AI is running out of electricity, and we can't build our way out. By 2030, U.S. data centers will consume more power processing data than manufacturing all aluminum, steel, and cement combined. The grid is sold out with years-long queues, while AI's killer apps remain impossible under current power budgets. The real bottleneck: moving data between memory and compute accounts for 60-80% of AI's power consumption, and incremental improvements won't cut it. The race to build architectures that eliminate data movement entirely is on.
