Paul Gilbert

This talk focuses on the infrastructure required to build and operate AI data centers, specifically addressing the networking challenges and solutions for training and inference. It highlights the significant differences between traditional data center networking and the demands of AI workloads, emphasizing the need for high bandwidth, low latency, and specialized traffic management. The core thesis is that building effective AI data centers requires a fundamental shift in network design and implementation to handle the unique, high-intensity demands of GPUs.