Dylan Patel, founder and CEO of SemiAnalysis, and Stacy Rasgon, Managing Director at Bernstein, stated at the 35th annual CASPA conference held in the US on October 23rd that AI infrastructure investment is facing multiple challenges. Patel pointed out that rising financing costs (such as the US 10-year government bond yield breaking 5%, and unsecured bond yields for some emerging cloud service providers approaching 15%) and semiconductor supply chain bottlenecks (substrates, HBM memory, optical interconnect components) are limiting AI hashrate expansion. Rasgon added that while NVIDIA holds a performance advantage, the in-house chip initiatives of cloud vendors like Google TPU, Amazon Trainium, and OpenAI are challenging its market share and pricing power. In the memory market, strong HBM demand has led to gross margins as high as 88%-90%, but new capacity is expected to be concentrated for release after 2028, at which point the supply-demand relationship may change, and memory shortages are likely to ease.