Rubin Ultra设计调整--价值量重新分配

Core Viewpoint - The cancellation of the Rubin Ultra 4-die design highlights the challenges in manufacturing complex chips, but the engineering team's adjustments reflect a pragmatic approach to problem-solving rather than a rigid adherence to initial plans [1][3]. Design Configuration - GFHK suggests a potential shift from a native 4-die design to a 2+2 configuration, though this is still under development, making it difficult to finalize the solution [2]. - The specific design choice is less critical as cloud service providers purchase complete computing systems rather than individual chips, meaning the overall system specifications remain unchanged [2]. - If the 2+2 configuration can still deliver approximately 1TB of HBM4e memory per high-end unit, the overall memory capacity of the system will not be affected [2]. Industry Impact - Adjustments in the Rubin Ultra design will alter the value distribution within the hardware supply chain. A native 4-die solution centralizes complexity and pricing power in advanced packaging, while a 2-die module approach reduces pressure on the most advanced packaging layers [4]. - As die aggregation shifts upstream, more integration work will be required at the PCB and module design levels, increasing their importance [4]. - The focus will shift towards system integration and scalability at the tray and rack levels, rather than solely on the scale of individual packaging [4]. Hardware Demand - The demand for hardware remains robust, driven by the AI boom, which has led to shortages across various components, including CPUs and transformers [6]. - A report from Nomura indicates that AI hardware demand is expected to continue growing from 2026 to 2030, with a projected increase in new data center demand from 7 GW in 2025 to 27 GW in 2026 and 28 GW in 2027 [7][10]. - The construction of data centers in North America is ongoing, with numerous projects announced, indicating a strong infrastructure development trend [8]. Capacity Deployment - Incremental capacity deployment is forecasted to rise significantly, with projections of 26.67 GW in 2026 and 28.48 GW in 2027, driven by major players like OpenAI and top cloud service providers [10]. - The demand for CoWoS wafers is expected to increase, with an additional 450,000 wafers needed in 2026 and 600,000 in 2027 [7][10].