GTC前夜：光模块，正在成为AI算力最被低估的主线

Core Viewpoint - The AI hardware investment focus is shifting from GPU performance to the efficiency of data flow between chips, servers, and data centers as the limits of computational power are approached [1][2]. Group 1: Transition of Computational Bottlenecks - The bottleneck in computational power is transitioning from computation to communication, particularly in large-scale AI training where data exchange between GPUs is exponentially increasing [6]. - In AI clusters, network bandwidth, latency, and power consumption are becoming critical variables for training efficiency, indicating a fundamental change in the core logic of AI computing networks [6][7]. Group 2: Emergence of Optical Communication - Traditional data center networks are designed with excess bandwidth, but AI clusters require high-frequency collaboration among GPUs, maintaining network utilization above 80%, making bandwidth bottlenecks and latency fluctuations detrimental [6][7]. - The upcoming NVIDIA GTC conference is seen as a pivotal moment for AI interconnect technology, with a focus on network architecture upgrades for both Scale Up and Scale Out strategies [6][7]. Group 3: Innovations in Optical Modules - The limitations of traditional optical module architectures are becoming apparent, including high power consumption, bandwidth constraints, and significant signal loss over long distances [9]. - New technological routes, CPO (Co-Packaged Optics) and NPO (Near-Packaged Optics), are being discussed as solutions to these issues, with CPO expected to reduce interconnect power consumption by 30-50% [10][11]. Group 4: NVIDIA's Strategic Moves - NVIDIA's recent $4 billion investment in optical communication companies Coherent Corp and Lumentum is viewed as a supply chain locking strategy to secure optical engine supply amid anticipated demand surges [17]. - The expected introduction of the Rubin Ultra architecture could significantly increase the number of optical engines per GPU, from approximately 1.5 in the H100 architecture to about 5.5, indicating a shift in the role of optical engines from auxiliary components to core bottlenecks [18][19]. Group 5: Market Implications - If the GTC conference confirms the new architecture, the valuation framework for the optical module supply chain may need to be re-evaluated, as traditional metrics may underestimate the technological premium and concentration in the CPO era [20]. - The AI investment narrative is evolving, with a potential shift from GPU-centric strategies to recognizing the critical role of optical communication infrastructure in AI hardware [21][22].