Power-Aware Scheduling of Distributed AI Training

The rapid expansion of AI infrastructure has led to the emergence of massive data centers with power demands exceeding 500 MW. Traditional data centers, originally designed as cloud storage solutions, have been fundamentally reshaped by the AI boom, both in their internal architecture and their external grid requirements. This project explores an underexamined consequence of this shift: the electrical "grid swings" caused by synchronous AI training workloads, and proposes a power-aware scheduling framework — built on grid-operator transparency combined with decentralized training architectures — as a path toward grid-compatible AI infrastructure.