Industry analysis identifies AI networking equipment as a critical growth vector as intra-cluster data movement emerges as the primary performance bottleneck.

Artificial intelligence continues to scale rapidly, but the primary constraint for AI companies may soon shift from raw chip power to interconnect speed. As systems grow more complex, the ability of processors to communicate efficiently is emerging as the critical bottleneck.

This conclusion follows Citi analysts’ review of presentations at the Hot Chips conference, which highlighted networking as the main performance limiter for increasingly sophisticated AI architectures. The shift carries significant implications for AI infrastructure equities. According to Citi, Nvidia, Broadcom, Arista Networks, Lumentum, Coherent, Marvell Technology, and Astera Labs are well-positioned to capitalize on rising demand for high-speed, efficient data movement.

This evolution signals a broader industry transition. While the initial AI infrastructure race focused on acquiring more powerful GPUs and accelerators to train increasingly complex models, the next phase will prioritize efficient processor interconnection. Early development cycles were straightforward: maximize compute capacity. However, as systems scale, complexity multiplies. Modern models feature massive parameter counts, expansive context windows, and sophisticated AI agent workloads, all of which demand continuous, high-volume data transfer across system components. A high-performance chip delivers minimal value if idle while awaiting data. Consequently, networking has become critical. AI architectures require constant data exchange between GPUs, memory, servers, and racks, often spanning multiple zones within large-scale data centers.

Presentations from Nvidia, Broadcom, Meta, Google's Alphabet, and Samsung reinforce this trajectory. “The future AI winner is not necessarily the company with the fastest processor, but the company that can move data most efficiently throughout the entire system,” the analysts said. This paradigm shift may redirect investor focus toward the next wave of AI infrastructure leaders. Nvidia remains a dominant force in AI computing, with its GPUs central to most modern data centers. Yet the company is simultaneously expanding its emphasis on the networking technologies that interconnect these chips. Citi specifically highlighted Nvidia’s concept of “AI factories”—next-generation facilities engineered not merely to house servers, but to optimize data flow across compute, memory, networking, storage, and security layers. This architecture elevates the strategic importance of firms like Broadcom and Arista Networks. Broadcom supplies core networking and connectivity solutions for large-scale deployments, while Arista Networks specializes in high-speed switching infrastructure. Additional players stand to benefit across complementary segments: Lumentum and Coherent provide optical technologies enabling ultra-high-speed data transmission; Marvell Technology develops connectivity and data infrastructure products; and Astera Labs specializes in interconnects between processors, memory, and peripheral components.

Memory integration is equally critical. Though distinct from networking, the two challenges are deeply intertwined. Locating data closer to the processing unit reduces reliance on congested interconnects, lowers communication overhead, and enhances overall system efficiency. Firms including Samsung, XCENA, and Cerebras are actively developing technologies to co-locate data near computation engines.

This reality requires investors to evaluate AI infrastructure as an integrated ecosystem rather than a fragmented set of components. Peak GPU performance alone will not dictate the efficacy of large-scale AI deployments. Memory capacity matters. Memory bandwidth matters. Networking speed matters. Optical connections matter. The software controlling all of it matters too. Citi’s analysis underscores this as a pivotal transformation in the AI infrastructure landscape. Consequently, the next phase of the AI expansion will likely diverge from its predecessor. Rather than prioritizing raw processor speed, capital allocation will increasingly favor firms that enable seamless coordination across thousands of processors, eliminating latency caused by data bottlenecks. As models expand and data centers evolve from scattered server clusters into highly optimized information-moving machines, AI networking is poised to emerge as one of the most consequential infrastructure themes ahead.