DeepSeek’s AI Model Just Upended the White-Hot US Power Market

(Bloomberg) -- It took just a single day's trading for Chinese artificial intelligence company DeepSeek to upend the US power market’s yearlong hot streak premised on a boom in electricity demand for artificial intelligence.

AI’s energy needs have led companies such as OpenAI, Alphabet Inc. and Microsoft Corp. to seek new sources of power, such as shuttered nuclear plants. It has also complicated their ambitious climate goals. DeepSeek’s model appears to be more efficient and can achieve the same results for a fraction of the energy use, which may mean AI will have a smaller climate impact than thought. 

The DeepSeek development “calls into question the significant electric demand projections for the US," analysts led by Julien Dumoulin-Smith at Jefferies wrote in a note on Monday. AI represents about 75% of overall US power demand forecasts through 2035 in most projections, Jefferies said. 

The stock selloff was driven by “expectations for how many data center deals can get signed,” said Dumoulin-Smith in an interview. The market now questions the timeline and cadence of those deals, reflected in the abrupt end to a 20% year-to-date rally in power stocks, Jefferies noted. Vistra Corp. sank by a record 28% and Constellation Energy Corp. fell 21%. (Companies that make chips, particularly Nvidia Corp., have also been part of the rout.)

Just last week, President Donald Trump told the World Economic Forum that he would use his energy emergency declaration to fast-track the building of fossil fuel power plants next to data centers.

Some researchers believe AI power demand projections were already too high even before DeepSeek threw cold water on them publicly this week. The “gargantuan” estimates were more reflective of utilities’ desires to pump up investment in grid infrastructure than reality, said Stephen Jarvis, an environmental economics professor at the London School of Economics. 

“It felt like the beginnings of a tulip craze,” said Eric Gimon, a senior fellow at clean energy think tank Energy Innovation, who pointed out that efficiency increased amid previous surges in data center construction, blunting energy growth.

Still, even if DeepSeek can drive down AI electricity consumption, the industry will continue to need power. And while AI’s needs are now less certain, power demand is still going to climb from homes and factories that are increasingly running on electricity, said Nikki Hsu, a utilities analyst for Bloomberg Intelligence.

Big tech companies have attempted to find clean power to keep data centers running around the clock. The market has been especially strong for nuclear energy. Notable examples include a deal Microsoft inked with Constellation last year to restart the Three Mile Island plant and Meta Platforms Inc. announcing it was seeking up to 4 gigawatts of nuclear power — enough to provide electricity for millions of homes — to help power its data centers in December. 

Those moves are partly in response to companies being off track to meet their climate goals. Microsoft’s emissions were 30% higher last year compared to 2020 while Google’s emissions were up 48% compared to 2019, in both cases largely due to AI. Energy use projections for the coming decades are even more extreme and largely driven by AI: in Sweden, it’s expected to double this decade while in the UK, demand was forecast to rise 500% over that same period.

DeepSeek relies on a much lighter model, which should offer a pathway to lowering AI emissions. However, Hsu cautioned that many questions remain about the potential impact of the company’s technology on the climate. "Nobody knows if the chips are really more efficient. But it seems like there's a lot of concern," she said. 

More competition and innovation in AI may also only spur greater power demand, something known as the Jevons paradox. 

“The market generally likes to do knee-jerk reactions,” said Susan Su, clean energy investor and partner at Toba Capital. AI is “infinitely scalable,” creates new use cases for itself and is still in its commercial and technological infancy. She believes that AI’s energy demand will grow exponentially with efficiency gains.  

“When I say exponential, I truly mean multiple orders of magnitude: more use cases and more demand, because now, all of a sudden, it's going to be much more accessible,” Su said.  

The AI arms race could further spur more widespread use. The US was already pushing to beat China, with Trump last week touting $500 billion investments as part of Stargate. DeepSeek’s emergence could redouble those efforts. 

"If the US wants to be a leader in AI going forward, we need to make sure we'll be able to power all those data centers," said Benton Arnett, senior director for policy and markets at the Nuclear Energy Institute trade group. He said companies developing nuclear projects are unlikely to cancel them based on today's news. 

Energy Innovation’s Gimon drew parallels between the AI data center boom and the dotcom crash of the early 2000s. Back then, telecom company Global Crossing Limited spent billions on fiber optic cables, betting that all the companies of the era would need more bandwidth. What they didn’t bet on was that engineers would figure out how to make those cables more efficient. “It was an over-optimism of market demand and underappreciation of the ability to do more with less,” Gimon said. 

This is a learning opportunity for utilities to think more deeply about risk management, said Gimon, particularly the risk of leaving ratepayers saddled with paying for costly projects if those AI businesses crash. 

“You don't want to be carried away by the fervor of the moment,” Gimon said.

©2025 Bloomberg L.P.