Summary of Key Points from Conference Call Industry Overview - The conference call discusses the transition in AI network interconnect technology from Scale Out to Scale Up, which is driving the development of copper connections and orthogonal board technology. Future discussions will focus on in-cabinet optical connection solutions, such as NVIDIA's Ruby Ultra and Google's I/O Superpod supernode designs [1][4]. Core Insights and Arguments - **High-Density Cabinets**: NVIDIA's Ruby Ultra high-density cabinet enhances performance by increasing card density, utilizing orthogonal boards and copper wires for internal interconnects. The design splits 576 cards into four frames, achieving efficient data transmission through a Dragonfly network architecture, with a potential future shift towards optical connections [1][6]. - **Scale Up vs. Scale Out**: Scale Up focuses on streamlining protocol stacks for high-speed interconnects between AI computing cards, suitable for parallel computing. In contrast, Scale Out handles more complex tasks and typically uses public or semi-public protocols for broader communication [2]. - **Bandwidth Elasticity**: Different manufacturers exhibit significant differences in SKA optical interconnect bandwidth elasticity. For instance, NVIDIA's SKU to OUT rate ratio is approximately 9 times, while Huawei's Cloud Matrix is 7 times, and Alibaba's is 8 times. This reflects variations in UP and OUT layer structures and average selling prices (ASP) [3][9]. - **Market Trends**: The optical connection technology is expected to see significant advancements in AI networks, with a shift towards Scale Up designs anticipated to drive the development of optical connections by 2025. This includes discussions on in-cabinet optical connection solutions and supernode designs [4][5]. - **Cost and Performance Comparison**: The cost of optical modules is decreasing, with silicon optical solutions around $0.5/Gbps and EML solutions at approximately $0.6/Gbps. In contrast, copper connections, while initially cheaper, will see rising costs with upgrades [10][11]. Additional Important Insights - **Future of Optical Modules**: The optical module market has a promising outlook, with expectations for continued ASP decline. Emerging technologies like NPO and CPO are projected to lower costs further, impacting the industry's value and participation levels [12]. - **CPU's Role in Ecosystem**: The CPU's position in the ecosystem is debated, with its market share potentially affecting net margins. The transition to optical engines like NPU or future CPUs could enhance profitability [13][14]. - **Market Size Differences**: The market size for Scaler is currently larger than for Skyout, but as costs decrease, this gap may narrow. SkyUp is seen as a significant blue ocean market that is gaining attention [15]. - **Inflation Logic in Optical Module Industry**: The optical module industry exhibits significant inflation logic, with Sparse optical connections providing new opportunities. However, short-term EPS adjustments are not expected to be substantial until 2026 and beyond [16]. - **PE Expectations**: The industry is currently undervalued with a PE of around 20 times. Given the super inflation logic, a future PE of at least 30 times is anticipated, potentially reaching 40 times depending on valuation timing [17].