In fifth paragraph, corrects from “100 million” to “1 million” and changes wording from growth in users to growth in usage. The author is a Reuters Breakingviews columnist. The opinions expressed are her own.

By Yawen Chen

LONDON, Feb 3 (Reuters Breakingviews) - DeepSeek has pulled the rug from under energy providers. The Chinese startup’s nimbler models may mean both training artificial intelligence and using it no longer needs so much power. While groups like $57 billion Vistra VST.N and $103 billion GE Vernova GEV.N have recouped some of their initial 30% and 20% share price slumps, they’re still rightly below their previous valuations.

Up until last week, AI had made usually staid utilities look like growth stocks. Big Tech players like Meta Platforms META.O are splurging on data centres that cost $10 billion each, to make their large language models (LLM) smarter. Goldman Sachs analysts estimate data centre power demand alone will jump by at least 160% by the end of the decade, to over 1,000 terawatt-hours (TWh) – roughly equivalent to the annual electricity consumption of Japan. Shares in U.S. nuclear power player Constellation Energy CEG.O, a potential AI winner, had soared 200% in the last year.

DeepSeek messes with these expectations. The company’s V3 model only had to train for 2.8 million chip hours, less than one-tenth of the level required to develop a comparable version from Meta. That could lower the overall need for AI training servers, or encourage AI players to use the older ones that powered DeepSeek. According to Goldman, a 10% cut in high-power server shipments could knock 5 percentage points off its assumptions for AI-related power growth by 2030, versus a 2023 base.

Vistra and rivals’ partial recovery by the end of last week likely reflects two factors. In publicly backing the $500 billion Stargate programme unveiled recently by OpenAI, SoftBank Group 9984.T and Oracle ORCL.N, President Donald Trump has signalled he wants to maintain the U.S. lead on the technology over China. Separately, the Jevons Paradox – whereby making an energy source more efficient and therefore cheaper increases its use – may apply to AI. Big Tech bosses certainly hope it does, and they may be right if AI usage by customers mushrooms.

Yet this user consumption would have to motor for energy providers to achieve the high volumes they might have hoped for, especially in the short term. Right now, developing AI models consumes more energy than deploying them for consumers. Training and fine-tuning Meta’s LLaMA 7B model, which is already much smaller than some of OpenAI’s, devours 55 megawatt-hours (MWh) of power, according to the Association of Data Scientists. Letting 1 million users interact with it once consumes just 0.1 MWh. If training now requires 10 times less power, then there needs to be 500 times as much usage to consume the same amount of energy.

Maybe that will happen. And whether AI usage explodes or not, electricity demand from sectors like new energy cars means data centres still only represent under half of the 2.4% annual power growth rate Goldman assumes up to 2030. That may explain why Vistra post-DeepSeek trades at 24 times its 2025 earnings, still comfortably above that metric’s 10-year average. But the S&P 500 Utilities .SPLRCU index’s level 2% below where it was beforehand also reflects that AI is less of a sure thing than it was.

Follow @ywchen1 on X

CONTEXT NEWS

Power groups slumped on Jan. 27 following the emergence of DeepSeek’s low-cost artificial intelligence models. U.S.-listed Vistra shares fell 28% while GE Vernova shares fell 22% and Constellation Energy lost 21%. In Europe, Siemens Energy fell 20%.

As of market close on Jan. 31, Vistra was down 12% from the Jan. 24 level to $168 per share. Similarly, GE Vernova closed 11% below Jan. 24 prices at $373 per share. Constellation Energy had lost 14% to trade at $300 a share. Siemens Energy is down 8% since Jan. 24 to 55 euros per share as of 1052 GMT on Feb. 3.

Blackstone, which has $80 billion worth of leased data centres, said on Jan. 30 its massive investments in data centres would not be undermined by the low-cost artificial intelligence models from China’s DeepSeek, as the need for physical infrastructure was still vital for AI.

CEOs of Microsoft and Meta Platforms also defended massive AI spending during earnings calls on Jan. 29 saying it was crucial to staying competitive in the new AI field.

(Editing by George Hay and Streisand Neto)

((For previous columns by the author, Reuters customers can click on CHEN/yawen.chen@thomsonreuters.com))