Powering AI is an architecture problem
MIT Technology Review Ricardo De Azevedo
Virginia’s AI data centers just knocked gigawatts off the grid twice. The real problem isn’t power supply — it’s that the old setup can’t handle AI’s wild swings.
Based on reporting by MIT Technology Review, Ricardo De Azevedo — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
On July 22, 2026, a transmission line fault in Ashburn, Virginia sent more than 3 gigawatts of load off the grid in seconds. It wasn’t a one-off. Two years earlier, a single failed surge arrester cut roughly 60 Virginia facilities and about 1,500 megawatts at once. The pattern is ugly: these sites can all react the same way, all at once, to the same grid problem.
That’s why the AI power fight is missing the point. People keep arguing about generation — more turbines, more solar, more transmission, more electrons. But the Virginia outages weren’t caused by a lack of supply. They were caused by the way the load is built. The grid was designed for steel mills, refineries, and homes at dinner time. AI campuses are different beasts. One can swing 70% of its load in milliseconds during training, then shut itself down just as fast if it senses trouble upstream.
The standard power stack inside data centers was never built for that kind of behavior. Power comes in at medium voltage, gets stepped down, passes through a UPS, and ends up at the racks. Push that setup to AI scale and it starts breaking in familiar places. The batteries inside the UPS are only meant to bridge a short outage, not smooth constant, violent swings. Legacy converters waste enough power that operators often run in eco-mode, which leaves the racks effectively tied straight to the grid. And the protection logic was written for a world where “large load” meant 50 megawatts, not a cluster this size.
So when trouble hits upstream, the system often does the wrong thing and drops out. In the 2024 Virginia event, most of the lost load came from protection schemes that were doing exactly what they were designed to do: count voltage dips and disconnect on the third one. That was supposed to protect equipment. At this scale, it just makes the grid look weaker.
The fix is to move the power layer up, out, and into the path. Up from 480 volts to medium voltage, out of the data hall and near the substation, and into a system every electron runs through all the time instead of one that only wakes up when something goes wrong. The point is simple: absorb the swing, hand the grid a flat load, and stay online through faults. In early 2026, a full-scale test at the National Laboratory of the Rockies hit both the compute side and the utility side, including a full zero-voltage event, and the system cleared ERCOT’s large-load voltage ride-through requirements with room to spare.
That matters because the next wave of AI campuses is already being planned at this scale. A medium-voltage inline design makes permitting easier, cuts out some interconnection pain, and can turn backup power into something that earns money in grid programs like peak shaving and demand response. The industry keeps treating this as a grid problem. It’s also a hardware problem, and one the old stack plainly can’t solve.
My take — AI-written commentary, not fact-checked reporting
The industry loves to talk about generating more power, because that sounds grand and expensive and safely someone else’s problem. But a lot of this mess is self-inflicted by data centers built like they’re still serving ordinary IT, not giant, twitchy AI loads. The boring answer is usually the right one: fix the architecture first, then complain about the grid.
Read more about this at: MIT Technology Review