By Dean Seal

Electrical and mechanical engineering firms that help build data centers are caught in a broad market downturn after Chinese upstart DeepSeek said it developed top-performing AI models at a fraction of the time and cost of U.S. tech giants.

The revelation sparked investor fears that U.S. companies are overspending on a massive buildout of AI-related data centers, which created a wave of demand for companies such as Emcor Group and Comfort Systems USA that do electrical and mechanical work in data-center construction.

Shares of Emcor Group, which more than doubled last year, fell 17%, to $441.87, in early Monday trading. Comfort Systems USA's stock, which also doubled last year, was down 21%, at $428.01.

The stock of Everus Construction Group, which offers specialty contracting services in E&M, dropped 10%, to $68.80, in the early session. Limbach Holdings provides building solutions for mechanical, electrical and plumbing infrastructure; its shares tumbled 12%, to $90.05, after the opening bell.

All four companies have been tagged by analysts as beneficiaries of rising demand for data centers that power AI. About 60% to 70% of data-center-construction costs are tied to wiring, HVAC and other E&M work, compared with less than 30% of the costs for other non-residential buildings, according to Stifel analysts.

On top of that, AI-related data centers generate about four times the heat of their more traditional counterparts. As AI becomes a larger part of data-center workloads, spending on E&M should grow over time, Stifel analysts said in a research note last November.

The demand wave could start to break now that DeepSeek's advances have investors reconsidering how much computing power is actually needed to develop AI systems. The startup's models have earned high performance ratings without using the high-end computer chips that the U.S. hyperscalers, including Alphabet and Meta Platforms, have spent heavily on.

Write to Dean Seal at dean.seal@wsj.com

(END) Dow Jones Newswires

January 27, 2025 11:26 ET (16:26 GMT)

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