Powering AI is an architecture problem
The article opens with two concrete grid events in Ashburn, Virginia, the world's largest data center cluster. On July 22, 2026, a transmission line fault knocked more than 3 gigawatts of load off the grid in seconds; two years earlier, a single failed surge arrester dropped roughly 60 Virginia facilities and 1,500 megawatts at once. The striking part, the author says, is that no one anticipated so much uniform load responding to grid faults the same way at the same time. The AI power debate is usually framed around generation — more turbines, more solar, more transmission — but the Virginia outages were not supply failures, they were architecture failures, and a large wave of new interconnections is arriving on that same architecture, putting reliability at risk.
The mismatch is explained by how differently AI loads behave. The grid was built around predictable loads — steel mills, refineries, houses at dinnertime — which draw power smoothly, misbehave occasionally, and recover gracefully. An AI campus can swing 70% of its load in milliseconds during a training run, then trip offline just as fast at the first sign of trouble upstream to protect billions of dollars in compute. Each response is rational in isolation, but together at gigawatt scale they create a problem the grid has never solved, and the next wave of data center campuses is planned at exactly that scale.
The standard data center power stack — medium-voltage power arriving, transformers stepping it down, low-voltage UPS units conditioning it, and delivery to the racks — has not changed in decades and cracks in three places at AI scale. First, the UPS sits deep inside the building near the racks, but its batteries are an undersized "spare tire" meant for a few minutes of outage, not for absorbing load swings this fast and volatile around the clock. Second, the UPS spends most of its life in bypass: legacy converters waste enough power that operators run in eco-mode, with a static switch feeding racks directly from the grid and nothing filtering in either direction, so compute swings go out raw and sub-millisecond grid transients arrive too fast for any switch to catch. Third, the protection logic was written when "large load" meant 50 megawatts and cannot see the grid it is now part of, so when trouble hits upstream it does exactly the wrong thing and drops out — in the 2024 Virginia event, most of the lost load traced to protection schemes that count voltage dips and disconnect on the third one, operating as designed at the worst possible moment.
The author stresses this is not sloppy engineering but careful engineering that the load has outgrown. The proposed remedy is to move power protection up the voltage stack — outside the building and into the power path — which the piece says would not just solve outages but also change density, permitting timelines, and backup power economics.