A transmission line fault in Ashburn, Virginia, on July 22, 2026, knocked more than 3 gigawatts of load off the grid in seconds. This incident, part of a growing trend, underscores the grid's struggle to handle the volatile power demands of AI data centers.
The grid was designed for predictable loads like steel mills and houses at dinnertime. However, AI data centers can swing 70% of their load in milliseconds during training runs, then trip offline at the first sign of trouble to protect compute resources.
The standard data center power stack, unchanged for decades, struggles with AI's scale. UPS systems, designed for short outages, can't absorb the fast, volatile load swings. Legacy converters waste power, leading to eco-mode operation where the grid feeds the racks directly without filtering.
"No one could anticipate so much uniform load responding to grid faults the same way, at the same time," said a source. This unpredictability is a systemic issue as AI campuses scale to gigawatt levels, putting grid reliability at risk.
The solution involves three key upgrades: moving power up to medium voltage, moving it out to modular enclosures near substations, and moving it into the path to absorb load swings. These changes rewrite every line item downstream, improving density and reducing permitting time.
Source: mittr
