By Avi Salzman

It has become conventional wisdom that the more powerful artificial intelligence becomes, the more space it will take up and the more electricity it will use. Meta Platforms CEO Mark Zuckerberg said last week that his company is building a data center "that is so large it would cover a significant part of Manhattan." Oracle, OpenAI, and SoftBank said last week that they're partnering on a data center project called Stargate that analysts think could use five times as much power as the city of Philadelphia.

On Monday, the "bigger is always better" thesis was thrown on its head. Chinese programmers developed an AI application called DeepSeek that they say uses a fraction of the computing power of the models developed by companies like Meta. It uses an older version of Nvidia chips, and fewer of those chips than the models being produced by big American tech companies, the company said.

Nvidia stock fell 17% on the news. But some power stocks fared even worse. Vistra, Constellation Energy, and Talen Energy each fell more than 20%. Those three are known as independent power producers, because their power production and profits are not regulated by state utility boards. They have been the biggest beneficiaries of AI expansion plans because they have access to a scarce resource -- electricity. Their stocks have all more than doubled over the past year.

Several regulated utilities have also gotten in on the trade over the past year. While most utilities are held back by regulations that cap their profits, some utilities have made side deals with big tech companies to provide them power for their data centers. The Manhattan-size data center that Zuckerberg was talking about is being powered by Entergy, a Louisiana utility. Entergy stock was down 4.5% on Monday.

"The bull thesis on independent power producers and most integrated utilities is entirely dependent on data centers," wrote Jefferies analyst Julien Dumoulin-Smith.

DeepSeek has thrown a wrench in that bullish thesis. If it can get by with smaller data centers, then the Big Tech names may be overspending on electricity for their AI plans.

"The DeepSeek model is more energy- and capital-efficient, which calls into question the significant electric demand projections for the U.S.," Dumoulin-Smith wrote. "AI represents about 75% of overall U.S. demand forecasts through 2030-35 in most projections."

Jefferies' has a $274 price target on Constellation, which is near where the stock was trading on Monday. But that price is premised on the idea that Constellation is likely to make special deals with tech companies for 75% of the power from its 22 nuclear reactors, Dumoulin-Smith wrote. The math only works if the company makes deals that value the power at a significant premium to the average market price of electricity. Constellation has made deals to provide nuclear power to tech companies -- including one to restart its Three Mile Island reactor -- but it's not close to signing deals for 75% of its power.

Dumoulin-Smith is "in the middle" about whether Monday's selloff is a buying opportunity or the bursting of a bubble in power stocks. He does still like at least one name -- Pennsylvania-based Talen Energy, "which is pricing in less robust data center contracting scenarios than its peers," he wrote.

Morgan Stanley analyst Stephen Byrd also was not ready to weigh in on whether Monday's selloff in power stocks was overdone. He has generally been bullish on the power companies, though he appears to be attributing less of their stock value to data centers specifically. For Vistra, for instance, about $20 of his $172 price target is premised on data centers. Vistra was trading at $137 on Monday.

The power stocks have flourished on the "bigger is better" data center thesis. If smaller is just as good, their momentum may cool off for a while.

Write to Avi Salzman at avi.salzman@barrons.com

This content was created by Barron's, which is operated by Dow Jones & Co. Barron's is published independently from Dow Jones Newswires and The Wall Street Journal.

(END) Dow Jones Newswires

January 27, 2025 18:41 ET (23:41 GMT)

Copyright (c) 2025 Dow Jones & Company, Inc.