NEW YORK/HOUSTON (Reuters) - Shares of U.S. power, utility and natural gas companies sold off on Monday in some of the biggest recorded one-day drops, as new AI technology from Chinese start-up DeepSeek cast doubt on a projected surge in U.S. electricity demand and tech spending.

Power producers were among the biggest winners in the S&P 500 last year on expectations of ballooning demand from the energy-guzzling data centers needed to scale Big Tech's artificial intelligence technologies.

The wider adoption of AI models like the one developed by DeepSeek, which it says it built in under two months and is cheaper than models currently used by U.S. companies, could result in less electricity demand overall and result in a smaller power build-out, analysts and economists said.

"If proven true, the efficiencies used within DeepSeek's open-source model can be applied by the hyperscalers to their models, which would result in a more moderated demand," analysts with Evercore ISI said in a note.

Big Tech firms, which are also known as hyperscaling data center developers, have devoted tens of billions of dollars in AI data center development over the last year.

In the U.S., data centers consumed roughly 4.4% of electricity in 2023 but are anticipated to use 6.7% to 12% of all power by 2028, according to a report produced by the Lawrence Berkeley National Laboratory.

Independent power provider Constellation Energy, whose shares had shot up about 100% in 2024 largely on its ability to sell nuclear and gas-fired power to U.S. data centers, sunk by about 20% in trading on Monday after news of DeepSeek's advancements.

Vistra was down 30% and rival Talen Energy Corp was down 22%.

DeepSeek AI could also threaten the dominance of current AI leaders, which are based in Silicon Valley, and slow their deployment of data centers. DeepSeek's AI assistance had overtaken U.S. rival ChatGPT in downloads from Apple's app store on Monday.

But with the wider adoption of AI, even with more energy-efficient models, power demand could surge everywhere, said Ed Hirs, an energy economist at the University of Houston. He cautioned that a sell-off of power stocks could be short-sighted and short-lived.

"In this instance, if DeepSeek turns out to be what everybody wants, and they sell to U.S. companies, and the U.S. companies change their algorithms to adopt to it, it just means a greater, faster broader development," Hirs said.

Still, electricity companies, and even producers of feedstocks related to power generation, were under pressure.

Earlier this month, Constellation acquired private natural gas producer Calpine Energy for $16.4 billion in one of the largest U.S. power industry deals ever, a sign of rising expectations that demand for gas will grow as a generation source for AI.

Shares of publicly-traded producers of natural gas, which makes up the biggest share of fuels used to generate electricity in the United States, also slumped.

EQT Corp was off 9%. Midstream operator Energy Transfer, which said it has received connection requests from dozens of data centers, was down about 7%.