On July 22, 2026, a fault in a power transmission line in Ashburn, Virginia—a region known for hosting the world’s largest concentration of data centers—suddenly removed over 3 gigawatts of electricity from the grid. This is equivalent to the power needs of about 2 million homes. This wasn’t an isolated incident. Two years earlier, a single failed electrical component caused a blackout affecting 60 facilities and 1,500 megawatts of power. These outages highlighted a growing concern: the way data centers now interact with the power grid is creating new, unpredictable risks. Most discussions about the power needs of artificial intelligence (AI) focus on generating more electricity—more wind turbines, more solar panels, more transmission lines. But the Virginia outages weren’t about a lack of power supply. They were about the structure of the grid itself. As AI data centers grow in size and number, they are placing new demands on the grid’s design, which was built for more predictable and slower-changing power needs. Traditional power systems were designed for loads like steel mills and homes, where electricity use changes gradually and predictably. AI data centers, however, have very different behaviors. During intensive AI training, a single campus can rapidly increase or decrease its power use by 70% in milliseconds. These sudden shifts can cause problems for the grid, especially when many such facilities are interconnected. The grid has never had to handle such massive, fast-moving loads before, and the next wave of data centers is being built with exactly these kinds of power demands in mind. Current data center power systems are outdated for this new reality. These systems typically bring in medium-voltage electricity, step it down using transformers, and then use low-voltage uninterruptible power supply (UPS) units to provide clean, stable power to the servers. However, at the scale of AI centers, these systems fall short. The UPS units, which are usually located inside the building, are not designed to handle rapid and continuous load swings. They are also often in bypass mode, allowing raw power to flow directly to the servers without filtering, which can lead to instability. Additionally, the grid’s protection systems—designed for smaller loads—can misinterpret sudden power changes and shut down critical infrastructure at the worst possible moment. To address these issues, engineers are proposing three key upgrades. First, move the voltage up to medium levels (13.8 kilovolts or higher), which is what large facilities typically draw from the grid. Second, relocate the power systems to modular enclosures near the substation, freeing up space inside the building for computing and cooling. Third, integrate the power system directly into the grid’s path, so every electron flows through a system that continuously manages load, rather than relying on batteries or switches to react after the fact. These changes transform data centers from a potential grid liability into an asset. When thousands of graphics processing units (GPUs) are in use, the system absorbs the load fluctuations, providing a steady power draw to the grid. When disturbances occur, the data center’s equipment remains unaffected, and the grid is not destabilized. These upgrades also simplify interconnections with the utility, reduce permitting time, and increase the efficiency of construction. Additionally, the new systems can qualify for tax incentives and generate revenue through grid programs like demand response, where data centers help balance power use during peak times. In early 2026, engineers tested a full-scale version of this system at the National Laboratory of the Rockies, a U.S. Department of Energy facility. The test involved simulating both real AI load profiles and grid faults, including a complete zero-voltage event. The system performed flawlessly, meeting and exceeding the grid reliability standards set by the Electric Reliability Council of Texas (ERCOT). This success shows that the new design not only meets regulatory requirements but also enhances grid resilience. The shift toward medium-voltage power systems in AI data centers represents a critical evolution in infrastructure. By adapting the power architecture, these facilities can become a source of strength rather than a strain on the grid. The changes not only improve reliability and efficiency but also open up new economic opportunities. As the next wave of AI data centers is built, the choice to adopt this new layer of infrastructure will determine whether these facilities are a burden or a benefit to the power grid.