Powering AI is an architecture problem

A recent report from MIT Technology Review highlights that the massive energy demands of AI data centers are creating critical infrastructure challenges. Rather than just a supply issue, the problem lies in the outdated power architecture of these facilities. AI workloads are highly volatile, with power consumption capable of swinging significantly in milliseconds. Current data center designs, which rely on legacy uninterruptible power supply (UPS) systems and outdated protection logic, are failing to handle these rapid fluctuations. This mismatch has led to grid instability, including major load drops in Virginia. Experts propose a fundamental redesign of the power stack, moving from low-voltage systems to medium-voltage modular enclosures located near substations. By implementing systems that process power continuously rather than reacting to faults, data centers can transform from unpredictable grid neighbors into stable, reliable partners for utility providers, ultimately preventing future large-scale outages.