A Gigawatt of Grid Relief Needs Workloads That Can Stop, in a Bite

OpenAI’s Grid Protectors Program promises compute campuses that cut their power when the grid strains. Its Georgia site commits up to 1,000 megawatts on a 3,200 megawatt campus, and the arithmetic behind that number is where it gets interesting.

• Ember expects 87 percent of data center capacity in 2030 to run cloud and AI inference, workloads that break a service level agreement if they stop. Training is the other 13 percent, and 25 to 55 percent of a training campus can flex.
• Run those assumptions on Effingham and the pausable slice is 104 to 229 megawatts, against a promise worth 31 percent of the campus. That calculation is ours, on Ember’s published range.
• The proof point is a 96 GPU Blackwell Ultra cluster that shed about 30 percent in 30 seconds. A 3,200 megawatt campus needs a checkpoint schedule, not a switch.
• Ember’s Texas case puts on site generation at 2.7 gigawatts of relief against 0.55 gigawatts from shifting load, so much of the promised capacity may never involve an idle GPU.

The number to watch is the pause budget, the share of a campus that can stop mid job. Read the full analysis.