Google says 48V power standard hits limits as AI racks scale to megawatts

At the OCP APAC Summit held in Taipei on August 11, Google, through executive Gregory Moore, presented a keynote focused on AI data center infrastructure challenges. The core argument centers on a fundamental limitation of the 48V power architecture that has been standard across the data center industry: as artificial intelligence server racks grow toward megawatt-scale power consumption, the electrical components required by the 48V design—specifically converters that step voltage down from the main distribution and busbars that route power across the rack—consume increasing amounts of physical space. This space consumption becomes a practical constraint when every inch of a data center rack is needed for compute, memory, and cooling systems. Google's framing suggests that the transition to megawatt-scale AI racks is not merely a matter of adding more power capacity to existing infrastructure, but rather a fundamental architectural shift. The 48V standard, while proven and widely adopted across the industry, was designed for earlier generations of server density and power consumption. AI workloads operating at the scale Google and other hyperscalers are now deploying exceed the practical headroom of this older standard. The company's decision to raise this issue publicly at an industry summit, rather than in closed vendor discussions, indicates a belief that the problem requires broad industry recognition and coordination to solve. Future AI data center designs may require higher-voltage power distribution systems, alternative power delivery architectures, or other innovations currently under development to accommodate the power and thermal demands of next-generation AI infrastructure.