Three gigawatts vanish in a second — the grid, not the fab, is the real compute ceiling

The compute conversation has been about chips for three years. A fallen power line dropping 3 GW of data-center load off PJM in one event is a reminder that the binding constraint moved, and that the new one behaves badly under fault.

More than 3 gigawatts of data-center demand disappeared from PJM simultaneously after a single fallen line. Not degraded — disappeared. That is a frequency-stability event, and the grid operator did not choose to carry that risk; it arrived with the buildout.

Loads are supposed to fail gracefully

Grid engineering assumes load that decays, varies and fails independently. Data-center load does none of those things. Protective equipment drops it fast, and it drops together because it is electrically and operationally correlated in ways that a city's worth of air conditioners is not.

With 190 GW of hyperscale capacity announced across 777 projects, this stops being an incident and becomes a design parameter. Generators have always had to specify ride-through behaviour. Loads at this scale will now have to as well.

The cost signal arrives separately

Virginia began taxing data-center electricity at $0.011/kWh on 1 July — the first serious attempt to make consumption carry a jurisdiction-specific operating cost. In the same month AirTrunk committed $21 billion to a 3 GW campus in Maharashtra. Capacity is becoming more expensive to host in some places and is moving to others, which is what a market does when a constraint binds.

The read

Chip supply was a constraint you could buy your way out of with enough capital and patience. Interconnection is not, because the thing being rationed is somebody else's stability margin. Expect the next round of compute scarcity stories to be about substations rather than wafers.

Quartz — AI data centers pass 1 gigawatt and strain the U.S. power grid → · Data Center Knowledge — New Data Center Developments: July 2026 →