Summary: This article from Microsoft Research identifies memory and network bottlenecks as fundamental limits on AI system performance, reducing utilization of GPU clusters in large‐scale AI workloads. It explains the traditional trade-off between power, reach and reliability in copper vs optical links and introduces MOSAIC, a novel optical link design using many parallel low-speed channels and microLED transmitters to provide low power, high reliability, and extended reach simultaneously. This research aims to overcome current datacenter connectivity constraints and enable more efficient scaling of AI infrastructure. 
Source: https://www.microsoft.com/en-us/research/blog/breaking-the-networking-wall-in-ai-infrastructure/