《广发海外电子通信》 ☄GTC 2026前瞻：LPX、CPO及PCB关键亮点 ☀LPX机架强化英伟达在推理领域的产品组合：如先前所述，LPX（又称LPU）预计将采用基于SRAM的片上内 存，提供快速Token生成及超低延迟，提升英伟达在推理领域的地位。 在英伟达于2025年12月与Groq达成非独家许可协议前，该机架设计将搭载64个Groq LPU，采用RealScale芯片间互 连。 《广发海外电子通信》 ☄GTC 2026前瞻：LPX、CPO及PCB关键亮点 ☀LPX机架强化英伟达在推理领域的产品组合：如先前所述，LPX（又称LPU）预计将采用基于SRAM的片上内 存，提供快速Token生成及超低延迟，提升英伟达在推理领域的地位。 在英伟达于2025年12月与Groq达成非独家许可协议前，该机架设计将搭载64个Groq LPU，采用RealScale芯片间互 连。 对于GTC 2026，我们预计增强型LPX机架将搭载256个LPU，采用多层52LM9 Q-glass PTHPCB，单个LPU对应PCB 价值量约200美元。 因此，我们预计PCB、液冷（冷板）公司将受益。 我们预计英伟达将于GTC 2026 ...

Core Viewpoint - Starting from 2025, supernodes will become a significant technological innovation direction in the AI computing network, with increasing competition among AI chip manufacturers in both chip performance and Scale up network [1][5]. Group 1: Supernode Development - Nvidia has launched mature supernode solutions, with plans to release GH200 NVL72, GB200/GB300 NVL72, and VR200 NVL72 from 2024 to 2026 [1][3]. - The Blackwell architecture standardizes Scale up with GB200 NVL72 stabilizing the scale at 72 GPUs per cabinet, consisting of 18 Compute Trays and 9 Switch Trays [2]. - The Rubin architecture will enhance bandwidth, with the NVLink 6 Switch increasing single GPU interconnect bandwidth to 3.6 TB/s, up from 1.8 TB/s [2]. Group 2: Nvidia's Competitive Advantage - Nvidia maintains a leading position in the supernode market, with a projected shipment of approximately 2,800 units of GB200/300 NVL72 by 2025 [3]. - Future plans include the introduction of Vera Rubin NVL144 and Rubin Ultra NVL576, expanding interconnected GPUs from 72 to 576 [3]. - Innovations such as NVLink and NVLink Switch are crucial for achieving high bandwidth and low latency in AI training clusters, with NVLink 5 Switch supporting a total bandwidth of 130 TB/s for 72 GPUs [4]. Group 3: Industry Landscape and Investment Strategy - The global supernode competition landscape is still being established, with Nvidia currently in a leading position [6]. - The report suggests monitoring Nvidia's supernode supply chain, including components like PCB backplanes, high-speed copper cables, optical modules, and cooling systems [6]. - Chinese manufacturers are actively participating in the supernode and Scale up network sectors, with potential for domestic firms to gain a competitive edge [6].

Summary of Key Points from the Conference Call Transcript Industry Overview - **Industry**: China Data Center Equipment - **Key Focus**: Power and cooling upgrades, particularly related to NVIDIA's Vera Rubin platform launch in 2026 [2][3] Core Insights and Arguments - **NVIDIA's Rubin Platform**: - The Rubin platform is set to double the rack power compared to the previous Blackwell platform, moving from 80% to 100% liquid cooling [2] - Rubin is currently in full production [2] - **VR200 NVL72 Specifications**: - Delivers approximately 3.5x to 5x the training/inference AI computing power compared to GB300 NVL72, significantly increasing rack-level power demand [3] - Upgrades power shelves to a 3*3U 110 kW configuration, compared to the GB300's 8*1U 33 kW shelves [3] - Introduces a 3+1 redundancy design for power shelves [3] - **Future Developments**: - Anticipation of transitioning to the next-generation Kyber rack design, which will support 800V HVDC to meet rising power requirements driven by AI compute scaling [3] - Potential launch of Rubin Ultra in 2027 could further unleash HVDC demand [3] Investment Opportunities - **Stock Picks**: - **Kstar and Kehua**: Expected to benefit from stronger UPS demand and potential upgrades to HVDC and SST [4] - Kstar is anticipated to have strong order intake from US hyperscalers due to strategic partnerships [4] - Kehua's domestic GPU ramp-up and H200 shipment could enhance order intake from domestic hyperscalers [4] Risks and Valuation - **Downside Risks for Data Center Equipment Sector**: - Slower-than-expected AI data center capacity growth [6] - Slower penetration of high-power density products [6] - Challenges in gaining market share in the overseas AIDC equipment supply chain [6] - **Kehua's Price Target and Risks**: - Price target based on DCF methodology; downside risks include slower IDC capacity expansion and lower overseas shipment expectations [7] - Upside risks include faster IDC capacity expansion and stronger relationships with hyperscalers [7] - **Kstar's Price Target and Risks**: - Similar DCF methodology for price target; downside risks include slower IDC capacity expansion and new entrants in the market [8] - Upside risks include faster IDC capacity expansion and higher overseas shipments [8] Additional Important Information - **Analyst Team**: The report was prepared by UBS Securities Asia Limited, with analysts Yishu Yan, Anna Yuan, and Ken Liu involved [5] - **Valuation Methodology**: The report emphasizes the use of DCF methodology for price targets and highlights the importance of understanding risks before making investment decisions [7][8] This summary encapsulates the critical insights and data from the conference call, focusing on the developments in the China data center equipment industry, particularly regarding NVIDIA's advancements and the implications for investment opportunities in Kstar and Kehua.

AI server assembler Wiwynn’s recent 4Q25 gross margin miss (7.2% vs. consensus estimates of 8–8.3%) has triggered a share price correction and reignited investor concerns. While rising component costs can mechanically dilute reported gross margins, it is more critical to examine the underlying profitability trends in server assembly through the lens of structural changes in AI server design.Nvidia continues to increase the level of design integration in AI servers to boost token output per unit of space and ...

近期AI伺服器組裝廠商緯穎公布低於預期的4Q25毛利率 (7.2% vs. 市場預期的8-8.3%)，導致股價下跌，此事又引起投資人關注。雖零組件漲價也會稀釋帳面上的毛利率，但更重要的是從AI伺服器設計的本質上去檢視組裝的獲利能力趨勢。Nvidia持續提升AI伺服器的設計整合程度，以提升每單位空間與電力的Token產出，來因應資料中心有限空間與稀缺電力挑戰。此外，高度整合的設計也有助於改善生產效率，進而有利供應鏈管理與降低維修成本。VR200 NVL72就是設計高度整合的例子，我的產業調查顯示，其Compute tray導入無線化設計（Cable-less），零組件項目顯著減少約40% (vs. GB300 NVL72)。但隨整合程度提高，伴隨而來的是Nvidia指定用料比重提升，與客製化空間也遭到壓縮，這些都不利組裝廠商的獲利能力。我的產業調查顯示，VR200 NVL72 compute tray「不允許客製化規格」的零組件項目佔比提升至20-22%，遠高於GB300 NVL72的5-7%，且VR200 NVL72更直接取消「ODM設計」的零組件項目。事實上，提高整合程度的設計趨勢自GB300 NVL72就 ...

Regarding recent discussions among CCL-focused investors on whether the key CCL spec for VR200 NVL72 could be downgraded from M9, below are my latest surveys and views:1. Nvidia previously began testing 896K2 and 892K2 and prototyping PCBs, which likely gave rise to the market rumors. Test results are expected by the end of 1Q26.2.Nvidia’s current mass-production target remains 896K3. While this has not been finalized, the uncertainty alone is sufficient to impact stock prices that have already priced in th ...

Key Updates on NVIDIA Vera Rubin/VR200 NVL72(Integrating my latest supply chain checks and NVIDIA CEO Jensen Huang’s CES 2026 keynote)1. NVIDIA has renamed the AI server VR200 NVL144 (die-based) to VR200 NVL72 (package-based). My supply chain checks indicate VR200 NVL72 will be offered in two power profiles: Max Q and Max P.➢ Jensen referenced NVL72, rather than the previously used NVL144 naming.➢ Max Q and Max P share the same hardware design.➢ Max Q GPU/rack power (TGP/TDP): ~1.8/190 (kW).➢ Max P GPU/rack ...

Nvidia Vera Rubin/VR200 NVL72重點更新，整合我的最新產業調查與Nvidia CEO黃仁勳的2026 CES演講1. Nvidia將AI伺服器VR200 NVL144 (die-based) 改名為VR200 NVL72 (package-based)。我的調查指出，VR200 NVL72將分為Max Q與Max P兩種power profile。➢ 黃仁勳提到了NVL72，而非先前的NVL144➢ Max Q與Max P硬體設計相同➢ Max Q GPU/機櫃功耗 (TGP/TDP)：約1.8/190 (kW)➢ Max P GPU/機櫃功耗 (TGP/TDP)：約2.3/230 (kW)➢ 兩者均顯著高於GB300 NVL72的1.4/140 (kW)➢ 因應資料中心供電規格，不同power profile提升安裝彈性。此改變也意味著Nvidia開始注意到物理世界對建置AI伺服器的限制，並將其反映在產品規格上2. VR200 NVL72 Max Q與Max P的GPU散熱設計升級。➢ 兩者均採用微通道冷板 (micro-channel cold plate；MCCP) 搭配鍍金散 ...