Metas engineering division published a summary of its annual @Scale: Networking conference on September 26, 2025, which this year was entirely dedicated to building network infrastructure for large-scale AI computing. The event featured leading global companies, including NVIDIA, Google, AMD, and Cisco. The key theme was the architecture of "network fabrics"—specialized networks designed to connect tens of thousands of GPUs into a single cluster. Discussions focused on technologies that enable ultra-low latency and maximum throughput, which are critical for the efficient training of giant AI models. The review emphasizes that as AI clusters grow, the complexity and importance of network solutions are becoming comparable to the computing chips themselves, and this is where key innovations that will define the future of AI are happening now.
Metas @Scale: Networking 2025 Recap — Networks for AI Clusters
