AI is moving into heavy industry
Recent market commentary argues AI is no longer just software — the buildout is shifting attention to the physical backbone: power, cooling, networking, site construction and heavy equipment. That matters for investors because second‑order winners may live in industrial and utility suppliers rather than only GPUs and cloud providers, changing where future capex will flow (podcast and market analysis highlighted this shift). ( )
The hottest part of the artificial intelligence boom is no longer the chip. It is the room around the chip: the substation, the chiller, the switchgear, the pipes, the backup generators, and the crews that can build all of it before the next lease starts. (goldmansachs.com) That shift shows up in the power bill first. Goldman Sachs Research said United States spending on data center construction has tripled in three years, and the bottleneck is no longer just getting more graphics processing units but getting enough electricity and cooling into each building. (goldmansachs.com) An artificial intelligence server works like a sports car engine stuffed into a small garage. The more graphics processing units you pack into one rack, the more heat you have to remove every second or the machines slow down or shut off. (vertiv.com) That is why old air conditioning is giving way to liquid cooling. Schneider Electric said in September 2025 that it was rolling out a full liquid-cooling lineup for high-density artificial intelligence sites, including coolant distribution units, rear-door heat exchangers, chillers, and cold plates. (se.com) Vertiv made the same bet from the other side. In June 2025 it launched three new coolant distribution units for liquid-cooled data centers, which is the plumbing-and-pump layer that sits between the building and the servers. (datacenterdynamics.com) Power is the harder problem, because you cannot cool a server that never turns on. The Electric Power Research Institute said last month that data centers are now the fastest-growing source of United States electricity demand and could consume 9% to 17% of U.S. power generation by 2030 in its scenarios. (epri.com) The U.S. Department of Energy put a shorter clock on the same trend. Its 2024 report said data center load growth has tripled over the past decade and could double or triple again by 2028 as artificial intelligence use expands. (energy.gov) Once that happens, the winners stop looking like pure software companies and start looking like utilities, turbine makers, electrical-equipment suppliers, and industrial contractors. BlackRock said in its infrastructure outlook that data centers and the grid energy needed to run them are moving into the same investment bucket. (blackrock.com) The clearest sign is that technology companies are now signing power deals that would have sounded absurd two years ago. Constellation said on September 20, 2024 that Microsoft agreed to a 20-year contract supporting the restart of Three Mile Island Unit 1, adding about 835 megawatts to the grid. (constellationenergy.com) Amazon went even bigger in Pennsylvania. Talen said Amazon Web Services signed a power purchase agreement tied to the Susquehanna nuclear plant that can ramp to 1,920 megawatts by 2032, which is utility-scale power contracted for cloud and artificial intelligence demand. (utilitydive.com) Google is placing a longer-dated bet on new reactors instead of restarted ones. Google and Kairos Power said in October 2024 that they plan a fleet of advanced nuclear projects totaling 500 megawatts by 2035 for future electricity supply. (kairospower.com) So the artificial intelligence buildout is starting to resemble a railroad boom more than a software launch. The companies selling the shovels now include firms that make transformers, switchgear, cooling loops, diesel backup systems, and the heavy equipment that turns an empty field into a 100-megawatt machine room. (blackrock.com)
