Why Power-Flexible AI Just Became Table Stakes

AI infrastructure's real constraint: power, not chips. Aurora AI Factory (NVIDIA/Emerald, Virginia, 96MW, 2026) implements interruptible compute—software layer throttles training during grid stress, delivers 20-30% reductions, maintains SLAs. Enables faster permitting, lower capacity charges, wholesale market participation. Trade-off: longer training times for cheaper power. Key question: will two-tier market emerge (fixed for inference, flex for training)? 100GW capacity unlock claim assumes perfect coordination—directionally correct but optimistic. Power flexibility now mandatory for deployment. Due diligence questions: demand response capability, interruptible/fixed split, interconnection impact.