Microsoft's Partha Maji discusses full-stack co-design, custom silicon, and optimizing AI infrastructure beyond GPUs to solve memory bottlenecks and reduce total cost of ownership.