Powering AI is an architecture problem

AI data centers require massive amounts of electricity, often measured in gigawatts (GW), to power their high-performance computing hardware. For example, a single transmission line fault in Ashburn, Virginia—a major hub for data centers—suddenly removed over 3 GW of load from the grid in seconds in 2026. Two years earlier, a single failed component caused 1,500 megawatts (MW) to drop across 60 facilities. These incidents were not due to a lack of power supply but because the grid's design cannot handle the sudden, synchronized power demands of AI data centers. Traditional grids were built for predictable loads like factories or homes, which draw power smoothly and recover from disruptions gradually. AI data centers, however, can change their power consumption by up to 70% in milliseconds during training runs or shut down instantly to protect equipment, creating unpredictable and extreme fluctuations that the grid was not designed to manage.

The standard power system for data centers has not evolved in decades and fails when scaled to AI demands. The system involves medium-voltage power entering the facility, transformers stepping it down to lower voltages, and uninterruptible power supply (UPS) units conditioning the electricity before it reaches the computing racks. At AI scale, this system breaks in three ways. First, UPS batteries are too small to handle rapid, large-scale power swings and are only meant to provide backup for a few minutes during outages. Second, UPS units often operate in bypass mode, running directly from the grid to save energy, which means they do not filter power fluctuations or protect against grid disturbances like voltage dips or spikes. Third, the protection logic in these systems was designed for smaller loads (e.g., 50 MW) and cannot properly respond to modern grid conditions. For instance, in 2024, many data centers in Virginia disconnected after the third voltage dip, following their outdated protection schemes, worsening grid instability.

To address these issues, a three-part architectural solution is proposed. First, move it up by increasing the voltage from 480 volts to medium voltage (13.8 kilovolts or higher), which is the level at which large sites draw power from the grid. This reduces energy loss during transmission and makes the system more efficient. Second, move it out by relocating power conditioning equipment from inside the data center to modular enclosures near the substation, freeing up space inside the facility for more computing hardware or cooling systems. Third, move it into the path by replacing traditional UPS batteries with a system where every electron of power passes through a central medium-voltage unit continuously, eliminating the need for detection or switching mechanisms. This design ensures the grid receives a stable load profile even when AI data centers experience rapid power changes, and it prevents disruptions from affecting the grid or the data center.

The proposed architectural changes offer multiple benefits. By stabilizing power demand, AI data centers become predictable neighbors for the grid, reducing the risk of outages. Utilities benefit because they only need to certify one medium-voltage box per facility instead of managing complex interconnections for every transformer, UPS, and switchgear. This streamlines permitting processes, potentially cutting months off project timelines. Inside the data center, space previously used for UPS rooms can now house additional computing or cooling equipment, increasing efficiency and reducing construction costs. Economically, medium-voltage energy storage systems may qualify for tax credits and can generate revenue through grid programs like peak shaving (reducing power use during high-demand periods) and demand response (adjusting power use based on grid conditions). Backup power systems shift from being a costly insurance policy to a revenue-generating asset.

In early 2026, a full-scale test of the new medium-voltage system was conducted at the National Laboratory of the Rockies, a U.S. Department of Energy facility capable of simulating real grid faults and AI-scale load swings simultaneously. The system was tested under extreme conditions, including a full zero-voltage grid event and real AI load profiles at full medium voltage. The results showed that neither the compute side nor the grid side was affected, and the system met the large-load voltage ride-through requirements set by the Electric Reliability Council of Texas (ERCOT), a major grid operator. These requirements exist because operators no longer accept large facilities on trust alone, and more AI data centers are being built. The new architecture meets these requirements by design, making compliance a built-in feature rather than an added challenge.