Jensen Huang Says AI Needs More Fossil Fuel Power for Now
Nvidia CEO Jensen Huang says the United States has not built enough fossil-fuel plants to meet power demand from AI. He made the comments while discussing energy and AI's effect on the planet on The Ezra Klein Show. The Verge
The episode, titled 'Jensen Huang Thinks A.I. Alarmism Has Gone Too Far,' was published on September 23, 2026. The New York Times The New York Times also published an edited transcript of the conversation. The New York Times
Huang argued the U.S. will need to use fossil fuels over the next several years because there is not enough sustainable energy, such as solar, wind and other low-carbon sources, to meet demand. His order was step by step. Use fossil-fuel electricity now, then shift to cleaner sources later. The Verge
He described the tradeoff in medical terms. AI is like surgery, he said. To save people, you have to hurt them first. His phrase was that AI must cause "an enormous amount of pain and suffering" to save them. The Verge
Unlike many leaders of frontier AI labs, the large groups building the most advanced models, Huang does not think AI could wipe out humanity. Apple Podcasts
Earlier in September, Donald Trump called Huang while Huang was onstage in Los Angeles. During that call, Trump criticized calls to slow down and regulate A.I. The New York Times
The broader context here is a mismatch in timelines. New AI systems are being rolled out on business schedules measured in quarters, or months. Power plants, grid upgrades and permits work on schedules measured in years. Anyone who plans data-center capacity will recognize the problem. It forces short-term choices about current power sources while new supply is still being built.
In my view, Huang's comments are best read as an argument about infrastructure, not about AI research. He is not debating model design or safety methods. He is saying the limit on AI progress in the United States is electricity. That is a practical point, which matters because it spreads responsibility beyond AI labs to utilities, regulators and builders.
Looking ahead for people who build and run these systems, the path Huang describes is uncomfortable but clear. If clean power cannot yet support new AI demand at this scale, operators must choose between slowing deployment and using higher-carbon power for a time. Over the long arc, engineering has often learned to do more computing per watt once the incentives are clear. The open question is whether efficiency gains and new clean supply arrive fast enough to shorten the fossil-fuel period Huang describes.
