The conversation begins by questioning the return on investment for AI, pointing to evidence like GitHub Copilot's pricing changes and AI's absence from GDP statistics. The hosts reveal that initial AI token pricing was subsidized, masking true costs, and that companies now face real expenses as subsidies are removed. They criticize AI companies for using inflated 'run rates' to disguise revenue figures and highlight a severe year-over-year deflation in token prices, creating a structural problem for data center ROI. The discussion then turns to how small language models are commoditizing the market, driving up costs and creating a situation similar to 19th-century railroad booms. The hosts argue that AI investments lack a clear return, comparing expensive, power-hungry GPUs to abandoned railroads and criticizing faith-based arguments for their eventual payoff. Finally, they note that AI coding tools are flooding markets with low-quality output, and the broader economy is not benefiting, with financial strains on hyperscalers signaling a potential market top.