Google's AI Infrastructure Chief, Amin Vahdat, on the Physics & Economics of Frontier AI

The massive capital expenditure build-out for AI infrastructure centers on the shift toward highly specialized, purpose-built data centers. Unlike traditional facilities designed for 30-year lifespans, modern AI data centers require co-designing buildings with hardware to handle extreme power densities, often reaching megawatts per rack. Performance measurement has shifted from theoretical chip-centric metrics like FLOPs to "goodput," which accounts for real-world reliability and failure recovery at the 100,000-accelerator scale. Google’s TPU program leverages deep partnerships with DeepMind to co-optimize model architectures and hardware, allowing for in-flight adjustments during the design pipeline. As power emerges as the primary binding constraint, the industry faces the challenge of balancing grid-connected utility partnerships with the need for massive, reliable energy, while exploring future innovations like orbital data centers to overcome terrestrial limitations.