The physics behind wafer-scale AI chips, why they make LLM decode faster, and why capacity, software, and unit economics may still decide the company’s fate.