Summary of Conference Call Notes Company and Industry Overview - The conference call discusses the upcoming launch of NVIDIA's LPU (Logic Processing Unit) cabinet, which is positioned in the inference market, competing against Google's TPU v6e. The LPU is expected to begin small-scale supply in Q1-Q2 of 2027, utilizing an ASIC architecture and SDRAM for cost-effectiveness [1][2][3]. Key Points and Arguments LPU Cabinet Specifications - The LPU cabinet will feature a total of 9 computing units, with 8 dedicated to inference calculations and 1 as a redundancy. Each unit will house several LPU inference chips, which will not use HBM but will instead utilize SDRAM [2]. - The liquid cooling system for each computing unit is projected to have a value of $1,000 to $1,200, which is higher than the existing Rubin and GB300 systems due to enhanced specifications [5]. Liquid Cooling Technology - The LPU will adopt a "light liquid cooling" solution, which is simpler than previous models, with an increase in the number of components like liquid cooling plates and quick connectors. The cooling system's design will be less complex due to the lower power consumption of the LPU chips [2][3]. - The introduction of microchannel technology in Rubin's cooling system has led to an increase in the number of quick connectors, enhancing cooling efficiency [11][12]. Market Positioning and Supply Chain Strategy - The LPU is strategically positioned to compete with Google's TPU, with a focus on cost reduction. NVIDIA plans to increase the share of domestic suppliers in components like cooling plates and quick connectors to 20%-30% [8]. - The supply chain for LPU will likely follow the existing supplier framework used for the GB300 series, minimizing the need for re-certification of established suppliers [8]. Production and Shipping Expectations - Rubin's production is set to begin in July 2026, with an expected shipment of approximately 9,500 units in the latter half of 2026. The primary focus for NVIDIA in 2026 will remain on the GB300 series, with an anticipated shipment of around 55,000 units [9][10]. Competitive Landscape - The high-end Rubin and GB series are currently dominated by Taiwanese and Western suppliers, while domestic manufacturers are making inroads in the cooling plate segment. However, the CDU segment still requires breakthroughs for domestic suppliers [1][11]. Other Important Insights - The CDU (Cooling Distribution Unit) segment is expected to see a decrease in demand due to the lower overall power consumption of the LPU cabinet compared to training GPUs like Rubin [6]. - The introduction of two-phase cooling plates is anticipated to enhance cooling efficiency significantly, with a value increase of 50%-60% compared to single-phase plates [17]. - The upcoming GTC conference will focus on advancements in liquid cooling technology for both Rubin and LPU, highlighting the importance of these products in NVIDIA's portfolio [20]. This summary encapsulates the critical aspects of the conference call, providing insights into NVIDIA's strategic direction, product specifications, and market positioning within the liquid cooling and inference computing landscape.