Go nuclear to keep the lights on as AI guzzles electricity

Recently, I was in a conversation with MIT researchers on artificial intelligence (AI) and nuclear energy.

- NISHANT SAHDEV

While discussing the subject, we saw a video clip of a data centre that looked like a giant fridge but buzzed like a jet engine. Inside, thousands of AI chips were training a new language model—one that could write poems, analyse genomes or simulate the weather on Mars. What struck me wasn't the intelligence of this machine. It was the sheer energy it was devouring. The engineer said, "This one building consumes as much power as a small town." That's when the magnitude of the challenge hit me: If AI is our future, how on earth will we power it?

All that intelligence takes energy. A lot of it. More than most people realize. And as someone who's spent years studying the physics of energy systems, I believe we are about to hit a hard wall. To be blunt: AI is growing faster than our ability to power it. And unless we confront this, the very tools meant to build our future could destabilize our energy systems—or drag us backward on climate. One solution has been pinpointed by the AI industry: nuclear energy. Most people don't associate AI with power plants. But every chatbot and image generator is backed by vast data centres full of servers, fans and GPUs running day and night. These machines don't sip power. They guzzle it. In 2022, data centres worldwide consumed around 460 terawatt-hours.