meet the growing demands of artificial intelligence.

In June, Core Scientific signed a 12-year contract with AI firm CoreWeave to provide infrastructure with a capacity of 200 MW for hosting Nvidia graphics processors.