The Grid-Aware Facility
A data center used to promise the grid that it would never move. Some AI work can move, because training pauses at checkpoints, and utilities have started to value that. Google's 2025 agreements with Indiana Michigan Power and the Tennessee Valley Authority, the Phoenix demonstration by Emerald AI and its partners, and Texas Senate Bill 6 all point the same way. The same automation carries risk: NERC's review of the July 2024 Virginia event showed many independent controls acting together in a way no one planned. Faraday's read is that proven flexibility will first pay off as faster access to power, and that the right test of any control-plane claim is the evidence ladder, from design to demonstration to contract to operating record.
Knowledge check
What property of AI training workloads makes demand response feasible, and what does the control plane have to do to honor a utility's request?
Show the answer
Training jobs can checkpoint and resume, so they can be slowed or paused for hours at a known cost, unlike live serving. To honor a request, the utility signal must reach the scheduler, which decides which jobs to slow or pause; the accelerators reduce power; and electrical monitoring confirms the reduction back to the utility.