AI's Next Bottleneck Is Networking, Not Just Compute, Citi Says

The primary bottleneck for advancing artificial intelligence is shifting from raw computing power to network interconnectivity, according to analysts at Citi. In a note following the Hot Chips technology conference, the bank's analysts concluded that as AI systems grow in scale and complexity, the ability to move data efficiently is becoming more critical than processing speed alone.

From Compute to Connectivity

The initial phase of the AI infrastructure build-out was heavily focused on increasing the speed and scale of GPU accelerators for training models. However, as AI evolves toward trillion-parameter models, expanded context windows, and autonomous agent workloads, the challenge has moved to efficiently transferring data between chips, servers, racks, and entire data centers.

Presentations at the conference from major tech players—including Nvidia (NASDAQ: NVDA), Broadcom (NASDAQ: AVGO), Meta Platforms (NASDAQ: META), and Alphabet's Google (NASDAQ: GOOGL)—all pointed toward this trend, Citi noted. "The winners in AI going forward may not be the ones with the fastest processors, but the ones who can move data most efficiently through the entire system," the analysts stated.

Market Implications and Beneficiaries

This shift stands to benefit companies that provide the essential components for high-speed data transfer. Citi identified several companies as clear positives from this trend:

Ad

• Nvidia (NASDAQ: NVDA)
• Broadcom (NASDAQ: AVGO)
• Arista Networks (NYSE: ANET)
• Marvell (NASDAQ: MRVL)
• Astera Labs (NASDAQ: ALAB)
• Optical component makers Lumentum (NASDAQ: LITE) and Coherent (NYSE: COHR)

Memory is also becoming a key part of the interconnect challenge. Technologies showcased by Samsung and others aim to keep more data located near the processing units, reducing the amount of information that must travel through congested interconnect pathways.

The 'AI Factory' Concept

Citi highlighted Nvidia's concept of the future data center as an "AI Factory." Unlike traditional server-centric architectures, these new facilities are designed around data flow, integrating compute, memory, networking, storage, and security into a cohesive system.

The bank believes this architectural evolution could shift more of the value in AI infrastructure toward companies that can optimize the entire data transfer link, rather than those who supply only standalone processors. Every byte of data stored and processed locally reduces communication overhead, making memory capacity, bandwidth, and efficiency integral to network performance.