最火芯片研究机构! SemiAnalysis创始人：算力瓶颈从CoWoS转移到EUV，存储吃掉30%资本开支

Core Insights - The core argument of the article is that the bottlenecks in AI computing power expansion are constantly shifting, with the semiconductor manufacturing segment becoming the primary constraint as other infrastructure components like data centers and power supply expand [2][5][6]. Group 1: Bottlenecks in AI Computing Power - The bottlenecks in the AI supply chain have changed almost every year, with previous limitations being CoWoS packaging, power supply, and data centers [4][5]. - As these issues are addressed, new bottlenecks emerge, indicating that the demand for AI is growing faster than the supply chain can expand [5][6]. - Currently, the core limitation is returning to semiconductor manufacturing, specifically in logic chip capacity, high-bandwidth memory (HBM), and wafer fabrication capabilities [6][8]. Group 2: Future Constraints - If AI computing power continues to grow rapidly, future bottlenecks may shift further downstream to semiconductor equipment capacity, particularly focusing on extreme ultraviolet (EUV) lithography machines produced by ASML [9][10]. - The current global production of EUV machines is about 70 units per year, with potential increases to 80 units, but even with aggressive expansion, it is unlikely to exceed 100 units by the end of the decade [11][12]. Group 3: Impact on Consumer Electronics - A significant shortage of memory chips is expected to be a core trading theme in the next couple of years, with predictions that about 30% of capital expenditures from tech giants will flow into memory chips by 2026 [17]. - The demand for high-bandwidth memory (HBM) will lead to a reduction in consumer electronics memory production, potentially increasing costs for devices like smartphones [18][19]. - The global smartphone shipment volume, originally projected at 1.4 billion units annually, may drop to 800 million this year and could halve to 500-600 million next year due to rising memory costs [20]. Group 4: Power Supply Considerations - The article argues that power supply will not be the ultimate constraint for AI computing, and alternative energy solutions can be implemented to support data centers [22][23]. - The concept of space-based data centers is dismissed as economically unfeasible due to high failure rates of chips and expensive communication costs [24].