On October 15, The New York Times published an interactive report using successive generations of GPT models as examples to demonstrate the rapidly expanding computing demands behind the advancements in AI capabilities. The report estimates that by 2026, the training computations for GPT-6 Astra will approach 10²⁷, more than 50 million times that of the original GPT. The article uses visualizations to illustrate how large-scale model training relies on massive amounts of data and computation, constantly adjusting trillions of parameters to improve prediction and generation capabilities. As model size increases, computing power and electricity demands also continue to grow. The report cites research predicting that by 2030, the electricity required to train the largest AI models could approach the current electricity load of New York City; during the same period, data centers in the United States may account for approximately 12% of the nation's total electricity consumption. This trend highlights the increasingly close link between advancements in AI technology and the expansion of energy infrastructure.
