The author argues that high semiconductor capex and R&D for technologies like HBM, HBM4e, photonic interconnects, and advanced packaging are justified primarily by the profitability of training superior AI models at major labs. While the post expresses a negative sentiment toward Anthropic, it suggests that a shift toward cheap, interchangeable open-source models could disincentivize large training runs and negatively impact the semiconductors market. The author claims to know methods for shrinking models for low-powered devices but warns this could lead to a market crash or consolidation.