How networking became the critical layer of AI

As AI models grow larger and more complex, Citi analysts say the limiting factor is shifting from raw compute to how efficiently data moves across the AI system. Speaking after the second day of Hot Chips, they pointed to a first wave that emphasized faster GPUs, larger accelerators, and more training capacity. Now, with systems expanding toward trillions of parameters, larger context windows, and autonomous agent workloads, networking and data transfer between chips, servers, racks, and data centers are becoming a bigger bottleneck. Citi sees positive implications for Nvidia, Broadcom, Arista Networks, Lumentum, Coherent, Marvell, and Astera Labs, and noted memory is increasingly part of the networking challenge.