FlexCast predicts grid power a week ahead and plans datacenter campuses around it. Cheaper for the datacenter. Better for the American grid.
The red line is FlexCast calling Texas demand a day ahead through all of 2025, a year the model never trained on; the black line is what actually happened.
FlexCast plans around whatever site you give it. This is the demo: a 500 MW AI campus near Abilene, in ERCOT West. Slide anything and the optimizer re-solves the week for your version, live.
The grid connection ERCOT has to approve. Above 75 MW, SB6 says you must be able to shed for the grid.
Inference, serving real users around the clock. FlexCast never touches it.
The tank, and how hard it can pour. 200 MWh at 100 MW is about two hours at full blast.
How hot the GPU fleet runs. More work means a bigger bill and less slack to move around.
python -m engines.serve
One campus, one week of real prices and grid load. Pick any hour to see what the model expects and what the plan does about it.
Hover any hour for the readout. Click any shape to see what it is.
The optimizer's schedule, one column per day. Click any block to read it, or press Play to watch the week unfold.
Texas bills big users on the four busiest hours each summer, unknowable until month end. The real ones:
The levers, the obligations, the assumptions.
Load sheds in steps, cheapest first:
Each job owes hours of compute by a deadline. The plan moves when they run, never whether: