Networking capacity has overtaken raw processing speed as the primary constraint on artificial intelligence performance, according to Citi analysts following the second day of the Hot Chips conference.
The shift reflects the growing complexity of AI workloads, which now demand data movement across trillions of parameters, larger context windows and autonomous agent operations. Traditional data center designs, organized around servers, are giving way to facilities modeled on Nvidia’s vision of "AI factories"—integrated systems where compute, memory, networking, storage and security are optimized for data flow rather than standalone hardware.
Citi identified seven companies positioned to benefit from this transition: Nvidia, Broadcom, Arista Networks, Lumentum, Coherent, Marvell Technology and Astera Labs. Presentations at the conference also highlighted the trend among Nvidia, Broadcom, Meta Platforms, Alphabet’s Google and Samsung, alongside discussions of memory and data-locality technologies such as XCENA and Cerebras.
The first wave of AI infrastructure prioritized faster GPUs, larger accelerators and expanded training capacity. However, as models scale and workloads diversify, the bottleneck has shifted to the efficiency of data movement within and between systems. Citi analysts emphasized that the future leaders in AI may not be those with the fastest processors, but those capable of moving data most efficiently across the entire system.
The concept of AI factories underscores the industry’s pivot toward holistic system design, where networking infrastructure becomes the critical layer enabling performance gains that raw compute alone can no longer deliver.
