This is Heatmap Daily, a weekday digest written by our executive editor. Subscribe here.

My colleague Emily Pontecorvo has a fascinating story today about Net Power, an energy and power development company whose technology once seemed key to decarbonizing the country’s electricity grid.

When Net Power went public in 2023, it was working on a natural gas-burning power plant that would generate little to no emissions in part by using carbon capture and storage. Its invention seemed to have big implications, and not just for the company’s bottom line: Back then, the Environmental Protection Agency was drafting climate rules that would have required new natural gas plants built in the 2030s to capture their carbon dioxide emissions. Net Power’s technology seemed like one of the best available ways to meet that standard.

Cut to 2026, and the world looks different. Thanks to the artificial intelligence boom, electricity demand is surging. But Net Power hasn’t successfully commercialized its proprietary power plant technology. Even if it had, few companies are willing to pay the high premium required by Net Power’s technology, especially when cheaper options are available. And the government certainly isn’t going to require it. Emily’s story reveals what happened next.

When I helped start Heatmap more than three years ago, I never thought I would write this much about data centers. I believed that electricity demand growth was inevitable, sure, and I thought the U.S. power grid had to expand for the country to decarbonize. But I didn’t foresee such a staggering near-term surge in electricity demand. (To be fair, even long-time students of the computing industry’s energy use didn’t see the coming increase, either.)

What I didn’t anticipate — or understand — were two fast-moving trends. The first was that the AI models were responding to scale: the more text or images they consumed, or training they received, or (crucially) computational power they had access to, the better they got. The second was that Wall Street was happy to pay to continue that speedy scale-up. As long as models kept getting better, investors were willing to keep lending or investing more money to the companies making them. Through the power of those two scaling functions, the data center boom became one of the biggest infrastructure buildouts in American history.

In recent days, the second of those trends has started to falter. After a surprisingly hot inflation report last week, and a disclosure on Wednesday evening that Meta has gone free cash flow negative to finance its ongoing AI buildout, stocks with any proximity to AI or data centers stumbled on Wall Street today. The data center developers Coreweave and Vistra were each down about 10%, while Microsoft, Alphabet, and Nvidia all fell in daily trading. Meta’s stock lost almost 7% of its value; GE Vernova, which produces gas turbines, was down 10%.

These declines look more significant when you zoom out. Since late June, a good chunk of the AI portfolio has sagged into the red. Coreweave’s stock is down 54% over the past six months. Alphabet and GE Vernova have also posted declines since the summer. (The trend isn’t uniform: Nvidia and Microsoft are still very much in the green.)

None of this, to be sure, guarantees that the AI boom is ending or even slowing down. But this finance-applied friction is what such a slowdown could look like. Investors aren’t (yet) saying that AI is a business dud, to be clear, just that companies perhaps don’t need to go and construct all this stuff so fast. And given what the data center buildout has meant for energy construction — not to mention greenhouse gas emissions — that’s a trend nobody in climate or energy can ignore.