DeepSeek Wipes Out a Year of AI Power Gains in a Day

For a year, the market treated U.S. power producers as AI stocks. On January 27 it sold them like AI stocks too. Vistra closed nearly 30% lower, erasing its gains for 2025, according to CNBC. Constellation Energy, Talen Energy and GE Vernova each fell more than 20%. Gas producer EQT lost nearly 10%, and pipeline companies Kinder Morgan and Williams fell more than 8%. Advanced nuclear developers Oklo and NuScale dropped more than 20%, Utility Dive reported.

DeepSeek released a model on Christmas Day and followed it last week with a reasoning model, DeepSeek-R1, that competes with OpenAI's o1. The companies that fell hardest were the ones whose valuations had come to rest most heavily on data center demand. Before Monday, Constellation, Vistra and GE Vernova had led the S&P 500 as investors bet that AI data centers would need enormous amounts of electricity. Utility Dive noted that Vistra and Talen remained more than twice as expensive as a year earlier, and that Constellation was still up about 127% year on year.

Jefferies' power and utilities team wrote that DeepSeek's success "calls into question the significant electric demand projections for the U.S.," because AI represents about 75% of overall U.S. demand forecasts through 2030 to 2035 in most projections. Bank of America analysts said DeepSeek was challenging the notion of U.S. leadership in AI and raising doubts about high expectations for cloud capital spending, chip growth and power requirements, according to CNBC. Utility Dive cited a blog post by technology investor Jeffrey Emanuel estimating that DeepSeek's models may have cost 45 times less to train than leading U.S. products.

ICF's Himali Parmar told Utility Dive that it is too early to tell whether projections account for model efficiency gains. For grid operators, the event is a reminder that data center demand is the least certain part of every load forecast.