When you drive down the right highway in rural Texas or northern Virginia, you’ll notice something difficult to ignore: enormous, windowless buildings stretching across cleared land, humming with the low industrial drone of cooling fans operating nonstop. There are no visible employees, no signs, and no clear reason. These data centers, which are subtly emerging as the decade’s defining infrastructure, are so hungry that the American electrical grid was never built to support them.
Even though the public is still not fully aware of the numbers, they are startling. Between 2023 and 2028, AI data centers alone could be responsible for 44% of the increase in the US electricity load. They are predicted to account for between 11 and 12 percent of all U.S. power consumption by 2030, up from the current 3 to 4 percent. Philadelphia’s electricity consumption is comparable to that of a single large facility. The equivalent of 50,000 homes can be powered by a single data center—not for a year, but continuously, every hour of every day. After you sit with the math for a while, it becomes uncomfortable.
More data centers aren’t the main cause of this; rather, it’s what those centers are currently doing. Conventional facilities needed 25 to 40 kilowatts per server rack to run cloud storage and e-commerce platforms, which was already a substantial load. AI workloads fall into a completely different category. Large language model training and operation require chips that produce heat on an almost violent scale, necessitating cooling systems that use massive amounts of power. Once thought of as a specialized solution, liquid cooling is now a key component of contemporary infrastructure design, with businesses creating direct-to-chip water loops and immersion tanks to prevent processors from self-destructing. Cooling infrastructure may become as expensive to construct and run as the computers it is meant to safeguard in a few years.
Earlier this year, Elon Musk stated bluntly that chipmakers might soon be producing more chips than the grid can actually power. That is not a far-off speculation. Nvidia is pushing processors that require power densities that no data center was designed to handle, and the development of facilities that can support them is lagging behind the rate at which chips are being produced. The physical world—transmission lines, substations, and generation capacity—does not scale at the rate of a semiconductor roadmap, despite trillions of dollars being poured into the development of AI.
The energy sector appears to be genuinely in disarray. Because they are unsure if renewable capacity can cover the time gap, utilities that had secretly planned to retire coal plants are delaying those retirements. Unpredictable power swings from AI clusters are already being reported by grid operators in a number of areas as a modeling issue, which is compelling engineers to reconsider how they model grid behavior during disruptions.
The carbon consequences are mounting in the meantime. According to a significant analysis, data centers produced over 105 million tons of carbon emissions, with fossil fuels providing more than half of their electricity. Data center emissions could triple by 2030, according to Morgan Stanley’s projections. The public pledges made by the majority of large tech companies regarding sustainability are incongruous with that trajectory.
The irony at play here is difficult to ignore. Infrastructure that consumes energy at a scale that the grid was not intended to support is being built by the same companies that are creating tools to help the world become more efficient. Some are making nuclear investments, such as Microsoft, which has agreements with nuclear operators, while others are considering small modular reactor projects that might or might not be completed in a timely manner. Geothermal energy is gaining popularity again. Nevertheless, the issue seems urgent and the solutions seem early.
The problem with data centers is not going away. Something will have to give somewhere between the aspirations of the AI boom and the physical constraints of the electrical infrastructure, and it’s still genuinely unclear which side bends first.
