America’s Power Grid May Not Keep Up with AI Data Centers

Summary

AI data centers in the United States are projected to increase their power demand from 5GW in 2025 to 50GW by 2030. If the same growth continued, they could require 500GW by 2035, exceeding the electricity currently purchased by all American households, businesses, and other users in a year. The article argues that the main constraint is not only chip production or data-center construction, but the physical infrastructure needed to supply electricity. AI companies are estimated to spend $800 billion annually on data centers, while some projects are using portable gas turbines to bypass slow grid connections; these generators create noise, local air pollution, and carbon emissions. Political resistance is growing, including a New York state moratorium on new data centers and a proposed federal moratorium on facilities above 20MW. Large projects are also being delayed: of 12GW of campuses planned for 2026, Bloomberg reported that only one-third were under construction, while large gas turbines and high-voltage transformers can take five years or more to deliver. The article says solar panels and batteries may support smaller facilities, but they cannot readily replace bespoke transformers and other components needed by multi-gigawatt campuses. Space-based or ocean-based data centers remain too small and costly for frontier-model training, and distributing training across many 1MW sites would reduce performance despite being technically feasible. Near-term measures such as dynamic line ratings and allowing data centers to reduce power use for a few hours each year could increase available capacity, but would not remove the long-term physical bottleneck. The author concludes that AI companies may need advance purchase commitments or government guarantees to encourage manufacturers to expand transformer production, although their current finances make such commitments difficult. Without action, energy infrastructure could slow AI development and economic growth, even as that delay might give society more time to adapt to increasingly capable models.